Legged balance on moving table by reinforcement learning

Abstract

Balancing is one of the most essential ability for legged robots. From maintaining posture on unstable surfaces to withstanding sudden force disturbances, balancing is widely used in rescue robots. Among these various balancing issues, this study focused on balancing on moving plates. The study is applicable when the robot needs to perform an operation on an earthquake or shaking ship environment. In addition, since ZMP(Zero Moment Point) is used for the control, the performance is exhibited not only when an external force is applied such as wind force, but also an inertial force is exerted. In this study, reinforcement learning was used so that the proposed algorithm can avoid redundancy issues and kinematic singularities which cause problems in traditional method. The simulation results show how this new approach can improve the robot's balancing ability.