Agile Wheel-Bipedal Loco-Manipulation With Active Force Rejection

Recent research advances in wheel-legged robots underscore their potential for versatile real-world deployment. Integrating a robotic arm with such a platform enables loco-manipulation capabilities, significantly expanding its range of potential applications. However, robust whole-body control and coordination for loco-manipulation tasks are critically challenged by external forces inherent to contact-rich environments. This letter addresses the whole-body loco-manipulation problem for wheel-bipedal robots operating under such conditions. We propose an Extended Kalman Filter (EKF)-based explicit force estimator that synergistically integrates rigid-body dynamics with a data-driven neural network. This hybrid approach combines the real-time predictive power of neural networks with the interpretability of model-based dynamics. Leveraging these force estimates, we also introduce a force-responsive whole-body controller capable of dynamically adapting to external forces. This integrated active force rejection whole-body control framework facilitates real-time estimation and adaptive compensation for external forces, and therefore enable the robot's agile loco-manipulation in contact-rich environments. Extensive experimental validation in contact-rich loco-manipulation tasks demonstrates the effectiveness of our approach. Results confirm that the proposed framework improves the robot's performance in force-intensive scenarios.