Online Domain Adaption for Sim2Real Transfer of High-Precision Manipulation with Visuotactile Sensing

Visuotactile sensors can provide rich and high-resolution contact information, having great potential in contact-rich manipulation tasks with reinforcement learning (RL). Sim2Real techniques address the challenge of RL's reliance on a large amount of interaction data. However, for tasks that require high-precision actions, the domain gap between simulation and real-world restricts the performance. In this work, we propose an online adaption framework for Sim2Real transfer of high-precision contact-rich manipulation tasks. Specifically, we use a recurrent network to learn the correlation between the robot proprioception and tactile sensing signals, and use such correlation to adapt the policy. Experimental results show that our method can achieve a high success rate for zero-shot Sim2Real transfer of peg insertion policy with a clearance of only 40 $\mu \mathrm{m}$.