Summary
Anthropic says NASA JPL engineers used Claude to plan an approximately 400-metre route for the Perseverance rover. Engineers checked the proposed waypoints in simulation and reviewed them before sending the commands to Mars.
Anthropic says NASA Jet Propulsion Laboratory engineers used its Claude model to plan an approximately 400-metre route for the Perseverance rover on Mars. The plans covered sols 1707 and 1709, corresponding approximately to December 8 and 10, 2025. After simulation checks and human review, the commands were sent to Mars and the rover completed the planned route. Anthropic describes it as the first AI-planned drive on another planet.
Claude helped prepare the route; it did not steer Perseverance in real time. The demonstration used the rover’s existing process for planning and sending drives, with engineers responsible for reviewing the AI-generated plan before it was transmitted.
How Claude produced the route
Signals take about 20 minutes to reach a rover from Earth. Operators therefore plan a drive in advance rather than controlling each turn as it happens. Before this demonstration, human rover drivers laid out a sequence of waypoints—sometimes called a “breadcrumb trail”—using orbital and rover images. Perseverance’s AutoNav system can help it avoid obstacles between those points, but it cannot plan far ahead on its own.
JPL engineers set up a process in Claude Code and supplied it with data and experience gathered from years of rover operations. Claude used overhead images to plot waypoints in roughly ten-metre segments, then iterated on the route by critiquing and revising its proposed points. It wrote commands in Rover Markup Language, an XML-based language developed for NASA’s Mars Exploration Rover mission.
Simulation and human review
Before sending the plans, engineers ran the commands through a simulation used in Perseverance operations. The simulation modelled more than 500,000 variables to project the rover’s position and identify potential hazards along the route.
JPL engineers then reviewed Claude’s waypoint plans and made minor adjustments. In one narrow corridor, images from cameras at rover level showed sand ripples more clearly than the overhead images Claude had analysed, so the drivers divided that part of the route more precisely. The rover subsequently traversed the planned path.
A planning assistant, not an autonomous rover
The test put Claude to work on a specific planning task within an established human-supervised process. That matters because route planning is a demanding part of rover operations: a poor path can risk the vehicle getting stuck, tipping or losing traction, while commands must be prepared well before their effects can be observed from Earth.
JPL engineers estimate that using Claude this way could halve route-planning time and make planning more consistent. That is a projected efficiency gain; the account does not report it as a measured reduction from routine operations. If the workflow proves useful beyond this demonstration, less time spent preparing routes could give rover teams more time for other mission work. For now, the reported achievement is narrower: an AI-assisted plan passed operational checks, received human review and was driven successfully on Mars.