Fusing Vision and Force: A Framework of Reinforcement Learning for Elastic Peg-in-Hole Assembly

Elastic Peg-in-Hole assembly has a wide range of applications in both industrial and home environments. However, accurately representing the infinite continuum of states and obtaining an accurate deformation model of the elastic deformable peg remains a challenge. Reinforcement learning (RL) has demonstrated its ability to learn manipulation skills from interactive experiences without the need for an exact physical model. Nonetheless, current RL methods rely on complex multimodal representations obtained using neural networks and lack clear physical interpretations. This article proposes a practical framework based on RL to interpret vision and force feedback through a concise, reformulated action space. The vision and force-based guidance are fused in the action space, and any continuous RL method can be adopted to learn the fusion pattern. Several experiments were conducted on an elastic peg-in-hole assembly platform to validate the effectiveness of the learned fusion scheme and its comparison to existing control and RL methods.