Extraction of Non-Regular Pegs Using Tactile Sensing and Reinforcement Learning from Demonstrations

As robots increasingly take on human tasks, effective human-robot knowledge transfer becomes crucial, particularly in assembly and disassembly tasks where visual occlusion poses significant challenges. Pegs inserted in parts become occluded, and their extraction requires trajectory updates that humans perform on the fly based on their tactile perception. However, this remains challenging for robotic agents performing the same task. The present work highlights the potential of leveraging human demonstrations to enhance robotic manipulation skills, particularly peg extraction scenarios. The research focuses on how pretraining Reinforcement Learning (RL) models with human demonstrations affect a manipulator's ability to infer tactile signals and improve successful peg extraction rates. Through real-world experiments using a manipulator equipped with tactile sensors, this study finds that pretraining of a reinforcement learning agent results in fewer episodes required to extract various differently shaped pegs from their corresponding holes successfully. Overall, human demonstrations and pretraining reduced the episodes needed to extract pegs successfully. The strategy of pretraining using different pegs with more intricate pegs also accelerates the learning process for simpler objects. The average number of episodes required for the peg removal was reduced by 95.33 % for the vertical peg, 40.06 % for the slanted peg, and 36.96 % for the curved peg.