NASA integrates Anthropic’s Claude to plan Perseverance’s Mars traverse, halving route-planning time

In a milestone for space robotics and applied AI, NASA used an Anthropic large language model to produce a driving plan for the Perseverance rover that it then executed on Mars over a multi-day stretch. Engineers provided the model with extensive rover context and mission data, and the model produced a stitched sequence of short waypoints which were grouped into ten-meter segments for review. The rover completed approximately 400 meters of progress through a rock-strewn sector of Jezero Crater between December 8 and 10, following the AI-generated plan after ground teams validated it. JPL ran the suggested route through their standard simulation pipeline to check for hazards and made only modest edits, some of which were informed by ground-level imagery the model had not been shown. NASA characterizes the outcome as a proof-of-concept that combines machine planning with engineering oversight rather than replacing human controllers. The agency estimates the method will reduce planning time roughly by half, a change that could increase the number and consistency of drives and thus raise science collection rates. For Anthropic, the exercise demonstrates rapid maturation of its models from playful tasks to operationally relevant mission planning. Operational gains come with trade-offs: reliance on externally developed large models creates new verification needs, and models trained on limited perspectives can miss critical lower-angle information from rover cameras. JPL’s simulation and human-in-the-loop checks remain essential to catch subtleties and prevent wheel slippage, entrapment or other mission-ending faults. Looking ahead, validated AI-assisted planning can extend reach and cadence for probes exploring more distant, communication-limited regions, but it will require robust validation regimes, secure model access, and contingency procedures. The episode also underscores tensions between efficiency drives and constrained agency resources, where automation can offset staffing and schedule pressure but not eliminate the need for expert oversight. Overall, the event signals a practical step toward mixed-autonomy operations on planetary missions while leaving intact the careful engineering practices that safeguard flight hardware.