Sensory acquisition for emergent body representations in neuro-robotic systems

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content SENSORY ACQUISITION FOR EMERGENT BODY REPRESENTATIONS IN
NEURO-ROBOTIC SYSTEMS
by
Jasmine Ana s Berry
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulllment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(COMPUTER SCIENCE)
December 2020
Copyright 2020
Dedication
This dissertation is dedicated to those who have contributed to both my personal and professional
growth through the years as a Doctoral scholar. Your unwavering support and encouragement
over the years have made this journey worthwhile.
To Dwayne, Gwendolen, Jaymes, Yvette, Li'Ma
and especially to God.
ii
Acknowledgements
I would like to rst acknowledge my friends and family for their continual support throughout my
time in graduate school. The Ph.D. is yours just as much as it is mine. I want to thank Dr. Fran-
cisco Valero-Cuevas (BBDL), my PI, for his advisement towards developing the research methods
for my dissertation. Our many discussions helped pave the way for the material grounding of
self-modeling and my professional dedication towards becoming a great scientist and engineer.
Much appreciation is due to Dr. Alice Parker (BioRC Group), my research advisor and mentor,
for her guidance, technical insight, and encouragement throughout my Ph.D. journey. She is
credited for sparking my interest in the topics of Consciousness and Self-Awareness. I am grateful
to Dr. Michael Arbib (USC Brain Project), my advisor, for introducing me to the concept of
Brain Theory and its principles within. I also acknowledge the contribution of my examiners and
qualications committee, Dr. Bhaskar Krishnamachari, Dr. Wei-Min Shen, and Dr. Paul Rosen-
bloom, for their valuable feedback towards the betterment of the thesis. This Ph.D. was made
possible through the combination of several disciplines, including Computer Science, Biomedi-
cal Engineering, Neuroscience, and Psychology. The dissertation would not have been possible
without the support and contributions from various funding agencies who deserve recognition.
To Dr. Timothy Pinkston, an eternity of gratitude for having always having an open oce for
me. You were one of the initial reasons I joined USC and embarked on this great opportunity.
To Dr. Aftab Ahmad, my undergraduate professor and mentor, I am indebted to you for your
guidance and shared intellect since NSU. I am grateful to all international collaborators, Ph.D.
and Master's students for their knowledge and shared expertise. Thank you all, and I hope to
give back throughout my career in ways to re
ect my gratitude.
iii
Table of Contents
Dedication ii
Acknowledgements iii
List Of Tables vi
List Of Figures vii
Abstract ix
Chapter 1: Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Research Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4.1 Self-recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4.2 Sense . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4.3 State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4.4 Sensory-Motor Gestalt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Dissertation Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Chapter 2: Background and Related Work 11
2.1 The Brain Represents the Body: Neuroanatomical and Articial Shaping of the Self 11
2.2 Tendon-Driven Neuromechanics: Sensorimotor Control in Muscle Redundancy . . . 15
2.3 Proprioception and Its Role in Bodily State Estimation . . . . . . . . . . . . . . . 16
2.4 Body Representations for Self-Awareness in Animals and Machines . . . . . . . . . 17
2.5 Implications of Physiological Subjective Experience to the Emergence of Machine
Subjective Experience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.5.1 Ineable Consciousness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.5.2 Explanatory Gap Dilemma . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.5.3 Unraveling Self-Awareness Toward Augmentation . . . . . . . . . . . . . . . 23
2.5.4 Proposed Transition to Machine Consciousness . . . . . . . . . . . . . . . . 26
2.5.5 Challenges Moving Forward . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.6 Models and Applications of Body Representations in Robotics: A Review . . . . . 27
Chapter 3: Self-Recognition: Extension of Mirror Neuron System II for Agency 30
3.1 Introduction to Mirror Self-Recognition . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.1 Agent Self-Recognition Test . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1.2 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
iv
3.1.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.2 Extension of Mirror Neuron System, 2 (MNS2) for Agency . . . . . . . . . . . . . 44
3.2.1 Learning with Operant Conditioning . . . . . . . . . . . . . . . . . . . . . . 47
3.2.2 Experimental Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Chapter 4: Sense: Quantifying High Dimensional Feasible Sensory Sets 59
4.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2 Introduction: How the Body Builds the Brain . . . . . . . . . . . . . . . . . . . . . 60
4.3 Experimental Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.4.1 Kinematics Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.4.2 Aerent Signaling Dependent on Muscle Velocity . . . . . . . . . . . . . . . 67
4.4.3 Sensory Bounds According to Task Constraints . . . . . . . . . . . . . . . . 67
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Chapter 5: State: Sensory Aerent Organization to Classication of Actionable
States 72
5.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.3 Experimental Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.3.1 Kinematic Model Structure and Parameters . . . . . . . . . . . . . . . . . . 75
5.3.2 Trajectory Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.3.3 Muscle Spindle Aerent Data Collection . . . . . . . . . . . . . . . . . . . . 79
5.3.4 Comparison of Inter-class Trajectory Context . . . . . . . . . . . . . . . . . 80
5.3.5 Spatial, Spatio-Temporal, Pre-processing of Muscle Spindle Aerent Data . 81
5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.4.1 Raw Multi-Dimensional Aerents Are Value Bound, But State-Indiscriminable 83
5.4.2 Pre-processing Suggests Observable Correlations in Sensory and Motor Maps 86
5.4.3 Correlation Index Reveals Markers of Action Discriminability, Classication 88
5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Chapter 6: Sensory-Motor Gestalt: Sensation and Action as the Foundations of
Identity, Agency, and Self 94
6.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.2 Introduction of Sensory-Motor Gestalt: Origin and Denition . . . . . . . . . . . . 95
6.3 Sensory-Motor Gestalt: Applying Gestalt Laws to Sensorimotor Function . . . . . 98
6.3.1 Law of Proximity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.3.2 Law of Similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.3.3 Law of Closure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.3.4 Law of Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.3.5 Law of Pr agnanz (Good Form, Clarity) . . . . . . . . . . . . . . . . . . . . 104
6.3.6 Supplementary Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.4 Functional Utility of the Sensory-Motor Gestalt . . . . . . . . . . . . . . . . . . . . 105
6.5 Abstracting Self from Sensorimotor Experiences for Neuromuscular Systems . . . . 106
6.6 Role of Self and Identity in Autonomous Robotics and Synthetic Biological Agents 110
6.7 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Chapter 7: Conclusion 113
Bibliography 118
v
List Of Tables
3.1 Brain Operating Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2 Neuron Firings for Execution and Observation Tasks . . . . . . . . . . . . . . . . . 46
4.1 Simulated limb and musculotendon parameters. . . . . . . . . . . . . . . . . . . . . 62
4.2 Velocity Signicance in Aerents (ANOVA P-values) . . . . . . . . . . . . . . . . . 68
5.1 Simulated limb and musculotendon parameters. . . . . . . . . . . . . . . . . . . . . 76
5.2 Bounded ranges of Ia aerent activity, measured in pulses per second (pps), for
muscle groups averaged across Forward and Reverse directions. . . . . . . . . . . . 85
5.3 Bounded ranges of II aerent activity, measured in pulses per second (pps), for
muscle groups averaged across Forward and Reverse directions. . . . . . . . . . . . 86
vi
List Of Figures
1.1 Research Phase Plan for Building Body Representations . . . . . . . . . . . . . . . 8
2.1 Framework for human self-knowledge levels versus computational self-awareness levels 19
2.2 Extension of self-representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.1 Inverse and Forward Models of Cortical Activity . . . . . . . . . . . . . . . . . . . 33
3.2 Bidirectional circuit map for coordinating visual body image and motor body image. 36
3.3 Agent versus Adversary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.4 Sampled training set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.5 Action and Inaction Paradigm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.6 Augmented MNS2 model for Action and Inaction Task . . . . . . . . . . . . . . . . 52
3.7 Mirror Neuron System II (MNS2) Graphical User Interface . . . . . . . . . . . . . 55
3.8 Obstacles for Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.1 Static Case: All possible x-y coordinates for q
1
and q
2
degree ranges. . . . . . . . . 61
4.2 Cartesian space and Conguration space of kinematic arm movement . . . . . . . . 63
4.3 Six-dimensional representation of change in muscle length along four trajectories
of the Dynamic Case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.4 Velocity speeds versus aerent signals in Group Ia and II. . . . . . . . . . . . . . . 66
4.5 Primary and secondary aerent space for Trajectories across 50,000 time samples . 70
5.1 Limb kinematics Derived from Distinctive Trajectory Types. . . . . . . . . . . . . 75
5.2 Spindle aerent population data for ve distinct trajectories . . . . . . . . . . . . . 84
vii
5.3 PCA Dimensions of Aerents Reveal Distinct Shapes . . . . . . . . . . . . . . . . . 87
5.4 Spread of Discriminability Within Cross Correlation Scatter . . . . . . . . . . . . . 89
5.5 Confusion matrices of raw and pre-processed spindle aerent data. . . . . . . . . . 92
6.1 Published article count per year that indicate association between subjective expe-
riences and sensory modalities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.2 Gestalt laws of perceptual organization. . . . . . . . . . . . . . . . . . . . . . . . . 98
6.3 Evolutionary high-dimensional time-varying manifold of spike trains of spindle af-
ferents. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.4 Formal representation of minimal self . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.5 Neuromechanical perspective sensory inputs transforming to motor outputs . . . . 107
6.6 Test-bed applications for Sensory-Motor Gestalt implementation in tendon-driven
systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
viii
Abstract
An ongoing engineering challenge is achieving agility, information processing, and
exibility in
robotic systems. Building neuromorphic robots called NeuRoBots (i.e., robots that imitate the
mechanisms of neural sensorimotor processing in animals) is one approach to accomplishing this
goal. NeuRoBots oer several advantages over traditional robots and also serve as test beds for
understanding the sensorimotor dynamics of mammalian neuromuscular physiology. The notion
of how the anatomical brain builds a sense of self and how neuro-robotic agents can utilize body
schemas (or representations) to build a sense of self have not been particularly successful due to
varied and often contradictory accounts. In this dissertation I present a critical step in forming
self-identied body schemas, based on physiological simulation of proprioceptive aerent signals,
to determine the plausibility of whether biological signals can be used to inform the operation of a
state machine. First, I demonstrate that a given movement gives rise to a distinct sensory manifold
embedded in the 12-D space of muscle spindle information that is largely independent of the choice
of muscle coordination strategy. Given that muscle lengths and velocities are fully determined by
joint kinematics, such manifolds provide a rich set of information to use in its control. Secondly,
I show that high-dimensional multi-muscle proprioceptive ensembles can usefully discriminate
limb states and be utilized as a suitable classier for inter-trajectory comparisons|but only
after minimal pre-processing. Lastly, I present the concept of Sensory-Motor Gestalt, which
provides a unifying framework for constructing body states into useful behaviors to understand
the foundations of sense of self in hybrid robots and synthetic biological agents.
ix
Chapter 1
Introduction
1.1 Motivation
Our motivation for this dissertation is to assist in the eorts towards building robotic systems
that can acquire unique movement capabilities online that lead to classifying possible actions amid
encounters of changing external circumstances. Such systems should have the capability to learn
how to perform movements given their unique set of anatomical and physiological constraints.
Achieving this goal requires us rst to survey the main challenges within the eld of computing
that our intermediate goals will address.
As conventional computing reaches practical limitations for performance [172], additional com-
puting methods are sought after for the next generation of autonomous devices and systems. A
longtime technological goal of articial intelligence (AI) and robotics is to create computational
systems with functions similar to biological brains with the prospect of machines behaving and
thinking like humans or mammals. Neuromorphic computing is increasingly becoming one plau-
sible approach for accomplishing this goal as it introduces fundamental architectures that can
1
potentially perform similarly to anatomical neurons. This essentially attempts to emulate in sili-
con what the biological cortex does in vivo. One application of neuromorphic computing is found
in biorobotics and bio-inspired machines for enhanced information processing and hosting software
models of neural processes that typically address problems of memory, perception, motor control,
and multisensory integration [11]. A signicant problem for both traditional robots and biorobots
is the complexity of building general-purpose autonomy for executing various cross-domain tasks
while learning new skills without catastrophic forgetting [173]. These systems do not entirely exist
yet, as the eld is still heavily challenged with single-task robotics. However, a closer examination
of how our brains and bodies are designed can provide the means to make incremental strides in
this direction.
The brain's primary mode of operation is to assure its host's survival, especially amid adverse and
unpredictable conditions. In many ways, so should our autonomous machines behave. Biorobotics
ought to be constructed with the basic physiological needs in mind, too. Llin as [111] and Carter
[24] suggest our brains must execute the several objectives to properly maintain a sucient level
of autonomy:
1. Generate internal sensory signals that indicate our bodies' primal needs (i.e., food, rest,
safety, and security).
2. Form a map of the world to direct us to locations to satisfy our needs.
3. Produce the appropriate actions and energy to move us there.
4. Alert us of opportunities and threats (both present and foreseeable).
5. Lastly, tailor our actions based on the requirements and constraints of our current state,
goals, and environment.
2
At the core of these faculties is the concept of the physical self, body representations, and bodily
awareness, which all together form the basis of this thesis. One can also extrapolate from this
list the basis for how an organism uses its body to generate actions (according to specics of the
system's needs, mechanical makeup, and physical constraints) and what underlying physiological
mechanisms are responsible for the selection of these actions. Without a functional body repre-
sentation, there will be disruptions in the sense of agency, therefore negatively aecting actions
towards goal-directed behaviors. Actions, choices, and decisions that an agent should be mostly
identity-congruent as it has been observed in humans [136].
For brain-based systems like NeuRoBots, we seek to determine the scope of parameters neces-
sary to comprise an articial self? Consequently, we can answer address the implications for
future biorobotic systems? We are particularly interested in how a NeuRobot can improve its
"capabilities by being able to automatically synthesize, extend, or adapt to a model of its body"
[82] from action-oriented body representations. These are the primary factors motivating this
research. By examining sensory-motor contingencies, we can explore how bodies and their models
are cognitively encoded and decoded to produce meaningful behaviors that are tailored to the
system's physical constraints. The development of new hardware materials and manufacturing
schemes gives way to an advanced generation of robots that seek to become increasingly power-
ecient, multifunctional, adaptable, and autonomous in ways similar to biological organisms.
The present work aims to provide a cognitive architecture to enable these robots to optimize their
actions for decision making, ecient locomotion, and planning through the use of self-modeling
techniques and body representations. Through computational methods of sensory acquisition in
simulated tendon-driven limbs, we are then motivated to build a system that determines unique
movement capabilities online, leading to classifying possible actions amid encounters of changing
circumstances. How neuromorphic systems and tendon-driven robotics are capable of performing
movements given its anatomical and physiological constraints will be further explored.
3
1.2 Problem Statement
Generally, in practice, robotic agents are pre-programmed to perform a pre-assigned set of specic
tasks in a controlled environment. When systems are programmed to learn in such a way, they
typically do so through imitation [42] [100], exhaustive iterations of execution, and simulation
of the motor-to-sensory maps. In contrast, vertebrate animals usually learn by limited trial-and-
error interactions with the physical world [5]. The biological approach supports learning new tasks
that overlay existing capabilities, essentially demonstrating that novel behaviors can evolve and
emerge through trials. To achieve this learning for our robotic agent (i.e., NeuRoBot), I propose
for the acquisition and redundant integration of sensory information to be used as the driver of
motor map development instead of the consequence of motor behavior; a methodological process
we will refer to as Motor Learning by Active Sensing.
The cortical-motor-physical-sensory feedback loop, which is made possible by the NeuRoBot,
should be capable of supporting independent exploration in the physical world. Including a
model of self that emerges from the formalized construction and classication of sensory aerents
can address several issues. One is ensuring biorobotics continue to closely align their mechanisms
with biological systems, namely mapping sensorimotor representations of the body for action.
Another is improving closed-loop control with sensory feedback that can be predicted in advance
before it is perceived | thus being useful for systems with feedback delays. Our implementation
should serve as a fundamental layer to self-modeling systems while helping launch robots capable
of continuous, autonomous, and cumulative learning. For NeuRoBots, this level of functionality
is ideal for deployment in environments that are not completely observable. Our assumptions are
based on the following Fundamental Premises:
4
Fundamental Biological Premise: The nervous system constantly assesses and enforces
its current experienced-based estimate of body model against incoming sensory input and
feasible motor actions [166].
Fundamental Robotic Premise #1: The brain's body-model constantly assesses and
enforces its current \hypothesis of body representation" against incoming sensory input and
feasible motor actions.
Fundamental Robotic Premise #2: A system's physical actions and interactions with
the environment provide sucient information to build a minimal representation of a robot's
physical properties.
Then, the minimal body representation (implicit) of a robot is dened as a repertoire
of physical actions, resulting movements and possible transitions between the actions.
Here, a minimal body representation may alternatively be referred to as the minimal
sense of physical self.
1.3 Hypothesis
We hypothesize muscle spindle and Golgi tendon organs (GTO) signals that are available to
the mammalian nervous system are also useful enough for classifying dierent movements with
sucient discriminability. Preliminary experimental observations have demonstrated how the
evolution of feasible sensory sets (FSS) can provide variables of interest to be extracted for col-
lection and statistical analysis. We then ultimately seek to show the validity in utilizing FSS for
goal-directed motor mapping of neural-driven limbs. Anatomical bodies, such as the proposed
neuromorphic and neuromechanical system [130, 85], can be modeled as a time-invariant control
system that channels sensory inputs into actionable states for task classication. The aerent
5
data collected from our analyses form topological manifolds. Therefore, by denition, we can
expect the entirety of the model's state space to be maintained as a non-linear representation in
the form of topological manifold maps that satisfy countability and separability conditions such
that in N-dimensional, each point p2 X in the topological space M has a local neighborhood
that is homeomorphic to Euclidean spaceR
N
for some n 0 [76].
The combination of linear and non-linear manifold mapping methods can delineate the feasible
transitions between adjoining states. Aggregating the states into a unied whole will then form
the minimal computational frame a body representation (BR) that emerges as a byproduct of
experiences may be specied as a self or identity that is constructed from the agent's subjective
experiences throughout the runtime. When mathematically expressed, the emergent BR entails
knowledge of how an agent can move within its anatomical constraints and how it will classify
the internal dependencies (e.g., muscle lengths, muscle velocity moment arms, etc.) controlling
each task action.
Consequently, this state classication method is a prerequisite for determining how individual
body representations (i.e., body schemas) for neuromuscular-driven robotic systems can be self-
generated from physiological sensory signals that result from a range of immature to skilled motor
actions, prompting learning of useful actions and feasible behaviors. The resulting model of the
neural-driven system will be produced by demonstrating that a system can learn a dynamic model
of its neuromuscular behavior, self, and identity from aerent constraints (proprioception and
somatosensory stimulation). Applying intrinsic motivation within the robotic agent's exploration
of the state space may be accomplished with randomly selected movements coded in the form, or
neural ring rates could be implemented for undirected curiosity or exploration of the state space
[4].
6
1.4 Research Approach
Our end-goal is to equip a neuro-robot (i.e., NeuRoBot, a robot with an articial nervous system)
that forms a repertoire of physiologically-inspired sensory and motor couplings to explore and
exploit physical actions and transitions among them. This dissertation's foundation is grounded
in the fundamental premise that the brain's body-model continuously assesses and enforces its
current \hypothesis of a body representation" against incoming sensory input and feasible motor
actions. We further hypothesize that computationally modeling the biological central nervous
system's bottom-up method for extracting sensory aerents to form unique motor maps can
build functional action-oriented body representations for our NeuRoBot system. Revealing this
coupling can show how sensory synergy is intertwined with motor combinations, thus enabling
improvements to the biorobots' design and control.
To demonstrate this, I will use an alternative approach to self-modeling that incorporates su-
pervised and unsupervised learning approaches. We acknowledge that a machine learning (ML)
approach does not fully align with the brain's methodical process of forming online body schemas.
Simultaneously, a manifold description more intuitively utilizes the biological data distribution in
a multidimensional space [180]. Although useful for many applications, ML techniques may be
prone to inaccuracies but can still be somewhat reliable when revealing the dimensional structure
of the raw data manifolds collected. Combining the mathematics of manifolds and Gestalt law
principles can provide a more intuitive expression of the intrinsic structure, shape, and space
complexity. The dissertation deliverable is a system that performs arbitrary,
exible tasks using
the topologically rened sensory-to-motor maps. Fig. 1.1 depicts the phases of this research
plan that will be implemented in order to achieve the desired goal. The phases incrementally
build on top of each other for the body representation of our NeuRoBot. Upon completion, we
will evaluate the implementation's success against three main criteria that encompass the four
7
phases. The system's nal result should provide a computational foundation for constructing
body representations, with contributions from each of the following phases.
1. Self-recognition: Extending the Mirror Neuron System, II (MNS2) for Agency in Reaching
and Hand Grasping
2. Sense: Quantifying High Dimensional Feasible Sensory Sets
3. State: Organization of Sensory Aerents to Classication of Actionable States
4. Sensory-Motor Gestalt: Exploring a perceptual continuum for constructing an articial self
via Gestalt Laws
Figure 1.1: Research phase plan for building unique body representations. The research begins
with a study on Self-recognition (Chapter 3) then on to Sense (Chapter 4), State (Chapter 5),
and lastly Sensory-Motor Gestalt (Chapter 6).
1.4.1 Self-recognition
The rst phase of our research examines the Mirror Neuron System II (MNS2). We provide a
theoretical extension of the model that identies principal neural correlates and Brain Operating
Principles (BOPs) that are useful for functions of agency in autonomous systems. We use several
8
BOPs in two tasks: 1) a simulation of self-recognition using and 2) a hand reaching and grasping
in an interactive user interface environment.
1.4.2 Sense
We transition from the identication of brain theory principles of agency in the Self-Recognition
phase to the Sense phase, where we analyze the usability of physiological proprioceptive signals
in the muscle spindle. The Sense phase of our research focuses on quantifying high-dimensional
Feasible Sensory Sets (FSSs), through geometric interpretation, that can detect and categorize
functional movements and tasks. It justies the practicality of recording plausible movements for
the quadruped robotic platform through a single-limb inspection of a 6-muscle, 2-link experimental
human arm.
1.4.3 State
The second observable functionality of our system will be its ability to generate and identify feasi-
ble transitions among the functional movements detected in the Sense milestone. The State mile-
stone uses simulated physiological proprioceptive signals to organize sensory aerents and classify
them into actionable states. Here we determined that high-dimensional multi-muscle proprio-
ceptive ensembles can usefully discriminate limb states|but only after minimal pre-processing.
Importantly, this nding may explain the documented subcortical pre-processing of aerent sig-
nals, such as cutaneous signals processing by the cat's cuneate nucleus.
1.4.4 Sensory-Motor Gestalt
The nal metric of my dissertation's success is the demonstration of sensory-to-motor-sensory
maps as useful body representations of actionable states. We propose the Sensory-Motor Gestalt
as a perceptual continuum for constructing self via Gestalt Laws. This phase is expected to assist
9
in autonomously updating the sensory-to-motor maps of proprioceptive aerents and spindle
information by forming and constraining the topology and geometry of manifold shapes.
1.5 Dissertation Outline
The remaining content of this dissertation is summarized as follows. Chapter 2 presents the neces-
sary background information, current knowledge, and related literature for framing the right per-
spective of the research on constructing body representations (i.e., body schemas). This chapter
will additionally acknowledge the core contributions from specialists in the eld of body schemas
while highlighting their competing views. Chapters 3-6 describe the experimental methodology
for the four phases of work (i.e., Self-recognition, Sense, State, and Sensory-Motor Gestalt, respec-
tively) and their relevant contributions to forming emergent body representations in neuro-robotic
systems. Lastly, in Chapter 7, we interpret our results, their signicance and discuss the new in-
sights obtained.
10
Chapter 2
Background and Related Work
This chapter introduces our fundamental understanding of the neuroscience, physiology, and bi-
ology involved in the behavioral and cognitive phenomenon of constructing body representations.
I begin with a primer on past and contemporary views on how the brain forms representations of
the body. Next, I will provide insight into how the shaping of a self (i.e., body schema) may form
from the mathematics of the neuromechanical perspective. Following is an understanding of how
the sensory and motor interactions in muscle redundancy (i.e., the fact that we have `too many'
muscles). Lastly, I will analyze comparable models and applications in the eld of robotics and
cognition to present future improvements or trade-o for the use of our proposed model.
2.1 The Brain Represents the Body: Neuroanatomical and
Articial Shaping of the Self
Body representations are valuable as functional utilities in both biological agents and hybrid
systems. The neuroanatomical basis of shaping body representations into an embodied self has
been observed through empirical research, therefore providing potential computational analogs
for shaping an articial self in bio-inspired robots. Engineering
exible bio-inspired robots require
11
an understanding of the neural interactions and computations between the brain and body. We
choose to investigate the brain-body connection through the perspective of neuromechanics for
vertebrates. Neuromechanics examines the nervous system's functions from within the body's
mechanical constraints and anatomical structure. Within the context of this research, neurome-
chanics addresses the neural computational complexities that occur when the brain controls body
movement. The implementation of neuromechanics presented here will cover the span of several
concepts that include neuroscience, computational geometry, muscle mechanics, and anatomy
[178]. Several neural computational problems arise within the brain's circuitry while it controls
the body's mechanical movement. One such problem is that of spatiotemporal representations in
the brain and its resulting control abilities. The brain is a self-organizing and self-repairing cir-
cuit. Its plasticity allows dynamic construction of the generated maps when external and internal
changes are experienced and observed.
We argue that the nature of modeling the individual self with all of its attributes is a necessary
precursor for biological agents to control their actions directly. However, some have interestingly
undervalued its role and made counter-claims against it being a requirement, namely in collective
systems [124]. Others have argued that a primal sense of self in animals and humans develops due
to internal prospective foraging in the environment, also known as exploration and exploitation,
which we will take advantage of in our implementation [81]. Next, let us consider what attempts
have been made to build a model of the self [105, 191, 79] and address the self's representation
within biological neural correlates [95]. We argue that the self is a by-product of the formation of
sensory manifolds made available through perceptual learning. Unfortunately, the precise neural
correlates for forming the self and achieving self-awareness are mostly unidentied. Gallagher's
[55] initial step to this approach is constructing a primitive version of the self that ignores irrelevant
features, called the \minimal" self. He assumes this model of self is most pertinent to robotic
models and is reinforced by neurocognitive disorders [46] such as schizophrenia that aects the
12
prefrontal cortex, an area thought to be critical for the formation of self. Gallese [57] provided
insight on how the body in action is necessary for the building of self, which is a nding we agree
with and aligns with the Exploration and Exploitation concept. The neural mechanisms that
constitute the dierentiation between self and other were determined by some to be the link for
self-actualization [143, 182, 184]. However, several other approaches have suggested for identifying
the brain regions involved: using imaging techniques (PET, MRI, and fMRI) to evaluate healthy
brains and making contrasts with studies of impaired brain patients [91].
The same pattern of a lack of consensus in dening self-awareness follows in the construction
of a neural basis of the self. Morin [126] challenges a common stance in neuroscience that self-
awareness is based in the right hemisphere, particularly in the anterior insular cortex, which has
been noted for its integration of interoceptive and exteroceptive signals in the body [34, 161, 35].
This region was observed to activate with the detection of mismatches and discrepancies between
predicted signals and interoceptive signals that were perceived [68]. Morin [126] also assessed
the hemispheric activity in the mirror self-recognition (MSR) and theory-of-mind (ToM) tests,
amongst other self-awareness evaluation tests (e.g., self-description, autobiography). One of the
conclusions drawn from this study was that both hemispheres were active during the tasks, which
insinuates a distributed network of connections in self-referential activities rather than in localized
brain regions; thus, debunking the right hemispheric claim. Another study involving a patient
with severe brain damage to three specic regions that were once considered imperative for self-
awareness development surprisingly exhibited no signs of mental degradation during self-awareness
tasks. Philippi et al. [138] agree with Morin [126] in stating that we cannot pinpoint self-awareness
processes to a single brain area or lobule, but instead rather distributed neural networks. Others
have theorized the brainstem, posteromedial cortices, thalamus, and spindle cells in the anterior
cingulate in the frontal lobe are responsible for self-awareness development [114].
13
The sense of agency [25] (i.e., subjective ownership and control for one's actions) is the next
trait realized after establishing a model of the implicit self [86]. How agency develops is also
a controversial matter. However, ownership and intent may be anatomically represented in the
brain and subsequently used for the dynamic model of self. During the cognitive assessment of
interoceptive signals resulting from eerent motor intentions, intentional action was observed to
cause amplied activity in the anterior insular region[19]. Although there are competing theories
and evidence on the matter, the prevailing theme is the distribution of neural activity across the
brain is more likely for self-shaping than localized activity.
We recommend having emergent self models of the neuromuscular systems and proposing a dy-
namic property that emphasizes plasticity according to experiences, which is an aspect our self-
model implementation will feature. Other researchers [104] have previously demonstrated robots
that can build and calibrate themselves according to their subjective properties. Bongard et al.'s
[15] model showed how the self could develop via movements made under the exploration of its
current locomotive capabilities. \Injuries" to the four-legged robot that render a particular limb
ineective would prompt the robot to update the various models of its morphology. Consequently,
compensating behavior was observed via the system's inference of its topology and parametric
changes. In contrast to this implementation, which uses an actuation-sensation method to reason
its own structure, we will obtain a holistic mathematical representation of the self as it evolves.
Instead of forming opposing robot internal models and \generating actions to maximize disagree-
ment between predictions of these models," our approach will primarily look at all the physically
possible states of transition that occur within the manifold space.
14
2.2 Tendon-Driven Neuromechanics: Sensorimotor
Control in Muscle Redundancy
Determining the role of sensory information in the body is an underappreciated area of study
within sensorimotor research. Sensorimotor control research, both past and ongoing, has made
eorts to predominately provide evidence for how the brain in
uences the body's actions and
perceptive capabilities [72] [162]. However, the counter to these works (i.e., how the body's
perception of sensory aerent shapes the brain [28] [5] [135]) is not as extensively considered until
recent years as shown in Fig. 6.1. Often not taken into account are sensory states and their
eects on building the brain's body composite model necessary for involuntary and voluntary
behaviors. Such behaviors serve as a form of self-expression to evaluate the ecacy of one's use of
functional behaviors and practical actions. Arguments have been made both for and against the
view of whether sensory information's presence and quality is a necessary condition for implicit
self-awareness [98]. Does the existence of a self-modeled body schema or self-awareness hinge
upon the availability of sensory aerents? Most have answered this question in the armative
and provided empirical data to support their claim. The importance of established manifolds
is found in the coherence of sensory signals for kinetic energy optimization of arm movements
during object manipulation tasks [47]. Platek et al. [140] hypothesized self-related information
that emanates from the various sensory domains (e.g., visual, auditory, and olfactory) aects self-
face recognition in such a way that enhances the priming of a model of the self, and also models
of familiar faces and strangers. Incorporating dierent sensory domains leads to the discussion of
multisensory integration [169] and how such can be a sensory set representation of an action.
15
2.3 Proprioception and Its Role in Bodily State Estimation
There is insucient research on sensory paradigms in sensorimotor neuroscience compared to the
numerous works solely on motor activity. However, the two should not be separated so widely.
Motor planning is indeed happening in sensory space and should be taken into account. Motor
control is often viewed as a direct outcome of neural activity descending from the brain's motor
areas: brain stem, basal ganglia, cerebellum, and the primary motor cortex (M1). Nevertheless,
we need to take it a step further. Much less attention is given to how an agent would react from an
inverse of this activity. In other words, how would motor control look instead as the consequence
of sensory input? We start by examining this issue by inspecting proprioception, a sense usually
associated with body awareness. Proprioception can aect our learning, focus(attention), and
behavior. Such a sensory system of receptors located in our muscles, ligaments, and joints is
designed to articulate where our body is in space without visual stimuli. It is evident in many
studies that the mind (brain) shapes the body. Conversely, how the body shapes the mind (and
in turn, aects our behavior) is a dynamic that is left without sucient understanding. These
conceptual streams coincide with debates on the vitality of cohesive perception and action for
eective sensorimotor control (Iberall and Arbib 1990, Mechsner et al. 2001). In this thesis,
we seek to showcase how far precisely one can get with their motor control abilities by primarily
targeting the senses of vision, somatosensory signals, and proprioception, as they all have the most
signicant bearing on output motor control. Additionally, our sensory states will initially span
proprioception of limb position, joint torques, spindle signals, skin sensation, Golgi tendon organs,
and kinematic frames of reference. If successful with these modalities, we can move onward to
examine other senses, such as tactile feedback and auditory signaling.
16
2.4 Body Representations for Self-Awareness in Animals
and Machines
Before explaining what self-awareness means for machine intelligence, let us ascertain how it is
dened. The importance of a body representation and its features should not be overlooked when
designing the architecture for autonomous individual and collective systems. It is a critical feature
that will become progressively vital as technologies continue to advance in the coming years.
However, the questions \What is a body representation?" and \What do body representations
oer autonomous systems?" evokes answers that are neither well-dened nor understood, and are
often subjectively characterized by the disciplines that dene them. The perspective we chose to
inspect body representations from is from the context of self-awareness. For example, Morin [125]
from the eld of psychology, denes self-awareness \as the capacity to become the object of one's
attention." Nagel and Searle, philosophers of the mind, identied three features necessary for the
formation of self-awareness in an agent (or self): subjectivity, unity, and intentionality [127, 53].
Subjectivity denotes the awareness of the self as a private and distinct experience of sensations.
Unity in self-awareness conveys the unied singular experience that an individual or agent may
have instead of separate sensory modalities. Lastly, intentionality directs consecutive moments
that occur within our self-aware state to a designated goal. Damasio [37], Koch [96], and Crick [33]
also agree that these features must be attended to for full comprehension of self-awareness. Of the
three features, we have chosen to evaluate subjectivity, which relates to building an appropriate
introspective model of the self.
Within the scope of computer science, self-awareness is fundamentally viewed as a combination
of the ability to possess information about one's internal state (private self-awareness), possess-
ing knowledge about one's external environment for insight on how it is perceived by others
(public self-awareness), and maintaining information about future actions and decisions it could
17
potentially make [110]. In biology, self-awareness is self-directed behavior guided by external en-
vironmental factors [10]. Some cognitive scientists portray self-awareness as the embodiment of
a sense of agency and a sense of ownership [55]. McGeer [118] claims that the target of atten-
tion within any experience is the true meaning of self-awareness, but also further classies the
dierence between an agent simply experiencing something and then actually having an alerted
meta-awareness of such experience.
