Multi-modal vision-guided reinforcement learning control for autonomous capture of non-cooperative space targets

Autonomous capture of non-cooperative targets by space manipulators remains challenging because of uncertain target motion, limited perceptual information, and complex system dynamics. Existing vision-based methods typically rely on low-dimensional image features, while conventional control approaches have limited adaptability under nonlinear and time-varying conditions. Reinforcement learning (RL) provides a promising framework for handling such complexity, but many existing approaches still do not explicitly coordinate perception-derived cues with structured control. To address this issue, this paper proposes a multi-modal vision-guided reinforcement learning control framework for autonomous capture of non-cooperative space targets. A multi-modal visual perception model is first developed to jointly estimate image-space position, relative depth, time-to-contact, orientation, and uncertainty from visual observations, forming a structured and control-oriented state representation. Based on this representation, a learned actor is integrated with a structured image-space controller, so that perception-derived cues provide a compensation term within a perception-informed RL-ISC control loop. Simulation results show that the proposed method improves tracking accuracy, convergence speed, and control smoothness under the tested dynamic target motion and initial-position perturbation conditions, providing a structured route for vision-guided control of space manipulators.