Robust Quadruped Locomotion via Reinforcement Learning with Deep Generalized-Momentum-Based Kalman Filter

Robust quadruped locomotion in real-world environments remains challenging because external disturbances, sensor noise, and model uncertainties are coupled with intermittent foot–ground contact. Reinforcement learning has shown strong capability in generating agile locomotion, but many existing methods handle unobserved disturbances through implicit latent representations or domain randomization. This paper presents a disturbance-aware locomotion framework that integrates state and disturbance estimation with learning-based control. The core component is a deep generalized-momentum-based Kalman filter, which combines generalized momentum disturbance modeling with adaptive covariance inference to estimate the base velocity and external disturbance force. These physically meaningful estimates are incorporated into the policy observation space, reducing the gap between privileged simulation states and deployable onboard observations. The framework was evaluated in a simulation and on a quadruped robot platform under disturbance and outdoor locomotion scenarios. Compared with the baseline and ablated variants, the proposed method reduced estimation and tracking errors, limited impact-induced torque peaks, and improved locomotion success rates under the evaluated conditions. The results suggest that explicit disturbance estimation can complement a latent adaptation for quadruped locomotion under impact-rich conditions.