Implementation of assembly task based on guided policy search algorithm

At present, safely solving complex and high- precision assembly tasks in an unstructured environment is still an unresolved challenge. The development of artificial intelligence technology provides new ideas for robots in unstructured scenes to autonomously solve the problem of contact-rich peg-in-hole tasks. In this paper, we construct the shaft hole assembly task as a reinforcement learning problem, explore the change of the convergence rate when adding force and torque information in the state space, and evaluate the performance of the guided policy search algorithm on the shaft hole assembly task. We also compared the changes in the force and torque feedback from sensors at the beginning and end of learning. The experiment proves that we can complete the specific shaft hole assembly task by learning, and also shows the effectiveness of using the force and torque information when the peg and hole parts are in contact.