SDARL: Safe Deep Adaptive Representation Learning for High-DoF Non-Linear System Manipulation in Space

Ye Zhang, Rui Liu, Linyue Chu, Zhengjian Kang, Qi Chen, Kangtong Mo

Abstract: Space robotic manipulation, particularly satellite servicing and repairing missions under uncertain and dynamically evolving orbital conditions has posed significant challenges in recent years due to inherent uncertainties and environmental disturbances. Precise and safe control of high degree-of-freedom (DoF) nonlinear robotic systems remains a significant challenge, particularly in uncertain and dynamic space environments. To address this, we introduce a novel deep adaptive control framework. Our approach integrates representation learning with a model predictive control (MPC) architecture to achieve robust and precise manipulation. The core contribution of our approach lies in the hierarchical architecture that incorporates deep representation learning with rigorous control-theoretic principles, enabling rapid online adaptation to unforeseen disturbances and environment-dependent nonlinearities. Motivated by the complexities encountered in robotic satellite manipulation tasks, our method learns shared representations that generalize across varying operational conditions to facilitate robust and efficient adaptation during manipulation. We rigorously integrate deep online representation learning methods with MPC, achieving a unified approach that provides both theoretical performance guarantees and practical control effectiveness to keep the whole system safe during the manipulation process. We provide rigorous theoretical analyses establishing explicit non-asymptotic convergence and stability guarantees, ensuring safe and reliable operation. Extensive simulations validate the effectiveness and superiority of our method, demonstrating significant improvements in trajectory tracking accuracy, disturbance rejection capability, and overall system stability compared to existing state-of-the-art approaches.