Most approaches to achieving self-awareness should be an interdisciplinary eort due to its inher-
ently subjective nature and bias. A computational view and denition of self-awareness allow for
scalability [110] and highly complex integration of nodes in a network, with the choice of even
implementing self-awareness directly or as an emerging property [60]. Amongst the varied deni-
tions, there are also many categorical types of self-awareness. Moreover, those types are discrete
levels to gauge how much an agent is self-aware. Prominent researchers who have made level-type
distinctions include Rochat [151], Neisser [129], Piaget [139], and Lewis et al. (2015). Rochat
[151] was motivated to observe children's behavior in what has been deemed the conventional
self-awareness test, the mirror self-recognition test. He questioned how the self develops over time
and at what stage of development does one view themselves as a separate entity in relation to
the world. He concluded that there was a range of ve levels needed to describe the maturation
a child experiences to reach the self-aware state:
Level 0 { Confusion
Level 1 { Dierentiation
Level 2 { Situation
Level 3 { Identication
Level 4 { Permanence
18
Level 5 { Self-consciousness or \meta" self-awareness
Rochat's approach reveals that we need to be cautious of having a dualistic view of an agent
either possessing self-awareness or not, with no intermediary stages.
Figure 2.1: Neisser's [129] appraisal of the various levels of self-related knowledge one must attain
to reach self-awareness (left). Lewis [109] juxtaposes this with their own framework (right) of
a computational perspective beginning with stimuli awareness and concluding with meta-self-
awareness.
Neisser [129] proposed ve dierent types of selves that we gradually become knowledgeable in
early development to get to the self-aware state. The selves span from the ecological self in
which the self is perceived with respect to the physical environment to the conceptual self, where
one forms a concept of self in a social-like structure. Lewis et al. [109] took on the challenge
to convert these levels, formerly based in psychology, into engineering for architecture design of
computational systems Fig. 2.1. Starting with stimulus awareness, the agent is capable of using
incoming stimuli to respond to events. Interaction awareness and Time awareness prompt the
agent to form interactions with other systems in the environment and procure knowledge related
to past and potential future events, respectively. Goal awareness preserves information about
objectives and system constraints. Lastly, a meta-self-aware system maintains knowledge about
19
its awareness. My opinion on why this framework is insucient is its inability to serve as a
physiologically realistic model according to the brain's neural underpinnings. Additionally, there
should be a convergence towards a central idea and unied perspective on the manner, which
appears to be lacking. Consequently, this presents a dilemma that may prevent some aspects of
the eld from advancing forward in the proper direction.
2.5 Implications of Physiological Subjective Experience to
the Emergence of Machine Subjective Experience
In Berry and Parker [7], we gave a succinct primer of consciousness and self-awareness(SA). We
issued a proposition as to how brain augmentation can in
uence the arrival of machine agency and
self-awareness. Overall, we stated our opinion for (i) why self-awareness must be systematically
examined in conjunction with brain augmentation approaches and (ii) how such a merger could
become a tool for investigating subjective experiences, namely consciousness. This section will
review related works that reinforce our proposal for physiological subjective experience to machine
subjective experience.
The successes of the articial retina and cochlea have lent encouragement to researchers in the
general eld of brain augmentation [61, 36]. However, in order for brain augmentation to progress
beyond conventional sensory substitution to comprehensive augmentation of the human brain,
we believe a better understanding of self-awareness and consciousness must be obtained, even
if the \hard" problem of consciousness [26] remains elusive. Here we propose that forthcoming
brain augmentation studies should insistently include investigations of its potential eects on self-
awareness and consciousness. As a rst step, it is imperative for comprehensive augmentation to
include interfacing with the biological brain in a manner that either distinguishes self (biological
brain) from other (augmentation circuitry) or incorporates both biological and electronic aspects
20
into an integrated understanding of the meaning of self. This distinction poses not only psycho-
logical and physiological issues regarding the discrepancy of self and other. However, it raises
ethical and philosophical issues when the brain augmentation is capable of introducing thoughts,
emotions, memories, and beliefs in such an integrated fashion that the wearer of such technology
cannot distinguish his biological thoughts from thoughts introduced by the brain augmentation.
A consideration of self begins with the conventional mirror self-recognition test (MSR) [60] that
has been successfully executed with Eurasian magpies [142], bottlenose dolphins [145], orca whales
[40], human infants typically between 18 and 24 months [2, 151], and notably the Asian elephant
(Plotnik et al., 2006). The only primate species reported to pass the Gallup Mirror Test, albeit
controversially, were orangutans and chimpanzees [168]. For years, MSR has been the designated
litmus test for determining whether a species possesses self-awareness (SA), ultimately raising the
question of whether the animal is then a conscious entity due to passing this test [39]. \Mirror
self-recognition is an indicator of self-awareness," proclaims Gallup et al. [59]. If indeed so, then
the subsequent query to raise is whether self-awareness, the ability to dierentiate oneself among
others, is a precursor to or derivative of consciousness and whether the mirror test is necessary
and sucient [126].
In light of brain research like the Blue Brain Project[78], BRAIN Initiative [89], and the devel-
opment of neural prosthetics, the interest in consciousness is steadily growing. Here, we not only
encourage the study of and suggest methods for addressing science's \elephant in the room," which
asserts consciousness is neither physical nor functional, but also place the Elephas maximus in our
proverbial mirror to obtain a perspective toward forming a cohesive alliance between philosophical
studies of consciousness and neural engineering's augmentative innovations. As MSR is purposed
to grant the animal subject personal physical inspection from an objective viewpoint, resulting
in self-cognizance, so shall we take the approach to examine our modern scientic methods in
21
conceptual mirrors, to appraise our consciousness dilemma and propose an assertion for progres-
sion in augmentative technologies. Following here is a succinct primer of consciousness and SA.
We also issue a proposition as to how brain augmentation can in
uence the arrival of machine
consciousness. Overall, we state our opinion for (1) why SA must be systematically examined in
conjunction with brain augmentation approaches and (2) how such a merger could become a tool
for investigating consciousness.
2.5.1 Ineable Consciousness
The rst pitfall encountered with consciousness is the inability to derive a functional explanation
for what it means to experience. Chalmers [26] lists the \easy" problems of consciousness as \the
ability to discriminate, categorize, and react to environmental stimuli; the integration of informa-
tion by a cognitive system; the reportability of mental states; the ability of a system to access its
internal states; the focus of attention; the deliberate control of behavior; the dierence between
wakefulness and sleep." These phenomena are relatively feasible to exploit and can be described
in computational model terms and neural operation derivations. Chalmers then counteracts them
with the \hard" problem of lacking competency to explain why and how we have phenomenal
experiences when being entertained by a movie, exhibiting a sensation toward classical music,
or having feelings when watching a sunset. Explaining how the brain processes visual and audi-
tory signals is trivial compared to how those signals translate to qualia, subjective phenomenal
experiences.
2.5.2 Explanatory Gap Dilemma
The term explanatory gap, coined by philosopher Joseph Levine [107], notes our inability to
connect physiological functions with psychological experience, thus creating the gap. Although
Levine synonymizes consciousness with subjective feelings, the explanatory gap also alludes to
reasoning, desires, memory, perception, beliefs, emotions, intentions, and human behavior/action.
22
Correlating physical brain substrates to thoughts and feelings is the base of a dispute between two
parties: materialist reductionists and non-reductionists [156]. Materialists' prevailing view, repre-
sentative of most neuroengineers, on the matter involves the belief that \when the brain shuts o,
the mind shuts o," and the brain is the sole causative driver for consciousness. However, non-
reductionists (typically philosophers) embrace a holism approach of mandating that the brain's
cortical components are insucient in capturing consciousness, undertaking the possibility of su-
pernatural properties. It is an inquiry of necessity and suciency. The brain may be necessary for
mental functions, but is it sucient? Earlier analytical inspections on conscious experience have
implied that an exclusive reductive justication is not satisfactory in delineating its emergence
[29, 93, 30, 48]. A novel approach is needed to explain such experience. Our explanatory gap
needs an explanatory bridge.
2.5.3 Unraveling Self-Awareness Toward Augmentation
Although many facets of consciousness are dicult to investigate, the development of objective
tests for SA could be utilized for brain augmented technologies. With SA comes the sense of
agency. Agency imparts a sense of who is the owner of an action/trait, the self, and who represents
any entities excluding self, the other(s). Self-other dichotomy processing in the brain is essential
to consciousness due to the necessary implications the embodiment of \self" must have to form
body ownership. Once an agent gains the ability to discern when its own body is the source of
sensory perceptions, it will form body awareness that entails proprioceptive information. We can
look to working experiments that attempt to showcase how the brain augments the \self" when
necessary to complete a task (Fig. 2.2). Perceptual parametric information builds a premeditated
awareness of (1) body part locations and (2) the manipulation of those same parts in space. Body
awareness was demonstrated by a machine via Gold and Scassellati [64] who built a robot named
23
Nico that successfully distinguished its own \self" from \other." Nico observably achieved self-
recognition by completing mirror-aided tasks expending inverse kinematics. Nevertheless, it is
believed that Nico lacked consciousness.
Figure 2.2: Extension of self-representation. Here are two depictions of macaque monkeys
that exhibit a body using tools as an extension of the \self." If given a task to retrieve an object
(yellow hexagonal shape) that is outside the peripersonal space and the immediate reach of an
extended limb (left macaque), the body relies on its physical limitations to dene the \self" and its
aptitude for the success of the task. However, when an apparatus is introduced (right macaque)
that can help achieve the task's goal, the brain's neural correlates can augment themselves to
psychophysically merge tools that were formerly considered to be of \other" classication into the
\self" body schematic and permit optimal behavioral actions to take place [80, 22]. The paradigm
for \self" is malleable to accept the dynamic interplay necessary to achieve an aim for a biological
function that was once previously unattainable. As tool-use changes the brain's representations
of the body and alters proprioception, we subsequently believe it parallels how enriched brain
augmentation can alter an individual's self-awareness and consciousness.
Before the sense of agency becomes fully rened through experiences over time, there must be a
repertoire built for perceptions and actions. Whether action and perception are interdependent or
each fundamentally isolated has been the focus of another ongoing debate. It's not yet concretely
understood how the representation of self forms during the initial stages of life. Either an agent
rst uses perception to motivate their actions in the world or directs their actions to help drive
perception of the sensory world, or both occur simultaneously. In either method, bodily awareness
24
is eventually acquired, which contributes to dening subjective cognitive attributes. Two oppos-
ing views attempt to solve this problem: the action-oriented theory of visual perception, which
suggests that perception results from sensorimotor dynamics in an acting observer [63, 131, 115],
and the dual-visual systems hypothesis, which advocates independent streams of perception and
action [158, 67, 84, 123]. Self-awareness uses the expectation of impending perceptions and ac-
tions to gauge the assimilation of inner experience and external reality. Building a self-aware
framework in augmentative technologies requires integrating an expectancy intuition, which can
critique based on dierences between reality and internal experience. This is our tactic for creating
systems with faculties for using perception and action to make predictions of self-sensory states,
become self-adaptable to new environmental stimuli, and set objectives for self-improvement.
Crucial for understanding agency is determining how the embodied senses fuse to form self-
referential experience [50, 51]. It is our opinion that future advances of brain augmentation
hinge on the application of such knowledge. Once we bridge this gap of the unknown, we will
be challenged to use computational intelligence to create consciousness articially and integrate
synthetic qualia with that produced in the brain. Presently, articial devices can create various
aspects of consciousness. Articial perception is made available via cochlear, retinal, and tactile
implants. However, they work alone as replacements for sensory organs with consciousness and
SA arriving later in the brain's neural processing. Applications for augmenting consciousness
would contribute to studies relating to emotions, attention, supplementing memory capacity,
personality alteration, experience enrichment, sensory perception enhancement, and hypernormal
brain plasticity for self-repair.
25
2.5.4 Proposed Transition to Machine Consciousness
The marvel of human intelligence is its ability to eclipse physical limitations and overcome our
biological constraints to form an ever-evolving existence [87]. One primary goal for reverse-
engineering the human brain is to recreate the same functional mechanisms that underlie human
consciousness in our software infrastructures, neurorobotic agents, and computational systems.
However, prosthetic memory, sensory implants, neurofeedback (EEG Biofeedback), and brain
computer interfaces (BCIs) are all working examples of fusing such \intelligent" systems with the
brain, leading to conceivable prospects for consciousness-altering devices. Although BCIs com-
monly target disability treatments and brain function recovery from a lesion, the amalgamation of
computational devices with the cortical brain itself [52] may even prompt increasing developments
of an operational \exobrain" [12] for the purposes of better understanding how our brain works.
For example, in a scenario where a split-brain condition is present within a subject, we now have
the option to look toward interfacing articial exobrains with the cerebrum; such an interface can
either serve as a replacement for neurological issues or supplement features the brain does not
naturally comprise. If these exobrains have a modicum of manipulability, then we can explore the
plausibility of mind transfer from device to organ and vice versa; thus, providing speculation for a
conscious machine that can aect how we can perceive, act, express emotion, feel, and adapt. This
poses ethical concerns as it opens the door for alterations of an individual's SA when augmen-
tation can modify reasoning skills and subjective judgment. Successful augmentation of the sort
might render the individual powerless in discriminating actual characteristics and thoughts from
those that are mock and introduced articially outside the cortex. Combining the precision and
information processing speed of a computer with the intrinsic non-computational attributes of a
human may provoke discoveries of the mind (e.g., consciousness) that we as humans are currently
incapable of resolving. We suggest eorts made toward an augmentative interface between brain
26
and machine that prompts the human mind to think beyond its unknown limits for constructing
our explanatory bridge.
2.5.5 Challenges Moving Forward
Many people view an in-depth exploration into consciousness and its emergence as a gamble,
considering decades already spent on the matter with a void of consensus [41, 88, 71, 164, 33, 174,
38, 160]. Before we attempt to create another hypothesis, our approach needs to change; it is our
suggestion to further rene the constructs and emergence of SA and to use brain augmentation as
an instrument for inspection. We need to dene an objective test for determining whether an entity
is a sentient being. This test, in addition to advances in neural engineering, provides optimism
that disputes within the consciousness eld can be resolved. Augmentation has a promising future
as an enhancement to our brains and will hopefully in
uence our centuries-old methods of thinking
about consciousness toward an answer for science's greatest mystery.
2.6 Models and Applications of Body Representations in
Robotics: A Review
Now we will discuss previous attempts made to build body schemas and the applications they
were designed for. Beginning with Lewis et al.'s [109] Reference Architecture framework for
computing systems, we observed a common approach taken in the development of self-modeling
systems; that is, forming an engineered architecture directly from psychology. Without taking
into account the neuroscience, this may permit some dilemmas in accurately encompassing all of
the facets related to the formation of body schemas. Their attempt to create a model in such
a way was purposed to bring explicit structure to the design of self-aware systems in general,
paralleling Neisser's [129] levels of self-appraisal. It categorizes various levels of self-awareness
capabilities as system benchmarks (e.g., stimulus awareness, time awareness, goal awareness).
27
The agents' and host systems' tasks and goals determine the benchmark complexity chosen for
implementation. However, their template will prove inadequate in one fundamental area: self-
modeling to incorporate new features of the system. In support of the model, it does accentuate
self-awareness as an ongoing going process of online learning, which we agree with to some extent.
It will classify a version of this process as emergence. Lewis et al. [109] also implies that action
selection directly aects the agent's ability to learn. We would like to further advance this notion
by suggesting that decision making, then subsequently action selection, aects the learning and
mapping of the manifold space within the self-model. The Reference Architecture additionally
assigns the same goals and methodology for public and private self-awareness. Although the two
domains are not mutually disjoint, both physiology and psychology sciences identied them as
having dierent trade-os, especially considering adaptation. Although Lewis et al.'s [109] self-
aware framework proved to be eective in one case-study for a service-selection cloud computing
platform, it does not provide the means for the system to learn and adapt at runtime to changing
conditions. It was not explicitly stated that their implementation involved an emergent model of
the self.
Through mirror perspective-taking, a makeshift humanoid robot named Nico was assembled to
demonstrate self-awareness as an emergent property [75] through the goal of developing an ar-
chitecture that permitted the robot to pass the classical Mirror Test [60]. The architecture is
composed of six sub-models describing various levels of self-knowledge that could be obtained
from the robot to complete the task: end-eector model, perceptual model, perspective-taking
model, structural model, appearance model, and functional model. Many have supported this
body of work with armative claims that this exhibited the rst \self-aware" robot to pass the
mirror test. However, opposing views countered those claims with the argument that the system
was lacking introspection. Instead of being self-aware, Nico was classied as the rst machine
with the ability to reference the location of its body part in three-dimensional space by using a
28
re
ection. Our perception of this model aligns with the latter opposing claims that this model
mainly demonstrated visual recognition. The robot was instructed to maintain three designated
arm postures to achieve this recognition, each having 50 training sets and 100 sample tests of
dierent positions in space. The self-knowledge procured through the training sets is then used
to make predictions about the body's whereabouts via calibrated kinematics and a stereo vision
system. Results indicated that the robot successfully developed a model of its arm based on its
visual point of view. Self-observation, rather than self-awareness, as we noted earlier, appears to
be the running theme here. It is evident Nico observed, but it did not obtain awareness of its
experiences.
Self-aware frameworks have also been implemented in collective host systems like autonomous
multi-camera networks to coordinate object tracking [147]. What interests us about this frame-
work is the attention given to topological learning for resource adaptation among the cameras,
which we feel is necessary and will incorporate in our implementation of the simulated neuromus-
cular system. Continuous topology monitoring will create an enduring self-model and not solely
create a temporary model representation from online learning. Bongard and Lipson [14] further
discuss the concepts of self-modeling in robots as self-re
ection is stated as a vital aspect for
robustness when encountering unexpected changes in the body.
29
Chapter 3
Self-Recognition: Extension of Mirror Neuron System II
for Agency
\. . . the human brain is peppered with mirror neurons and they activate in us exactly
what we see in the other person: Their emotions, their movements, and even their
intentions." -Daniel Goleman
3.1 Introduction to Mirror Self-Recognition
We conducted a study to identify and simulate the brain's minimal neural correlates for achiev-
ing proprieties of self-recognition and agency, a trait once deemed be a unique characteristic of
only humans but has been disproven [3, 81]. In modeling the classical Mirror Test performed on
Asian Elephants, we sought to provide a solution to an ongoing inquiry. In addition to the mir-
ror neuron system (known for responding to performed actions of the self and observed actions
of the other), which neural patterns are responsible for making an agent aware of its physical
characteristics and behaviors? We proposed a framework, based on article entries in the Brain
Operating Database System (formerly located at http://bodb.usc.edu), which linked systems neu-
roscience data to testable models and designs for the generalization of high-level concepts required
30
for making a self-other distinction. The framework was applied in a simulation involving three-
dimensional shapes, representing gurative anatomical bodies. Each was categorized as either the
Agent or Adversary. Over a pre-dened set of runtime iterations of the simulation, the Agent
trains itself with an adaptive network to optimize its ability to dierentiate between the physical
characteristics preserved by itself and those of others. This framework's execution served as a
working example of identifying and executing the minimal components required for an agent to
suciently reach one of the early levels of self-recognition through social interactions. Our rst
step towards the dicult goal of creating computational self-awareness by way of self-recognition
was a software simulation of self versus other based on the \gold standard" test for self-awareness
in animals and human infants, the Gallup Mirror Test [60]. We sought to successfully implement
a system that can pass the Mirror Test from a biological perspective (targeting and modeling
specic neurons in the brain). The mirror test assesses an animal's ability to discern its social
and behavioral responses based on its re
ection in a mirror. Only a select group of non-human
species were reported to pass the mirror test and achieve this particular level of self-awareness:
orangutans [168], gorillas, dolphins [145], elephants [141], orcas [40], macaques, Eurasian magpies
[142], and bonobos. Human infants are unable to pass the test until brain nerves, and supportive
tissues develop at an average age of 18 months [9, 2]. Examining this phenomenon in further detail
has assisted greatly in forming the basis for this research, which extends towards forming body
representations, self-modeling, and self-awareness. DARPA's recommendation [1] for designers of
self-aware systems is to contemplate the architecture of a self-aware computer system from three
distinct perspectives: (1) an autonomous agent view, (2) an information processing view, and
(3) a biological view. A combination of all three perspectives likely serves as the most benecial
instead of only looking into one perspective alone.
The framework was designed to be extensible according to each new brain component linked to
self-awareness. One of the initial questions proposed at the beginning of this research was whether
31
the mirror neuron system alone was sucient to solve self-recognition? Considering that the brain
is highly labyrinthine and dense with neural networks, the simple answer is no. If not, then we were
left to learn the additional elements that play a role in allowing the brain-mind-body complex
to achieve self-awareness. Approaches taken to discover the neural circuitry associated with
awareness/recognition connections are not scarce. One is the neural correlations of consciousness
(NCC), which are dened as the minimal neuronal mechanisms cooperatively adequate for any
specic conscious percept [33]. In this study, we reviewed recent works relating to NCC and
applied some of the computational mechanisms found to be aspects of aordance extractions for
the recognition of objects. Another supplementary method used is to examine disease conditions in
which self-recognition and self-awareness are degraded or perturbed. Autism [187], schizophrenia
[155], and psychopathy [150] can serve as for disease models in assisting with characterizing the
NCC paradigm. The mirror neuron system (MNS) [148] operates as the central groundwork for
our blueprint. Basic self-recognition tasks and mirror self-recognition are jointly the core of the
self-awareness complex. Our proposal that the MNS is central to our architecture is founded on
the notion that the MNS has been highly functional in self-other distinction recognition. The
MNS has been primarily observed in experimental studies involving macaque monkeys as subjects
performing various visuomotor tasks. Within the MNS are mirror neurons in the premotor area F5
that re when a monkey performs a set of actions and observes another monkey performing that
same action, if not very similar [149, 58, 13]. To understand the workings of the MNS, we must
note the internal models that connect motor control to cognitive perception. Similar to feedback
loops and standard robotic motor control, the framework patterned two forms of internal models
that play a role in recognizing actions and adjusting the motor system accordingly: inverse and
forward [119]. Inverse models were responsible for activity during the observation of actions. A
mapping of the intended action and motor commands that encode an action is created, while
forward models are responsible for activity during the execution of imitated actions.
32
A. B.
Figure 3.1: (A) Inverse Model: Cortical activity routes during observation of actions. The circuit
linking STS, PF, and F5 (solid arrows) acts as an inverse model. The cerebellum has this function
(dashed arrows). (B) Forward Model: Cortical activity routes during the execution of imitated
actions. The circuit linking F5, PF, and STS (solid arrows) acts as a forward model to generate a
prediction of movement outcome. Alternate routing is made available with the cerebellum (dashed
arrows). Red Arrows: Prediction error coding through empirical Bayesian inference [92]. This
gure was adapted and enhanced from Miall's concept for linking mirroring and modeling [119].
When a subject (e.g., monkey) performs an action such as throwing a baseball or views another
subject perform that same action, neural processes are reported to take place in the posterior
parietal cortex (PPC). Fig. 3.1 displays the inverse and forward models incorporating the F5
mirror neurons in their signaling pathways for information transfer. Mirror neurons in the PF
area are shown to code for somatosensory components of the observed action [23]. Mirror neurons
in the superior temporal sulcus (STS) are tasked with signaling for the visual response of biological
motion, body parts (e.g., appendages and faces), and for grasp movements [137]. STS is also noted
for perspective-taking. The primary motor cortex (MI) in Area IV serves as a control for voluntary
movements. In Fig. 3.1A, the pathway for ring is directed from the STS to PF, and then F5.
What was visually seen or conceived in the mind is recorded in the STS, then processed through
PF for matching against instantiated goals. A signal is subsequently directed to F5 to induce
ring if a match is made for the observed action or the subject performing the intended action.
Fig. 3.1 B of the forward model contains links that are in reverse of the inverse model. The motor
plan from F5 is converted back to STS, where a sensory action is demonstrated, likely visual.
33
Prediction error coding is used to rene motor movement to become more accurate through time
and experience.
In both models, there are additional connections made to the cerebellum (CB). It was discovered
to be an alternate route in the models, yet it performs essentially the same processing functions
[120, 189], suggesting diused connectivity among the cortical regions. For our simulation imple-
mentation of this particular brain mechanism, there will likely be a representation of the internal
models. Suppose either of the general inverse or forward pathways are obstructed by way of a
lesion, for example. In that case, the cerebellum should be able to take over and continue cus-
tomary overall processing. It has been proposed that self-awareness depends on specic brain
regions: the insular cortex, the anterior cingulate cortex (ACC), and the medial prefrontal cortex
(mPFC) [138]. The insula is presumed as the necessary substrate for nerve impulse awareness
[31, 32]. Emotional awareness, facial self-recognition, and the overall conscious experience have
been linked to the ACC [35, 94]. Self-referential, self-re
ective thought processing, and the projec-
tion of future self is associated with the mPFC [132, 167, 157]. On the contrary, many neurologists
believe self-awareness is a product of a disseminated assortment of networks in the brain. In a
study conducted on a human patient that suered from herpes simplex encephalitis [138], it was
revealed that the insular cortex, ACC, and mPFC are not a requirement for most properties of
self-awareness. The patient, given the name R, had brain damage extending the basal forebrain,
anterior inferior parietal lobe, medial temporal lobe, amygdala, and hippocampus. Brain damage
was not found in the hypothalamus, thalamus, basal ganglia, and occipital and parietal lobes.
After being probed with extensive tests, results concluded that Patient R maintained a sense of
self-agency, self-recognition, and judgment. Experimentally justied, self-awareness is likely to
transpire from the brain's distributed networks, including the thalamus, cerebellum, and brain
stem. This justication will make a neural inspired self-aware framework rather dicult. Addi-
tionally, there are other components to consider in our initial framework. BA5 cells of the PPC
34
should be implemented for the purpose of coding for non-dynamic kinematics. Also to be included
is the anterior cingulate, referred to as sensory pain neurons, because it signals when a subject
receives a touch stimulus such as being poked with a needle.
After identifying some of the key cortical regions, we looked towards encoding rst-person per-
spective and third-person perspective in the brain. One aspect of self-awareness is the Theory of
Mind (TOM), a principle used to recognize, predict, and justify the actions of both the self and
other subjects. When considering the self, a rst-person perspective is used and a third-person
perspective is used for the other. An external study involving subjects under PET investigations
revealed the brain regions activated when a subject imagines a frame of mind from either per-
spective [153]. The right inferior parietal, precuneus, and somatosensory cortex are involved in
distinguishing perspectives. More specically, a rst-person perspective will show a strong left-
hemispheric regional cerebral blood
ow(rCBF) increases in the inferior, precentral gyrus, superior
frontal gyrus (SMA proper), the occipitotemporal junction (MT/V5), and anterior insula. The
cerebellum and precuneus were activated in the right hemisphere. The third-person perspective
shows bilateral rCBF increases in the precuneus. On the left side, activation was detected in the
precentral gyrus, superior frontal gyrus (pre-SMA), and occipitotemporal junction (MT/V5). On
the right side, the inferior parietal lobule and frontomarginal gyrus were both activated [153].
The diusion of areas listed here is yet another testament to the disparity of the brain activity to
region ratio.
Now that we have established a few (relative to the scope of the brain as a whole) physiological
components that have been found to lead to self-recognition possibly, their functions and connec-
tions should be placed in a black-box model detailing how and where neural communication links
are made. Our framework from the conceptual high-level will ensure that the general concepts for
attaining self-recognition are met. Fig. 3.2 shows the dated platform for what was required for
self-recognition. Its objectives are to 1) recognize and generate the desired action from training,
35
Figure 3.2: Bidirectional circuit map for coordinating visual body image and motor body im-
age. Three phases were implemented: Learning, Observation, and Autonomous Action/Behavior
Generation.
and 2) successfully coordinate the visual body image and motor body image. The themes to
keep in mind when constructing this platform are cognition - the way the body is conceptual-
ized, visual - the way bodies are sensed and perceived, and motor - the actual control and active
sensing of bodily movement. The model consists of three coalescing phases of the main process
occurring when a subject's brain attempts to match what is perceived through sensory inputs
against internal representations. Black arrows indicate the information
ow transfer through
this bidirectional circuit map. Our map begins at the Learning-Phase (L-P) at the cerebellum.
The cerebellum receives information about the positions in the space of the joints and the body
from proprioceptors. Proprioceptive cues are sensory elements indicating factors about the body.
Such cues for a robotic agent would include acceleration sensors, temperature sensing, gyroscope,
touch sensor, etc. In this phase, the Agent can relate between the visual-body-image that it wants
to see achieved and the motor-body-image, as previously demonstrated by Steels and Spranger
36
[165]. Following is the Observation-Phase (O-P), where the self-other distinction is made, mainly
in the right inferior parietal lobe, precuneus, and somatosensory areas. The concluding stage is
Autonomous Action/Behavior Generation-Phase (AABG-P). This phase is not nite as the map
shows a feedback route for adaption by updating the system with more precise information in
reverse. AABG-P provides feedback information to previous phases on whether the success of the
action occurred or not. If goals involving motor plans can be accurately attained in the case of the
mirror test at AABG-P, then the Agent can distinguish itself in a re
ection. An additional list of
information collected to assist with our task is presented in Table 3.1. Brain Operating Principles
(BOPs), Summaries of Empirical Data (SEDs), and Summaries of Simulation Results (SSRs) are
metrics for ensuring that our architecture is complete in including the brain's conceptual activities
at the neuron level.
3.1.1 Agent Self-Recognition Test
Many tests can be used to gauge self-awareness. Two of those tests include tests of self-recognition
and self-agency. Self-recognition tests seek to gauge whether the Agent can interact with its
intermediate environment based on the visual scene of the mirror. An agent should have the ability
to use mirrors for spatial reasoning [75]. Our simulated Agent should also readily recognize itself
as a separate entity passing the classical Mirror Test. In self-agency, agents must dierentiate
between themselves and others that it resembles to reach the target goal location or successfully
single itself out among like members. Agency also needs to entail body ownership. There must be a
successful match between self, intended action, and perceived action. Here, the brain theoretically
creates a representation of itself while incorporating a conscious awareness of intentionality. How
the subject attributes an agent to himself or another agent will determine the level of self-agency
engaged. For the software implementation, we have used a fundamental version of the self-
recognition test.
37
Table 3.1: Brain Operating Principles
3.1.2 Experimental Design
A testing environment was created to showcase the eectiveness of our proposed framework in
replicating the mirror test. An Agent vs. Adversary approach was taken. The hypothetical sub-
jects created were ve three-dimensional shapes: square/cube, rectangle, sphere/circle, triangle,
and cylinder. The color of the Agent was always yellow, for simulation consistency and control.
The Agent's shape is selected by the user at initial runtime, leaving the total number of possible
Agents to be n=5, each having their characteristics and aordances. On the other hand, the
38
Figure 3.3: Agent versus Adversary. Top row: Five 3-D shapes were used as either the
Agent or Adversary, with the color of the Adversary bring randomly chosen among the 12 HSV
color hues. The Agent's color is always yellow. Shapes included were square, rectangle, sphere,
triangle, and cylinder. Bottom row: Activity Field. Square Agent { Purple Triangle Adversary
b) Square Agent { Blue Triangle Adversary c) X-axis view d) Y-axis view e) Z-axis view
Adversary's shape is randomly selected along with a random selection of 12 color hues. Fig. 3.3
gives a depiction of how the shapes are resembled in simulation and the colors chosen from the
360-degree HSV (hue, saturation, value) color wheel. The number of possibilities for the Adver-
sary subject is 60 (12x5). When selecting the Adversary for the Agent to match against, the
code is given a 30% chance to auto-generate a subject that resembles an exact representation of
the Agent in both aesthetics and movement. The remaining 70% chance is the likelihood of the
Adversary being one of the 60 possibilities. As shown in Figs. 3.3A-E, the environment for our
subjects is referred to as the Activity Field, featured with a three-dimensional axis space with
39
standard units on the x-y-z plane and a gray-colored margin that represents either a Mirror or a
Window depending on how an Agent perceived its Adversary. The Window will signify that the
Adversary's features and movement do not match from the Agent's perspective. Mirror signies
that the Agent visualizes an equivalence of its features and movements. The Agent is inevitably
tasked with discriminating the perception eld as Mirror or Window. Since the Agent's shapes
are code-generated and do not have an actual visual perception as a human would have eyes,
each Adversary shape's body parameters were created and passed to the Agent's domain as it
would perceive it through a similar method to that of the Pinhole Camera Model. Several kinds
of information are processed as an image for the Agent: 1) Geometric for axis positions, points,
lines, and curves, 2) Photometric for color intensity, and 3) Movement direction and speed.
3.1.3 Results
Self-awareness can make a computational system more robust and self-repairing over an extended
time period. This is, in retrospect, the expected hypothesis for our Agent behavior when prompted
to test for self-recognition. Just as an infant's and toddler's brain continuously evolves by going
through a process of pruning synapses for optimization, so should our system. As a result,
the simulation required a sort of adaptive network for learning. Hebbian learning and Perceptron
learning are common schemes for strengthening and weakening synapses for accurate neural ring.
Better suited methods are the incorporation of Simulated Annealing and Hill Climbing techniques.
These techniques facilitate our system to progressively meet their objectives in uncertain and
dynamic environments. To determine which algorithm to use for our self-recognition test, the
pros and cons were rst assessed. Hill-climbing search only looks one step ahead at a time to
decide if any successor is better than the current state the Agent is in. The disadvantage of
hill-climbing is its inability to allow backtracking since it does not have the capacity to recall the
previous state it has been in, due to its single state memory. Therefore, there was an issue with the
state getting stuck in local minima, and the system not fully optimizing. Simulated annealing had
40
slightly more momentum as it can escape the local minima. If the system was provided enough
runtime, then a complete and fully optimal solution can be given. With simulated annealing, the
system gradually degrades its repertoire of \bad" actions through a random search that accepts
adaptions to increase objective realization. Thus, making it the preferred algorithm. Three
main scripts were programmed to emulate our framework: Learning, Observation, and Action.
At the start of simulation runtime, the program is given a specic number of iterations as the
maximum opportunities for the Agent to correctly identify whether what it is currently viewing
is a representation of itself or the Adversary. With each iteration, a new Adversary is randomly
chosen for the Agent to examine. And also, with each iteration, the Agent is technically more
\aware" of its self-properties than before. Using the mechanisms mentioned throughout this
report, we should expect a familiar learning trend to an aging human infant when interacting
