Online Non-linear Centroidal MPC for Humanoid Robot Locomotion with Step Adjustment

Overview

This paper introduces a nonlinear centroidal Model Predictive Controller (MPC) for humanoid locomotion. Unlike classical approaches that fix the footstep plan in advance, the controller reasons directly about the centroidal dynamics and simultaneously optimizes the contact wrenches and the foot locations, allowing online step adjustment in response to disturbances.

Method

• The MPC operates on the reduced centroidal model, keeping the optimization tractable for real-time execution.
• Footstep positions are decision variables, so the controller can relocate contacts to recover balance instead of relying on a rigid, precomputed plan.
• The formulation integrates with a whole-body controller that tracks the resulting centroidal and postural references.

Results

The approach is validated in simulation on a humanoid robot, demonstrating robust walking and the ability to reject external pushes through automatic step adjustment. The full experimental setup is reproducible via the provided Docker image in the companion repository.