RL-augmented Adaptive MPC

Model predictive control is a common tool for bipedal locomotion, but rough and slippery terrain make its models and control settings harder to choose. This paper uses reinforcement learning to adapt the MPC system dynamics, swing-leg controller, and gait frequency.

Overview

A learned policy adjusts these three parts of the controller while MPC continues to plan the robot’s motion. The evaluation uses bipedal simulations in IsaacLab on stairs, stepping stones, and low-friction surfaces, comparing the resulting behavior with MPC and RL baselines.

Publication

Links

• Project page
• arXiv: 2509.18466
• PDF
• Video on YouTube