with a mirror. In the beginning, the Agent virtually knows nothing about itself and is not certain
of what traits and properties make up its being as an entity. With each passing iteration, the
Agent becomes increasingly reinforced of what is the self and can expeditiously recognize it.
The range for self-awareness to be achieved in a child is 15-24 months. Scaling this range down
to our IDE simulation time in seconds shows mimicry of brain development. The average life
span of a human is approximately 82.5 years, which is 990 months. Taking the quotient of the
months of initial awareness and the lifespan gives us a lower and upper bound value range for
our system to approximate at what moment in time we should expect it to become fully aware.
The lower bound is 1.5% and the upper bound is 2.4% of the system time span. For example, if
our system runs for 100 seconds, we should predict that self-recognition is wholly demonstrated
within 1.5 and 2.4 seconds from program start time, depending on the simulation processing rate.
Several executions of the simulation were conducted to observe the eect of trial runs on the
length of time full self-recognition occurs. As the Agent learns, observes, and acts in response
to the Adversary, the accuracy of the Agent's sense of self improves over time. Prediction with
41
regards to movement, rst-person perspective, third-person perspective, color, and proprioceptive
cues are the main aspects the Agent must closely identify to dierentiate the self from other. As
mentioned earlier, the method of approach uses simulated annealing. The next iteration of this
simulation test will use more conventional synaptic plasticity. With each trial run iteration, the
accuracy of the Agent's perception of self is fed to the annealing script, which starts with an
initial `temperature' of 100 and maximization factor {x
2
. Assuming that our Agent begins the
simulation without any correct information about itself, our rst iteration value would be -10000
(y-axis), as shown in Fig. 3.4. The stopping condition is at the threshold value level of 0, which
the system attempts to achieve. Here we examine cases where there are 100, 200, 400, and 800
training sets. In Fig. 3.4, the following learning curve plot symbols and their representations are
used:
blue dot: value is accepted because the new solution is better than the current solution
green circle: the new solution is not only better than the current solution, but also better
than the best overall curve solution
cyan dot: the new solution is somewhat worse than the current solution, but is accepted to
keep in temporary memory for checking later
red cross: the new solution is rejected because it is much worse than the current solution.
The trends we observed within our plots are 1) the closer a 1:1 ratio of green circles to blue dots
as seen over successive trials indicate earlier stages of self-recognition and 2) as the iterations
continue, the fewer green plots appearing on the curve show the system getting closer towards the
threshold value with less chance of nding another overall best solution. Noticeably, what is seen
in the latter parts of the plots, where an elongated plateau is observed, are more concentrations
of all other symbols besides the green marker. In Fig. 3.4a, the 100 training sets give us a best
42
Figure 3.4: Sampled training set results for trial iterations. Results simply indicated longer trial
durations produced a greater quantity of successful recognition solutions.
value of -3026.45, far from value 0. Fig. 3.4b's training set of 200 has the value -481.69. Fig.
3.4c with 400 training sets has a best value of -0.000018. And Fig. 3.4d, shows 800 training
sets has a best value of -0.000002. As expected, we concluded that the longer the system can
perform with trial runs to improve the accuracy of the Agent's self-other distinction, the more
likely it will approximately reach the threshold level, signaling self-recognition. In the future, we
seek to update this model with features that address the limitations (i.e., number of trials does
not describe the eects of rst exposure to a mirror) of the latest version. We are interested in
viewing the system's performance if it contained prior experience with theoretical conspecics.
If possible, we would like to compile available data sets on imitation by elephants and compare
them to the results of the updated version of this system that uses more neural computing over
machine learning principles.
43
3.2 Extension of Mirror Neuron System, 2 (MNS2) for
Agency
Building on the FARS (Fagg{Arbib{Rizzolatti{Sakata) (Fagg and Arbib 1998) and MNS2 [13]
models, we sought to incorporate additional neurophysiological and anatomical data that high-
lights a compilation of vital anatomical regions that are necessary to the mirror neuron circuit's
ability to predict and interpret actions that are both performed (Action) and withheld (Inaction).
The goal was to construct a neurobiological simulation of the operant conditioning method to train
our computational system for the Inaction and Action task in Bonini et al. [16, 17]. We address
the integration of reinforcement learning and temporal dierence learning in achieving results
showing the anticipation of an agent's action based on a sensory cue from both the environment
and intrinsic expectations. Interactions between the primate's (found in Macaca nemestrina and
Macaca mulatta) cortical and subcortical regions have been identied and simulated to achieve
the desired visuomotor sequences. Part I of this study particularly focused on how the system
learns to perform the task. This phase develops the desired synaptic weights and encoding for
each neural population identied in the Bonini experiments. Several examinations were conducted
on macaque mirror and motor neurons, while the primate was instructed to perform visuomotor
tasks of grasping an assortment of objects according to some external incentives. Bonini et al.
[16, 17] presented scenarios that have allowed us to take an introspective look into multiple cog-
nitive themes relating to action recognition, motor mapping for grasping, and prediction. Bonini
et al. [17] show explicitly that while the majority of macaque ventral premotor neurons are silent
(no discharge) when the monkey refrains from grasping an object, there are also other neurons
that re both when the monkey performs action and inaction tasks with an object. A movement
is classied as an action when any muscle re
exes have followed the cue stimulus to form the grasp
and an inaction when no observable movement in the monkey's hand or arm. The experiment
paradigm in Bonini et al. [17] can be explicitly depicted in Fig. 3.5 (Left). Depending on the
44
experiment's initial setup and the xation point's placement, the monkey can infer whether the
task context is an Execution or Observation. An action or inaction condition is randomly chosen,
which is signaled by the audio tone's frequency. High and low-frequency tones of 1200 MHz and
300 MHz, respectively, were the only audio options used in the experiment. Most of the trials
were completed with some form of light. The purpose of the Dark condition in the experiment
was to ensure that all the recorded motor responses were present in the dark as well, concluding
that they cannot be simply caused by the monkey's hand visual feedback.
Figure 3.5: Left: Action-Inaction Paradigm. Experimental structure of the Action-Inaction task.
Two task contexts were selectively chosen at the beginning of each trial. Within each context, a
random selection of the action and inaction condition is made. An execution task context. Right:
Epochs of interest. (*) indicates when the monkey becomes aware of 1) whether it will be acting
or not and 2) whether to voluntary grasp/not grasp.
The experiment presented several epochs of interest that assisted our eorts in formulating a
sound model: baseline, object presentation, pre-go/no-go cue signal, and lastly, post-go/no-go
cue signal. Fig. 3.5 (right) outlines the sequential steps that occur throughout a trial. After
the cue sounds, the third sequence branches o into two streams of either Action or Inaction.
Although 663 area F5 grasping neurons were recorded, there were advantages in showcasing
neurons individually. Only two types were selected, with a third in the supplementary paper.
Table 3.2 logs these two neurons and in what conditions they were reported to re. Neuron 1
and Neuron 2 were respectively showcased in the paper as a motor neuron, which red when
45
the monkey both grasped and refrained from grasping an object, and a mirror neuron that res
during grasping of execution and observation tasks and when there is refraining from grasping
during observation. In other words, out of the four total conditions that the neurons can operate
in, mirror neurons will not discharge when the Agent itself is performing the inaction. This is
the case where I vouched more individual neuron tests should have been done because it may
be challenging to claim that all mirror neurons do not re in Execution-Inaction based on the
performance of only one neuron. On the other hand, this behavior was seen at the population level
as well. This observation does not appear overwhelmingly convincing that every mirror neuron
here res the exact same way.
All of the total neurons observed by Bonini [17, 16] can be divvied into distinct categories:
All (n=663) neurons discharged during action execution
188 { 28.35% also red during action observation: Mirror Neurons
105 { 15.8% also became active during the inaction condition relative to both baseline and
the object presentation epoch
26 out of 105 were motor (purely) neurons and discharged exclusively during the inac-
tion condition of the execution task
79 out of 105 were mirror neurons. Most (42 out of 79) discharged exclusively during
the observation-inaction task.
Table 3.2: Neurons 1 and 2 red in only certain conditions out of the four designed in the
experiment.
46
3.2.1 Learning with Operant Conditioning
Examining why and how to bring about certain behaviors within the brain must involve using
the experimental psychology approach. To understand the full scope of why we have chosen to
use specic algorithms for this particular model, it would be best rst to explore the paradigm of
operant conditioning and its intended purposes. By now, we should already be familiar with the
action/inaction tasks that the monkey was trained to do. But how exactly it was trained appeared
to be a negated detail in the ocial publications of the study conducted by the University of
Parma's Brain Center for Social and Motor Cognition. This is a mark in our work where we
believe some novelty may result. Usually, in a visuomotor task, electrophysiology data are not
collected (or rather reported) before or while the primate is being trained. This brings up the
question as to how do mirror neurons (MNs) develop over time to code for changes an Agent
needs to make? Do MNs initially have the ability to encode specic action repertoires, or do
they eventually gain this trait over time? This is just one question of many we hope that future
electrophysiology tests will provide answers to. In the meantime, we have attempted to train
our computational system with neural subpopulations using the same method that the actual
primates in the experiment were trained on: operant conditioning (OC).
OC, somewhat dierent from the more widely-known classical conditioning method, resolves to
goal-oriented behavior. In OC, the Agent under observation learns to behave in a way as to obtain
rewards and avoid punishments. It ts the mold for learning by trial and error. OC is also viewed
as a more aggressive method. In contrast, classical conditioning is associated with passiveness
because it generates a behavior within the Agent that naturally emerges due to an unforeseeable
connection between a stimulus and reinforcement. But it can also be argued that some classical
conditioning traits are incorporated in the present experiment because the system is also learning
to predict important events and not just learning the outcome of behaviors. Prediction will be
a key theme throughout this study and the thesis topic. Delving more into the OC method
47
of Applied Behavior Analysis, we have determined that the Inaction/Action task uses positive
reinforcement (presenting a motivating stimulus, e.g., juice reward is given for touch start) and
negative punishment (removal of desired stimulus or something \good," e.g., juice). Since the
terms reinforcement and punishment will be used repeatedly, to distinguish them feasibly, it is
best to remember that reinforcement is a process used to help increase the probability of behavior
and punishment is the process of allowing a consequence to occur after a behavior to decrease the
frequency of that behavior in future trials. We have just established the type of training method
used in the Part I-Learning stage and its associated responses (i.e., operants) instrumental in the
action/inaction task. But there was another element to our training procedure that we considered
and implemented into the model: timing. The timing of reward and punishment also aects
learning in ways that can signicantly alter the Agent's ability to make denitive associations and
the rate of learning some arbitrary task. Our model incorporates a varying sequence schedule to
strengthen neural synaptic weights for associative learning. Schedules of reinforcement have been
shown to have dierent eects on an Agent's behavior. Two primary schedule types were examined
as potential time template structures to utilize. We have evaluated both continuous scheduling
and intermittent scheduling. The rst being continuous schedules, which reinforce every instance
of the desired behavior with a reward. Modeling a human's reaction with continuous scheduling
provides two main benets: 1) emphatic associations will be made between the Agent's desired
behavior and the reinforcements received, and 2) there will be a rapid increase in desired behaviors.
Once a behavior occurs to the desired frequency, then intermittent scheduling can be introduced.
Intermittent schedules, also referred to as partial reinforcement scheduling, further encompass four
other scheduling variations for behavior maintenance based on ratios: the number of responses,
intervals, and time. Fixed ratio (FR) schedules are touted as the most eective mechanism for
teaching new behaviors. It promotes a heightened behavior rate immediately before the expected
reinforcement is received. After receiving the reward, behavior pauses momentarily and will
48
eventually steadily increase as the time nears closer to receiving the reward again. Fixed interval
(FI) schedules consist of high and low rates of behaviors and a brief pause after reinforcements.
Still, it diers from xed ratio schedules in that the rate of behaviors is generally lower because
reinforcement is only given after some delayed quantum of time after the behavior occurs. One
detrimental consequence of xed intervals is that behaviors can eventually cease to exist if there
is a case when reinforcement is no longer present. Variable ratios (VR) is another intermittent
scheduler that is best suited for maintaining newly acquired behaviors. Here, the Agent is not
knowledgeable of the amount of responses that are required to receive reinforcement. As a result,
the Agent will repetitively engage in the behavior until the reward is given. One advantage of
doing such will make the behavior more resistant to obsolescence. And lastly, there are variable-
interval schedules. Similar to variable ratios, variable intervals (VI) also generate steady behavior
rates because the Agent is unaware how much time needs to pass for reinforcement. However,
the dierence is that behavior rates are low due to the structure being based on the amount of
time within the quantum instead of the number of reward responses. Simply put, VI behavior is
reinforced after an average amount of time passes. For example, a VI-15 seconds schedule will
reinforce the rst instance a response is given at the average length of 15 seconds.
We have concluded that out of all the schedule of reinforcements mentioned, one supersedes the
rest as it provides the most accurate template structure for our learning model. Fixed interval
scheduling gives us the capability to set up the reinforcement in a manner that replicates the timing
of reward/punishment delivery. Within FI, responses are reinforced after a xed amount of time.
The characteristic that sets FI apart from the rest is that it refrains the reinforcement from being
automatically obtained after the established quantum. FI only makes the reinforcement available
while the Agent is still required to respond appropriately to receive the reinforcement.
In addition to a temporal structure, we were also set up the experiment to recreate how associations
are built. OC is primarily based on three events occurring within a trial. The events include
49
the stimulus (S), response (R), and outcome (O). There are multiple ways in which they can
be arranged to suit the task constraints. OC produces a faster learning rate if the interval
space is relatively short in the beginning. But as the experiment progresses, the interval length
may increase if there is prior evidence of an association that is successfully learned. According
to the OC steps taken, as made explicit by Luca Bonini via external conversations, we have
compartmentalized the overall OC training into ve tiers to recognize and implement the phases
necessary for training.
Within each tier, the primate is trained to perform certain sub-sequences of the overall task. They
are the following:
Tier 1: Start Position.
{ Monkey randomly touches start position. Reward is given, quickly.
{ Allow monkey to increase holding time before delivering reward. Reward is delayed
Tier 2: Motor Part & Go-Tone
{ Monkey is trained to reach or grasp a target as soon as brief HIGH tone is played.
{ Reward is delayed until monkey grasps/holds object for desired time (.8s).
{ Other objects are introduced. Shaping procedure takes place for monkey to grasp
objects in appropriate way.
Tier 3: Sound Duration (a criterion of duration 0.8-1.2s is achieved)
{ Sound duration is progressively increased, making false start errors. Monkey does not
get rewarded. Monkey learns to stay while sound is played, go when it stops.
50
Tier 4: No-Go Tone
{ LOW tone introduced, Reward delivered as soon as sound stops. Monkey immediately
associates new sound with \reward for not acting"
{ Delay between the end of sound and reward delivery is slowly increased, up to 1.2s
(max time)
{ Alternate with Go-trials with nal parameter, to reinforce previous steps of training.
Tier 5: Sound Duration (optional)
{ Similar to Tier 3, trains system for sound duration of the LOW tone.
The tiers work together serially. Once the system has adequately learned tier 1, activity proceeds
to Tier 2, Tier 3, etc. The metric for determining whether a tier has been suciently learned
is dependent on the success rate. The learning process usually involves a type of examination
afterward to gauge how well the Agent learned it. This will be the Part II-Performance stage.
Instead of partitioning the experiment by the tiers mentioned above, Part II will alter the measure
of success by the formal experimental steps in which the neural ring patterns and data were col-
lected. The brevity of these experimental steps is expressed as follows: 1. Monkey subject, sitting
in a chair, focuses attention on an LED. 2. A ready signal (turning on the LED) indicates the
start of the trial. Monkey subject responds by placing its hand on a touchpad and xating on the
LED, indicating movement may begin. The LED is positioned such that it appears superimposed
on top of the object to be grasped. The object is lighted in such a way that it is also visible.
3. LED changes color (GO signal). Monkey subject responds by removing his hand from the
touchpad and reaching towards the object. Reaching phase: monkey subject preshapes his hand
in anticipation of making contact with the object. 4. Contact with the object is made, and the
51
monkey subject secures a grasp. Manipulatory movement is made (pulling or pushing the object
in some direction). The resulting position is held for a random delay period.
Figure 3.6: Augmented MNS2 model for Action and Inaction Task. Updates included 1) Emphasis
on Basal Ganglia for learning, 2) Action and Inaction Encoding, 3) Space-Dependent Representa-
tion, and neural pathways for 4) Self versus Other pathways. Red arrows denote new connections
from previous models. A new sensory modality for auditory input was added. And a Basal
Ganglia unit, including the dorsal and ventral striatum, was another feature added to enhance
decision making.
3.2.2 Experimental Implementation
The updated version of an MNS implementation, shown in Fig. 3.6, may be labeled \How the
brain generates predictive motor representations of action based on `decision' for Self inaction and
visual responses to Others' inaction." A possible theme for this study is in questioning whether
this problem is a matter of focus of attention: an IS vs. IF. \IS" places the Agent's attention in
the present moment of what is actually taking place currently. \IF" places the Agent's attention
on what possible actions can take place in the potential proximate future. In Bonini et al. [17],
52
experimental data showed that when the macaque viewed an inaction event by the human, the
same neural patterns red here as there is ring when the monkey views the grasp in motion.
In this upcoming version of MNS, using data from Bonini's study, we want to show that motor
representation can be predictively encoded when an agent executes or observes the negation of an
action. Rightly so, it is necessary to provide a thorough analysis of the neural processing taking
place in iterative steps of F5 neurons and other neurons that contribute. Other neurons than
the F5 neurons shall either be hypothesized or conclusive to be a part of this process analysis
presented in this report. In light of this, let's discuss what has to be tentatively added in F5 and
elsewhere to extend the F5 mirror and non-mirror (motor) F5 neurons mentioned in the study.
The objectives for the new additions are listed below. Two approaches for making model changes
are to 1) reveal new brain regions that play a role in the process, and 2) discover how the already
model-implemented brain regions might work dierently considering the context. With that said,
some proposed high-level model changes from both Bonini papers [16, 17] include:
Adding a Prediction class: Classify MNs as either Action or Inaction MNs. They are not
entirely distinct, but they do encode dierently through a reactive and predictive context.
Space-representation class: Implement region that is responsible for discriminating action
in peri-personal and extra-personal space.
Neural patterns are dierent for action and inaction conditions in the varied spaces [21]. The
objectives we felt needed to be implemented are based on the FARS and DAJ (`95) models. We
sought value in incorporating areas such as the caudate, striatum, SNr, and thalamus. Several
updated model proposals to consider should include the following:
1. Classify (mirror neurons) MNs as either Action or Inaction MNs|relay when they do and
don't re.
53
2. Model Inaction MNs (potential) inhibitory function.
3. Showcase Four Discharge Patterns among the F5 mirror neuron Predictive and Reactive for
both Action and Inaction conditions.
4. Implement Canonical-Mirror Neuron region, in addition to the separable Canonical and
Mirror Neuron [16]
5. Incorporating the IT cortex back in the model.
6. Incorporating the basal ganglia to deal with the motivation of motor control and learning.
BG works with the thalamus and cerebral cortex to help make decisions and shift between
activities. This will more than likely assist in our eorts in distinguishing pathways for
Action and Inaction.
7. Incorporate the Inferior Colliculus cortex as it is used in MNS for receiving and training for
auditory cues.
8. Incorporate the striatum (caudate nucleus, putamen), which receives inputs from all cortical
areas and, throughout the thalamus, project to frontal lobe areas (prefrontal, premotor, and
supplementary motor areas) which are purposed with motor planning. The neural circuits
involving these regions (i) provides information for both automatic and voluntary motor
responses (ii) assist in predicting future events, reinforcing wanted behavior and suppressing
unwanted behavior, and (iii) are involved in shifting attentional sets, movement initiation,
and spatial working memory.
9. Showcase the ring pattern for extra-personal and peripersonal space [16]
54
• Here, we can somehow show that predictive discharge occurs earlier and more frequently
when neurons are tested during observation of actions performed in the monkeys' ex-
trapersonal rather than peripersonal space.
Figure 3.7: Left: Graphical user interface for the simulation settings. Users have the option to
select the training weight les using Hebbian and backpropagation, backpropagation through time
(BPTT), and BPTT with Hebbian networks. Variation of type of arm movement can be made
for a Reach or Eat trajectory, with grasping actions ranging from Natural, Side, Power, Precision,
and Slap. Objects used for the target and obstacle span several basic shapes like a coin, box, and
rectangular sheet, to name a few. Right: Here are two frames of the eld of view of the simulation
in action. The top frame depicts the trained in-motion arm performing a precision grasp of the
box with the index and thumb ngers. The bottom frame shows an attempt for Power grasping
the Pent shaped object, following the blue-line trajectory.
In terms of the outlook of both simulation and visualization, a setup of trials shall consist of
an observation of what it is like for MNs to acquire the ability to distinguish self and other's
action/inaction, therefore needing to incorporate a learning mechanism. Several initiating stimuli
have been considered suitable in bringing about an agent's inaction (e.g., Not grasping a coee
cup). As you are in the process of reaching, you may be inhibited to grasp by:
1. Thought- You quickly recall the cup is empty of contents, therefore you cease your action
to get a drink.
55
2. Vision- You see that the cup is empty, therefore there is no need to pick up the mug to
drink. Grasp ceases.
3. Sensory- You feel that the cup is absurdly hot before touching. Grasp ceases to avoid burn.
4. Auditory- You hear someone yell \Stop" or \Don't" in reference to not touch the cup. Grasp
ceases.
Granted, Thought and Somatosensory may involve more in-depth simulation eorts, but it would
be worthwhile to mention what factors encompass the entire phenomenon of inaction. A potential
simulation and visualization of such stimuli would be to use Auditory and Vision in a \grasping"
scenario of some sort. The simulated monkey will have to learn to associate cue sounds with
inaction/action events with a high success rate, as mentioned in the experimental paper. Using
the model to show correct grasps can be learned to be predicted by an observer in the absence
of physical motion if enough appropriate stimuli are available to hint at the Agent's goal intent.
The accuracy of this prediction will gauge whether our model is sound or not. Fig. 3.6 shows
the initial implementation of the software based on MNS2 that taught the simulated arm how
to properly grasp various objects according to the size, shape, and aordances. The goal of this
portion of the model, as shown previously by Bonaiuto [13] was to demonstrate through dierent
training weight les how an object's physical characteristics can alter action types for reaching
grasping tasks.
3.3 Discussion
We are particularly interested in how our potential ndings might be practical to the routines
and architecture design used by autonomous systems to gather information (via senses) about
their environment and form perceptions that prompt ideal behaviors for a given task. If our
hypothesis shows the evidence that we foresee, there is an opportunity for an implementable
56
Figure 3.8: Obstacles for Interactions.
framework to be created for a diverse set of technologies to use that exploit the usage and benets
of sensory aerents. Machine learning and computational models will be used, as they are often
the fundamental bases for introducing vital and fundamental breakthroughs in neuroscience on
many topics. Interfacing the developed self-aware software framework with the lab's hardware
robotic limbs will be one of the nal results of demonstrating the working solutions. Lastly, this
thesis intends to show how aerent information traverses through the nervous system, is perceived
by the body, and integrates to create behavior. The experimental outcome will display how the
role of sensory aerents is indeed necessary toward the development of a `self' model that has the
capacity to construct predictive motor control capabilities for the contrast inner experience and
external reality.
It is not enough to solely implement self-awareness into a system. Another goal includes the
pairing of self-awareness to enable more eective behaviors and resulting actions via a process
called self-expression. To truly be autonomic, a computing system needs to know itself, its
limits, its constraints, its capabilities, and its constrictions. In bio-inspired applications, such
knowledge can be obtained and maintained through continuous high-dimensional sensory inputs
57
like proprioception, vision, auditory signals, and tactile senses. Then a model of the self can be
built using those same aerent values. What self-awareness means for machine and computational
means is not well-dened and remains to be an open eld with competing and contrasting theories.
As self-awareness is a concept and term that is inherently subjective in nature, it is understood
why there is such dissonance in the eld among experts. Many facets of the eld remain to be
ambiguous due to a lack of agreement on how to dene self-awareness and in which manner is
it appropriate to duplicate it. Furthermore, we'll contribute insights from the MNS to develop a
unied methodology and framework for creating the underlying principles for an emerging self-
aware system with more concentration on exploiting perceptions and in
ux of sensory information.
Following, we will address the (i) importance of sensory information in building the brain and the
representation of self as an object, (ii) why sensory information is coincidentally an unexplored
eld in both self-aware systems and sensorimotor research, and (iii) why self-awareness must be
holistically examined.
58
Chapter 4
Sense: Quantifying High Dimensional Feasible Sensory Sets
\The senses are gateways to the intelligence. There is nothing in the intelligence which
did not rst pass through the senses." -Aristotle
4.1 Abstract
We introduce the concept of trajectory-specic sensory manifolds. They are the unique multi-
dimensional and time-varying combinations of aerent signals that obligatorily emerge during a
limb movement. We use the example of muscle spindles (i.e., the muscle's proprioceptors for
length and velocity) that arise during movements of an arm (a planar 2-DOF 6- muscle model)
during the production of straight, curved and oscillatory hand movements. Through the use of
parallel coordinates, we visualize the high-dimensional evolution of the aerent signaling across
muscles and tasks. We demonstrate that a given movement gives rise to a distinct sensory mani-
fold embedded in the 12-D space of spindle information that is largely independent of the choice
of muscle coordination strategy. Given that muscle lengths and velocities are fully determined by
joint kinematics, such manifolds provide a rich set of information to use in its control.
59
4.2 Introduction: How the Body Builds the Brain
Sensorimotor control research, both past and ongoing, has made eorts to predominately provide
evidence for how the brain shapes the body [72, 162]. However, the counterpart to these works,
how the body shapes the brain [28, 5, 135] is not as extensively considered. Often not taken
into account are sensory states and their eects on building the brain's body awareness which
is necessary for involuntary and voluntary behavior. Here we investigate the
ow of information
underlying limb movements, and explore its signicance to perceptual learning. We begin our
work in this area by dening the concept of a feasible sensory set for a given movement. Our
study follows the work of [101, 176, 177] that developed a theoretical framework for all possible
body accelerations, activations and torques for a given tasks (e.g., feasible acceleration, activation,
torque or force sets). By extension, we can also speak of a feasible sensory set (FSS). In the case
of muscle spindle aerents|which sense muscle lengths and velocities|the FSS denes the set of
sensory signals that can emerge for a given limb posture or movement. In particular, given that
muscle lengths and velocities are geometrically dened by joint angles and angular velocities [179]
then a given posture or limb movement will also prescribe the spindle aerent signals. Given a
motor task, and a denition of available sensory information, we sought to dene the associated
manifold of spindle aerent signals that dene its FSS.
4.3 Experimental Methodology
We used a simplied tendon-driven arm model where simulated muscles pull on tendons that cross,
and therefore actuate, kinematic Degrees of Freedom (DOFs). The planar model had six mus-
cles, two links (upper arm and forearm) and two DOFs (Shoulder Flexion/Extension (SFE) and
elbow
exion/extension (EFE). A combination of paired antagonistic muscles formed the tendon
routing of a right arm appendage: deltoid anterior (monoarticular shoulder
exor), deltoidmpos-
terior (monoarticular shoulder extensor), biceps brachii (biarticular elbow
exor), triceps brachii
60
q
1
q
2
endpoint
Figure 4.1: For the Static Case, all possible x-y coordinates for q
1
andq
2
degree ranges. Parame-
tersq
1
andq
2
were constrained within ranges of motion 0-130° and 0-150°, respectively. Location
of the SFE joint remains xed at the origin (0, 0)
(biarticular elbow extensor), brachialis (monoarticular elbow
exor), and anconeus (monoarticular
elbow extensor). The study was partitioned in two parts beginning with kinematic calculations of
an arm during a task to derive limb joint angles and endpoint locations, and then applying those
metrics to the spindle model for observation of aerent signaling. Incorporating modeled muscle
parameters of optimal ber lengths L
o
, change in angle q, and constant moment arm values r
from upper extremity analyses [83] allowed the initial computation of tendon excursion (change
in length of musculotendon) values, as shown in Eq. 4.1.
s =rq (4.1)
61
Table 4.1: Simulated limb and musculotendon parameters.
Receptor Type Axon Fiber Fiber Name Transducer Modality
Muscle Spindle Primary
a
A Ia Muscle length and speed
Muscle Spindle secondary
b
A II Muscle strength
a. Classication of aerents and their respective function for detecting deformation of muscle
tissue and transducing those signals into electrical responses. Fastest conduction speed and ber
diameter, A (72-120 m/s) has the thickest myelination.
b. A (36-72 m/s) possess thinner myelinated axons. [90]
We investigated how limb movements aect two chief elements of muscle aerentation for muscle
length/contractile velocity. Using inverse kinematics [159], a Static Case was used to nd all
possible discrete positions our modeled arm can achieve. Variations in proprioceptive signaling
are shown to be dependent on task constraints as studies have revealed active movements tending
to report more accurate proprioceptive approximations [54, 193], which led us to incorporate a
case with continuous arm movement in dynamic settings. The Dynamic Case consists of specic
trajectories the arm follows over a set time frame that can be modied via the Speed Factor
parameter. As the Speed Factor increases so does the velocity of the movement under observation.
Within the Dynamic Case, we evaluated our arm limb moving in several tasks starting with the
Circle trajectory (in counterclockwise direction) as illustrated in Fig. 4.2A. The Straight Linear
trajectory task consisted of ve distinct pathways on a plane for the arm to follow, each perturbed
at the slope by a 0.1 decrement (Fig. 4.2B). The Oscillatory trajectory represented sinusoidal
movement with an angular frequency of 6, amplitude of .05m, and .35m vertical shift (Fig. 4.2C).
And lastly, we dened the symmetrical lobe Lemniscate (i.e., \gure of eight") trajectory [146]
(Fig. 4.2D) using the mathematical expressions shown in Eq. 4.2 and Eq. 4.3. It must be noted
that the derived conguration spaces only disclose exclusive values for q
1
and q
2
despite the
possibility of a multiplicity of joint angles producing the same end-eector position.
x =
cos(t)
(1 +sin
2
(t))
(4.2)
62
Figure 4.2: Cartesian space and Conguration space of arm movement in directions indicated by
the red cursors for (A) the arm limb in action during the Circle Trajectory task in the coun-
terclockwise direction. Conguration space illustrates the joint angles for 360 distinct postures.
(B) Arm limb in action during the Straight Linear Trajectory task in a left to right direction.
Line 1 trajectory, in blue, sustains a slope of .5. Line 2 trajectory, in green, sustains a slope of
.4. Lines 3-5 follow according with a negative .1 gradient. Conguration Space illustrates the
joint angles for 1,000 distinct postures. (C) Arm limb in action during the Oscillatory Trajectory
task in a left to right direction. Conguration Space illustrates the joint angles for 1,000 distinct
postures from the leftmost to rightmost point along the trajectory. (D) Arm limb in action for
Lemniscate Trajectory task with symmetrical lobes. The depicted path is partitioned according
to color scheme for mapping the end-eector location in Cartesian coordinates to the joint angles
illustrated in the Conguration Space, which illustrates for 1,000 distinct postures.
y =
sin(t)cos(t)
(1 +sin
2
(t))
(4.3)
After solving for the joint and limb kinematics, we utilized a computational sub-model to simulate
the biological spindle as observed in mammalian muscles, namely that of the cat [121, 122], which
has also been used in human simulations [163, 103]. Action potentials in pulses per second (pps)
were generated for primary (Ia) and secondary (II) aerents based on the interactions of the
63
intrafusal bers (chain, bag1, bag2). The rst analysis that we performed examined whether
aerent signals are dependent on muscle velocity throughout a task. We varied the Speed Factor
in the system by a combination of values ranging in ascending speed: 0.0005, 0.005, 0.05, 0.5,
and 1. One-way analysis of variance (ANOVA) of the measured spindle signals under these
varying velocities tested whether there was a signicant dierence between the group output
values. Velocities were categorical and set as the independent variable while the spindle rings
served as the continuous dependent variable. Our second analysis developed the high-dimensional
sensory space for Ia and II aerent signaling to extract the sensory aerent sets for the Dynamic
tasks.
64
Figure 4.3: Six-dimensional representation of change in muscle length along four trajectories of
the Dynamic Case. Color gradient depicts initial postures(yellow), intermediate postures (green),
and concluding postures (blue). (A) Muscle length values (meters) during Circle Trajectory task.
360 postures were examined ranging from Posture 1 at 0 radians to Posture 360 at 2 radians.
Direction of movement along the trajectory is counter- clockwise. (B) Muscle length values during
Line 1 Trajectory task. 50 postures were examined ranging from Posture 1 at the leftmost point
on the line to Posture 50 at the rightmost point. Lines 2-5 follow the same paradigm of movement
sequences just with an altered slope. (C) Muscle length values during Oscillatory Trajectory task.
1,000 postures were examined ranging from Posture 1 at the leftmost point on the sinusoidal wave
to Posture 1,000 at the rightmost point. (D) Muscle length values during Lemniscate Trajectory
task for 1000 postures.
65
4.4 Results
4.4.1 Kinematics Assessment
Parallel coordinates were used to clearly illustrate the multi-dimensional change in muscle lengths
for each posture during the tasks of the Dynamic Case. Such assessment was conducted to verify
the ecacy of our model to ensure isometric, concentric and eccentric contractions according to
physiological expectations [106, 183]. As shown in 4.3 we sampled n postures along each trajec-
tory (i.e., task) and integrateds from the initial posture. The three pairs of antagonistic muscles
showed the expected concentric and eccentric contractions along their respective continuous tra-
jectories. The muscle lengths were dierentiated to derive the obtain their respective velocities
and accelerations, which served as direct input parameters to each spindle model.
Figure 4.4: Velocity speeds versus aerent signals in Group Ia (left) and II (right). Five values
were used for the Speed Factor, with value 1 signaling the fastest speed across the task. Top
row: Aerent ring in the biceps muscle for the Oscillatory task shows slight oscillations with
increasing speed. Bottom row: Aerent Firing for the triceps muscle shows a uniform and
smooth signal throughout the span of Speed Factors. After
66
4.4.2 Aerent Signaling Dependent on Muscle Velocity
We expected that increased muscle velocities would aect spindles, and therefore, \body sense" in
a heavily nonlinear way. Fig. 4.4 provides a sample of our observations. We detected the presence
of perturbations in the Circle, Oscillatory, and Lemniscate; prospectively owing to the curvature
of the trajectories which can induce abrupt changes in velocities. These nding corroborate
observations from [121]. ANOVA tests revealed p-values for each muscle's Ia and II aerent
signals in each task of the Dynamic Case (Table 4.2). For those cases where p 0.05, we rejected
the null hypothesis that there was no dierence between the dened sets of velocities and the
resulting aerentation in each task. We detected signicance in only particular groups of muscles
in the Circle and Oscillatory trajectory tasks. The Straight Linear trajectory demonstrated no
signicant dierence across all six muscles for each Speed Factor value (Fig. 4.4), while the
Lemniscate trajectory showed evidence for all muscles having aected aerentation.
4.4.3 Sensory Bounds According to Task Constraints
We also used parallel coordinates to describe Group Ia and II signals in the Dynamic Case with
a reasonable duration of 5 seconds. Our sampling frequency (fs) was set at 10 kHz (10,000 sam-
ples/second), resulting in 50,000 time samples throughout each trajectory. Fig.4.5 presents the
high-dimensional correlated relationships among aerent signals. To read these parallel coordi-
nates, please note: 1) each axis is likely to have a dierent scale depending on the range of values
reported for that muscle, 2) adjacent dimensions are more easily interpretable than non-adjacent
dimensions, and 3) a web-based view provides the ability to interactively analyze subsets of ac-
tivities of single muscles (as shown in Fig. 4.5 for the Line 1 & 5 trajectories). Therefore, we
can explore the multivariate comparisons, patterns, and sequences that are unique to each muscle
and trajectory. For example, for the Group II aerents of the Line 5 trajectory, we isolated the
signals on the triceps muscle between 50-150pps. This revealed the associated ring rates for
67
Table 4.2: Velocity Signicance in Aerents (ANOVA P-values)

