Hybrid Real-Time State Estimation for Quadruped Robots With Proprioceptive Sensors

In degenerate environments where exteroceptive sensors fail, precise state estimation of quadruped robots depends critically on proprioceptive sensors. Traditional Kalman filters (KFs), which fuse data from inertial measurement units (IMUs) and encoders, are computationally efficient but struggle with disturbances from periodic ground-impact forces during dynamic locomotion. Extended-state KFs, while capable of estimating both state and disturbance, rely on Gaussian assumptions that are poorly suited for the abrupt, non-Gaussian disturbances inherent in legged locomotion. To address these challenges, this article proposes a novel framework integrating an error-state Kalman filter (ESKF) with a robust super-twisting algorithm (STA)-based observer. The ESKF fuses state data and filters noise, while the STA observer, coupled with the dynamics model, enables real-time reconstruction of both state and disturbances. These components are tightly coupled, with the ESKF utilizing continuously updated disturbance estimates from the STA. We evaluated this framework by controlling a quadruped robot across diverse scenarios—flat surfaces, uneven terrain, slippage conditions, and external force disturbances—in simulation and experimental settings. The results demonstrate superior attitude and position accuracy compared to traditional KF-based methods, while maintaining comparable processing times.