A reinforcement learning based control framework for robot gear assembly with demonstration learning and force feedback

Reinforcement learning (RL) has found extensive application in robotic manipulation tasks, particularly in scenarios involving contact-rich activities such as axis-hole assembly and spot welding. However, existing methodologies encounter several key limitations: difficulties in gathering real-world experimental data, slow reward convergence leading to suboptimal strategy performance, and a significant disparity between simulation and reality. This paper introduces a novel framework aimed at over-coming these limitations in the context of robotic gear assembly. Initially, we capture the assembly trajectory executed by a human expert through manual guidance, subsequently incorporating this expert experience into an experience pool. Leveraging the Deep Deterministic Policy Gradient (DDPG) algorithm, along with demonstrations, force feedback, and domain randomization, we train the policy until achieving satisfactory results in simulation. We establish the simulation environment within the MuJoCo simulator and validate the proposed method in real-world gear shaft assembly tasks. Experimental results showcase the robustness and efficacy of our framework in successfully completing assembly tasks.