Fiber
Type
Circle
Line 1-5
()
Oscillatory Lemniscate
Deltoid A.
Ia 0.049 0.249 <0.001 <0.001
II 0.830 0.861 <0.001 <0.001
Deltoid P.
Ia 0.538 0.999 <0.001 <0.001
II 0.999 0.999 0.087 <0.001
Biceps
Ia <0.001 0.981 <0.001 <0.001
II 0.213 0.999 <0.05 <0.001
Triceps
Ia <0.001 0.753 <0.001 <0.001
II <0.01 0.950 <0.001 <0.001
Brachialis
Ia 0.362 0.999 <0.001 <0.001
II 0.850 0.999 0.072 <0.001
Anconeus
Ia <0.001 0.474 <0.001 <0.001
II <0.001 0.967 <0.001 <0.001
other muscles as related to the triceps: deltoid anterior 0-5 pps, deltoid posterior 80-130 pps,
biceps 35-60+ pps, brachialis 0 pps, and anconeus 185-250+ pps. Similar introspections of signal
bandwidth can be made for all other muscles. We also have the ability to trace and correlate any
subset of physiological with kinematic variables such as velocities, accelerations, muscle lengths
and stretch, and limb position in space.
To gain insight into the robustness of spindle aerents, we performed Monte Carlo simulation
[144] for each of the six muscles with variation of the gamma static and dynamic fusimotor drive
values. Within 100 trial iterations, boundary limits on both
dynamic
and
static
were set to
inclusively span 70 and 150 pps. Maximum standard deviation between any given set of the
observed points approximated to 20pps, consequently resolving to a 10% deviation estimate of
the signal as
dynamic
and
static
were constrained at a constant rate of 100pps.
68
69
Figure 4.5: Primary and secondary aerent space for the Circle, Line 1, Line 5, Oscillatory, and
Lemniscate trajectory tasks marked by 50,000 samples in a time interval of 5 seconds from starting
position to ending position along the prescribed trajectory. Left side: Parallel coordinates
showing the activation of each group of muscle during a sampling range along the trajectory.
Right side: Spindle aerentation of each muscle according to the range 0.75-1.3 of the optimal
lengths. The parallel coordinate aerents for the Lemniscate trajectory is mapped with the color
segments used in Fig. 2D. It can now be observed which location along the Lemniscate trajectory
produces a certain aerentation value.
This allows us to quantify how fusimotor activation, naturally, aects both motor capabilities
and body sense. Furthermore, we present the importance of how each movement leads to a very
specic set of sensory information. We can then propose the concept of feasible sensory manifolds,
FSS, associated with each movement task.
4.5 Discussion
How do our sensory signals shape the motor choices we make in daily life? In this paper, we
addressed the body sense that arises from muscle spindle aerents, to enable future studies in-
vestigating how those same sensory signals aect the representation of our physical self and the
actions we make. As per equation 4.1, we know that muscle excursions and velocities are com-
pletely determined by the time history of joint angles (so long as muscle tone prevents any muscle
from being slack in any posture). Thus, every limb movement is associated with a unique set
of specic sensory states, the FSS. As such we must consider how the nervous system obtains
and processes sensory data to create a body sense that interacts with explicit or implicit internal
models of the body, and external in
uences on the body. There are several perspectives on how
sensory data (mostly visual) leads to perceptual states: the action-oriented theory of percep-
tion, which suggests that perception is the result of sensorimotor dynamics in an acting observer
[63, 115, 131] and the dual-visual systems hypothesis, which advocates for independent streams
of perception and action [67, 66, 84, 123, 158, 175]. Recognized predominantly as the Perception-
Action Cycle [128], various methods developed from this framework may be utilized to replicate
70
the decision making that occurs during the process of acquiring sensory modalities regarding the
external world [113]. In the context of neuromechanics, we posit that sensory data obtained in
any moment is dependent on the kinematic posture, position, and action task of the respective
limb producing the sensory stimuli. Our present study delved into the consequences to sensory
systems towards representation of high-dimensional observability to complement the controlla-
bility of muscle- driven limbs; specically, within the mammalian muscle spindle. Our methods
for obtaining these results can be employed towards systems such as robotics and brain-machine
interfaces (BMIs) that are optimized on the limits of simulated neural drive obtained from sensory
inputs. This concludes the rst phase towards focusing on the categorization of sensory states.
Results for this phase are primarily based on the results documented by Berry et al. 2017 [8].
71
Chapter 5
State: Sensory Aerent Organization to Classication of
Actionable States
\Thoughts make the plan. Actions make the man." -Unknown
5.1 Abstract
High-dimensional proprioceptive signals like those from muscle spindles are thought to enable
robust estimates of bodily states. Yet, it remains unknown whether spindle signals suce to
discriminate limb movements. Here, we used a 4-musculotendon, 2-joint limb model to simulate
muscle spindle II and Ia signals (length and velocity, respectively) during repeated cycles of ve
end-point trajectories in forward and reverse directions. We nd that cross-correlation of the 8D
time series of raw ring rates (four Ia signals, four II signals) cannot discriminate among most
movement pairs (only 29% by one measure). However, projecting these signals onto their 1st
and 2nd principal components greatly improves discriminability of movement pairs (82% by that
same measure). We conclude that high-dimensional multi-muscle proprioceptive ensembles can
usefully discriminate limb states|but only after minimal pre-processing. Importantly, this may
72
explain the documented subcortical pre-processing of aerent signals, such as cutaneous signals
processing by the cat's cuneate nucleus.
5.2 Introduction
Physical behavior in vertebrates is made possible by hierarchical neuronal systems that send
motor commands from the central nervous system to muscles on the basis of sensory information
coming from the peripheral nervous system. Motor function has received much attention given
the relative ease with which the activity of motoneurons and muscle can be measured and
associated with physical behavior. In contrast, the emergence of somatosensory `percepts' (i.e.,
the transformation from spike trains from an ensemble of mechanoreceptors to a neural impression
useful to the control of movement) has proven much more challenging to understand. This is
because the action potentials from mechanoreceptors on the skin, muscles and joints are not
easily isolated or recorded [170], and the somatosensory percepts they elicit in the central nervous
system cannot be readily inferred.
The lack of understanding of the physiological bases of somatosensory percepts is particularly
problematic to the study and theories of sensorimotor control [112]. In particular, the somatosen-
sory percept of proprioception, also called kinaesthesia, provides the sense of self-movement and
body conguration/position [49]. Rigorous neurophysiological work on mechanoreceptors has led
to the fundamental tenet of sensorimotor control that muscle spindles (whose II and Ia aerent
bers encode the length and velocity of each muscle) provide necessary, if not sucient, limb
conguration information for adaptable, accurate, and robust control of limb movement. This is
supported by the geometrically obligatory relationship between joint angles and muscle lengths,
but also challenged by the facts that muscles often span multiple joints and that spindle signals
can be modulated independently of joint angles by
motoneuron drive to their intrafusal bers.
We are not aware of conclusive evidence of this tenet, which is adopted to the point that other
73
mechanoreceptors also aected by joint angles (i.e., synovial capsule, ligaments and skin) and
Golgi tendon organs are considered secondary for reasons detailed in the Discussion. However,
this has not been demonstrated experimentally because spindle aerent recordings from numer-
ous limb muscles in peripheral nerves or dorsal root ganglia cannot be obtained during large limb
movements.
Therefore, we performed a computational experiment to assess the utility of muscle spindle af-
ferents to provide usable limb conguration information. A minimal requirement for utility is
the statistical notion of discriminability. Discrimination test are employed in sensory evaluations
and analyses. Discriminability has been used to test how raw and processed signals from skin
mechanoreceptors on the ngertips can be used to distinguish among dierent edges and textures
to inform manipulation [133, 154]. In our case, we performed pair comparison tests to evaluate
the extent to which raw and processed ensembles of Ia and II spindle aerents signals, during
ve distinct limb movements (Fig. 5.1), could discriminate among the ve limb movements that
produced them.
5.3 Experimental Methodology
The computational design of the simulated tendon-driven system, the trajectories selected for
inspection, and the modied spindle aerent model will be described. Then we'll detail the
methods of pre-processing and ltering used to reduce the dimensions of aerent signals. Our
pre-planned trajectories produced aerent signals that were compared in inter-class contexts in
then processed in data series estimation, pattern identication, and unsupervised machine learning
algorithms on the resulting aerents to reveal their spatial and temporal dynamics. Lastly, we'll
conclude with a review of how the feature selection and extraction techniques were implemented
to determine which relevant spindle model features maintained substantial eects in classifying
one trajectory from another within sensory space.
74
Circle Line Oscillatory
Lemniscate
Square
A E D C B
Knee
Flexion
Knee
Extension
Hip
Flexion
Hip
Extension
Anterior Biceps
Iliopsoas
Vastus Lateralis
Semitendinosus
Dorsal
Ventral
Posterior Anterior
Figure 5.1: - Limb kinematics were derived from distinctive trajectory types. The
2-joint kinematic model, which was tted with four muscles found in the cat musculoskeletal
hindlimb structure encompassed a muscle group that included anterior biceps, iliopsoas, vastus
lateralis, and semitendinsosus. The following ve planned trajectories, with the red arrow in-
dicating its direction of motion, were selected for comparative analysis. (A) Circle trajectory
maintained motion progression in clockwise direction. (B) Line trajectory completed a full cycle
by . (C) Oscillatory trajectory replicates the sine wave curve oscillations. (D) Lemniscate trajec-
tory is a polar curve that is usually referred to by the name "gure-eight". (E) Square trajectory
maintained motion progression in counterclockwise direction.
5.3.1 Kinematic Model Structure and Parameters
We constructed a simplied tendon-driven leg model, represented as the feline hindlimb, with a
pivot at the hip joint. In tendon-driven anatomies, tendons are responsible for permitting muscles
to act on vertebrate limbs and actuating the kinematic Degrees of Freedom (DOF) [181]. The
planar model consisted of four muscles, two links, and two DOFs (Hip Flexion/Extension and
Knee Flexion/Extension) connecting the thigh and shank, as shown in Fig. 5.1. For simplicity we
excluded actuation of the foot (i.e., paw), which is normally included in a feline hindlimb model
and would be more representative of the actual feline. Lengths of the thigh and shank segments
were set to 90 mm and 100 mm, respectively, with musculature comparable to the muscle-joint
interactions and parameter data resulting from system identication analyses [74] that were based
on mathematical properties of skeletal muscle formulated by Zajac [192].
To imitate the useful dynamics of the cat's hindlimb mobility, we captured the movements of
75
the leg as generated by 4 muscles: Anterior Biceps (AB), Iliopsoas (IL), Vastus Lateralis (VL),
and the Semitendinosus (SM). Table 5.1 summarizes the parameters we used in the musculo-
tendon structure which contained parameters of maximal length as L
max
, constant moment arm
values as r, and optimal length values (L
O
) per muscle at the reference angle. Fig. 5.1 depicts
the tendon routing of AB, IL, and VL as unifunctional joint muscles. AB and IL are acting in
paired antagonistic form on the hip. VL activates knee extension movements, while SM serves
as a bifunctional joint muscle acting on both the hip and knee. According to Harischandra and
Ekeberg [74], the resting (neutral) posture of the hip at 65° and the knee at 100° maintained
mono-articulated muscles at a length of 85% of L
max
and 75% for bi-articulated muscles.
Table 5.1: Simulated limb and musculotendon parameters.
Muscle Name L
max
(mm) Angle Movement Moment Arm (mm) Reference Angle L
O
Anterior Biceps 70 Hip Extension 30 85%
Iliopsoas 70 Hip Flexion -44 85%
Vastus Lateralis 50 Knee Extension 9 85%
Semitendinosus 70
Hip Extension
Knee Flexion
30
-38
75%
Optimal lengths (L
O
) of each muscle at the reference angle were set to 85% and 75% of L
max
for
unifunctional and bifunctional muscles, respectively.
5.3.2 Trajectory Planning
Arbitrary shapes were selected as pre-planned trajectories in two-dimensional planar space for the
end-eector limb positions. All trajectories were performed in closed loops and mathematically
expressed as parametric functions of time, t, to obtain the x andy coordinate locations. The cat
limb executed ve point-to-point movements that will be further referred to as task representations.
Each task representation contained a total of 200 equidistant points per cycle on the trajectory.
One full cycle lasted for a time frame of one cycle/second.
76
The rst task representation is the Circle trajectory (Fig. 5.1A) which prompted the limb to
perform uniform circular motion within a 5 mm diameter in the clockwise direction. The end-
eector's total distance traveled approximates to 15.71 mm. Next, the Line trajectory (Fig.
5.1B) positioned the end-eector on the path of a straight line to simulate smooth, uninterrupted
movement along a ramp. Relative to the horizontal plane, the line segment retained a 50% incline
at 26.57°steepness. Its midpoint position was at the 100 mm y-intercept on the Cartesian plane.
The total distance travelled for one cycle of the Line trajectory was 12 mm.
The Oscillatory trajectory (Fig. 5.1C) is a sinusoidal wave forming a path of a smooth periodic
oscillation. Using Eq. 5.1 as a function of time, the amplitude A was set to 20 mm with a
frequency, f, of 10Hz. The angular frequency, w, expressed in radians at run-time Eq. 5.2 along
with zero phase shift, '.
! = 2f (5.1)
y
n
(t) =A sin(!t +') (5.2)
The Lemniscate trajectory (Fig. 5.1D) created two symmetrical and uniform-sized lobes to form a
shape resembling the \gure-of-eight" curve [146]. The curve was formed using parametric curves
from Eq. 5.3 and Eq. 5.4.
x
n
= 4 10
2
sin(5 10
1
t) (5.3)
y
n
= 2 10
1
sin(t) + 11 10
2
(5.4)
Lastly, we prescribed the Square trajectory (Fig. 5.1E) as a proximity comparison to the Circle
trajectory movement. Considering squares and circles are topologically equivalent shapes, we
expected to view closer spatiotemporal similarities in the sensory space between these two shapes
over others. However, squares dier in their non-continuity and nite lines of re
ectional symmetry
77
which also might re
ect symmetry within the aerent manifolds. To what extent will the aerent
signals re
ect these features in the observed kinematics and make them muscle activities and joint
motions distinguishable is one facet of the experimental outcomes we sought to observe.
The limb joints on the planar limb actuate as revolute joints with links capable of rotating around
it. The 2 links comprise of an end eector which maintains the foot position at the end of the
shank link and also the end of the articulated body. While the hip position remained axed as the
root joint, we calculated the tracing of the end eector position across each of the ve trajectories
using inverse kinematics. For each trajectory, the 200 target positions in the Cartesian space were
selected as inputs for the inverse kinematics algorithm and the limb pose (i.e., state) required
for the target position were derived to determine the joint angles at the hip and knee, q
1
and q
2
respectively.
Inverse kinematic solutions are generally not unique, and are sometimes dependent on the initial
joint coordinatedq
0
, which typically defaults to value 0. However, the values forq
1
andq
2
of the
limb were successfully obtained despite the possibility of a multiplicity of joint angles producing
the same end-eector position. Given the desired limb's end-eector positions, for each time step
across the trajectory at instancei, the segment link lengths,l
1
andl
2
, and the coordinate positions,
x
1
andx
2
, were recorded to calculate variablesc ands in Eq. 5.5 and Eq. 5.6, respectively. Joint
anglesq
1
andq
2
for each segment were then iteratively derived using equations Eq.5.7 and Eq.5.8.
c =
(x
2
i
+y
2
i
l
2
1
l
2
2
)
(2l
1
l
2
)
(5.5)
s =
p
1c
2
(5.6)
q
1
= sin
1
y
i
(l
1
+l
2
c)x
i
l
2
s
x
2
i
+y
2
i
(5.7)
78
q
2
= cos
1
x
2
i
+y
2
i
l
2
1
l
2
2
(2l
1
l
2
)
(5.8)
Once the limb's joint angles are calculated, a Jacobian matrix can be generated to determine
the relationship between simulated limb's joint parameters and the end-eector velocities. The
change in joint angles are then used as inputs for the muscle spindle model to obtain raw sensory
aerents for each trajectory.
5.3.3 Muscle Spindle Aerent Data Collection
In a similar method that was used in Chapter 4, the joint and limb kinematics were solved using
a computational sub-model to simulate the biological spindle as observed in mammalian muscles,
namely that of the cat [121, 122], which has also been used in human simulations [163, 103]. Action
potentials in pulses per second (pps) were generated for primary (Ia) and secondary (II) aerents
based on the interactions of the intrafusal bers (chain, bag1, bag2). Fusimotor activation and
the property changes in induces within the spindle model is represented by contractile elements
(CE). The spindle model operates from a set of parameterized inputs that included that included
L
o
as optimal muscle lengths, L
ce
as muscle length normalized to L
o
, V
ce
as the rate of change
in muscle length (i.e., velocities), A
ce
as muscle length acceleration, Fs as sampling frequency,

