A NASA–Google collaboration tested a robot on locating and moving a medical bag, reaching an initial 85% success rate in early laboratory work, according to Google’s 27 October announcement. The proof of concept used Physical Intelligence’s Pi0 model and an external camera configuration. Google does not state the number of trials or a detailed test protocol, so the percentage cannot yet establish reliability across mission conditions.[1]
Google says the four-billion-parameter Pi0 vision-language-action model was fine-tuned on G4 cloud infrastructure in 2.5 hours and could then operate on local edge hardware. An eye-to-hand camera observes the workspace from outside the robot’s hand. MuJoCo simulation supported testing before physical demonstrations.[1][2]
What has been demonstrated, and what has not
| Question | Current evidence | Still unresolved |
|---|---|---|
| Why a robot is useful | Potentially reduces routine cargo, tool and equipment handling for a lean crew | The actual time saved during a representative missionSource 1 |
| Core manipulation test | 85% initial success locating and moving a medical bag in an early lab environment | Trial count, test protocol and performance beyond this task configuration were not disclosedSource 1 |
| Model and training | Open-source Pi0, about 4 billion parameters, fine-tuned in 2.5 hours | Onboard compute, radiation tolerance, thermal limits and long-duration maintenanceSource 1Source 2 |
| Medical role | Proposed future role: retrieving or preparing equipment such as an ultrasound machine | Diagnosis and treatment remain with the human crewSource 1 |
The mission problem is delayed help
Communication delay is the operational reason to pursue local assistance. Google cites an Earth–Mars round-trip delay of up to 44 minutes, making continuous ground control unsuitable for time-sensitive work. The proposed system would handle bounded physical tasks under crew supervision, with computation available onboard.[1]
Equipment handling is a proposed extension of the medical-bag demonstration. Google describes a future robot retrieving or preparing an ultrasound machine while a crew medical officer assesses the patient. The announcement does not report that clinical scenario as a completed test.[1]
Why the 85% figure needs context
The 85% figure needs a denominator and a definition of success before readers can compare it with other systems. Repeated trials across changed object positions and working conditions would help establish how broadly the result holds. The announcement provides a starting measurement rather than that fuller reliability record.[1]
NASA and Google used the MuJoCo physics simulator before moving to hardware. Simulation is valuable because it permits repeated practice and unsafe edge cases without risking equipment. Yet the gap between a simulator and a spacecraft is unusually consequential. Microgravity changes object behaviour; cameras can be occluded; restraints, gloves and confined geometry alter reach; and a system operating for months must handle drift rather than a short evaluation session.[1]
Autonomy has to fit the spacecraft
The model is only one layer of the engineering problem. A deep-space system must run with limited local compute, power and cooling. It must recover after software faults, preserve safe motion limits and remain useful when its preferred network service is unavailable. Google describes space as an edge-computing environment and says it has tested roughly 15 edge devices over two years, but the announcement does not establish which hardware would fly or how the full stack would be qualified.[1]
The demonstration connects a specific equipment-handling task with the practical need for local computation in space. The next useful evidence would include trial counts, operating conditions and failure categories, followed by repeated tests under representative constraints. The current announcement establishes a promising supervised prototype, not flight readiness.[1]
Sources & further reading
- Google Cloud and NASA: Testing robotic systems for deep space missions Google Cloud · 27 Oct 2026 · Accessed 30 Oct 2026
- Pi0: A Vision-Language-Action Flow Model for General Robot Control Physical Intelligence · 4 Feb 2025 · Accessed 30 Oct 2026