Autonomous Robust Skill Generation Using Reinforcement Learning with Plant Variation

摘要

This paper discusses an autonomous space robot for a truss structure assembly using some reinforcement learning. It is difficult for a space robot to complete contact tasks within a real environment, for example, a peg-in-hole task, because of error between the real environment and the controller model. In order to solve problems, we propose an autonomous space robot able to obtain proficient and robust skills by overcoming error to complete a task. The proposed approach develops skills by reinforcement learning that considers plant variation, that is, modeling error. Numerical simulations and experiments show the proposed method is useful in real environments.