NASA and Google test a robot that fetches medical gear for astronauts; Pi0 hits 85% in lab
Google Public Sector said in a blog post on Oct 27 that it built a proof-of-concept robotic system with NASA's Human Research Program, the Johnson Space Center robotics team and JPL to handle cargo and fetch medical equipment on deep space missions.
In a pick-and-place test of a medical bag, Physical Intelligence's open Pi0 model reached an initial 85% success rate; Google fine-tuned the 4-billion-parameter model in 2.5 hours and ran it on edge devices.
Google Public Sector said in a blog post on Oct 27 that it has built a proof-of-concept robot with NASA that is meant to take routine logistics off astronauts' hands and switch to medical support when a crew member falls ill. In the first lab demonstration, the robot had to find a medical bag and move it. Physical Intelligence's open-source Pi0 vision-language-action model did so in 85% of attempts, which Google called an initial baseline.
The post, written by Jim Kelly, vice president for federal at Google Public Sector, names three NASA partners: the Human Research Program, the Johnson Space Center robotics team and the Jet Propulsion Laboratory. It builds on the Crew Medical Officer Digital Assistant, a clinical decision support system Google described last year.
A cargo handler that doubles as a medic's helper
Google describes the system as a payload handler for a small crew. Day to day it would move cargo, organise tools and secure compartments. In a medical case, such as an astronaut with severe back pain that could signal kidney stones, it would fetch and prepare an ultrasound machine so the crew medical officer can concentrate on the scan. The long-term aim is a robot that takes spoken instructions under direct astronaut supervision, for example to capture vital signs or run a simple ultrasound, without replacing human decisions.
The reason is latency. A round trip of a signal between Mars and Earth can take up to 44 minutes, according to the post, so real-time guidance from the ground is not available.
How the stack was built
JPL is adapting situational-awareness tools it built to track objects in orbit so the robot can map its 3D surroundings from camera input. The team trained in Google DeepMind's MuJoCo physics simulator before testing at Johnson Space Center's iMETRO facility, using an eye-to-hand camera setup. Google said it fine-tuned the 4-billion-parameter Pi0 model in 2.5 hours on Google Cloud G4 instances and showed that the fine-tuned model runs locally on compact, low-power edge hardware. Over two years the team tested about 15 edge computing devices.
The post does not name the robot hardware, the number of trials behind the 85% figure or a date for flight testing.
MEKAWIRE analysis
The test is a reference case for open robot foundation models: a space agency picked a public checkpoint and a cloud fine-tune over a bespoke controller.
Physical Intelligence released Pi0 as open weights. MEKAWIRE's database lists two reported rounds for it, USD 400 million at USD 2.4 billion in November 2024 and USD 600 million at USD 5.6 billion in November 2025, a 2.3x step-up in a year (5.6 / 2.4); on Oct 27 Bloomberg reported talks to raise about USD 1 billion at about USD 20 billion. The company has confirmed none of these valuations. One day after the NASA post, Google Cloud opened Gemini Robotics ER 2 in public preview for enterprise robots, so Google now offers both its own robot model and infrastructure for a rival's. The weak point is the number itself: 85% on one pick-and-place task in a lab is a starting line, and a crew cannot accept one failed fetch in seven during an emergency. MEKAWIRE's explainer on VLA models covers how Pi0, GR00T and Gemini Robotics differ.
What to watch: a named robot platform, a repeatability figure from a larger trial set, and whether NASA moves the system into an analogue mission ahead of Artemis surface operations.
- Initial 85% success rate with Physical Intelligence's open Pi0 model
- NASA Human Research Program, Johnson Space Center robotics team, JPL
- 4-billion-parameter Pi0 fine-tuned in 2.5 hours on Google Cloud G4; MuJoCo simulation
- Google Public Sector blog post, Oct 27, 2026
- Locate and move a medical bag (pick-and-place), eye-to-hand cameras
- Runs locally on compact edge devices; about 15 devices tested over two years
- Robot hardware, number of trials, flight-test date