A CPG-Based Agile and Versatile Locomotion Framework Using Proximal\n Symmetry Loss

摘要

Humanoid robots are made to resemble humans but their locomotion abilities\nare far from ours in terms of agility and versatility. When humans walk on\ncomplex terrains, or face external disturbances, they combine a set of\nstrategies, unconsciously and efficiently, to regain stability. This paper\ntackles the problem of developing a robust omnidirectional walking framework,\nwhich is able to generate versatile and agile locomotion on complex terrains.\nThe Linear Inverted Pendulum Model and Central Pattern Generator concepts are\nused to develop a closed-loop walk engine, which is then combined with a\nreinforcement learning module. This module learns to regulate the walk engine\nparameters adaptively, and generates residuals to adjust the robot's target\njoint positions (residual physics). Additionally, we propose a proximal\nsymmetry loss function to increase the sample efficiency of the Proximal Policy\nOptimization algorithm, by leveraging model symmetries and the trust region\nconcept. The effectiveness of the proposed framework was demonstrated and\nevaluated across a set of challenging simulation scenarios. The robot was able\nto generalize what it learned in unforeseen circumstances, displaying\nhuman-like locomotion skills, even in the presence of noise and external\npushes.\n