dynamic
as dynamic gamma drive, and
static
as static gamma drive.
The model produced only two outputs, which were non-linear rings of the primary aerent
potential and secondary aerent potential modalities in the spindle, Ia and II respectively. As
stated in [121], the generation of aerent potential re
ects the stretch of the intrafusal ber
model's sensory zone. Aerent potential primary derived based on Eq. 5.9 where T=K
SR
is the
calculated stretch in the sensory region of each intrafusal ber,L
SR
N
is the sensory region threshold
length, L
SR
0
is the sensory region rest length, and G is a constant that indicates the numerical
relationship between intrafusal ber's sensory region to primary aerent ring. Aerent potential
79
secondary derived based on Eq. 5.10 where X is the percentage of the secondary aerent located
on the sensory region and L
secondary
is the secondary rest length.
AfferentPotential
Ia
=G

T
K
SR
(L
SR
N
L
SR
0
)

(5.9)
AfferentPotential
II
=G
(
X
L
secondary
L
SR
0

T
K
SR
(L
SR
N
L
0
N
)

+(1X)
L
secondary
L
PR
0

L
T
K
SR
L
SR
0
L
PR
N

) (5.10)
Both of the aerent ring model's output rings were collected as raw data to be to be statistically
analyzed for useful features that would indicated the current state of the limb.
5.3.4 Comparison of Inter-class Trajectory Context
In order to evaluate the discriminability of aerent signals against task-actions, the trajectory
types must be compared extensively. The ve trajectories selected for inspection are cycles of
shapes and curvatures that aren't typically associated with the natural gait of a feline hind limb:
Circle, Line, Oscillatory, Lemniscate, and Square. For this reason, there is an increased likelihood
for indisputably discern variations despite noise that may be present with a data set's dimen-
sionality, resolution, and sparsity. In our initial simulation executions, we observed that sensory
aerent outputs of the muscles varied signicantly depending on the initial conditions and the
direction the limb moves in to complete the cycle. Therefore, we ensured that the simulated limb
traversed each of the trajectories in two opposite directions: Reverse (REV) and Forward (FWD).
For example the Circle-FWD, which indicates the limb traversed the Circle trajectory moving in
the Forward direction, was compared in series to Circle-REV, Line-FWD, Line-REV, Oscillatory-
FWD, Oscillatory-REV, Lemniscate-FWD, Lemniscate-REV, Square-FWD, and Square-REV. All
possible combinations of trajectory comparisons totaled to 45 correlation pairs in both the raw
80
data set and pre-processed (i.e., PCA) data set. The combination set did not include pairs that
evaluated a trajectory-direction against each other.
5.3.5 Spatial, Spatio-Temporal, Pre-processing of Muscle Spindle
Aerent Data
To test the presence of discriminability across tasks, the aerent data sets were evaluated within
3 pattern constraints: spatial, spatio-temporal, and pre-processing from dimensional reduction.
Spatial Analysis
We rst evaluated the spatial patterns using the K-means++ algorithm. Since the standard K-
means algorithm does not guarantee to nd the optimum, an alternative, K-means++ chooses
initial centers on a justiable upper bound within cluster sum of squares objective. The approach
is initiated by separating the k initial cluster centers, spatially.
Overall the formal objective is to determine:
arg min
S
k
X
i=1
X
x2Si
kx
i
k
2
= arg min
S
k
X
i=1
jS
i
j VarS
i
(5.11)
where
i
is the mean of points in S
i
. This may also be shown to be equivalent to minimization
of the squared deviations of points, as shown by:
arg min
S
k
X
i=1
1
2jS
i
j
X
x;y2Si
kx yk
2
(5.12)
For an initial set of k means m
(1)
1
;:::;m
(1)
k
, the algorithm proceeds by alternating between the
assignment step and an update step, until convergence.
81
Spatio-Temporal Analysis
A useful statistical measure to use that identies signicant correlations among multiple trajec-
tories with spatial and temporal components is cross correlation. It compares the time-series of
aerent data across tasks, and is represented as the ratio in Eq. (5.13), wheren is the total number
of data point indices recorded per task cycle. This is suitable for measuring well two variables
move in relation to each other. Both x
i
and y
i
are the individual spindle aerent sets, Ia and II,
respectively. A temporal shift delay, phase lag , of the output cross correlation, R
xy
, measure is
applied to determine where the correlation of the data is maximized, as shown in Eq. (5.14).
R
xy
() =
P
n
i=1
(x
i
x)(y
i
y)
p
P
n
i=1
(x
i
x)
2
P
n
i=1
(y
i
y)
2
(5.13)

estimated
= arg max
2R
(R
xy
()) (5.14)
To retrieve the correlation coecients, local sums can be calculated in an alternative way to
normalize the cross-correlation. Using normalized cross correlation follows a general procedure
by [108, 73] in Eq. (5.15).

