Giulio Romualdi and Paolo Maria Viceconte contributed equally to this paper
This paper presents a three-layered architecture that enables stylistic locomotion with online contact location adjustment. Our method combines an autoregressive Deep Neural Network (DNN) acting as a trajectory generation layer with a model-based trajectory adjustment and trajectory control layers. The DNN produces centroidal and postural references serving as an initial guess and regularizer for the other layers.Β Being the DNN trained on human motion capture data, the resulting robot motion exhibits locomotion patterns, resembling a human walking style. The trajectory adjustment layer utilizes non-linear optimization to ensure dynamically feasible center of mass (CoM) motion while addressing step adjustments. We compare two implementations of the trajectory adjustment layer: one as a receding horizon planner (RHP) and the other as a model predictive controller (MPC). To enhance MPC performance, we introduce a Kalman filter to reduce measurement noise. The filter parameters are automatically tuned with a Genetic Algorithm. Experimental results on the ergoCub humanoid robot demonstrate the system's ability to prevent falls, replicate human walking styles, and withstand disturbances up to 68 Newton.
Integration of DNN and MPC for Locomotion: The paper introduces a three-layer architecture combining a Deep Neural Network (DNN) for trajectory generation and a Model Predictive Controller (MPC) for trajectory adjustment. This system bridges data-driven and model-based approaches, enabling stylistic human-like walking with online step adjustments.
MPC with Control Barrier Function for Step Adjustment: The MPC ensures kinematically feasible center of mass (CoM) motion while adjusting steps. A control-barrier function is used to keep the CoM within a safe region, ensuring stability and robustness during locomotion, even under external disturbances.
GA-tuned Kalman Filter for Noise Reduction: The paper introduces a Genetic Algorithm-tuned Kalman filter to reduce noise in CoM velocity and angular momentum measurements, improving the smoothness and stability of the robotβs movements. The system was validated to withstand disturbances of up to 68 Newtons on the ergoCub robotβ.
This interactive animation showcases two experiments where the robot is perturbed by an external force. In both experiments, the red circle under the foot represents the measured Zero Moment Point (ZMP), while the blue circle indicates the desired ZMP, calculated using the contact force output of the MPC. The green rectangle shows the nominal foot position as predicted by the DNN, and the yellow rectangle represents the actual footstep position generated by the trajectory adjustment layer.
In the left animation, the trajectory adjustment layer operates as a Receding Horizon Planner (RHP), while in the right animation, it functions as a Model Predictive Controller (MPC), allowing for a comparison of both approaches under external disturbances.
