OnAIR: Applications of the NASA On-Board Artificial Intelligence Research Platform

Evana Gizzi, Timothy Chase Jr., Christian Cassamajor-Paul, Rachael Chertok, Lily Clough, Connor Firth, Alan Gibson, Ibrahim Haroon, James Marshall, Patrick Maynard, Michael Monaghan, Hayley Owens, Daniel Rogers, Mahmooda Sultana, Jivko Sinapov, Bethany TheilingAAAI 2025 · DOI

Introduction

The aerospace industry faces a significant challenge in adopting modern artificial intelligence. While AI research progresses rapidly, the process of "flight-qualifying" software for space missions is arduous, often taking years due to stringent reliability requirements and extreme hardware constraints, such as low-power 100MHz processors. Historically, integrating AI into a mission required bespoke, one-off architectures, creating a high barrier for researchers.

NASA’s OnAIR (On-Board Artificial Intelligence Research) framework addresses this by providing an open-source, Python-based platform designed to bridge the gap between cutting-edge AI and space flight environments. By standardizing the interface between AI algorithms and flight hardware, OnAIR enables rapid prototyping and deployment across diverse assets, from ground robots to orbiting satellites.

OnAIR deployment across various mission assets

Cognitive Architecture and Methodology

OnAIR is built upon a modular, plugin-based cognitive architecture that implements a classic "sense-perceive-act" loop. The core of the system is a unified software pipeline that abstracts environmental data into actionable insights through four hierarchical interfaces:

  1. Knowledge Representation ($P_{KR}$): Converts raw sensor data into high-level symbolic knowledge (e.g., logical predicates).
  2. Learners ($P_L$): Facilitates sub-symbolic processing, such as training neural networks or Vision Transformers on live data streams.
  3. Planners ($P_P$): Utilizes outputs from lower layers to perform high-level decision-making, such as task and motion planning.
  4. Complex Reasoners ($P_{CR}$): Acts as the executive function, synthesizing outputs from all plugins to determine final, optimized actions.
The OnAIR cognitive architecture and software pipeline

To ensure hardware independence, OnAIR utilizes Data Adapters. These components allow the same AI logic to interface seamlessly with various environments—including CSV files, high-fidelity simulators (e.g., AirSim, Gazebo), or live flight software like NASA’s Core Flight System (cFS)—without requiring modifications to the underlying AI code.

Mission Applications and Datasets

The versatility of OnAIR has been demonstrated across a wide range of data environments and missions:

  • Simulated Environments: Utilizing PyBullet and AirSim for drone and rover mission modeling.
  • Terrestrial Field Operations: The NAMASTE mission in the Alaskan permafrost used OnAIR to process methane sensor and GPS data.
  • Space-Based Operations: The SCENIC mission on the International Space Station (ISS) involved processing real-time telemetry packets (40 points at 0.25 Hz).
  • Imaging: Training contaminant detection models using soil sample imagery.

Key Results and Impact

OnAIR has achieved significant milestones in real-world deployments:

  • NAMASTE Mission: The platform enabled a fleet of autonomous drones to navigate wilderness areas, identifying methane concentrations and optimizing flight paths in real-time. Interoperability was achieved in just two months, a fraction of traditional integration timelines.
  • SCENIC Mission (ISS): OnAIR was successfully deployed on the SpaceCube payload. Despite the hardware's limited 100MHz processor, the system ran a Kalman filter continuously for four days without error, demonstrating the platform's reliability in space environments.
  • Community Growth: The project has fostered a community of approximately 30 active contributors across NASA, academia, and industry, lowering the barrier for AI researchers to contribute to aerospace.

Limitations and Future Directions

While OnAIR demonstrates high portability, the overhead of a Python-based pipeline remains a consideration for the most extreme resource-constrained environments. Future development aims to expand the platform's capabilities in:

  • Swarm Intelligence: Standardizing architectures for heterogeneous multi-agent systems.
  • New Scientific Domains: Applying the framework to snow hydrology, space weather monitoring, and exoplanet detection.
  • Community Engagement: NASA intends to host hackathons to further encourage global AI community participation in space exploration.

Summary updated

BibTeX
@inproceedings{gizzi2025onair,
  author = {Evana Gizzi and Timothy Chase Jr. and Christian Cassamajor-Paul and Rachael Chertok and Lily Clough and Connor Firth and Alan Gibson and Ibrahim Haroon and James Marshall and Patrick Maynard and Michael Monaghan and Hayley Owens and Daniel Rogers and Mahmooda Sultana and Jivko Sinapov and Bethany Theiling},
  title = {{OnAIR: Applications of the NASA On-Board Artificial Intelligence Research Platform}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35156},
}
Figure