Object insertion under tight tolerances (< 1mm) is an important but challenging assembly task as even slight errors can result in undesirable contacts. Recent efforts have focused on using Reinforcement Learning (RL) and often de- pend on careful definition of dense reward functions. This work proposes an ef- fective strategy for such tasks that integrates traditional model-based control with RL to achieve improved accuracy given training of the policy exclusively in simu- lation and zero-shot transfer to the real system. It employs a potential field-based controller to acquire a model-based policy for inserting a plug into a socket given full observability in simulation. This policy is then integrated with a residual RL one, which is trained in simulation given only sparse, goal-reaching reward. A curriculum scheme over observation noise and action magnitude is proposed for training the residual RL policy. Both policy components use as input the SE(3) poses of both the plug and the socket and return the plug’s SE(3) pose transform, which is executed by a robotic arm using a controller. The integrated policy is deployed on the real system without further training or fine-tuning, given a visual SE(3) object tracker. The proposed solution and alternatives are evaluated across a variety of objects and conditions both in simulation and reality. The proposed approach outperforms state-of-the-art RL methods in this domain, as well as prior efforts in hybrid policies. Ablations highlight the impact of each component of the approach.
Comparison of the trained policy deployed in the real world on a task requiring elevated and convoluted force profile and the same task performed by a human (both in the assembly and disassembly track). The discomfort observed by the human demonstrates how challenging the task really is.