This paper, led by Christopher Dagher, presents our work on using motion planning to improve the convergence of reinforcement learning for continuous control strategies. The motion planning focuses on contact-rich manipulation, with an example of pushing a box over a step. Our key finding is that RL struggles to explore well, and the use of a contact-aware motion planner (Contact-Mode Guided Motion Planning) leads to convergence for a torque-level policy.

This work was supported by NSF grant 2330794.