Deep Reinforcement Learning with Shaping Exploration Space for Robotic Assembly

Abstract

Assembly is an essential part of complex product manufacturing. Due to the complexity of assembly operations and rapidly changing market demands, it is a labor-intensive industry with insufficient automation improvement. Force control methods are developed to allow industrial robots to do Contact-Rich manipulation with force sensors fixed at the end-effect. Deep Reinforcement Learning (DRL) provides a method to learn assembly skills by trial and error. Nevertheless, the performance of the framework combining DRL and force control is sensitive to the hyper-parameters, which limits the application in manufacturing. In this work, we analyze the force control based on primary impedance control and propose a method that designs hyper-parameters according to task information, thus speeding up learning and making the process safer by adding a constraint to the exploration space.