Google and NASA test a Mars cargo robot that runs on a 4B-parameter model

Google Cloud and NASA are testing a proof-of-concept robotic assistant for future Moon and Mars missions, Google said. In early lab runs at Johnson Space Center, the robot reached an 85% baseline success rate on a pick-and-place task.

The mirrored post carries only a one-line summary. The details come from a blog post by Jim Kelly, vice president for federal at Google Public Sector, published on October 27, 2026.

A cargo handler that can switch to medical support

Google describes the robot as a versatile payload handler for cargo, logistics and equipment. It would take over routine jobs such as organizing tools or securing compartments. It could also move into medical support. In Google’s example, a crew member with severe back pain may have kidney stones. The robot would fetch and prepare an ultrasound machine so the crew medical officer can concentrate on the scan and treatment.

The work was done with NASA’s Human Research Program, the Johnson Space Center Robotics Team and the Jet Propulsion Laboratory. It builds on the Crew Medical Officer Digital Assistant (CMO-DA), a decision-support tool Google Public Sector described last year. The stated motive is distance: Google says round-trip communication on a Mars mission could take up to 44 minutes, which makes real-time medical guidance from Earth nearly impossible.

On the technical side, Google lists several pieces:

• JPL adapted satellite situational-awareness software, originally built to track orbital objects, so the robot can understand its 3D surroundings from limited sensing such as cameras.
• Google DeepMind’s MuJoCo physics engine was used to practice movements in simulation before hardware runs.
• The control policy is the open-source Pi0 vision-language-action model from Physical Intelligence, used with an eye-to-hand camera setup.
• Google Cloud G4 instances fine-tuned the 4-billion-parameter model in 2.5 hours.
• Over two years the team tested about 15 edge computing devices to find low-power hardware suited to Mars missions.

Because the model links language to movement, Google says astronauts could eventually direct the robot with plain English text prompts. The post presents that as a goal, not a current capability. It also says the robot must assist “without ever replacing human decision-making.”

What the 85% figure does and does not show

The test was a pick-and-place maneuver to locate and move a medical bag, run at Johnson Space Center’s iMETRO facility. Google calls the 85% result “a strong foundation for the next phase of rigorous training and safety refinement.” The post gives no number of trials, no description of the failures and no test in flight-like conditions. It also does not name the robot’s manufacturer, body design or sensors beyond cameras, and it gives no date for further tests or a flight demonstration. These claims come from Google and have not been independently verified.

Nothing here is available to outside users. It is a research collaboration, not a product. The only public artifact is the Pi0 checkpoint that the blog links on Hugging Face.

What to watch as the robot leaves the lab

Three signals will show whether this moves beyond a demo. The first is the link to the earlier medical software. According to reports on last year’s CMO-DA test, three physicians, including an astronaut, rated its answers as likely correct 74% of the time for flank pain, 80% for an ear problem and 88% for an ankle injury. A physical assistant that sets up equipment for such a tool will face a similar trial: clinicians judging it task by task, not a single success rate.

The second is whether the Google model stack is involved at all. Google’s post names the open Pi0 model, not Gemini. We reported separately that Gemini Robotics ER 2 is opening in preview on Google Cloud, and the NASA post does not tie the two together. Any later move to Google’s own robotics models would be a notable change.

The third is NASA’s parallel work. The agency’s Dexterous Robotics Team at Johnson is developing humanoid robots meant to work alongside crews and, in NASA’s words, “enhance a crew’s performance and productivity while reducing risk and expanding the capabilities for space exploration.” Whether the cargo-handling approach, humanoids or both end up on the Moon will depend on testing that has not been announced yet.

Google has also been pushing its models into NASA’s paperwork side. Our earlier piece covers how Gemini 4 Argon turned 80,000 NASA files into a deck. Together the two stories show a growing set of Google projects with the agency, from documents to hardware. The robot work is the least mature of them.