Page 123 |
Save page Remove page | Previous | 123 of 151 | Next |
|
small (250x250 max)
medium (500x500 max)
Large (1000x1000 max)
Extra Large
large ( > 500x500)
Full Resolution
All (PDF)
|
This page
All
|
Learning Reward functions for Model-based Reinforcement Learning
In the MBRL example, the forward model of the dynamics is represented in both cases by a neural
network, the input to the network is the current state and action, the output is the next state of the
environment. The tasks consist of a free movement task of a point mass in a 2D space, we call this
environment PointmassGoal, and a reaching task with a 2-link 2D manipulator, which we call the
ReacherGoal environment. The Pointmass state space is four-dimensional. For PointmassGoal
(x; y; x_ ; y_) are the 2D positions and velocities, and the actions are accelerations (x ; y ). The
ReacherGoal environment for the MBRL experiments is a lower-dimensional variant of the MFRL
environment. It has a four dimensional state, consisting of position and angular velocity of the
joints [ 1; 2; _1; _2] the torque is two dimensional [ 1; 2] The dynamics model P is updated once
every 100 outer iterations with the samples collected by the policy from the last inner optimization
step of that outer optimization step, i.e. the latest policy. The task distribution p(T ) consists of
different target positions that either the point mass or the arm should reach. During meta-train
time, a model of the system dynamics, represented by a neural network, is learned from samples
of the currently optimal policy. The task loss during meta-train time is LT ( ) = E ;P [R( )],
where R( ) is the final distance from the goal g, when rolling out new in the dynamics model
P. Taking the gradient © E new;P [R( )] requires the differentiation through the learned model
P (see Algorithm 3). The input to the meta-network is the state-action trajectory of the current
roll-out and the desired target position. The meta-network outputs a loss signal together with the
learning rate to optimize the policy. Fig. 5.4a shows the qualitative reaching performance of a
policy optimized with the meta loss during meta-test on PointmassGoal. The meta-loss network
was trained only on tasks in the right quadrant (blue trajectories) and tested on the tasks in the left
quadrant (orange trajectories) of the x; y plane, showing the generalization capability of the meta
108
Object Description
| Title | Data scarcity in robotics: leveraging structural priors and representation learning |
| Author | Molchanov, Artem |
| Author email | a.molchanov86@gmail.com;molchano@usc.edu |
| Degree | Doctor of Philosophy |
| Document type | Dissertation |
| Degree program | Computer Science |
| School | Viterbi School of Engineering |
| Date defended/completed | 2020-05-11 |
| Date submitted | 2020-08-11 |
| Date approved | 2020-08-11 |
| Restricted until | 2020-08-11 |
| Date published | 2020-08-11 |
| Advisor (committee chair) | Sukhatme, Gaurav Suhas |
| Advisor (committee member) |
Ayanian, Nora Culbertson, Heather Gupta, Satyandra K. |
| Abstract | Recent advances in Artificial Intelligence have benefited significantly from access to large pools of data accompanied in many cases by labels, ground truth values, or perfect demonstrations. In robotics, however, such data are scarce or absent completely. Overcoming this issue is a major barrier to move robots from structured laboratory settings to the unstructured real world. In this dissertation, by leveraging structural priors and representation learning, we provide several solutions when data required to operate robotics systems is scarce or absent. ❧ In the first part of this dissertation we study sensory feedback scarcity. We show how to use high-dimensional alternative sensory modalities to extract data when primary sensory sources are absent. In a robot grasping setting, we address the problem of contact localization and solve it using multi-modal tactile feedback as the alternative source of information. We leverage multiple tactile modalities provided by electrodes and hydro-acoustic sensors to structure the problem as spatio-temporal inference. We employ the representational power of neural networks to acquire the complex mapping between tactile sensors and the contact locations. We also investigate scarce feedback due to the high cost of measurements. We study this problem in a challenging field robotics setting where multiple severely underactuated aquatic vehicles need to be coordinated. We show how to leverage collaboration among the vehicles and the spatio-temporal smoothness of the ocean currents as a prior to densify feedback about ocean currents in order to acquire better controllability. ❧ In the second part of this dissertation, we investigate scarcity of the data related to the desired task. We develop a method to efficiently leverage simulated dynamics priors to perform sim-to-real transfer of a control policy when no data about the target system is available. We investigate this problem in the scenario of sim-to-real transfer of low-level stabilizing quadrotor control policies. We demonstrate that we can learn robust policies in simulation and transfer them to the real system while acquiring no samples from the real quadrotor. Finally, we consider the general problem of learning a model with a very limited number of samples using meta-learned losses. We show how such losses can encode a prior structure about families of tasks to create well-behaved loss landscapes for efficient model optimization. We demonstrate the efficiency of our approach for learning policies and dynamics models in multiple robotics settings. |
| Keyword | robotics; machine learning; artificial intelligence |
| Language | English |
| Part of collection | University of Southern California dissertations and theses |
| Publisher (of the original version) | University of Southern California |
| Place of publication (of the original version) | Los Angeles, California |
| Publisher (of the digital version) | University of Southern California. Libraries |
| Provenance | Electronically uploaded by the author |
| Type | texts |
| Legacy record ID | usctheses-m |
| Contributing entity | University of Southern California |
| Rights | Molchanov, Artem |
| Physical access | 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 author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given. |
| Repository name | University of Southern California Digital Library |
| Repository address | USC Digital Library, University of Southern California, University Park Campus MC 7002, 106 University Village, Los Angeles, California 90089-7002, USA |
| Repository email | cisadmin@lib.usc.edu |
| Filename | etd-MolchanovA-8923.pdf |
| Archival file | Volume13/etd-MolchanovA-8923.pdf |
Description
| Title | Page 123 |
| Full text | Learning Reward functions for Model-based Reinforcement Learning In the MBRL example, the forward model of the dynamics is represented in both cases by a neural network, the input to the network is the current state and action, the output is the next state of the environment. The tasks consist of a free movement task of a point mass in a 2D space, we call this environment PointmassGoal, and a reaching task with a 2-link 2D manipulator, which we call the ReacherGoal environment. The Pointmass state space is four-dimensional. For PointmassGoal (x; y; x_ ; y_) are the 2D positions and velocities, and the actions are accelerations (x ; y ). The ReacherGoal environment for the MBRL experiments is a lower-dimensional variant of the MFRL environment. It has a four dimensional state, consisting of position and angular velocity of the joints [ 1; 2; _1; _2] the torque is two dimensional [ 1; 2] The dynamics model P is updated once every 100 outer iterations with the samples collected by the policy from the last inner optimization step of that outer optimization step, i.e. the latest policy. The task distribution p(T ) consists of different target positions that either the point mass or the arm should reach. During meta-train time, a model of the system dynamics, represented by a neural network, is learned from samples of the currently optimal policy. The task loss during meta-train time is LT ( ) = E ;P [R( )], where R( ) is the final distance from the goal g, when rolling out new in the dynamics model P. Taking the gradient © E new;P [R( )] requires the differentiation through the learned model P (see Algorithm 3). The input to the meta-network is the state-action trajectory of the current roll-out and the desired target position. The meta-network outputs a loss signal together with the learning rate to optimize the policy. Fig. 5.4a shows the qualitative reaching performance of a policy optimized with the meta loss during meta-test on PointmassGoal. The meta-loss network was trained only on tasks in the right quadrant (blue trajectories) and tested on the tasks in the left quadrant (orange trajectories) of the x; y plane, showing the generalization capability of the meta 108 |