(u;v) =
P
x;y
(f(x;y)f
u;v
)(t(xu;yv)t)
q
P
x;y
[f(x;y)f
u;v
]
2
P
x;y
[t(xu;yv)t]
2
(5.15)
We can treat the combined group of muscle modalities within the aerent data as a template
and image and calculating the cross-correlation in the spatial or the frequency domain. The
implementation closely follows the formula from [108], where f is the image, t is the mean of the
template, and f
u;v
is the mean of f(x;y) in the region under the template.
82
Data Pre-processing Analysis
Principal component analysis (PCA) is often used for dimensional reduction on multi-dimensional
data, which assists with visual interpretation. PCA can identify the principal components that
are able to distinguish the Ia and II modalities and which represent the most variations between
groups. However, PCA is not particularly useful in accurately dening clear boundaries between
dierent clusters in the data. The combined use of PCA with clustering methods helps us under-
stand the cluster size and distribution of the spike trains associated with each task.
5.4 Results
5.4.1 Raw Multi-Dimensional Aerents Are Value Bound, But State-
Indiscriminable
We rst evaluated the task-dependency of the raw spindle aerent distribution. By plotting the
averaged 200-point aerent distribution of all muscles during each of the ve trajectory cycles, we
can observe the spatial relationship among the Ia and II modalities. Fig. 5.2A displays the full
comprehensive view of the ve 8-dimensional sets of the spindle aerents for the average of the
trajectories in the Forward and Reverse directions. For example, the mean of the resulting aerents
were calculated between the time series of the Circle trajectory in the Forward and Reverse
direction to obtain a single representation for that specic task. For both the Forward and Reverse
plots, there were no spatially discernible clusters that could indicate a state association. However,
there was a recognizable aerent separation of the Iliopsoas group muscle from the remaining
group. This gave an indication that the sensory sets were likely to have a value boundedness
characteristic, or having nite limits based on the musculotendon stretch that occurs throughout
the designated type cycle in the performance of a task. K-means++ clustering analysis was
predictably unable to adequately dierentiate one trajectory from another. Although the data
83
Figure 5.2: Spindle aerent population data for ve distinct trajectories. (A) Full
comprehensive view of all sensory signals. The projections for principal components 1 and 2 for
each muscle are guided by the gray arrows. Overall patterns of dynamic stretch response for the
Ia modality, are shown the left side, with (B) and (D). Overall patterns of the static stretch
response for the II modality.
contained equal-volume clusters without outliers to represent each task, the spherical attribute
and overlapping cluster radii of the data were further evidence that K-means was undesirable for
84
spatial classication. Assuming the number of clustersK is initially unknown, the estimated value
for the number of clusters is K = 3 was a grossly underestimate of the true number of clusters
K = 10 (i.e., 5 trajectories in 2 directions). Since K-means clusters data points purely based on
their geometric closeness of Euclidean distance to the assigned cluster centroid, this analysis fails
at determining state discriminability among the raw aerent set. However, we observed the data
structure had unique boundaries of target values that were respective of the muscle group.
Across all trajectories the spatial distribution in the dynamic stretch response (Ia) for the Forward
and Reverse trajectories is shown in Figures 5.2B-E. Across both modalities, the Vastus Lateralis
spanned the minimal range of aerent spikes in contrast to the other 3 muscles, lending to the
expectation that the Vastus Lateralis may produce less accuracy overall for task classication in
the raw data set. The maximum, median, and minimum values for each trajectory were averaged
together in their respective muscle group to get a sense of the variations of nite space associated
with that particular muscle. Bounded ranges for the dynamic stretch response (Ia), Fig. 5.2B
and Fig. 5.2D, are explicitly listed in Table 5.2.
Table 5.2: Bounded ranges of Ia aerent activity, measured in pulses per second (pps), for muscle
groups averaged across Forward and Reverse directions.
Muscle Name Maximum (p) > Median > Minimum (q)
Anterior Biceps 249.07 > 199.4 > 142.22
Iliopsoas 199.33 > 125.22 > 64.15
Vastus Lateralis 191.18 > 165.5 > 145.77
Semitendinosus 257.17 > 200.55 > 150.50
Following is the static stretch responses (II), where the spatial distribution is shown in Fig. 5.2C
and Fig. 5.2E, are listed in Table 5.3.
85
Table 5.3: Bounded ranges of II aerent activity, measured in pulses per second (pps), for muscle
groups averaged across Forward and Reverse directions.
Muscle Name Maximum (p) > Median > Minimum (q)
Anterior Biceps 185.66 > 130.99 > 62.55
Iliopsoas 130.16 > 45.38 > .28
Vastus Lateralis 119.855 > 92.01 > 67.73
Semitendinosus 194.83 > 132.08 > 72.53
5.4.2 Pre-processing Suggests Observable Correlations in Sensory and
Motor Maps
In order to improve our analysis beyond the limitations of K-means, we invoked techniques for
dimensional reduction on the 8-D high dimensional space to a low-dimensional representation,
which we assumed would retain some meaningful properties of the original raw aerent data.
Interestingly, not only did we nd strong groupings in the pre-processed data sets but there were
observable correlations that exist within the spatiotemporal dynamics of sensorimotor space. In
Fig. 5.3A, the top 3 principal components are plotted the ve trajectories. The projections overlap
each other signicantly and are tightly clustered along the same plane. Most of the explained
variance is captured in the rst 2 principal components. PC1 captures the most variation at
70.22%, PC2 follows with 28.81%, and PC3 captures 0.58%. Fig. 5.3B shows the breakdown for
each component with their individual and then the overall cumulative values. When the principal
components for each individual trajectory were plotted separate from one another, we were able
to perceive discernible shapes that weren't visible, but possibly obscured, in the raw data set. In
Figs. 5.3C-G, PC1 and PC2 revealed projections that closely resembled the prescribed trajectories
and task in the joint kinematic space. Fig. 5.3C, associated with the Line trajectory, reveals a
non-straight line with slight curvature. Fig. 5.3E captures the full revolution of the Oscillatory
task. One half of the task's revolution does not completely trace over the other half, unlike
the Line, but overall aerent response still reveals the sinusoidal shape. The Square trajectory
roughly resembles the the planned trajectory, except the sides aren't quite equilateral and roughly
86
resembles a parallelogram. Some distortion is acceptable here and not indicative of any errors in
the reduction of data. In fact, the results of near-identiable shapes emerging from the principal
components were surprising and not expected, considering the raw data presented clusters oval-like
shapes.
C D
E F
G
Circle
Line
Oscillatory
Lemniscate
Square
−3 −2 −1 0 1 2 3 4
−1
−2
−1.5
−0.5
0
0.5
1
1.5
2
PC1
PC2
−3 −2 −1 0 1 2 3 4
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
PC1
PC2
−4 −3 −2 −1 0 1 2 3
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
PC1
PC2
−5 −4 −3 −2 −1 0 1 2 3
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
PC1
PC2
−4 −3 −2 −1 0 1 2 3
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
PC1
PC2
PC 1 PC 2 PC 3 PC 4 PC 5 PC 6 PC 7 PC 8
0
20
40
60
80
100
Individual
Cumulative
Explained Variance (%)
B
A
Figure 5.3: PCA Dimensions of Aerents Reveal Distinct Shapes. Principal component
plots of pre-processed data revealed shapes that are quantitatively correlated to the planned
trajectory cycles in kinematic space. (A) A three-dimensional PCA plot shows the cluster of
samples based on their similarity, revealing distinctive shapes in space. (B) PCA scree plot of
the variance explained by each of the 8 individual principal components are shown here in blue,
with cumulative percentages show in red. The rst 3 PCs explain 99:61% of the variance. The
two-dimensional plots of the (C) Circle, (D) Line, (E) Oscillatory, (F) Lemniscate, and (G)
Square shapes show more distinction in visual appearance of the trajectory when the PC1 and
PC2 variables were plotted together.
Since the pre-processed data was able to be visualized with 2-3 principal components there was no
need to consider other dimension reduction techniques such as T-distributed Stochastic Neighbor
Embedding (t-SNE) and multidimensional scaling (MDS). Two or three principal components
are usually sucient for our plotting purposes whereas for classication or modeling purposes,
the number of signicant components was can be properly determined using metrics such the
explained variance. Here, we were able to conclude that there is a presence of near-approximate
87
quantitative correlations of joint kinematic and sensory space of the muscle spindle. The next
experimental ndings further use these top three components to determine their usability for state
classication.
5.4.3 Correlation Index Reveals Markers of Action Discriminability,
Classication
Before pre-processing the aerent manifolds to detect useful features, cross correlation was per-
formed on the raw data set to retrieve the correlation coecient or index value that measures
similarity in movements of two time-series sets of data relative to each other. To our dissatisfaction,
cross-correlation analysis, as computed from Eq. (5.13), did not provide sucient discriminability
among the ve states when comparing the raw spindle manifolds. A positive 50%, the measure of
chance, was set as the threshold for verifying discriminability among the span of possible cross cor-
relation values where the value -1 indicates a perfect negative correlation, +1 indicates the perfect
positive correlation, and 0 is no correlation between the paired tasks. Essentially, R
xy
() 0:5
indicates less discriminability among the tasks and R
xy
()< 0:5 indicates more discriminability.
Assessments for cross correlation were divided into 5 sensory aerent groups: combined muscles
set (all four muscles combined), Anterior Biceps, Iliopsoas, Vastus Lateralis, and Semitendinosus.
For each of the n = 45 possible trajectory combinations and pairwise comparisons we plotted
their correlation coecients, R
xy
(), spatial scatter visualization for both the raw aerents and
pre-processed aerents as shown in in Fig. 5.4. All raw data correlations for the combined muscle
were set at the 8-D (i.e., 4 muscles x 2 aerents) high dimensional space while the individual
muscles were compared in 2-D space. All pre-processed data correlations for the combined muscle
were set at the 3-D space while the individual muscles were compared in 2-D space.
88
A
Anterior Biceps Iliopsoas
Vastus Lateralis
Circle-Line
Circle-Oscillatory
Circle-Lemniscate
Circle-Square
Line-Oscillatory
Line-Lemniscate
Line-Square
Oscillatory-Lemniscate
Oscillatory-Square
Lemniscate-Square
Semitendinosus
Combined Muscle Set
Less
Discriminability
More
Discriminability
B
C
D E
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
−1 −0.5 0 0.5 1
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
−1 −0.5 0 0.5 1
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
−1 −0.5 0 0.5 1
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
−1 −0.5 0 0.5 1
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
−1 −0.5 0 0.5 1
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
−1 −0.5 0 0.5 1
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
−1 −0.5 0 0.5 1
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
−1 −0.5 0 0.5 1
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
−1 −0.5 0 0.5 1
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
FF
FR
RF
RR
−1 −0.5 0 0.5 1
Cross Correlation
(Raw Data)
Cross Correlation
(Raw Data)
Cross Correlation
(Pre-processed Data)
Cross Correlation
(Raw Data)
Cross Correlation
(Pre-processed Data)
Cross Correlation
(Pre-processed Data)
Cross Correlation
(Raw Data)
Cross Correlation
(Pre-processed Data)
Cross Correlation
(Raw Data)
Cross Correlation
(Pre-processed Data)
Figure 5.4: Spread of Discriminability Within Cross Correlation Scatter. Cross Correla-
tions of all possible trajectory combinations (n = 45) for the raw information and pre-processed
aerent signal information, which are plotted on the left and right, respectively. Each trajectory
pairing has an assigned direction. An R label is for Reverse Direction and an F label is for Forward
Direction. For example, the Circle-Line pairing with label FF indicates Circle going Forward and
Line going Forward. The Line-Square pairing of RF indicates Line going in Reverse and Square
going Forward. Cross-correlations were plotted for (A) the combined set of the 4 muscles in the
cat limb, (B) Anterior Biceps, (C) Iliopsoas, (D) Vastus Lateralis, and (E) Semitendinosus.
89
Let be the span or bandwidth of the detected correlations in the cluster and be the total
space of possible correlation values, where = is the percentage covered by the correlations
values. Raw data correlations in the combined muscle set (Fig. 5.4A) show a tight cluster within
a 32.1% (i.e., where is .646 and is 2) of the full correlation range. However, that range
expands to 98% (1.958/2) in the pre-processed set as more pairings move away from being less
discriminable to more discriminable. This form of expansion was not only evident in the combined
muscle grouping but also in the individual muscle groups (Figs. 5.4B-E). Out of the four muscles,
the Vastus Lateralis (Fig. 5.4D) contained the most compact clustering in the raw set with the
maximal expansion, spanning a minimal 23% (.459/2) and expanding to 98% (1.96/2) in the
pre-processed set. We highlight the compact-to-expansion dynamic that occurs from raw to pre-
processed aerents to show the usefulness of pre-processing in giving each task more distinction
and separability to enhance classication. Furthermore, the usefulness of cross-correlation is
additionally investigated in this study in the context of a state classier. We nd the display of
confusion matrices as heat maps particularly useful here because of the ability to describe the
performance of our classication model. You can observe any patterns in value for one or both
variables by observing change in color gradients of cell colors change across each axis in Fig. 5.5.
The dierences in discriminability vary signicantly by each matrix. Our combined muscle set
reports 29% discriminability among the 10 possible trajectories in the raw 8-D set and drastically
increases to 82% in reduced 3-D pre-proccessed set, as shown in Fig. 5.5A. For the Anterior
Biceps muscle (Fig. 5.5B), cross correlation reports 60% discriminability in the raw 2-D set and
increases to 73% in 2-D pre-proccessed set. The Iliopsoas muscle (Fig. 5.5C) reports 66% and
73% discriminabilities, Vastus Lateralis (Fig. 5.5D) reports 0% and 78% discriminabilities, and
Semitendinosus (Fig. 5.5E) reports 49% and 89% discriminabilities, for the raw and pre-proccessed
set, respectively. The dierence in correlation between the two groups (i.e., Raw and PCA) was
determined to statistically signicant (p = 0.001), according to the Wilcoxon signed-rank test, for
the combined muscle group and individual muscle sets.
90
91
Figure 5.5: Confusion matrices of raw and pre-processed spindle aerent data. For (A)
the combined set of muscles and spindle modalities (four tendons as shown in Fig. 5.1), cross
correlation analysis was performed on the raw aerents across all ve trajectories; CI: circle, LI:
line, OS: oscillatory, LE: lemniscate, SQ: square. The direction of each task is label either fwd for
forward and rev for reverse. The lower half of matrix entries correspond to the raw collection of
pps signals. Upper half entries represent the resulting principle components from PCA, labeled
as pca. Three principal components were selected here. While in individual muscle analysis such
as (B) Anterior Biceps, two principal components were selected. The same cross correlation was
performed for the remaining muscles which include (C) Iliopsoas, (D) Vastus Lateralis, and (E)
Semitendinosus.
5.5 Discussion
In this chapter, we focused on using high-dimensional proprioceptive signals from muscle spindles
to enable robust estimates of trajectories as pre-dened tasks and bodily states. Before our trials
and analysis, it remained unknown whether spindle signals suce to discriminate limb movements.
We obtained three main ndings that we discovered in the post analyses. The rst is identifying
that raw multi-dimensional sensory sets of the muscle spindle are value bound, but are still able to
maintain state-indiscriminability. Secondly, pre-processed data shows high correlation of spatio-
temporal maps between sensory and motor space. Thirdly, the correlation index revealed markers
of sucient discriminability and classication among spindle aerents. Our ndings closely match
with similar results from Rongala et al. [152], where biological data on cuneate nucleus neuron
recordings in adult cats were obtained and modeled to study generalizable tactile representations.
Their work highlights that the cuneate nucleus forms the rst interface for the sense of touch
in the brain. We conjecture this would be similar for proprioceptive sensory aerent pathways.
Triangle matrices of correlations, similar to our analysis, demonstrated how weighted learning
in the cuneate nucleus resulted in decorrelated responses between neurons of the same stimu-
lus. Essentially this means the data were less "confused" with another and more discriminable.
Altogether, our ndings indicate that sensory aerents from the muscle spindle can adequately
supply the nervous system with features of discrimination to distinguish one task from another,
only if there is suitable forms of pre-processing or ltering to reduce the overwhelming amount
92
of sensory manifolds
ooding the nervous at a given time during the performance of an action or
task. This is a desirable and necessary result for our dynamic model of body representations or
body schemas in neuro-robotics systems.
93
Chapter 6
Sensory-Motor Gestalt: Sensation and Action as the
Foundations of Identity, Agency, and Self
\To understand is to perceive patterns." -Isaiah Berlin
6.1 Abstract
Body movement and proprioception are inextricably linked. Movement produces continuous high-
dimensional ensembles of aerent information that provide an internal proprioceptive body rep-
resentation and its relationship to the environment. Motor function is amenable to recording
and interpretation and has been relatively well studied. However, we do not yet understand how
physiological proprioceptive aerents contribute to internal body representations, neuromuscular
control, and even a sense of agency and self. Proprioceptive and motor signals have often been
seen as separate, and to be combined mainly to close feedback loops for neuromuscular control.
In contrast, `active sensing,' is an emergent concept for dynamically blending sensory and motor
signals. We extend and formalize active sensing into an integrative approach|{born out of a
neuromechanical perspective|that sees proprioceptive and motor signals as integral parts of the
same functional and perceptual continuum we call the Sensory-Motor Gestalt. The Sensory-Motor
94
Gestalt combines formalisms of physics, state estimation, biomechanics, dierential geometry, and
physiology to understand the emergence of the self in the context of proprioception and motor
actions in the physical world. Proprioception, by dening body state, denes feasible (continuous
or discrete) motor actions compatible with that state and the environment. Conversely, motor
actions produce subsequent, often predictable, body states. This syntactical relationship leads to
an epistemological continuum that spans body state, feasible behavior, agency, identity, and sense
of self in organisms and robots.
6.2 Introduction of Sensory-Motor Gestalt: Origin and
Denition
Our computational model of the self begins with Gestalt Theory. Gestalt (pronounced g@-'sht alt),
a concept originating in Austria and Germany, roughly translates to shape, form, conguration,
and unied whole. XX-century German psychologist Max Wertheimer utilized this denition to
originally present the Gestalt laws (or principles) of grouping for pictorially detailing how the
human eye perceives visual elements [186]. These laws are fundamental rules illustrating how
humans recognize elements and objects in their visual scene as organized patterns with meaning.
The Gestalt theory of the mind and brain intends to form an understanding of how humans and
animals 1) comprehend what they are perceiving and 2) obtain meaning from the world with
disordered visual stimuli.
In its original formulation, Gestalt laws aim to reduce complex visual scenes into simpler, less
complex shapes to can convey an image's meaning in a single formation instead of disparate
smaller elements. Hence by being a critical aspect of the perception of patterns into a coherent
whole for context and meaning, Gestalt plays an important role in combining epistemology (i.e.,
study of knowledge and how does one come to \know") and ontology (i.e., study of what is the
95
nature of the self) [70]. This paper proposes underlying mechanisms for brain-body dynamics to
merge proprioceptive and motor elements into an epistemological continuum from sensory and
proprioceptive input, to state of the body, to feasible motor action, to useful behavior, to the
sensory consequences of action| and then on to more abstract notions of agency, identity and
sense of self in organisms and robots.
In this study, we focus on proprioception as spike trains from muscle spindles (II and Ia) and
Golgi tendon organs (GTOs). They primarily encode muscle ber length and velocity, and ten-
don tension, respectively. These proprioceptive signals are known to inform various perceptual
modalities of body state (e.g., postures, movements, forces, limb stiness, alertness). Recent
exponential growth in literature relating proprioception to subjective experiences (Figure 6.1)
may suggest that the debate about the emergence of the self is advancing. We hypothesize that
Gestalt laws can be applied to organizing these physiologically-tenable proprioceptive signals to
construct a totality of what is perceived as the active body. We seek to do the same for motor
actions by developing a mathematical description of the set of plausible motor actions conditional
on proprioceptive signals.
Our prior simulation work characterized the high-dimensional, non-linear, time-varying manifolds
of muscle spindle aerents (Ia and II encoding, roughly, muscle ber contraction velocities and
lengths, respectively) that emerge during movement of a planar multi-muscle limb [8]. We provided
initial evidence that high-dimensional muscle spindle proprioception denes limb states that re
ect
the consequences of motor actions. We now extend that work by emphasizing that each limb state,
by its physical nature, only has a well-dened set of feasible motor actions. This results in a formal
sensory-to-motor-to-sensory map that denes both the current proprioceptive states and feasible
motor commands (i.e., plausible motor actions) that will lead to new (but expected) proprioceptive
signals. The Sensory-Motor Gestalt applies to both biological and engineered agents where the
96
Figure 6.1: Published article count per year that indicate association between subjective experi-
ences and sensory modalities. This chart similarly models a previous search conducted by [45], in
which the number of articles published mentioned the words \awareness" or \consciousness" in
conjunction with words denoting each sensory modality: \Visual" or \vision" (magenta), \propri-
oception" or \movement" (yellow), \auditory" or \audition" (red), \touch" or \tactile" (cyan),
\olfaction" or \olfactory" (green), and \multisensory" (black). Our PubMed search extended the
year range from 1950-2019 and added proprioception. Along with vision, proprioception showed
a signicant increase in documented work.
concepts of state, observability, and controllability are intimately related; therefore, providing a
basis for constructing an articial core of state, agency, identity and ultimately self.
To our knowledge, this is a rst attempt to formally apply the laws of Gestalt to the encoding
of the sense of agency, identity and self via proprioception. This article rst builds the concept
of Sensory-Motor Gestalt from the generic Gestalt theory. We then interpret the Sensory-Motor
Gestalt in the form of mathematical encoding for each of the core laws, which may be integrated
to form the sensorimotor self. How the self is directly related to sensorimotor experiences of
neuromuscular systems is then explored while providing sample platform applications to support
Sensory-Motor Gestalt, both biological and in bioinspired robots.
97
(a)
(d)
(e)
(b)
(c)
Law of Continuity
Law of Proximity Law of Similarity
Law of Closure Law of Prägnanz
Figure 6.2: Gestalt laws of perceptual organization for topological manifold data can be applied
using these core laws: (a) Law of Proximity aims to group elements together based on spatial
closeness. Black dots, red dots, and green dots are perceived as separate groups due to the nearness
of columns. (b) Law of Similarity groups the elements of black dots and red squares as separate
sets, although spatial distance between each element is consistent. Shape, orientation, and color
are the distinguishing factors. (c) Law of Closure prompts pattern perception of a green square
and black oval despite the non-continuous outline and presence of gaps. (d) Law of Continuity
perceives the gure as a green dotted line and a separate black dotted line due to the observed

uid connection of continuity and direction. (e) Law of Pr agnanz (i.e., pithiness, conciseness,
or Good Form) takes the abstract shape, as depicted on the left, and perceptually reorganizes
them into a simple, more recognizable forms as is depicted on the right with the colored circle,
triangle and square.
6.3 Sensory-Motor Gestalt: Applying Gestalt Laws to
Sensorimotor Function
\The whole is greater than the sum of its parts" is the popular adage Gestalt psychology is best
known for. It emphasizes the fact that although a sensory experience can be disassembled into
individual components (i.e., stimuli), the way in which those components coalesce together gen-
erates properties and qualities of the whole that only exist independently of their components.
Stimuli patterns presented as a whole often prompts a more meaningful perceptual response. As
alluded to earlier, Gestalt theory is typically associated with the visual sense and visual percep-
tion (e.g., object and shape recognition, coloring, arrangement of parts) that is used to process
98
graphic designs and images (Figure 6.2). Rarely has Gestalt theory been applied to other sensory
modalities such as haptic [27], auditory, and olfactory senses which can all be topographically rep-
resented on a multi-dimensional space in the depiction of manifolds (i.e., coherent and continuous
lower-dimensional subspaces embedded in a higher-dimensional space). Interestingly, the func-
tional mechanisms of Gestalt laws are active in other cortical areas of the brain and not solely in
the visual processing centers. As the brain's neural processing is responsible for stitching together
the visual scene of the external world in the primary sensory cortices and also seamlessly binding
raw multisensory information to project a single unied experience, we theorize there are benets
in extrapolating ideas of Gestalt laws of perceptual grouping from vision to other modalities of
the body such as the somatosensory system. Gestalt theory typically consists of ve core laws
that govern the fundamental organization of perception: Laws of Proximity, Similarity, Closure,
Continuity, and Pr agnanz.
6.3.1 Law of Proximity
Within visual perception, objects in space or points on a plane that are near or proximate to each
other have a tendency to be grouped together in a single unied set. Conversely, points that are
further apart have a lesser likelihood to be viewed as conjoined (see Figure 6.2a). This law is
useful for organizing information with increased speed and eciency. There are several ways in
which proprioceptive information can also be processed to yield proximity metrics. One of the
earlier attempts to address the Law of Proximity is the Pure Distance model [99], which attempts
to quantify visual proximity grouping in dot lattices with an attraction function that measures
the probability distribution of grouping.
Several algorithms can process proprioceptive stimuli in this manner. Consider our prior work
[8] on the simple case of spindle model output of a single muscle ber, which is 2-D Ia and II
aerent spike trains over time. Figure 6.3 shows a higher-dimensional case of a simulated human
99
arm. When examining proprioceptive signals that are encoded as spike trains in units of pulses
per second (pps), we are presented with unlabeled sample points (x
1
;x
2
;:::;x
n
), where n is the
set of observations, that can be further mathematically expressed to form representations. Since
the notion of proximity is to associate observed points by measurement of Euclidean distance,
then a standard unsupervised algorithm such as the K-Means clustering (i.e., a simplied version
of vector quantization) proves to be sucient for revealing underlying data structure.
Figure 6.3: Spike trains from spindle aerents produce an evolutionary high-dimensional time-
varying manifold of raw aerent information that is distinct for dierent arm movements. Using
parallel coordinates, we show the Ia Group Aerent in 50,000 time samples for the case for a 6-
muscle, 2-joint simulated planar arm performing the Lemniscate (gure-of-eight) trajectory with
the end point [8]. The coordinates are colored according to the segmented locations within the
duration of the Lemniscate trajectory. The shadow boxes to the left and right of the manifolds
are scaled-down sample snapshots of the data for the Deltoid Anterior and Anconeus muscles,
respectively; ultimately revealing their specic cluster ranges.
6.3.2 Law of Similarity
Elements (e.g., points) that are similar in visual appearance in at least one degree with alike
components are more likely to be grouped and organized together perceptually. The Law of
Similarity generally spans the attributes of orientation, texture, color, and shape (see Figure 6.2b).
There are ways to apply this law to the manifolds produced by proprioceptive signals. Considering
the contours and curves that emerge from the collection of proprioceptive manifolds (e.g., Figure
6.3), shape is the most applicable attribute when measuring for similarity. Shape dimensions, such
as curvature and elongation, can be perceived as integral dimensions and also used for comparison
for similarity. In a similar fashion that [18] quanties geometric similarity of anatomical surfaces
100
and morphological identication, we can apply statistical analysis when viewing the Ia and II
stimuli as a collection of discrete or continuous points on an anatomical surface. Measures of
similarity of each aerent signal across various tasks can be applied across the collected time-
series data using signal processing. Comparable to the Law of Proximity, K-means clustering may
also be used here if measuring 'similarity' of clusters by its relation to Euclidean distance of data
points.
cos =
!
x
!
y
k
!
xkk
!
yk
=
P
n
1
x
i
y
i
p
P
n
1
x
2
i
p
P
n
1
y
2
i
(6.1)
In our example, let x and y be two vectors of aerent spike trains, Ia and/or II. The cosine
similarity function is a measure of similarity that can be used to compare aerent signals in
the inner product space. Using the cosine measure, we have Eq. (6.1) where
!
x
!
y =
P
n
1
x
i
y
i
=
x
1
y
1
+x
2
y
2
++x
n
y
n
is the dot product of the two vectors. A cosine similarity, cos, value closer
to 1 indicates a higher propensity for perceptual clustering along the manifold. The convolution
function would be another choice that quanties similarity over time for all possible lags between
signals.
Another option is cross correlation. It compares the time-series of aerent data across tasks, and
is represented as the ratio in Eq. (6.2), wheren is the total number of data point indices recorded
per task cycle. Both x
i
and y
i
are the individual spindle aerent sets, Ia and II, respectively.
A temporal shift delay, phase lag , of the output cross correlation, R
xy
, measure is applied to
determine where the correlation of the data is maximized, as shown in Eq. (6.3).
R
xy
() =
P
n
i=1
(x
i
x)(y
i
y)
p
P
n
i=1
(x
i
x)
2
P
n
i=1
(y
i
y)
2
(6.2)
101

estimated
= arg max
2R
(R
xy
()) (6.3)
Magnitude-squared coherence is similar to correlation except that signals are compared in fre-
quency !, instead of time space, as shown in Eq. (6.4), which values satisfy 0 C
xy
(!) 1.
S
xy
(!) represents the cross-sprectral density between x and y, while S
x
(!) and S
y
(!) are the
autospectral densities for their respective signals.
C
xy
(!),
kS
xy
(!)k
2
S
x
(!)S
y
(!)
(6.4)
S
xy
(!) =
Z
1
1
R
xy
(t)e
j!t
dt
=
Z
1
1
Z
1
1
x()y( +t)d

(6.5)
Lastly, Kullback{Leibler (K-L) divergence is a means to quantify the likelihood that the statistics
of a given process are similar to that of another, Eq. (6.6). Probability distributions P and Q
are measured in comparison to reveal the relative entropy. This is particularly useful because it
measures how much information is lost when we approximate distributions.
D(PkQ) =
X
x2X
P (x) log

P (x)
Q(x)

