Maxime Kawawa-Beaudan
EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2022-146
May 19, 2022
http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-146.pdf
This thesis presents methods for training intelligent robotic systems to navigate challenging, diverse environments. In particular it focuses on legged locomotion for hexapods rather than well-studied bipedal or quadrupedal robots. We employ hierarchical reinforcement learning to integrate proprioception with a low-level gait controller, and train on-policy algorithms entirely in simulation before transferring to a real robot. We develop command-conditioned policies in PyBullet that learn to walk at commanded target velocities, and switch to Isaac Gym to train policies with multi-objective rewards both in rough and flat terrains. A comparison is established between gaits with motion priors and prior-free gaits. Both methods successfully surmount joist obstacles typically seen in attics in the real world. We propose novel approaches to the sim-to-real problem designed to address limitations of affordable hardware. We demonstrate methods in the Isaac Gym simulation for integrating vision.
Advisor: Avideh Zakhor
BibTeX citation:
@mastersthesis{Kawawa-Beaudan:EECS-2022-146, Author = {Kawawa-Beaudan, Maxime}, Editor = {Zakhor, Avideh}, Title = {Learning to Walk: Legged Hexapod Locomotion from Simulation to the Real World}, School = {EECS Department, University of California, Berkeley}, Year = {2022}, Month = {May}, URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-146.html}, Number = {UCB/EECS-2022-146}, Abstract = {This thesis presents methods for training intelligent robotic systems to navigate challenging, diverse environments. In particular it focuses on legged locomotion for hexapods rather than well-studied bipedal or quadrupedal robots. We employ hierarchical reinforcement learning to integrate proprioception with a low-level gait controller, and train on-policy algorithms entirely in simulation before transferring to a real robot. We develop command-conditioned policies in PyBullet that learn to walk at commanded target velocities, and switch to Isaac Gym to train policies with multi-objective rewards both in rough and flat terrains. A comparison is established between gaits with motion priors and prior-free gaits. Both methods successfully surmount joist obstacles typically seen in attics in the real world. We propose novel approaches to the sim-to-real problem designed to address limitations of affordable hardware. We demonstrate methods in the Isaac Gym simulation for integrating vision.} }
EndNote citation:
%0 Thesis %A Kawawa-Beaudan, Maxime %E Zakhor, Avideh %T Learning to Walk: Legged Hexapod Locomotion from Simulation to the Real World %I EECS Department, University of California, Berkeley %D 2022 %8 May 19 %@ UCB/EECS-2022-146 %U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-146.html %F Kawawa-Beaudan:EECS-2022-146