Adviser–Actor–Critic: A Precision-Oriented Reinforcement Learning Framework for Space Robotics Control

High-precision control is critical for autonomous space robotics tasks, such as space-pointing observation, debris removal, and on-orbit assembly, where even submillimeter or subdegree errors can jeopardize mission safety. Classical feedback controllers, such as proportional–integral–derivative, can effectively eliminate steady-state error but are for nonlinear multi-input multioutput systems, whereas deep reinforcement learning (DRL) offers strong real-time optimal decision-making and adaptability to complex dynamics but suffers from significant steady-state error due to neural approximation errors. This article proposes an adviser–actor–critic (AAC) framework that couples traditional control theory with reinforcement learning (RL) through a dual-loop architecture featuring a lightweight proportional-integral-based adviser. During deployment, the adviser generates virtual goals that proactively compensate accumulated tracking errors, while a pretrained goal-conditioned actor handles complex dynamics without modification. A control-theoretic analysis shows that under mild assumptions, AAC can eliminate steady-state error for a broad class of systems. Simulations across standard manipulation, dexterous-hand benchmarks, and space-relevant docking scenarios demonstrate over 80% average steady-state error reduction across diverse RL backbones. Hardware experiments on a quadrotor platform validate that AAC improves attitude regulation from 1$^{\circ }$ to 0.03$^{\circ }$ under realistic disturbances, highlighting its practicality for precision-critical space applications.