(6.6)
6.3.3 Law of Closure
The Law of Closure is the tendency to complete unnished or partially obscured objects. Here,
incomplete gures are seen as complete or whole as depicted in Figure 6.2c. Warshall's Algorithm
[185] may address this through its approach in computing the transitive of a node relation in
a graph. We can envision, that as clusters are being formed via other laws, state nodes will
102
eventually emerge from the aggregate data. To establish state transitions from one aerent cluster
to another, the Warshall algorithm can determine whether a vertex j is 'reachable' from another
vertex i for all vertex pairs within the graph. This measure of reachability will serve as the
transitive closure, indicating directions and where paths exist for point-to-point movement across
the manifolds.
This law states that, given available information, there is the expectation (based on prior personal
experience) of closure when a fragmented version is presented. Bayes' Rule is a formal way to
represent such expectation in the case of visual information, visuomotor perception [97], and now
proprioception. Bayes' rule states that we can obtain the posterior distribution (the probability
of a given body state given current proprioceptive input p(x
true
jx
sensed
) by taking into account
the likelihood distributions of the prior (i.e., the cumulative information from prior experience)
and the evidence (i.e., the current proprioceptive input x
sensed
):
p(x
true
jx
sensed
) =p(x
sensed
jx
true
)
p(x
true
)
p(x
sensed
)
(6.7)
where p(x
sensed
jx
true
) is the likelihood of a particular proprioceptive input x
sensed
when the
perceived body state really is true. This then allows the inference of the current body state given
past experience and incomplete or polluted proprioceptive inputs.
6.3.4 Law of Continuity
Objects and points that are co-linear and follow the same direction will be grouped together as
a whole (see Figure 6.2d). We can construct procient continuations between neighboring local
environments. Density-Based Spatial Clustering of Applications with Noise (DBSCAN) identies
outliers as noises. The Mean-shift algorithm, Eq. (6.8), actually includes them in the cluster
despite dierences of the data point. DBSCAN also does not require a pre-set number of clusters,
103
and discovers arbitrarily shaped clusters. These are key facets for analysing proprioceptive data.
Given the manifolds of aerent information for natural movements are usually continuous, then
the Law of Continuity would naturally apply as the manifold during a movement continues along
a particular path, even if temporarily disrupted or occluded by a perturbation. In practice, Bayes'
Rule is a way in which such expectation of continuity can be quantied.
6.3.5 Law of Pr agnanz (Good Form, Clarity)
The Law of Pr agnanz focuses on simplicity and will prompt visualizations according to the simplest
way of grouping items. We perceptually organize shapes to simple forms, as in pithiness. The Law
of Pr agnanz is the tendency to interpret ambiguous images as simple and complete vs. complex
and incomplete. An example is how shapes overlapping each other can cause ambiguity, as shown
in Figure 6.2e. A potential resolution is an iterative method such as Mean-shift Clustering,
Eq. (6.8), where N(x) is the neighborhood of the set of points, x. Depending on the Gaussian
kernel bandwidth, Eq. (6.9), the Mean-shift algorithm iteratively shifts points until there is a
convergence of partitioning the clusters into semantically meaningfully groups. This is probable
to work well with proprioceptive aerents as it may account for the noise in signals which is
expected, and necessary for physiological function.
m(x) =
P
xi2N(x)
K(x
i
x)x
i
P
xi2N(x)
K(x
i
x)
(6.8)
K(x
i
x) =e
ckxixk
2
(6.9)
Dimensionality reduction is probably the most commonly applied (and potentially misinterpreted)
analysis of high-dimensional motor signals [102]. It is simply a way to quantify whether a high-
dimensional ensemble of signals evolves (i.e., has variance) along all dimensions equally, or inhabits
a lower-dimensional subspace. Conceptually, it is just the singular value decomposition of a
104
covariance matrix, where the number of `large' singular values (principal components) quanties
the rank of the covariance (the eective `dimensionality' of the data), and the left singular vectors
(principal vectors) form a basis for those dominant variances (the basis for the eective subspace
the data inhabit). Independent Component Analysis and Nonnegative Matrix Factorization is a
variations on this idea that do not require orthogonality of the basis vectors, and the latter also
imposes a non-negative constraint on the elements of the basis (as neural signals are conceptualized
as intensities or spiking frequencies that are > 0). It is good to see that some work is beginning
to be done on dimensionality reduction in tactile aerents, which are famously dicult to record
from even in animal preparations [152]. Our current work is beginning to apply dimensionality
reduction to higher-dimensional simulated proprioceptive signals [6].
6.3.6 Supplementary Laws
Other Gestalt grouping laws that can be applied to sensory stimuli integration include the Laws
of Focal Point, Symmetry, Common Fate, Common Region, Synchrony, Convexity, Isomorphism,
Parallelism, Unity, Element Connectedness, and Figure vs. Ground.
6.4 Functional Utility of the Sensory-Motor Gestalt
Figure 6.4 describes our working hypothesis of the Sensory-Motor Gestalt in operation. At any
time point, proprioceptive (and other sensory) information dene a state of the body that lies
within a particular manifold of like inputs (Laws of Proximity and Similarity) and feasible next
states (Laws of Continuity and Closure). Such body state allows feasible transitions to `next'
proprioceptive states via feasible motor actions that will lead to a, usually predicted, new body
state (Laws of Continuity and Closure).
105
Figure 6.4: We envision the representation of minimal self as a collection of categorized states in
R
N
space formed from sensory and motor maps, and made useful by the agency they provide.
Our data-driven projection method categorizes the set of feasible inputs from muscle spindles
for each specic task performed (i.e., arm reach, sit, squat, standing) as a manifold. Transitions
from one state to another occur through point-to-point transitions along the manifold. The high-
dimensional space of aerent modalities has an underlying structure given by the anatomy of the
body and the physical transitions it can undergo such as changing postures via self movement.
6.5 Abstracting Self from Sensorimotor Experiences for
Neuromuscular Systems
Now let's examine how the foundations of neuromuscular systems can provide context to con-
structing the minimal operative self via proprioceptive signals. In Nature, proprioception pro-
vides animals with awareness of the state of their body and of their relation to the environment.
106
Figure 6.5: The neuromechanical perspective of how sensory inputs are transformed to motor
outputs (adapted from [178]). A Feasible Sensory Set (FSS) denes the aerent stimuli that
are plausibly detectable for a given state of the body (i.e., joint posture, force production, and
kinematic task). By incorporating the in
uence of proprioceptive space via neural spike rings,
an under-constrained mapping of transformations can be reinforced from neural motor commands
in the Feasible Activation Sets (FAS) to mechanical outputs (limb movements).
Proprioceptive signals arise from mechanoreceptors that re
ect the state of tissues, which are
driven by muscle forces, joint and body postures, and skin deformations. When integrated with
other sensory modalities, this re
ection of body state at any given moment in time and space
provides the nervous system with an overall representation of bodily position, actions, and task
experiences. Neuroscientists have long been intrigued with how the brain represents the body and
forms models of bodily states through proprioception [69]. However, there is still no consensus
regarding how these representations, facilitated by multi-muscle control, compartmentalize and
process high-dimensional aerent information as continuous feedback for ongoing tasks.
The fundamental formulation of a control law for a linear system (without loss of generality) is
_ x =Ax +Bu (6.10)
where the outputsy (and therefore sensory and proprioceptive signals) are a function of the state
x and the control signals u
y =Cx +Du (6.11)
By denition, the equations of motion (i.e., _ x =Ax) are an important determinant of the feasible
transitions away from any given state. Moreover, changes in sensory and proprioceptive signals
107
(a) (b)
n n
Tendon
Routing
Figure 6.6: Test-bed applications for Sensory-Motor Gestalt implementation in tendon-driven
systems. The Gestalt can provide state cases and their feasible transitions along the manifold
for agents. (a) Proposed neuromorphic cat-like robot and limb in hardware. Image adapted
from [116]. (b) Various hand and nger states depicted using the American Sign Language. The
nervous system registers the proprioceptive feedback generated with each hand shape is unique
to each letter signed with gestures.
are driven by changes in state (i.e., y =Cx). This is a formal way to conceptually anchor some
aspects of the Sensory-Motor Gestalt. Please note we do not claim or endorse that the engineering
concept of `state' applies to biology. But the Sensory-Motor Gestalt is a formal way to describe
how the stream of sensory and proprioceptive signals is useful to biological behavior in a way that
is agnostic to how those signals are processed.
We can conjecture how the nervous system processes incoming aerents (e.g., proprioceptive sig-
nals) by observing how neural activation commands mathematically map to mechanical outputs,
as shown in Figure 6.5. Neural commands simply refer to the nervous system's distribution of
excitatory impulse signals to activate muscle tissue. For tendon-driven limbs, [178] emphasizes
that the nervous system's primary function is to use (i.e., learn, explore, and exploit) the set
of feasible neural commands from the optimized activation space with dimensionality of vector
108
a2 R
N
, where N is the number of independently controlled muscles. From activation space,
vector spaces are successively mapped to muscle force space, to joint torque space, then lastly to
output wrench space to produce a set of feasible mechanical outputs (forces and movements).
In prior work [8], we have extrapolated this perspective of muscle redundancy to feasible sets of
proprioceptive signals, called Feasible Sensory Sets (FSS). These are dened by a body's anatom-
ical structure and the mechanical tasks being performed. Here, we rst introduced the concept
of trajectory-specic proprioceptive manifolds, which are the unique multidimensional and time-
varying combinations of aerent signals that obligatorily emerge during a limb movement. We
demonstrated that a given movement gives rise to a distinct sensory manifold embedded in the
12-D space of spindle information that is largely independent of the choice of muscle coordination
strategy. These are referred to as manifolds because they are a systematic collection of points
(i.e., spindle neural spikes) that provide information for its control.
Following this work it remained unknown whether spindle signals suce to discriminate limb
movements. We used a 4-musculotendon, 2-joint cat hindlimb model to simulate muscle spindle
length and velocity signals (II and Ia, respectively) during repeated cycles of ve distinct end-point
movements, similar to the manifolds in Figure 6.3. In [6], we concluded that proprioceptive infor-
mation can usefully discriminate limb states|but only after conducting minimal pre-processing of
high-dimensional multi-muscle ensembles to low-dimensional subspace components. This nding
may this explain the documented subcortical pre-processing of aerent signals of various mam-
mals [152]. It is this resulting set of constrained sensory signals that we believe could suce as
a minimal representation of the articial self and should be incorporated into the Sensory-Motor
Gestalt paradigm. We project the usefulness of Sensory-Motor Gestalt to be a suitable core to
execute on dierent applications that utilize neuromuscular dynamics, incorporate neuromorphic
and bio-inspired architectures, and classication of human bodily states (Figure 6.6).
109
6.6 Role of Self and Identity in Autonomous Robotics and
Synthetic Biological Agents
A semblance of selfhood, identity, and agency should be expected outcomes for constructing a
dynamic sensorimotor representation [171]. For robots, concepts of identity are typically viewed
as a necessity for interactions in social environments [43]. For humans, the self and identity
combination are purposed for storing the traits, stereotypes, characteristics, and roles they play
in social settings. What features constitute a person's self? How do disparate sensory perceptions
cohesively fuse together to form a singular experience of self? Although these questions are
typically addressed within the human scope, we can also apply these inquires to autonomous
robots that are bioinspired and create their own experiences with action.
It is our opinion that sensorimotor contingencies (discussed in Related Work) do not achieve their
full potential if solely used to estimate error signals in closed-loop controllers. We believe one can
ask the extent to which these contingencies facilitate and embody a self, re
ect an identity, and
activate agency needs to be thoroughly explored. Self and identity are often used interchangeably
to encapsulate the entirety of individual's behavior, character, and the restricted contextual con-
straints in which they operate within. However, it is important to clearly know the distinctions
of these terms if we're determined to adequately construct models that emulate their functions.
Agency is known as the control of intentional actions and volition; leading to the ability to
plan and action ownership. For the purpose of our study we distinguish the self and identity
according to Oyserman's [136] conceptualization. It is thought best to consider self, self-concept,
and identity as nested elements: self is the top-tier construct, self-concepts reside within the self,
and identities reside within self-concepts. Oyserman denes self as the ability to consider oneself
as an object. The self maintains re
exive capacity that is able to direct an agent to what is \me";
it is the focal point of personal account and a reference for anchoring temporal sequences of events
110
(e.g., memory recall). Identities are \content and readiness to act and employ mindsets to make
meaning." Personal identities are the traits, characteristics, values and goals belonging to the
agent. Altogether self and identity are mental concepts, social products, and forces of action. As
Oyserman states, what makes this nested unit interesting is that they appear to predict behavior
over time. What is not fully understood by many in literature is how this happens.
6.7 Related Work
Further research into the topic of dynamic sensorimotor representations led us to original work on
the sensorimotor contingency theory [134], which has motivated an assortment of studies in the
area of human perception as it relates to understanding the nature of actions and their sensory
eects. Sensorimotor contingencies derived from the notion that vision should be treated as an
environmental exploratory activity. According to [77], sensorimotor contingencies spawned multi-
disciplinary projects that investigated how to model the action-sensory relationship of robotic
systems, which spanned the manipulation, classication, and categorization of external objects.
The researchers view the goal for most of these studies as autonomous robots learning skilled
behaviors via learning the structure of complex sensorimotor spaces and how actions aect the
environment. Despite these contributions, [20] believed there have been few attempts to formally
dene sensorimotor contingencies, which they view as a prerequisite for testing this approach via
models and empirical study. The sensorimotor contingencies view on perceptual awareness have
also been criticized for lacking a suitable foundation in the biology of autonomous agency. Prior
work on building computational approaches of body representations, self, identity and Gestalt
have been attempted with a grounding in minimal embodied, psychological and cognitive aspects
[56]. One drawback of past implementations is that they're unencumbered with understanding the
manifold of feasible transitions, which therefore leads to the incorrect perspective that any action
is permissible. Our approach addresses how the sensorimotor self should constrain one's agency
111
and perceptual space to feasible tasks. Attempts to create sensory-to-motor maps as a body
representation (i.e., body schema) have been accomplished by achieving robotic self-recognition
using a dynamic Bayesian network [65], online learning of arm reaching motor maps for humanoid
robots using open and closed loop control [62], body representations as cross-modal map learning
of invariance in multi modal sensory data [190], and estimation of a kinematic model for serial
robots [117]. To our knowledge, our approach is the rst that considers the inherent link among
the feasible capabilities of the body, the feasible sensory information that will emerge, and the
physics of the world as the manifold dening agency, and therefore delineating the concept of self.
6.8 Conclusion
The emergence of self and its role in biological and articial agents continues to be a subject
of debate across many disciplines; leading to the perception that there is a lack of congruence
among perspectives. The absence of a unied concept of self presented us with an opportunity
to propose a sensorimotor mechanism by which the self can emerge, via the Gestalt laws of
perceptual organization, in the context of articial systems operating in the physical world (i.e.,
robots). This enables us to investigate the foundation of self, identity, learning, and agency as the
multifaceted interplay of proprioception and action while exploring their implications to autonomy.
The emergence of self through sensorimotor interactions has applications ranging from a self-other
distinction to `social' systems for robot-human and robot-robot interactions. Traditionally, self
and identity are considered to be theoretical concepts, social constructs, and therefore enablers
of agency. We visit these concepts in reverse order to propose that it is sensorimotor agency that
can enable the emergence of self and identity|which is an evolutionarily plausible order of events
[188]. The Sensory-Motor Gestalt provides a solid foundation to enable such cross-fertilization
to move towards the creation of truly autonomous and versatile robots, and promote advances in
articial intelligence.
112
Chapter 7
Conclusion
This work's long-term goal is to equip a robot with an emulated nervous system (i.e., a Neu-
RoBot) that forms a repertoire of physiologically-inspired sensory and motor couplings to explore
and exploit physical actions and transitions among them. For this thesis, we have focused on
building the software framework of the sensorimotor couplings, which can then be merged with
the hardware plant of the model. From our initial hypothesis, we anticipated that the physiolog-
ical sensory signals that result directly from a range of immature to skilled motor actions would
suce to create a self-generated body representation that prompts learning of useful, functional
actions and potentially evolving behaviors. We believe the work successfully met the criteria for
forming useful body representations for tendon-driven systems such as the NeuRobot.
Using various computational methods, we demonstrated the plausibility that an individual body
representation of the self (i.e., body schema) for neuromuscular-driven robotic systems can be
self-generated from:
Dening feasible behaviors and muscle activation patterns that produce task-oriented limb
mechanics.
113
Using computational integration of multimodal, spatiotemporal sensorimotor contingencies
of physiological realistic (articial) proprioceptive aerents.
Quantifying somatosensory modalities: Muscle spindle for posture, GTO for muscle force.
Encoding multisensory mathematical formalisms for system-specic state space and actions.
In this thesis, we addressed the body senses that arise (e.g., muscle spindle aerents and visual
cues) to enable future studies investigating how those same sensory signals aect the representa-
tion of our physical self and the actions we make. First, I demonstrated that a given movement
gives rise to a distinct sensory manifold embedded in the 12-D space of muscle spindle information
that is largely independent of the choice of muscle coordination strategy. Given that musculoten-
don lengths and velocities are fully determined by joint kinematics, such manifolds provide a rich
set of information to use in its control. Secondly, I show that high-dimensional multi-muscle pro-
prioceptive ensembles can usefully discriminate limb states and be utilized as a suitable classier
for inter-trajectory comparisons|but only after minimal pre-processing. Lastly, we present the
Sensory-Motor Gestalt, which through computational approaches, demonstrates how the syntac-
tical relationship between motor actions and bodily sensory states can lead to an epistemological
continuum that streamlines the body state into useful behavior for constructing the foundations
of agency, identity, and sense of self in hybrid robots and synthetic biological agents.
To conclude, let's discuss some of the challenges that were undertaken and how we intend to map
out contingency plans for future research directions. We sought to address an underlying question:
how sensory signals shape the motor choices we make throughout daily life activities or in simple
task actions? From our past research experiments, we have learned that every limb movement
is associated with a unique set of specic sensory states, which we call the Feasible Sensory Set
(FSS). We must consider how the nervous system obtains and processes sensory data to create
a body sense that interacts with explicit or implicit internal models of the body and external
114
in
uences on the body. Distinguishing oneself from the environment and having an introspective
representation of self is a fundamental, biological challenge that the CNS of animals and humans
must encounter and solve daily. Within that self-representation is a sense of ownership, that
our actions, behaviors, and bodies belong to a certain individual, which is key for survival and
performance. Although there were once only relatively few studies on the awareness of one's body
[44], increasing volumes of research over the years has allowed us to make a commitment to this
thesis and fathom the realistic plausibility of creating self-aware devices.
The novelty in our work focused on addressing the remaining important challenges and limitations
in self-modeling based on the spatial constraints of states, the context of multisensory integration,
and control for exploration. Also, the distinction in our work is that we are explicitly analyzing
the unavoidable, physiologically basic, high-dimensional FSSs that accompany each movement.
This can be made possible by the use of interactive parallel coordinates and graphical animation
tools. Another key feature is the methodical design for autonomic self-knowledge that will equip
the system with the ability to express the causative relationships between sensory inputs and the
eerent motor events that are activated. Prediction of expected aerentation outcomes will be a
consequence of this relationship. From the perspective of neuromuscular systems, the mathemat-
ical and geometric perspective we applied served to ground the arguments of muscle abundance
and the feasibility of task constraints. Our methods for obtaining these results can be employed
towards robotics, and brain-machine interfaces (BMIs) optimized on the limits of simulated neu-
ral drive obtained from sensory inputs. Machines that possess self-aware computing will relieve
engineers, such as ourselves, of the need to assess system constraints and resource availability at
design-time. This need has the potential to be either be signicantly reduced or eliminated. The
goal with our proposed framework was to not directly install a full-body representation or model
into a suitable hardware host but to instead provide it with the rules, principles, and fortitude
115
to form a personal schema on its own merit...from the basis of proprioceptive signals. We be-
lieve our software implementation will provide valuable insight for engaging a new generation of
systems featured with computational self-modeling that will showcase a level of robustness and

exibility that tendon-driven systems in the robotics eld. At this time, the framework design is
tentative and open for modication as the project develops in the future. Our past research in
the 1) Mirror Neuron System for reaching arm tasks, 2) Agency according to self-other dichotomy
processing, 3) translating psychological self-awareness to machine subjective experience by way of
neuro-physiology, and 4) the forming of feasible sensory sets has primed us for making progress
in the phase of implementing the actualization of self-aware computing.
This thesis helped us make to make an eective contribution to the eld of self-model computing
and assist in providing rm answers to the following inquiries:
How do we understand and interact with the world around us (solely via aerentation)?
How do we display that sensory understanding with motor control?
What processes occur to allow us to make sense of the world via our senses?
What are the sensorimotor dynamics that lead to articial self-modeling and self-awareness?
A self-aware system, which we discussed in the Background section, should be able to comprehend
and facilitate the overheads and tradeos associated with the act of learning information about
itself. Subsequently, we can make fundamental progress in interpolating sensorimotor schemas of
self-impressions and experiences and their eects on the acquisition of knowledge. If developed
correctly, the system's ability to self-generate its states can contribute to the goal of transforming
autonomous systems into devices that have the means to form self-knowledge of experiences (dis-
cover conditions and patterns that occur during operation). Through useful body representations,
116
systems can obtain the ability to evaluate their eectiveness while improving their action plan-
ning and decision-making capabilities. Overall, we hope to have contributed to the aim of robot
capabilities to be more easily extended and adapted to novel situations and lifelong learning.
117
Bibliography
[1] Eyal Amir, Michael L Anderson, and Vinay K Chaudhi. Report on darpa workshop on
self-aware computer systems. Technical report, 2007.
[2] Beulah Amsterdam. Mirror self-image reactions before age two. Developmental Psychobi-
ology: The Journal of the International Society for Developmental Psychobiology, 5(4):297{
305, 1972.
[3] James R Anderson and Gordon G Gallup Jr. Which primates recognize themselves in
mirrors? PLoS Biology, 9(3), 2011.
[4] AF Barto, S Sutton, and Charles W Anderson. Neuronlike adaptive elements that can solve
dicult systems. IEEE Trans. Sys. Man Cyber, 13(5):834{846, 1983.
[5] Nicol o F Bernardi, Mohammad Darainy, and David J Ostry. Somatosensory contribution
to the initial stages of human motor learning. Journal of neuroscience, 35(42):14316{14326,
2015.
[6] Jasmine Berry, Ali Marjaninejad, and Francisco J Valero-Cuevas. Minimal pre-processing
of multi-muscle ensembles of spindle signals improves discriminability of limb movements,
2020. In Preparation.
[7] Jasmine A Berry and Alice C Parker. The elephant in the mirror: Bridging the brain's
explanatory gap of consciousness. Frontiers in systems neuroscience, 10:108, 2017.
[8] Jasmine A Berry, Robert Ritter, Akira Nagamori, and Francisco J Valero-Cuevas. The
neural control of movement must contend with trajectory-specic and nonlinearly distorted
manifolds of aerent muscle spindle activity. In Neural Networks (IJCNN), 2017 Interna-
tional Joint Conference on, pages 1188{1194. IEEE, 2017.
[9] Bennett I Bertenthal and Kurt W Fischer. Development of self-recognition in the infant.
Developmental psychology, 14(1):44, 1978.
[10] Lynda IA Birke and Ruth Hubbard. Reinventing biology: Respect for life and the creation
of knowledge. Georgetown University Press, 1995.
[11] Kwabena Boahen. A neuromorph's prospectus. Computing in Science & Engineering,
19(2):14{28, 2017.
118
[12] Tamara Bonaci, Jerey Herron, Charlie Matlack, and Howard Jay Chizeck. Securing the ex-
ocortex: A twenty-rst century cybernetics challenge. In 2014 IEEE Conference on Norbert
Wiener in the 21st Century (21CW), pages 1{8. IEEE, 2014.
[13] James Bonaiuto, Edina Rosta, and Michael Arbib. Extending the mirror neuron system
model, i. Biological cybernetics, 96(1):9{38, 2007.
[14] Josh Bongard and Hod Lipson. Evolved machines shed light on robustness and resilience.
Proceedings of the IEEE, 102(5):899{914, 2014.
[15] Josh Bongard, Victor Zykov, and Hod Lipson. Resilient machines through continuous self-
modeling. Science, 314(5802):1118{1121, 2006.
[16] Luca Bonini, Monica Maranesi, Alessandro Livi, Leonardo Fogassi, and Giacomo Rizzolatti.
Space-dependent representation of objects and other's action in monkey ventral premotor
grasping neurons. Journal of Neuroscience, 34(11):4108{4119, 2014.
[17] Luca Bonini, Monica Maranesi, Alessandro Livi, Leonardo Fogassi, and Giacomo Rizzolatti.
Ventral premotor neurons encoding representations of action during self and others' inaction.
Current Biology, 24(14):1611{1614, 2014.
[18] Doug M Boyer, Yaron Lipman, Elizabeth St Clair, Jesus Puente, Biren A Patel, Thomas
Funkhouser, Jukka Jernvall, and Ingrid Daubechies. Algorithms to automatically quantify
the geometric similarity of anatomical surfaces. Proceedings of the National Academy of
Sciences, 108(45):18221{18226, 2011.
[19] Marcel Brass and Patrick Haggard. The hidden side of intentional action: the role of the
anterior insular cortex. Brain Structure and Function, 214(5-6):603{610, 2010.
[20] Thomas Buhrmann, Ezequiel Alejandro Di Paolo, and Xabier Barandiaran. A dynamical
systems account of sensorimotor contingencies. Frontiers in psychology, 4:285, 2013.
[21] Vittorio Caggiano, Leonardo Fogassi, Giacomo Rizzolatti, Peter Thier, and Antonino Casile.
Mirror neurons dierentially encode the peripersonal and extrapersonal space of monkeys.
science, 324(5925):403{406, 2009.
[22] Thomas A Carlson, George Alvarez, Daw-an Wu, and Frans AJ Verstraten. Rapid assim-
ilation of external objects into the body schema. Psychological Science, 21(7):1000{1005,
2010.
[23] Laurie Carr, Marco Iacoboni, Marie-Charlotte Dubeau, John C Mazziotta, and Gian Luigi
Lenzi. Neural mechanisms of empathy in humans: a relay from neural systems for imitation
to limbic areas. Proceedings of the national Academy of Sciences, 100(9):5497{5502, 2003.
[24] Rita Carter. Mapping the mind. Univ of California Press, 1999.
[25] Emilie A Caspar, Andrea Desantis, Zoltan Dienes, Axel Cleeremans, and Patrick Haggard.
The sense of agency as tracking control. PloS one, 11(10):e0163892, 2016.
[26] David Chalmers. The hard problem of consciousness. The Blackwell companion to con-
sciousness, pages 225{235, 2007.
119
[27] Dempsey Hsiu-Ju Chang. The gestalt principles of similarity and proximity apply to both the
haptic and visual grouping. In Eighth Australasian User Interface Conference (AUIC2007),
pages 79{86. Flinders Press, 2007.
[28] Hillel J Chiel, Lena H Ting,

Orjan Ekeberg, and Mitra JZ Hartmann. The brain in its
body: motor control and sensing in a biomechanical context. Journal of Neuroscience,
29(41):12807{12814, 2009.
[29] Patricia Smith Churchland. Reduction and the neurobiological basis of consciousness. 1988.
[30] Philip Clayton. Mind and emergence: From quantum to consciousness. 2004.
[31] Arthur D Craig. Human feelings: why are some more aware than others? Trends in cognitive
sciences, 8(6):239{241, 2004.
[32] Arthur D Craig and AD Craig. How do you feel{now? the anterior insula and human
awareness. Nature reviews neuroscience, 10(1), 2009.
[33] Francis Crick and Christof Koch. A framework for consciousness. Nature neuroscience,
6(2):119{126, 2003.
[34] Hugo D Critchley and Neil A Harrison. Visceral in
uences on brain and behavior. Neuron,
77(4):624{638, 2013.
[35] Hugo D Critchley, Stefan Wiens, Pia Rotshtein, Arne

Ohman, and Raymond J Dolan. Neural
systems supporting interoceptive awareness. Nature neuroscience, 7(2):189{195, 2004.
[36] Gislin Dagnelie. Retinal implants: emergence of a multidisciplinary eld. Current opinion
in neurology, 25(1):67{75, 2012.
[37] Antonio Damasio and Raymond J Dolan. The feeling of what happens. Nature,
401(6756):847{847, 1999.
[38] Avinash De Sousa. Towards an integrative theory of consciousness: part 1 (neurobiological
and cognitive models). Mens sana monographs, 11(1):100, 2013.
[39] Monique W De Veer and Ruud Van den Bos. A critical review of methodology and in-
terpretation of mirror self-recognition research in nonhuman primates. Animal Behaviour,
58(3):459{468, 1999.
[40] Fabienne Delfour and Ken Marten. Mirror image processing in three marine mammal
species: killer whales (orcinus orca), false killer whales (pseudorca crassidens) and california
sea lions (zalophus californianus). Behavioural processes, 53(3):181{190, 2001.
[41] Daniel C Dennett. Consciousness explained. Penguin uk, 1993.
[42] Yan Duan, Marcin Andrychowicz, Bradly Stadie, OpenAI Jonathan Ho, Jonas Schneider,
Ilya Sutskever, Pieter Abbeel, and Wojciech Zaremba. One-shot imitation learning. In
Advances in neural information processing systems, pages 1087{1098, 2017.
120
[43] Brian Duy. Robots social embodiment in autonomous mobile robotics. International
Journal of Advanced Robotic Systems, 1(3):17, 2004.
[44] H Henrik Ehrsson, Charles Spence, and Richard E Passingham. That's my hand! activity
in premotor cortex re
ects feeling of ownership of a limb. Science, 305(5685):875{877, 2004.
[45] Nathan Faivre, Anat Arzi, Claudia Lunghi, and Roy Salomon. Consciousness is more than
meets the eye: a call for a multisensory study of subjective experience. Neuroscience of
consciousness, 2017(1):nix003, 2017.
[46] MN Fargeau, N Jaafari, S Ragot, JL Houeto, C Pluchon, and R Gil. Alzheimer's disease
and impairment of the self. Consciousness and Cognition, 19(4):969{976, 2010.
[47] Ali Farshchiansadegh, Alejandro Melendez-Calderon, Rajiv Ranganathan, Todd D Mur-
phey, and Ferdinando A Mussa-Ivaldi. Sensory agreement guides kinetic energy optimization
of arm movements during object manipulation. PLoS computational biology, 12(4):e1004861,
2016.
[48] Todd E Feinberg. Neuroontology, neurobiological naturalism, and consciousness: a challenge
to scientic reduction and a solution. Physics of Life Reviews, 9(1):13{34, 2012.
[49] WR Ferrell, SC Gandevia, and DI McCloskey. The role of joint receptors in human ki-
naesthesia when intramuscular receptors cannot contribute. The Journal of physiology,
386(1):63{71, 1987.
[50] Andrew A Fingelkurts, Alexander A Fingelkurts, and Tarja Kallio-Tamminen. Long-term
meditation training induced changes in the operational synchrony of default mode network
modules during a resting state. Cognitive Processing, 17(1):27{37, 2016.
[51] Andrew A Fingelkurts, Alexander A Fingelkurts, and Tarja Kallio-Tamminen. Trait lasting
alteration of the brain default mode network in experienced meditators and the experiential
selfhood. Self and Identity, 15(4):381{393, 2016.
[52] Andrew A Fingelkurts, Alexander A Fingelkurts, and Carlos FH Neves. \machine" con-
sciousness and \articial" thought: An operational architectonics model guided approach.
Brain research, 1428:80{92, 2012.
[53] American Association for Research into Nervous, Mental Diseases, and JR Searle. How to
study consciousness scientically. Philosophical Transactions of the Royal Society of London.
Series B: Biological Sciences, 353(1377):1935{1942, 1998.
[54] Christina T Fuentes and Amy J Bastian. Where is your arm? variations in proprioception
across space and tasks. Journal of neurophysiology, 2009.
[55] Shaun Gallagher. Philosophical conceptions of the self: implications for cognitive science.
Trends in cognitive sciences, 4(1):14{21, 2000.
[56] Shaun Gallagher. A pattern theory of self. Frontiers in human neuroscience, 7:443, 2013.
[57] Vittorio Gallese et al. How the body in action shapes the self. Journal of Consciousness
Studies, 18(7-8):117{143, 2011.
121
[58] Vittorio Gallese, Luciano Fadiga, Leonardo Fogassi, and Giacomo Rizzolatti. Action recog-
nition in the premotor cortex. Brain, 119(2):593{609, 1996.
[59] GG Gallup and SHILLITO DJ ANDERSON JR. The cognitive animal: Empirical and
theoretical perspectives on animal cognition, 2002.
[60] Gordon G Gallup. Chimpanzees: self-recognition. Science, 167(3914):86{87, 1970.
[61] Bruce J Gantz, Christopher Turner, Kate E Gfeller, and Mary W Lowder. Preservation
of hearing in cochlear implant surgery: advantages of combined electrical and acoustical
speech processing. The Laryngoscope, 115(5):796{802, 2005.
[62] Chris Gaskett and Gordon Cheng. Online learning of a motor map for humanoid robot
reaching. Proc. 2nd Int. Conf. Computat. Intell., Robot. Autonom. Syst. (CIRAS 2003),
2003.
[63] James Jerome Gibson. The senses considered as perceptual systems. 1966.
[64] Kevin Gold and Brian Scassellati. A bayesian robot that distinguishes" self" from" other".
In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 29, 2007.
[65] Kevin Gold and Brian Scassellati. Using probabilistic reasoning over time to self-recognize.
Robotics and autonomous systems, 57(4):384{392, 2009.
[66] Melvyn Goodale and David Milner. Sight unseen: An exploration of conscious and uncon-
scious vision. OUP Oxford, 2013.
[67] Melvyn A Goodale, A David Milner, et al. Separate visual pathways for perception and
action. 1992.
[68] Marcus A Gray, Neil A Harrison, Stefan Wiens, and Hugo D Critchley. Modulation of
emotional appraisal by false physiological feedback during fmri. PLoS one, 2(6):e546, 2007.
[69] Michael SA Graziano. How the brain represents the body: insights from neurophysiology
and psychology. Common mechanisms in perception and action, 2000.
[70] Nicola Guarino, Daniel Oberle, and Steen Staab. What is an ontology? In Handbook on
ontologies, pages 1{17. Springer, 2009.
[71] Stuart Hamero and Roger Penrose. Orchestrated reduction of quantum coherence in brain
microtubules: A model for consciousness. Mathematics and computers in simulation, 40(3-
4):453{480, 1996.
[72] Takashi Hanakawa, Ilka Immisch, Keiichiro Toma, Michael A Dimyan, Peter Van Gelderen,
and Mark Hallett. Functional properties of brain areas associated with motor execution and
imagery. Journal of neurophysiology, 89(2):989{1002, 2003.
[73] Robert M Haralick and Linda G Shapiro. Computer and robot vision, volume 1. Addison-
wesley Reading, 1992.
122
[74] Nalin Harischandra and

