Demo

Abstract：

We present SpaceMind, a hybrid agent architecture combining Large Vision Models with specialized tools via Model Context Protocol (MCP) for On-Orbit Servicing. Current space operations face challenges from LVM limitations in precision-critical tasks, where specialist models consistently outperform general vision models. Our architecture addresses this by separating high-level cognitive planning from precise execution through MCP’s unified interface. The LVM handles task decomposition and strategic reasoning, while specialized tools execute domain-specific operations including pose estimation, part segmentation, and spacecraft control. Comprehensive validation in high-fidelity UE5/AirSim simulation environments demonstrates significant improvements: 87.5% vs 12.5% success rate compared to pure LVM approaches in autonomous approach tasks, with Claude Sonnet showing superior systematic tool usage. The modular design enables scalable integration of mission-specific tools without modifying the core architecture.