Robotic Peg-in-hole Assembly Based on Generative Adversarial Imitation Learning with Hindsight Transformation

In recent years, autonomous robotic peg-in-hole assembly has emerged as a prominent area of research. Due to the challenge of modeling and determining reward functions for peg-in-hole assembly scenarios, imitation learning is a more appropriate option as it enables learning policies from expert demonstrations without relying on environmental reward functions. However, traditional generative adversarial imitation learning (GAIL) methods are not efficient as they generate low-quality samples, resulting in high costs of interaction between the agent and the environment. This paper proposes a novel approach called hindsight transformation generative adversarial imitation learning (HT-GAIL), which employs hindsight experience to transform the samples generated by the generator into expert-like data. This expert-like data are then combined with expert demonstrations for adversarial network training. The proposed approach effectively addresses the issue of low utilization efficiency of generated samples, resulting in faster convergence of the algorithm. To validate the proposed algorithm, we constructed an autonomous peg-in-hole assembly platform using a six-degree-of-freedom manipulator. Comparative experiments were conducted to demonstrate the superior performance of the algorithm. The trained policies achieved success rates of 96%, 88%, and 61% for fit clearances of 1.12mm, 0.80mm, and 0.52mm, respectively, while maintaining assembly forces within 1.18N, 2.21N, and 9.10N, respectively.