SpaceMind: A Modular and Self-Evolving Embodied Vision-Language Agent Framework for Autonomous On-orbit Servicing

A team of researchers led by Aodi Wu has introduced SpaceMind, a groundbreaking framework for creating embodied AI agents capable of autonomously servicing satellites in orbit. The system is built as a modular and self-evolving Vision-Language Model (VLM) agent, decomposing its capabilities into three independently extensible dimensions: skill modules, configurable Model Context Protocol (MCP) tools, and injectable reasoning modes. A key innovation is its MCP-Redis interface layer, which allows the same codebase to operate seamlessly across high-fidelity Unreal Engine 5 simulations and physical laboratory hardware without any modifications, dramatically simplifying the path from testing to real-world deployment.

SpaceMind's most significant feature is its Skill Self-Evolution mechanism. Unlike traditional systems that require full model fine-tuning, this process allows the agent to distill its operational experience—including failures—into persistent skill files. This enables continuous improvement. The framework was rigorously tested through 192 closed-loop runs across five different satellite models and three task types, deliberately including degraded conditions to stress-test robustness. Under nominal conditions, all operational modes achieved 90-100% navigation success. Crucially, in a real-world validation, the system achieved a 100% success rate for rendezvous tasks with a physical robot, proving the zero-modification transfer from simulation.

The research demonstrates SpaceMind's exceptional resilience and learning capability. In degraded scenarios, its 'Prospective' reasoning mode succeeded in search-and-approach tasks where other modes failed. A dedicated self-evolution study showed the agent could recover from failure in four out of six test groups after just a single failed episode, with one case jumping from complete failure to 100% success. Another metric saw inspection scores improve from 12 to 59 out of 100 through learned experience, showcasing its practical, on-the-job learning potential for long-duration, complex missions.