AI-Enabled Decision Support for EVA Operations

Onboard, explainable AI systems for human spaceflight operations under communication delay and limited ground support.

Future human spaceflight missions will operate under conditions where real-time support from Earth is no longer possible. For Mars missions, communication delays of up to 20 minutes remove the Mission Control feedback loop entirely, requiring crews to operate with significantly greater autonomy.

Our research develops AI-enabled decision support systems that replicate and extend the capabilities of Mission Control within the spacecraft or spacesuit. These systems integrate domain knowledge, physical models, and multi-agent reasoning to support safe, explainable decision-making during extravehicular activity (EVA).

We develop a multi-agent architecture that decomposes EVA decision-making into specialized roles, mirroring the structure of Mission Control. Each agent is responsible for a specific function such as planning, systems monitoring, physiology, or operations, and operates within defined constraints and authority.

The system integrates three core components:

• Retrieval: access to validated procedures, flight rules, and mission documentation with full traceability
• Simulation: deterministic models of physiology, environment, and vehicle systems to evaluate feasibility
• Agentic reasoning: coordinated multi-agent workflows that produce structured, explainable decisions

Multi-agent architecture for onboard EVA decision support, replicating Mission Control functions through distributed reasoning and tool integration.

This architecture is implemented in the RASAGE (Retrieval and Simulation Augmented Guidance Engine) framework, enabling real-time onboard decision support under communication delay and computational constraints.

We evaluate the system using both historical mission reconstruction and structured benchmarking.

Reconstruction of Apollo 14 EVA operations demonstrates that a multi-agent architecture can replicate key Mission Control functions while maintaining traceability to source documentation.

Prototype EVA decision-support interface showing integrated timeline, navigation, and physiological monitoring during Apollo 14 reconstruction.

To assess foundation model capability in this domain, we developed EVA-Bench, a scenario-based benchmark comprising over 600 tasks grounded in NASA EVA operations. Results show strong performance in procedural knowledge retrieval but significant degradation in multi-step planning and contingency handling.

EVA-Bench scenario examples illustrating structured evaluation of procedural reasoning, anomaly response, and safety-critical decision-making.

• A multi-agent framework for Earth-independent EVA decision support
• Integration of physiological and operational models for safety-critical reasoning
• A domain-specific benchmark for evaluating AI systems in spaceflight operations
• Demonstration of low-hallucination, evidence-backed reasoning in mission scenarios

This work supports a transition from Earth-dependent mission operations to autonomous onboard decision support. The resulting architecture enables safer EVA execution, reduced crew cognitive load, and scalable operations for future lunar and Mars missions.