Online DNN-driven Nonlinear MPC for Stylistic Humanoid Robot Walking with Step Adjustment

Overview

This work presents a three-layered architecture that enables stylistic locomotion with online contact-location adjustment. It bridges data-driven and model-based control: an autoregressive Deep Neural Network (DNN) generates centroidal and postural references, which then serve as an initial guess and regularizer for a model-based trajectory adjustment layer and a trajectory control layer.

Because the DNN is trained on human motion-capture data, the resulting robot motion exhibits a natural, human-like walking style while remaining dynamically feasible and robust to disturbances.

Contributions

• DNN + MPC for locomotion. A learned trajectory generator is combined with a model predictive controller for online step adjustment, unifying data-driven style with model-based feasibility.
• MPC with control barrier function. The MPC keeps the center of mass within a safe region through a control-barrier function, ensuring stability under external disturbances while adjusting footsteps.
• GA-tuned Kalman filter. A Genetic-Algorithm-tuned Kalman filter reduces noise in the center-of-mass velocity and angular-momentum measurements, improving smoothness.

Results

Experiments on the ergoCub humanoid robot show the system can prevent falls, replicate human walking styles, and withstand external disturbances of up to 68 Newtons. The paper compares two implementations of the adjustment layer — a receding horizon planner (RHP) and a model predictive controller (MPC).