NASA tests CMO-DA, an AI doctor for deep space when Earth calls arrive too late

NASA is testing CMO-DA, the Crew Medical Officer Digital Assistant, an AI system meant to help astronauts diagnose and treat medical symptoms when they are too far from Earth for a real-time doctor call. The core problem is built into the mission math: as crews venture farther from home, communication delays can rule out live consultation, and an early return may no longer be practical. NASA has already signaled that the medical threshold matters, bringing Crew-11 back from the International Space Station (ISS) early earlier this year because of a medical concern.

The next step is operational, not futuristic. CMO-DA is powered by RamaLama, a Red Hat-backed open source tool designed to simplify how developers run, pull, and serve AI models. In the current test setup, the assistant runs on a terrestrial twin of the HPE Spaceborne Computer aboard the ISS, so NASA and its partners can validate behavior on Earth before any potential deployment up there. Red Hat frames the intent plainly: once validated on Earth, CMO-DA will be demonstrated to NASA leadership so they can evaluate its further use. Translation: this is an engineering and governance milestone, not a science demo.

Under the hood, the system is built for a specific kind of clinical realism, where relevant information is not only text. Red Hat says RamaLama provides the engine to run both large language models (LLMs) for complex medical reasoning and Vision Language Models (VLMs) for image-based symptom analysis. That multimodal approach matters because astronauts are not just reporting symptoms with perfect documentation. They may need to describe observations, upload or capture images of areas of concern, and get consistent, structured guidance without waiting for a clinician hundreds of milliseconds or seconds away. Red Hat also says this enables CMO-DA to process both text and visual data without needing a massive infrastructure footprint.

The “offline” part is the real constraint win. CMO-DA runs locally on the device, which means responses do not depend on a connection to Earth. That design choice directly addresses the failure mode that drives deep-space mission risk: even if a doctor exists on the other end, the network might not deliver in time. Red Hat says CMO-DA started life as a proof of concept, then moved from a cloud-dependent model to a fully disconnected edge deployment. That shift is not a small refactor. In regulated, high-stakes environments like crewed spaceflight, you typically need predictable compute, repeatable deployment, and a clear boundary between what’s “decision support” and what’s “human responsibility.” Running on the edge helps bound the system’s dependencies.

There is also a staging strategy visible in the testing plan. CMO-DA has yet to leave Earth, and testing on the Spaceborne twin is positioned as refinement before any potential ISS deployment. This is a familiar pattern in regulated tech rollouts, even if the setting is more dramatic than a typical hospital pilot. The system is being refined in an environment that mirrors the constraints of the actual platform, which is what executives should look for when evaluating whether an AI program is ready to scale beyond its first prototype.

The hardware and iteration cadence are worth noting because they tell you how quickly AI products can become stable enough to trust. HPE's Spaceborne project is on its third iteration aboard the ISS, built from off-the-shelf components. It is based on HPE Edgeline and Proliant servers and is more than capable of machine learning and AI workloads. In the future, the team plans to integrate Red Hat Enterprise Linux AI for the next iteration of the CMO-DA. That “next iteration” phrasing is a quiet signal that CMO-DA is on a roadmap, where operating system and AI runtime choices are treated as part of reliability, not just performance.

So what is the strategic stake for executives and boards watching this space, beyond curiosity? NASA is effectively running a prototype for an offline, multimodal, edge-based clinical decision support system under the harshest possible communication constraints. If this approach holds up in testing, it becomes a template other mission profiles could follow, and it accelerates the broader legitimacy of enterprise AI stacks in extreme environments. If it fails, the lesson is just as valuable: disconnected deployments and multimodal reasoning are not “ship it later” features, they are readiness features. Either way, the decision-making clock is ticking for anyone funding or governing AI systems that must operate without perfect connectivity, perfect data, or real-time human oversight.