Tensegrity Robot Locomotion Under Limited Sensory Inputs via Deep Reinforcement Learning

摘要

Tensegrity robots are composed of rigid rods connected by elastic cables, and their unique light-weight yet compliant structure makes them an appealing choice for space exploration. However, locomotion control for these robotic systems remains difficult due to their nonlinear dynamics and high-dimensional state space. We demonstrate that in the domain of tensegrity robotics, it is possible to efficiently learn end-to-end locomotion policies using mirror descent guided policy search (MDGPS) even with limited sensory inputs. We compare learned neural network policies with other locomotion control policies in various testing environments; and results show that neural network policies consistently outperform others. We also shed light to the policy learning process by analyzing different choices of observation inputs to the robot. Moreover these findings motivate exploration of deep reinforcement learning algorithms in the domain of tensegrity robotics. We show preliminary results with one such locomotion example on discontinuous rough terrains.