Reinforcement Learning for Legged Robots

Ayush Agrawal, Prasanth Kotaru, CS289A Introduction to Machine Learning, Spring 2019

Legged robots have the promise to solve some of today’s most challenging problems. However, controlling these robots is a challenging task due the high dimensionality of their state space, nonlinear and underactuated dynamics. Current approaches either lack the robustness required for deployment of these robots in the real world or are computationally very expensive to run in real time. Deep Reinforcement Learning on the other hand is a promising approach that can produce robust policies while being simple and computationally inexpensive to run on the robot. This work explores the use of Deep Reinforcement Learning techniques to a bipedal robot Cassie in a complex simulated environment to find robust control policies. In particular, we looked at the simple case of balancing while being robust to external perturbations and random initial conditions.