Cost-Matching Model Predictive Control for Efficient Reinforcement Learning in Humanoid Locomotion

TL;DR

This paper introduces a cost-matching MPC-based reinforcement learning method for humanoid locomotion, improving efficiency and robustness by approximating action-value functions from high-fidelity data.

Contribution

It presents a novel cost-matching approach that trains a parameterized MPC to efficiently approximate value functions, enhancing humanoid locomotion control.

Findings

Enhanced locomotion performance in simulation

Increased robustness to disturbances and model mismatch

Efficient gradient-based learning without repeated MPC solves

Abstract

In this paper, we propose a cost-matching approach for optimal humanoid locomotion within a Model Predictive Control (MPC)-based Reinforcement Learning (RL) framework. A parameterized MPC formulation with centroidal dynamics is trained to approximate the action-value function obtained from high-fidelity closed-loop data. Specifically, the MPC cost-to-go is evaluated along recorded state-action trajectories, and the parameters are updated to minimize the discrepancy between MPC-predicted values and measured returns. This formulation enables efficient gradient-based learning while avoiding the computational burden of repeatedly solving the MPC problem during training. The proposed method is validated in simulation using a commercial humanoid platform. Results demonstrate improved locomotion performance and robustness to model mismatch and external disturbances compared with manually tuned baselines.