NASA Eyes Edge AI for Autonomous Moon, Mars Missions

As the Artemis Program moves toward returning astronauts to the Moon, NASA’s Science Mission Directorate is betting on artificial intelligence for capabilities including autonomous navigation and fault detection, even as it works to close a major hardware gap.

Chief Technologist Carolyn Mercer said the agency envisions edge AI applications across most of its mission directorates, enabling spacecraft to process data and make decisions without relying on ground-based systems.

By processing data directly on devices rather than through cloud networks, edge AI could reduce the need to transmit telemetry and commands between spacecraft and Earth, accelerating response times and enabling greater autonomy.

Potential applications include autonomous fault management, intelligent data downlinking, onboard decision-making, autonomous navigation, surface mobility, multi-agent coordination and in-space assembly.

“NASA organizations have identified the need for reactive science, which is a capability that has historically been difficult to achieve due to larger response time given time-critical events,” Mercer told GovCIO Media & Research. “AI-enabled multi-agent coordination will facilitate reactive science as space-grade AI-enabled edge computing capabilities enter the avionics ecosystem.”

To support AI at the edge, NASA is working with industry partners to develop computing infrastructure capable of operating in the harsh conditions of space, including radiation, extreme temperatures and limited maintenance opportunities. Mercer said these efforts require integrating computing capabilities with the power systems needed to sustain them.

“There is a significant gap between what currently exists and what is needed to carry out AI-enabled autonomous space missions,” said Mercer. “This necessitates substantial investments to achieve the right capabilities for enabling fully autonomous missions using edge AI, and NASA recognizes the value of partnerships to leverage capabilities developed elsewhere and expedite development.”

Mercer added that AI is particularly important for deep-space missions to the moon and Mars, where reducing ground intervention and optimizing communications through NASA’s Deep Space Network are critical.

The Deep Space Network, NASA’s global collection ground-based spacecraft communication facilities, requires a more intelligent approach to data downlinking rather than simply transmitting all data to Earth systems. Mercer said scientists are using AI to identify only high-value data to transmit to ground stations, which enables NASA to support missions simultaneously.

“Things like autonomous navigation also significantly benefit deep-space missions. With the priority of establishing a lunar base, surface mobility is a key capability that will be enabled by edge AI, which are demonstrated by rovers like Perseverance, Endurance and the Lunar Terrain Vehicle,” said Mercer.

Navigating Roadblocks on the Journey to Space

NASA faces significant barriers to deploying AI capabilities in space. Mercer said the most pressing challenge is the lack of space-grade avionics capable of supporting modern AI workloads.

“NASA’s standard space-grade computer, the RAD-750, is not capable of running modern AI capabilities,” she said. “NASA has been working with industry to overcome this roadblock by developing an edge computer as well as AI-capable neuromorphic co-processors that can operate reliably in deep space.”

Beyond hardware, NASA must address challenges related to the maturity, verification and validation of AI systems and the datasets used to train them. AI systems can produce unpredictable outputs when encountering new conditions, complicating efforts to verify their reliability in mission-critical environments. Mercer said NASA is drawing on its aeronautics expertise to strengthen its approach to testing and validation.

“We’re looking into verification and validation capabilities developed by the agency’s aeronautics experts to leverage the experience we do have,” said Mercer. “NASA is taking a phased approach, adopting component-level autonomy first before directly implementing system-level autonomy.”

Data quality presents another obstacle. AI models depend on representative training data, but collecting sufficiently large datasets that capture the variability of space environments remains difficult.

NASA often relies on synthetic or hybrid datasets to train AI models. However, differences between simulated environments and real-world conditions can undermine model performance once systems are deployed.

“The key challenge with this approach is the significant gap between synthetic and real data. AI models are trained using synthetic data or hybrid datasets if some real-world data is available,” said Mercer. “However, space introduces unpredictable visual artifacts like stark shadows, hot pixels and extreme sun glare, as well as physical dynamics like lunar regolith that often cause models to fail when they transition from simulation to flight hardware.”

Balancing Cloud and Edge AI for Mission Needs

NASA is evaluating cloud, edge and local computing architectures to support different AI applications, with mission requirements determining which approach is most effective.

Mercer said AI training typically requires substantial computing power, memory and energy, making cloud infrastructure better suited for developing models and processing large datasets. Once trained, those models can be deployed to lower-power edge devices aboard spacecraft.

NASA also uses cloud-based AI to support ground operations, including mission planning and data analysis.

“Many NASA missions, for example, have used AI in ground segments for path planning, such as Mars rovers, and other perception capabilities that are being developed for technical demonstrations and future missions,” said Mercer. “Mars Perseverance’s autonomy-based onboard planner reduced power consumption and increased mobility rate tenfold.”

NASA’s approach also reflects the need to protect sensitive mission data. Mercer said the agency uses on-premises cloud infrastructure for sensitive ground-based information and partitioned, localized edge computing aboard spacecraft to maintain control over critical data.

The agency is also developing additional security capabilities for its flight software, including AI-enabled anomaly detection and zero-trust protections.

“NASA is developing a highly sophisticated security solution to enable its widely used flight software framework to create security-hardened flight software for NASA missions,” said Mercer. “This implementation is being developed for onboard processors to provide an extra layer of security that includes AI-based anomaly detection and enables zero-trust architecture.”