Google Cloud and NASA test a Pi0-driven robot for deep space crews, hitting 85% in the lab

What Google and NASA announced

Google Public Sector said on 27 October 2026 that it has developed a proof-of-concept robotic system with NASA's Human Research Program, the Johnson Space Center Robotics Team and the Jet Propulsion Laboratory. The post is by Jim Kelly, vice president for federal at Google Public Sector.

The robot is meant as a general payload handler for a small crew on the Moon or on the way to Mars. In routine use it would move cargo, organise tools and secure compartments. When needed it would switch to medical support. The example given: if an astronaut has severe back pain that could signal kidney stones, the robot fetches and prepares an ultrasound machine while the human crew medical officer reads the scan and treats the patient.

The work builds on the Crew Medical Officer Digital Assistant, a clinical decision support system the two organisations described in 2025.

The model and the numbers

The robot runs the open-source Pi0 vision-language-action model from Physical Intelligence, which turns English text prompts into physical actions. The team fine-tuned the 4 billion parameter model in 2.5 hours on Google Cloud G4 instances. In a technology demonstration at Johnson Space Center's iMETRO facility, the robot had to find and move a medical bag. With an eye-to-hand camera setup it reached an initial baseline success rate of 85% in early lab environments. Google calls this a foundation for further training and safety work, not a finished system.

Movements were first practised in Google DeepMind's MuJoCo physics simulator before running on hardware. JPL adapted tools first built to track objects in orbit to help the robot understand its 3D surroundings from limited camera input. Because a spacecraft cannot rely on Earth data centres, the team checked that fine-tuned models run on small, low-power edge devices, testing about 15 of them over two years.

What changed and why it matters

The test points to robots that take open-ended instructions in plain language, the same approach labs are pursuing for homes and warehouses. The reason is latency: the post notes that a Mars round trip can mean up to 44 minutes of communication delay, which rules out remote control from Earth.

An 85% success rate on one task in a lab is far from flight readiness, and the post is clear that this is an early phase. It does not say which robot arm or platform was used, how many trials produced the 85% figure, or whether NASA has assigned the system to a mission. Physical Intelligence also offers its newer models through a paid API beta, but this project used the open Pi0 weights.

What to watch next

Google says the goal is a robot that takes voice instructions under direct astronaut supervision and could help with tasks such as taking vital signs or simple ultrasounds, without replacing human decisions. Watch for a published success rate after further training, tests in a spacecraft mock-up, and any link to Artemis Moon Base planning. More physical AI news is under Robotics & Physical AI.

Questions people ask

What AI model does the NASA robot prototype use?

The open-source Pi0 vision-language-action model from Physical Intelligence, a 4 billion parameter model the team fine-tuned in 2.5 hours on Google Cloud G4 instances.

How well did the robot perform?

It reached an initial baseline success rate of 85% on a pick-and-place task with a medical bag in early lab tests, according to Google Public Sector.

Will the robot fly on an Artemis mission?

Not stated in the announcement. Google describes it as a proof of concept that could assist astronauts on future NASA missions.

Sources

1. Google Cloud and NASA: Testing robotic systems for deep space missions (Google Cloud, 27 Oct 2026)