Vision-Based Human–Robot Handover System with Reinforcement Learning

Handover control in human–robot collaboration remains a significant challenge. This paper proposes a three-step vision-based human–robot handover system (VHS). Vision inputs are used to perceive the environment and enable adaptive control of the robotic arm. Moreover, a three-step behavior cloning learning strategy is designed. Furthermore, a modified Temporal Difference (TD) loss function based on transfer models is proposed to train the algorithm to improve policy exploration and convergence. The proposed method results in substantial enhancements in comparative experimental validation in a simulation environment with a realistic dynamic hand model.