Geometric-Feature Representation Based Pre-Training Method for Reinforcement Learning of Peg-in-Hole Tasks

摘要

Recently, reinforcement learning (RL) is often used for learning the strategy of peg-in-hole tasks. However, traditional state representation of PiH RL might be either redundant or abstract, which leads to unnecessary learning steps and incompatibility with mathematical training optimization. To issue these problems, a geometric-feature (GF) state representation method for peg-in-hole DDPG (Deep Deterministic Policy Gradient) RL is proposed to get rid of the redundant states. Also, a network pre-training method for peg-in-hole DDPG based on GF representation is provided to help DDPG acquire basic assembly skills before training. Finally, we executed the simulation experiments in the environment with Open AI Gym + Pybullet to test the performance of the proposed GF representation based pre-training method. We also executed the lenrned strategy on the real-world Panda robot.