Google Cloud, NASA Test AI-Powered Robot for Deep-Space Missions
aiinsider ·
Google Cloud and NASA are collaborating on an AI-powered robotic system designed to assist astronauts during deep-space missions where communication delays with Earth make real-time support impractical. The proof-of-concept project involves multiple agencies, including NASA’s Human Research Program and the Jet Propulsion Laboratory, aiming to handle logistics and medical-support tasks autonomously. In early tests at Johnson Space Center, a Pi0 vision-language-action model successfully located and moved a medical bag with an 85% success rate. The 4-billion-parameter model was fine-tuned in approximately 2.5 hours using Google Cloud G4 instances and can operate locally on compact edge hardware. This local processing capability is crucial for future missions, such as those to Mars, which may experience round-trip communication delays of up to 44 minutes. Researchers utilized Google DeepMind’s MuJoCo physics simulator to virtually train robotic movements before deploying them on physical hardware. The system allows astronauts to issue simple spoken or written commands to direct the robot for tasks like retrieving equipment or preparing medical devices. Ultimately, the technology aims to reduce reliance on terrestrial computing while keeping human astronauts in control of critical decision-making processes.
aiinsider