Decentralized Repetitive Learning for Whole-Body Planning and Control of Humanoid Robots With Centroidal Momentum Dynamics

Humanoid locomotion remains a fundamental challenge in robotics due to the high degrees of freedom, strong coupling between centroidal and joint-level dynamics, underactuation during walking phases, as well as pervasive model uncertainties and external disturbances. To address these challenges, this paper proposes an advanced whole-body planning and control framework that enables compliant posture regulation and stable locomotion through optimized centroidal motion and force planning. In the proposed framework, the full-body humanoid dynamics are first decoupled into centroidal momentum dynamics and joint-level dynamics. Based on this decomposition, a whole-body planning module formulates an optimization problem to generate dynamically consistent center-of-mass (CoM) trajectories and desired ground reaction forces (GRFs), which serve as reference inputs for posture adjustment and balance stabilization during walking. These planned centroidal quantities are then tracked by a whole-body control layer, where the desired GRFs are mapped into joint torques through a torque distribution scheme that preserves compliance and accommodates contact constraints. To cope with model uncertainties and unmodeled dynamics, an adaptive robust decentralized repetitive learning control strategy is integrated into the tracking control framework. By exploiting task repetitiveness, the proposed learning scheme iteratively refines control inputs, effectively compensating for system uncertainties without highly accurate models. A Lyapunov-based stability analysis demonstrates that the closed-loop tracking errors are uniformly ultimately bounded under bounded uncertainties. Extensive experimental results validate the effectiveness of the proposed approach, demonstrating stable posture regulation, robust walking performance, and strong disturbance rejection capabilities.