Google Cloud, NASA Test AI-Powered Robot for Deep-Space Missions
By Greg Bock · · 3 min read
Insider Brief
- Google Cloud and NASA are testing an AI-powered robotic system designed to handle logistics and medical-support tasks during future deep-space missions where communications with Earth may be delayed.
- In an early test at Johnson Space Center, Google said the system used a Pi0 vision-language-action model to locate and move a medical bag with an 85% success rate.
- Google said the 4-billion-parameter model was fine-tuned in about 2.5 hours and can run locally on compact edge hardware, a capability intended to reduce reliance on Earth-based computing during deep-space missions.
Google Cloud and NASA are testing an AI-powered robotic system designed to help astronauts handle routine logistics and some medical-support tasks during future deep-space missions.
According to Google, the proof-of-concept project brings together NASA’s Human Research Program, the Johnson Space Center Robotics Team, Jet Propulsion Laboratory and Google Public Sector. The system is being developed for missions where long communications delays could make real-time support from Earth impractical.
Google said the robot is designed to handle tasks such as moving cargo, organizing equipment and retrieving medical devices, allowing astronauts to spend more time on mission-critical work. In an example described by Google, a robot could retrieve and prepare an ultrasound machine for an astronaut experiencing severe back pain while a crew medical officer concentrates on diagnosis and treatment.
The project builds on earlier work by Google Public Sector and NASA on the Crew Medical Officer Digital Assistant, a clinical decision-support system intended to provide astronauts with medical guidance in remote environments.
Teaching the Robot to Work in Space
The team is testing how AI models can help a robot understand its surroundings and respond to instructions inside confined spacecraft environments. JPL is adapting situational-awareness technology originally developed for tracking objects in orbit to help the system understand its three-dimensional surroundings using cameras.
Google said the team also used Google DeepMind’s MuJoCo physics simulator to train and test robotic movements virtually before moving them onto physical hardware. In one proof-of-concept test at Johnson Space Center’s iMETRO facility, the robot was asked to locate and move a medical bag. Google said an open-source Pi0 vision-language-action model achieved an 85% success rate in the early laboratory tests.
Vision-language-action models connect visual information and language instructions with physical robot movements. In this case, the approach could eventually allow astronauts to direct the robot using ordinary spoken or written commands.
Running AI Away From Earth
The project is also testing whether the models can operate without relying on continuous connections to terrestrial data centers. Google said it used Google Cloud G4 instances to fine-tune the 4-billion-parameter Pi0 model in about 2.5 hours. The resulting model was then able to run locally on compact edge-computing hardware.
According to Google, the team has evaluated roughly 15 edge devices over the past two years to identify hardware capable of running the models within the power and weight constraints expected on future missions. Google pointed out that a future Mars mission could experience round-trip communication delays of as much as 44 minutes, so that local-processing capability is important for deep-space operations.
The longer-term goal is a robotic system that astronauts can direct with simple instructions to perform a range of logistical and medical-support tasks. Google said the system would remain under astronaut supervision and is meant to assist rather than replace human decision-making.
