Vision intelligence-conditioned reinforcement learning for precision assembly

Robots that embrace human-level performance on precise, dexterous and dynamic assembly tasks can significantly enhance the efficiency in precision assembly but remain big challenges. This paper introduces a vision intelligence-conditioned method for precision assembly, enabled by human-in-the-loop reinforcement learning. Upon visual demonstrations collected and trained by a reward classifier, a data-efficient reinforcement learning algorithm trains and learns vision-based robotic manipulation policies under human-in-the-loop corrections. An impedance-based control strategy derived from policies and visual guidance achieves high-precision contact-rich assembly manipulations with near-perfect success rates (above 98%) and compliance behaviours. The effectiveness of the presented method is experimentally demonstrated with semiconductor assembly.