Summary: NASA JPL announced on January 30, 2026 that Perseverance completed the first entirely AI-planned drives on another planet, covering 456 meters on Mars on December 8 and 10, 2025. A vision-language model using Claude drafted routes from orbital imagery, and JPL’s digital twin checked the commands before onboard autonomy executed them. The milestone provides a qualified reference architecture for AI planning in high-delay, high-cost operations.
Perseverance has set another record on another world, and this one belongs to AI: NASA’s Jet Propulsion Laboratory announced on January 30, 2026 that the rover had completed the first drives on another planet planned entirely by artificial intelligence. The two runs fell on December 8 and December 10, 2025 — sols 1,707 and 1,709 of the mission — and together they covered 456 meters. No human mission planner drew these routes; a vision-language model generated them after reading orbital imagery. JPL ran the project, brought in Anthropic as a partner, and used the company’s Claude models.
For developers, the interesting part is not the romance of putting AI on Mars. It is that a complete AI planning pipeline has been fully qualified in an environment where communication delays are punishing and mistakes cost dearly: perception data flows in, structured action plans flow out, a high-fidelity simulation screens everything in advance, and the final leg is left to onboard autonomy.
Turning Orbital Images into Waypoints
The workflow operates in layers. Input to the vision-language model is orbital imagery from the HiRISE camera aboard NASA’s Mars Reconnaissance Orbiter, with slope data derived from digital elevation models laid over it. The model picks out terrain features — bedrock, outcrops, boulder fields, sand ripples — and produces one continuous path from start to goal, annotated with waypoints. It reads the very same imagery the human mission planners use; what changed is who goes first. Picking a route used to begin with an engineer studying the pictures, whereas the model now drafts the initial proposal and people review only the decisions that matter most. ScienceDaily’s report adds that the models were checked against a JPL surface mission dataset accumulated over decades of images and measurements gathered during surface operations on Mars and other worlds.
Two Drives, Cleared by a Digital Twin
Before upload, the commands for both drives had to clear JPL’s “digital twin” of the rover — a simulation that audits more than 500,000 telemetry variables — and only once they passed did the commands travel up over the Deep Space Network. The December 8 run covered 210 meters (689 feet); the December 10 run covered 246 meters (807 feet). Perseverance’s autonomous navigation software owned execution end to end, and the ground team’s involvement stayed minimal. NASA Administrator Jared Isaacman called the work a seminal contribution to roverspace and added that autonomy of this kind is returning new insights from Mars on near-real-time timescales.
The Case for Letting AI Pick the Route
The reasoning begins with physics. On average Earth and Mars sit about 140 million miles (225 million kilometers) apart, and the communication delays that follow rule out real-time remote control. The conventional response has been to hold waypoint spacing to roughly 100 meters (330 feet) and assign drives one short segment at a time, which keeps route planning a heavy and repetitive chore. JPL space roboticist Vandi Verma put it plainly: the tougher the terrain the mission takes on Mars, the more value AI-planned routes deliver. Hand the repetitive route engineering to a model, and human engineers regain time for science decisions.
Where Claude Enters the Picture
The demonstration also stands out for an uncommon pairing: a national-scale space mission running a commercial frontier model inside an operational milestone. JPL’s Rover Operations Center coordinated the work, which was carried out with Anthropic and built on Claude models — the family that writes code inside a developer’s API calls also plotted routes on Mars, 140 million miles away. For Anthropic, the exercise put Claude through a public test under extreme reliability requirements. For NASA, it confirmed a workable route for bringing commercial models into mission-critical workflows. The arrangement also carries a procurement signal: instead of training a bespoke navigation model from zero, JPL took a commercial model that already existed, adapted it, and wrapped it in the verification machinery NASA already trusts.
What Comes Next: Kilometer Drives and Onboard Models
JPL’s agenda for the capability stretches from single drives at kilometer scale, to AI that flags surface features on its own when they deserve a research stop, to models that eventually run at the edge — aboard future rovers, helicopters, and drones able to plan locally rather than wait on ground uplink. For product and infrastructure teams on Earth, the episode amounts to a reference architecture proven in the most unforgiving setting available: work through the risks in high-fidelity simulation first, then release the autonomous system. AI planning has crossed from demo into production, and Mars is today’s hardest production environment.