Multimodal spiking neural network for space robotic manipulation

The microgravity environment places stringent demands on autonomous manipulators for on-orbit assembly. While model-free reinforcement learning (RL) improves adaptability, its high computational cost hampers deployment on resource-limited platforms. We propose a multimodal control framework based on a spiking neural network (SNN) that enhances perception by integrating geometric states, tactile forces, and semantic cues. The system features a dual-channel, three-stage curriculum reinforcement learning (CRL) mechanism. Specifically, the dual-channel design shifts grasping reliance from spatial data to tactile feedback to ensure accuracy. Furthermore, the three-stage approach decomposes the challenge by gradually increasing complexity, which fosters robust control strategies. Compared with conventional artificial neural network (ANN) controllers, our method achieves significantly higher success rates and energy efficiency in robotic manipulation tasks, validating its suitability for resource-constrained space applications.