← Projects
RL-augmented Adaptive MPC
Model predictive control is the workhorse of bipedal locomotion, but it depends on a dynamics model that is rarely accurate on rough or unfamiliar terrain. This project augments adaptive MPC with reinforcement learning so the controller can adjust its own model online from interaction data.
Overview
The adaptive MPC layer maintains a parametric model of the robot’s contact and dynamics, and an RL agent learns to correct or rebias that model based on residual error. The combination keeps the predictability and constraint handling of MPC while letting the robot improve its footing on terrain that the nominal model alone cannot handle.
Publication
ICRA 2026