Exploration of a Space-based Intelligent Agent Foundation: The Aurora 1000 On-Orbit Computing Architecture and Applications

The evolution of space missions toward multimodal perception and autonomous decision making demands onboard systems capable of closed loop perception, decision, and execution. Conventional radiation hardened processors lack the capacity for large model inference, while commercial off the shelf (COTS) devices introduce reliability and thermal challenges in the space environment. To address these issues, we present Aurora 1000, an energy efficient and reliable onboard computing system for low Earth orbit. The system adopts a heterogeneous cluster architecture with a modular board level design for efficient resource coordination. Reliability is achieved through a multilayer protection framework combining physical isolation, hardware redundancy, and software fault tolerance, together with a low resistance conductive thermal path optimized for vacuum operation. For intelligent applications, we further develop an onboard model deployment framework incorporating quantization, LoRA incremental fine tuning, and dynamic loading to operate under limited onboard resources. This framework enables the first deployment of a vision--language foundation model, JigonGPT, in orbit for semantic understanding of remote sensing imagery. On orbit experiments demonstrate that Aurora 1000 significantly improves intelligent processing efficiency while maintaining robust operation in harsh space environments, supporting the transition from conventional satellites to intelligent space systems.