Actor-Critic Method-Based Search Strategy for High Precision Peg-in-Hole Tasks

摘要

In the field of 3C(Computer/Communication/Consumer Electronic) product assembly, Peg-in-hole task, such as fiber assembly, is widely used. However, it remains as a big challenge for robots to automatically execute peg-in-hole tasks. Building a contact model is the traditional idea, which requires lots of time and effort. However, the model suffers low accuracy in the situation with tighter clearance. Currently, the most learning-based methods do not take into account the particularity of such assembly tasks, which lead to slow convergence. In this paper, we propose a new search strategy based on reinforcement learning for high precision peg-in-hole assembly tasks. The assembly task is divided into two steps: search and insert. Afterwards, a Markov Decision Process (MDP) is designed for the two steps according to different assembly features and solved by an Actor-Critic method. The robot can learn how to choose the optimal action and accomplish peg-in-hole task with less training and execute steps, high success rate and smaller contact force. Moreover, the proposed method can be applied to the multi-hole task without retraining. The results of simulation and experiment demonstrate its fast and stable performance.