Orjan Ekeberg. System identication of muscle{joint interactions
of the cat hind limb during locomotion. Biological cybernetics, 99(2):125, 2008.
[75] Justin Wildrick Hart and Brian Scassellati. Mirror perspective-taking with a humanoid
robot. In Twenty-Sixth AAAI Conference on Articial Intelligence, 2012.
[76] Stephen W Hawking and George Francis Rayner Ellis. The large scale structure of space-
time, volume 1. Cambridge university press, 1973.
[77] Nicholas Hay, Michael Stark, Alexander Schlegel, Carter Wendelken, Dennis Park, Eric
Purdy, Tom Silver, D Scott Phoenix, and Dileep George. Behavior is everything: Towards
representing concepts with sensorimotor contingencies. In Thirty-Second AAAI Conference
on Articial Intelligence, 2018.
[78] Markram Henry. The blue brain project. Nat Rev Neurosci, 7(2):153^ a160, 2006.
[79] Louis M Herman. Body and self in dolphins. Consciousness and cognition, 21(1):526{545,
2012.
[80] Sayaka Hihara, Tomonori Notoya, Michio Tanaka, Shizuko Ichinose, Hisayuki Ojima,
Shigeru Obayashi, Naotaka Fujii, and Atsushi Iriki. Extension of corticocortical aerents
into the anterior bank of the intraparietal sulcus by tool-use training in adult monkeys.
Neuropsychologia, 44(13):2636{2646, 2006.
[81] Thomas T Hills and Stephen Butterll. From foraging to autonoetic consciousness: The
primal self as a consequence of embodied prospective foraging. Current Zoology, 61(2):368{
381, 2015.
[82] Matej Homann, Hugo Marques, Alejandro Arieta, Hidenobu Sumioka, Max Lungarella,
and Rolf Pfeifer. Body schema in robotics: a review. IEEE Transactions on Autonomous
Mental Development, 2(4):304{324, 2010.
[83] Katherine RS Holzbaur, Wendy M Murray, and Scott L Delp. A model of the upper extrem-
ity for simulating musculoskeletal surgery and analyzing neuromuscular control. Annals of
biomedical engineering, 33(6):829{840, 2005.
[84] Pierre Jacob and Marc Jeannerod. Ways of seeing: The scope and limits of visual cognition.
2003.
[85] Kian Jalaleddini, Chuanxin Minos Niu, Suraj Chakravarthi Raja, Won Joon Sohn, Gerald E
Loeb, Terence D Sanger, and Francisco J Valero-Cuevas. Neuromorphic meets neuromechan-
ics, part ii: the role of fusimotor drive. Journal of neural engineering, 14(2):025002, 2017.
[86] W James. The principles of psychology, vol. 2. henry holt and company, 1890.
[87] Harry J Jerison. Evolution of the brain and intelligence. Current Anthropology, 16(3):403{
426, 1975.
[88] Mari Jibu, Kunio Yasue, and K Yasue. Quantum brain dynamics and consciousness. John
Benjamins C., 1995.
123
[89] Eric R Kandel, Henry Markram, Paul M Matthews, Rafael Yuste, and Christof Koch.
Neuroscience thinks big (and collaboratively). Nature Reviews Neuroscience, 14(9):659{
664, 2013.
[90] Eric R Kandel, James H Schwartz, Thomas M Jessell, Department of Biochemistry, Molec-
ular Biophysics Thomas Jessell, Steven Siegelbaum, and AJ Hudspeth. Principles of neural
science, volume 4. McGraw-hill New York, 2000.
[91] Julian Paul Keenan, Hanna Oh, and Franco Amati. An overview of self-awareness. The
Oxford Handbook of Social Neuroscience, page 314, 2011.
[92] James M Kilner, Karl J Friston, and Chris D Frith. Predictive coding: an account of the
mirror neuron system. Cognitive processing, 8(3):159{166, 2007.
[93] Jaegwon Kim. Physicalism, or something near enough, volume 19. Princeton University
Press, 2007.
[94] Tilo TJ Kircher, Carl Senior, Mary L Phillips, Sophia Rabe-Hesketh, Philip J Benson,
Edward T Bullmore, Mick Brammer, Andrew Simmons, Mathias Bartels, and Anthony S
David. Recognizing one's own face. Cognition, 78(1):B1{B15, 2001.
[95] Stanley B Klein. The cognitive neuroscience of knowing one's self. 2004.
[96] Christof Koch. The quest for consciousness a neurobiological approach. 2004.
[97] Konrad P K ording and Daniel M Wolpert. Bayesian integration in sensorimotor learning.
Nature, 427(6971):244{247, 2004.
[98] Uriah Kriegel. Consciousness as sensory quality and as implicit self-awareness. Phenomenol-
ogy and the Cognitive Sciences, 2(1):1{26, 2003.
[99] Michael Kubovy, Alex O Holcombe, and Johan Wagemans. On the lawfulness of grouping
by proximity. Cognitive psychology, 35(1):71{98, 1998.
[100] Yasuo Kuniyoshi. Learning from examples: Imitation learning and emerging cognition. Hu-
manoid Robotics and Neuroscience: Science, Engineering and Society, G. Cheng, Ed.(CRC
Press, Boca Raton, FL, 2015), pages 223{250, 2015.
[101] AD Kuo and Zajac FE. Human standing posture: multi-joint movement strategies based
on biomechanical constraints. Progress in brain research, 97:349{358, 1993.
[102] Jason J Kutch and Francisco J Valero-Cuevas. Challenges and new approaches to proving
the existence of muscle synergies of neural origin. PLoS computational biology, 8(5), 2012.
[103] Christopher M Laine, Akira Nagamori, and Francisco J Valero-Cuevas. The dynamics of
voluntary force production in aerented muscle in
uence involuntary tremor. Frontiers in
computational neuroscience, 10:86, 2016.
[104] Christopher Landauer and Kirstie L Bellman. Self-modeling systems. In International
Workshop on Self-Adaptive Software, pages 238{256. Springer, 2001.
124
[105] Mark R Leary and Nicole R Buttermore. The evolution of the human self: Tracing the
natural history of self-awareness. Journal for the Theory of Social Behaviour, 33(4):365{
404, 2003.
[106] Michel A Lemay and Patrick E Crago. A dynamic model for simulating movements of the
elbow, forearm, and wrist. Journal of biomechanics, 29(10):1319{1330, 1996.
[107] Joseph Levine. Materialism and qualia: The explanatory gap. Pacic philosophical quar-
terly, 64(4):354{361, 1983.
[108] JP Lewis. Fast normalized cross-correlation, industrial light and magic. unpublished, 2005.
[109] Peter R Lewis, Arjun Chandra, Funmilade Faniyi, Kyrre Glette, Tao Chen, Rami Bahsoon,
Jim Torresen, and Xin Yao. Architectural aspects of self-aware and self-expressive computing
systems: From psychology to engineering. Computer, 48(8):62{70, 2015.
[110] Peter R Lewis, Arjun Chandra, Shaun Parsons, Edward Robinson, Kyrre Glette, Rami
Bahsoon, Jim Torresen, and Xin Yao. A survey of self-awareness and its application in
computing systems. In 2011 Fifth IEEE Conference on Self-Adaptive and Self-Organizing
Systems Workshops, pages 102{107. IEEE, 2011.
[111] Rodolfo Riascos Llin as. I of the vortex: From neurons to self, volume 50. MIT press
Cambridge, MA, 2001.
[112] GE Loeb, WS Levine, and Jiping He. Understanding sensorimotor feedback through optimal
control. In Cold Spring Harbor symposia on quantitative biology, volume 55, pages 791{803.
Cold Spring Harbor Laboratory Press, 1990.
[113] Gerald E Loeb and Jeremy A Fishel. Bayesian action&perception: Representing the world
in the brain. Frontiers in neuroscience, 8:341, 2014.
[114] HC Lou, JP Changeux, and A Rosenstand. Towards a cognitive neuroscience of self-
awareness. Neuroscience & Biobehavioral Reviews, 83:765{773, 2017.
[115] Pete Mandik. Action-oriented representation. Cognition and the brain: The philosophy and
neuroscience movement, pages 284{305, 2005.
[116] Ali Marjaninejad, Dar o Urbina-Mel endez, Brian A Cohn, and Francisco J Valero-Cuevas.
Autonomous functional movements in a tendon-driven limb via limited experience. Nature
machine intelligence, 1(3):144{154, 2019.
[117] Ruben Martinez-Cantin, Manuel Lopes, and Luis Montesano. Body schema acquisition
through active learning. In 2010 IEEE international conference on robotics and automation,
pages 1860{1866. IEEE, 2010.
[118] Victoria McGeer. Autistic self-awareness. Philosophy, Psychiatry, & Psychology, 11(3):235{
251, 2004.
[119] R Christopher Miall. Connecting mirror neurons and forward models. Neuroreport,
14(17):2135{2137, 2003.
125
[120] RC Miall, D Jo Weir, Daniel M Wolpert, and JF Stein. Is the cerebellum a smith predictor?
Journal of motor behavior, 25(3):203{216, 1993.
[121] Milana P Mileusnic, Ian E Brown, Ning Lan, and Gerald E Loeb. Mathematical models of
proprioceptors: I. control and transduction in the muscle spindle. Journal of neurophysiol-
ogy, 2006.
[122] Milana P Mileusnic and Gerald E Loeb. Mathematical models of proprioceptors: Ii. struc-
ture and function of the golgi tendon organ. Journal of Neurophysiology, 2006.
[123] David Milner and Mel Goodale. The visual brain in action, volume 27. OUP Oxford, 2006.
[124] Melanie Mitchell. Self-awareness and control in decentralized systems. In AAAI Spring
Symposium: Metacognition in Computation, pages 80{85, 2005.
[125] Alain Morin. Levels of consciousness and self-awareness: A comparison and integration of
various neurocognitive views. Consciousness and cognition, 15(2):358{371, 2006.
[126] Alain Morin. Self-recognition, theory-of-mind, and self-awareness: What side are you on?
Laterality, 16(3):367{383, 2011.
[127] Thomas Nagel. What is the mind-body problem. Experimental and theoretical studies of
consciousness, pages 1{7, 1993.
[128] U Neisser. Cognition and reality, 1976.
[129] Ulric Neisser. The roots of self-knowledge: Perceiving self, it, and thou a. Annals of the
New York Academy of Sciences, 818(1):19{33, 1997.
[130] Chuanxin M Niu, Kian Jalaleddini, Won Joon Sohn, John Rocamora, Terence D Sanger, and
Francisco J Valero-Cuevas. Neuromorphic meets neuromechanics, part i: the methodology
and implementation. Journal of neural engineering, 14(2):025001, 2017.
[131] Alva No e, Alva No e, et al. Action in perception. MIT press, 2004.
[132] Georg Northo, Alexander Heinzel, Moritz De Greck, Felix Bermpohl, Henrik Dobrowolny,
and Jaak Panksepp. Self-referential processing in our brain|a meta-analysis of imaging
studies on the self. Neuroimage, 31(1):440{457, 2006.
[133] Elizaveta V Okorokova, James M Goodman, Nicholas G Hatsopoulos, and Sliman J Bens-
maia. Decoding hand kinematics from population responses in sensorimotor cortex during
grasping. arXiv preprint arXiv:1904.03531, 2019.
[134] J Kevin O'Regan and Alva No e. A sensorimotor account of vision and visual consciousness.
Behavioral and brain sciences, 24(5):939{973, 2001.
[135] David J Ostry and Paul L Gribble. Sensory plasticity in human motor learning. Trends in
neurosciences, 39(2):114{123, 2016.
[136] D Oyserman, K Elmore, and GS Smith. Self, self-concept, and identity-handbook of self
and identity, 2012.
126
[137] David I Perrett, Mark H Harries, Ruth Bevan, S Thomas, PJ Benson, Amanda J Mistlin,
Andrew J Chitty, Jari K Hietanen, and JE Ortega. Frameworks of analysis for the neural
representation of animate objects and actions. Journal of experimental Biology, 146(1):87{
113, 1989.
[138] Carissa L Philippi, Justin S Feinstein, Sahib S Khalsa, Antonio Damasio, Daniel Tranel,
Gregory Landini, Kenneth Williford, and David Rudrauf. Preserved self-awareness following
extensive bilateral brain damage to the insula, anterior cingulate, and medial prefrontal
cortices. PloS one, 7(8):e38413, 2012.
[139] Jean Piaget. Piaget's theory of cognitive development. Childhood cognitive development:
The essential readings, 2:33{47, 2000.
[140] Steven M Platek, Jaime W Thomson, and Gordon G Gallup Jr. Cross-modal self-recognition:
The role of visual, auditory, and olfactory primes. Consciousness and cognition, 13(1):197{
210, 2004.
[141] Joshua M Plotnik, Frans BM De Waal, and Diana Reiss. Self-recognition in an asian
elephant. Proceedings of the National Academy of Sciences, 103(45):17053{17057, 2006.
[142] Helmut Prior, Ariane Schwarz, and Onur G unt urk un. Mirror-induced behavior in the magpie
(pica pica): evidence of self-recognition. PLoS biology, 6(8), 2008.
[143] Jennifer S Rabin, Asaf Gilboa, Donald T Stuss, Raymond A Mar, and R Shayna Rosenbaum.
Common and unique neural correlates of autobiographical memory and theory of mind.
Journal of Cognitive Neuroscience, 22(6):1095{1111, 2010.
[144] Samik Raychaudhuri. Introduction to monte carlo simulation. In 2008 Winter simulation
conference, pages 91{100. IEEE, 2008.
[145] Diana Reiss and Lori Marino. Mirror self-recognition in the bottlenose dolphin: A case of
cognitive convergence. Proceedings of the National Academy of Sciences, 98(10):5937{5942,
2001.
[146] Magnus JE Richardson and Tamar Flash. Comparing smooth arm movements with the two-
thirds power law and the related segmented-control hypothesis. Journal of neuroscience,
22(18):8201{8211, 2002.
[147] Bernhard Rinner, Lukas Esterle, Jennifer Simonjan, Georg Nebehay, Roman P
ugfelder,
Gustavo Fernandez Dominguez, and Peter R Lewis. Self-aware and self-expressive camera
networks. Computer, 48(7):21{28, 2015.
[148] Giacomo Rizzolatti and Laila Craighero. The mirror-neuron system. Annu. Rev. Neurosci.,
27:169{192, 2004.
[149] Giacomo Rizzolatti, Luciano Fadiga, Vittorio Gallese, and Leonardo Fogassi. Premotor
cortex and the recognition of motor actions. Cognitive brain research, 3(2):131{141, 1996.
[150] Philip Robbins. Consciousness and the social mind. Cognitive Systems Research, 9(1-2):15{
23, 2008.
127
[151] Philippe Rochat. Five levels of self-awareness as they unfold early in life. Consciousness
and cognition, 12(4):717{731, 2003.
[152] Udaya Bhaskar Rongala, Anton Spanne, Alberto Mazzoni, Fredrik Bengtsson, Calogero
Oddo, and Henrik J orntell. Intracellular dynamics in cuneate nucleus neurons support
self-stabilizing learning of generalizable tactile representations. Frontiers in cellular neuro-
science, 12:210, 2018.
[153] Perrine Ruby and Jean Decety. Eect of subjective perspective taking during simulation of
action: a pet investigation of agency. Nature neuroscience, 4(5):546{550, 2001.
[154] Hannes P Saal, Benoit P Delhaye, Brandon C Rayhaun, and Sliman J Bensmaia. Simulating
tactile signals from the whole hand with millisecond precision. Proceedings of the National
Academy of Sciences, 114(28):E5693{E5702, 2017.
[155] Louis A Sass and Josef Parnas. Schizophrenia, consciousness, and the self. Schizophrenia
bulletin, 29(3):427{444, 2003.
[156] R Keith Sawyer. Emergence in psychology: Lessons from the history of non-reductionist
science. Human development, 45(1):2{28, 2002.
[157] Daniel L Schacter, Donna Rose Addis, and Randy L Buckner. Remembering the past to
imagine the future: the prospective brain. Nature reviews neuroscience, 8(9):657{661, 2007.
[158] Gerald E Schneider. Two visual systems. Science, 1969.
[159] Lorenzo Sciavicco and Bruno Siciliano. Modelling and control of robot manipulators. Springer
Science & Business Media, 2012.
[160] William Seager. Theories of consciousness: An introduction and assessment. Routledge,
2016.
[161] Anil K Seth, Keisuke Suzuki, and Hugo D Critchley. An interoceptive predictive coding
model of conscious presence. Frontiers in psychology, 2:395, 2012.
[162] Reza Shadmehr, Maurice A Smith, and John W Krakauer. Error correction, sensory pre-
diction, and adaptation in motor control. Annual review of neuroscience, 33:89{108, 2010.
[163] Dan Song, N Lan, GE Loeb, and J Gordon. Model-based sensorimotor integration for
multi-joint control: development of a virtual arm model. Annals of biomedical engineering,
36(6):1033{1048, 2008.
[164] Henry P Stapp. Quantum theory and the role of mind in nature. Foundations of Physics,
31(10):1465{1499, 2001.
[165] Luc Steels and Michael Spranger. The robot in the mirror. Connection Science, 20(4):337{
358, 2008.
[166] Fabio Sterpetti. Models, brains, and scientic realism. In Model-Based Reasoning in Science
and Technology, pages 639{661. Springer, 2016.
128
[167] Donald T Stuss. Self, awareness, and the frontal lobes: A neuropsychological perspective.
In The self: Interdisciplinary approaches, pages 255{278. Springer, 1991.
[168] Susan D Su arez and Gordon G Gallup Jr. Self-recognition in chimpanzees and orangutans,
but not gorillas. Journal of human evolution, 10(2):175{188, 1981.
[169] Keisuke Suzuki, Sarah N Garnkel, Hugo D Critchley, and Anil K Seth. Multisensory
integration across exteroceptive and interoceptive domains modulates self-experience in the
rubber-hand illusion. Neuropsychologia, 51(13):2909{2917, 2013.
[170] Gregg A Tabot, John F Dammann, Joshua A Berg, Francesco V Tenore, Jessica L Boback,
R Jacob Vogelstein, and Sliman J Bensmaia. Restoring the sense of touch with a pros-
thetic hand through a brain interface. Proceedings of the National Academy of Sciences,
110(45):18279{18284, 2013.
[171] Evan Thompson. Sensorimotor subjectivity and the enactive approach to experience. Phe-
nomenology and the cognitive sciences, 4(4):407{427, 2005.
[172] Neil C Thompson, Kristjan Greenewald, Keeheon Lee, and Gabriel F Manso. The compu-
tational limits of deep learning. arXiv preprint arXiv:2007.05558, 2020.
[173] Sebastian Thrun. -a lifelong learning perspective for mobile robot control. In Intelligent
Robots and Systems, pages 201{214. Elsevier, 1995.
[174] Giulio Tononi. Consciousness as integrated information: a provisional manifesto. The
Biological Bulletin, 215(3):216{242, 2008.
[175] Colwyn B Trevarthen. Two mechanisms of vision in primates. Psychologische Forschung,
31(4):299{337, 1968.
[176] Francisco J Valero-Cuevas. An integrative approach to the biomechanical function and
neuromuscular control of the ngers. Journal of biomechanics, 38(4):673{684, 2005.
[177] Francisco J Valero-Cuevas. A mathematical approach to the mechanical capabilities of limbs
and ngers. In Progress in motor control, pages 619{633. Springer, 2009.
[178] Francisco J Valero-Cuevas. Fundamentals of neuromechanics. Springer, 2016.
[179] Francisco J Valero-Cuevas, BA Cohn, HF Yngvason, and Emily L Lawrence. Exploring the
high-dimensional structure of muscle redundancy via subject-specic and generic muscu-
loskeletal models. Journal of biomechanics, 48(11):2887{2896, 2015.
[180] Francisco J Valero-Cuevas, Heiko Homann, Manish U Kurse, Jason J Kutch, and Evan-
gelos A Theodorou. Computational models for neuromuscular function. IEEE reviews in
biomedical engineering, 2:110, 2009.
[181] Francisco J Valero-Cuevas and Marco Santello. On neuromechanical approaches for the
study of biological and robotic grasp and manipulation. Journal of neuroengineering and
rehabilitation, 14(1):101, 2017.
129
[182] Susanne J van Veluw and Steven A Chance. Dierentiating between self and others: an
ale meta-analysis of fmri studies of self-recognition and theory of mind. Brain imaging and
behavior, 8(1):24{38, 2014.
[183] HEJ Veeger, FCT Van Der Helm, LHV Van Der Woude, GM Pronk, and RH Rozendal.
Inertia and muscle contraction parameters for musculoskeletal modelling of the shoulder
mechanism. Journal of biomechanics, 24(7):615{629, 1991.
[184] Kai Vogeley, Patrick Bussfeld, Albert Newen, Sylvie Herrmann, Francesca Happ e, Peter
Falkai, Wolfgang Maier, Nadim J Shah, Gereon R Fink, and Karl Zilles. Mind reading:
neural mechanisms of theory of mind and self-perspective. Neuroimage, 14(1):170{181,
2001.
[185] Stephen Warshall. A theorem on boolean matrices. Journal of the ACM (JACM), 9(1):11{
12, 1962.
[186] Max Wertheimer. Laws of organization in perceptual forms. A source book of Gestalt
Psychology, 1923.
[187] Justin HG Williams, Andrew Whiten, Thomas Suddendorf, and David I Perrett. Imitation,
mirror neurons and autism. Neuroscience & Biobehavioral Reviews, 25(4):287{295, 2001.
[188] Frank R Wilson. The hand: How its use shapes the brain, language, and human culture.
Vintage, 1999.
[189] Daniel M Wolpert, R Chris Miall, and Mitsuo Kawato. Internal models in the cerebellum.
Trends in cognitive sciences, 2(9):338{347, 1998.
[190] Yuichiro Yoshikawa, Koh Hosoda, and Minoru Asada. Does the invariance in multi-
modalities represent the body scheme?-a case study with vision and proprioception. In
2nd Intelligent Symposium on Adaptive Motion of Animals and Machines. Citeseer, 2003.
[191] Dan Zahavi and Andreas Roepstor. Faces and ascriptions: Mapping measures of the self.
Consciousness and cognition, 20(1):141{148, 2011.
[192] Felix E Zajac. Muscle and tendon: properties, models, scaling, and application to biome-
chanics and motor control. Critical reviews in biomedical engineering, 17(4):359{411, 1989.
[193] Shagufta Zia, Frederick WJ Cody, and Donald J O'Boyle. Identication of unilateral elbow-
joint position is impaired by parkinson's disease. Clinical Anatomy, 15(1):23{31, 2002.
130

Asset Metadata

Creator Berry, Jasmine Anaíís (author)

Core Title Sensory acquisition for emergent body representations in neuro-robotic systems

Contributor Electronically uploaded by the author (provenance)

School Andrew and Erna Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Computer Science

Publication Date 09/28/2020

Defense Date 08/21/2020

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag brain theory,neuromechanics,OAI-PMH Harvest,proprioception,robotics,sensorimotor

Language English

Advisor Valero-Cuevas, Francisco (committee chair), Krishnamachari, Bhaskar (committee member), Parker, Alice (committee member)

Creator Email berry@alumni.usc.edu,jasminab@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-379377

Unique identifier UC11665946

Identifier etd-BerryJasmi-9020.pdf (filename),usctheses-c89-379377 (legacy record id)

Legacy Identifier etd-BerryJasmi-9020.pdf

Dmrecord 379377

Document Type Dissertation

Rights Berry, Jasmine Anaíís; Berry, Jasmine Anaiis

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Abstract (if available)

Abstract An ongoing engineering challenge is achieving agility, information processing, and flexibility in robotic systems. Building neuromorphic robots called NeuRoBots (i.e., robots that imitate the mechanisms of neural sensorimotor processing in animals) is one approach to accomplishing this goal. NeuRoBots offer several advantages over traditional robots and also serve as testbeds for understanding the sensorimotor dynamics of mammalian neuromuscular physiology. The notion of how the anatomical brain builds a sense of self and how neuro-robotic agents can utilize body schemas (or representations) to build a sense of self have not been particularly successful due to varied and often contradictory accounts. In this dissertation I present a critical step in forming self-identified body schemas, based on physiological simulation of proprioceptive afferent signals, to determine the plausibility of whether biological signals can be used to inform the operation of a state machine. First, I demonstrate that a given movement gives rise to a distinct sensory manifold embedded in the 12-D space of muscle spindle information that is largely independent of the choice of muscle coordination strategy. Given that muscle lengths and velocities are fully determined by joint kinematics, such manifolds provide a rich set of information to use in its control. Secondly, I show that high-dimensional multi-muscle proprioceptive ensembles can usefully discriminate limb states and be utilized as a suitable classifier for inter-trajectory comparisons—but only after minimal pre-processing. Lastly, I present the concept of Sensory-Motor Gestalt, which provides a unifying framework for constructing body states into useful behaviors to understand the foundations of sense of self in hybrid robots and synthetic biological agents.