Anthropic’s Claude machine learning model has boldly planned what no Claude has planned before – a path across Mars for NASA’s Perseverance rover.
Perseverance traveled about 400 meters on the Martian surface last month based on an AI-generated path. It did so with the blessing of engineers at NASA’s Jet Propulsion Laboratory (JPL), who decided to delegate the meticulous work of route planning to Anthropic’s AI model.
As Anthropic explains in its writeup of the milestone, the surface of Mars can be treacherous for rovers. No one wants to be responsible for getting pricey space kit stuck in the sand, as happened with the Spirit rover in 2009.
So the Perseverance team spends a fair amount of time on route planning. This involves consulting orbital and surface imagery of Mars in order to set a series of waypoints to guide the rover’s movements. Once plotted, this data gets transmitted about 140 million miles or 225 million kilometers – the average distance from Earth to Mars – where it’s received by Perseverance as a navigational plan. Live-driving via joystick isn’t feasible given the distance involved.
Perseverance has an AutoNav system that handles real-time decision making. “AutoNav allows the rover to autonomously re-plan its route around rocks or other obstacles on its way to a pre-established destination,” NASA explains.
The re-planning may not be needed if the pre-planning went well.
The pre-planning is “time-consuming” and “laborious,” as Anthropic puts it, so JPL researchers decided to let Claude – using its vision capabilities – have a go.
“Generative AI provided the analysis of the high-resolution orbital imagery from the HiRISE (High Resolution Imaging Science Experiment) camera aboard NASA’s Mars Reconnaissance Orbiter and terrain-slope data from digital elevation models,” JPL said in an online post. “After identifying critical terrain features – bedrock, outcrops, hazardous boulder fields, sand ripples, and the like – it generated a continuous path complete with waypoints.”
Claude generated Rover commands in Rover Markup Language (RML), which is based on XML. The version of Claude available on the web could not emit RML when asked and initially denied any knowledge of RML. When pointed to Anthropic’s statement on the matter, Claude responded, “You’re absolutely right, and I apologize for my initial response!” Nonetheless, Claude could not provide an example of RML, a shortcoming that the model attributed to the lack of a publicly documented standard.
But Claude evidently did generate RML when it had access to NASA’s data. And that’s when the humans took the opportunity to check the route plan. AI models make mistakes and, even if that weren’t a concern, that’s just the sort of thing one does when programming rovers on other planets. Using a simulator representing a virtual replica of the rover, JPL engineers checked more than 500,000 telemetry variables about the rover’s projected position and potential obstacles. And they made corrections.
“When the JPL engineers reviewed Claude’s plans, they found that only minor changes were needed,” Anthropic said. “For instance, ground-level camera images (which Claude hadn’t seen) gave a clearer view of sand ripples on either side of a narrow corridor; the rover drivers elected to split the route more precisely than Claude had at this point. But otherwise, the route held up well. The plans were sent to Mars, and the rover successfully traversed the planned path.”
On Martian days (sols) 1,707 and 1,709 (starting from the landing date 18 February 2021 at 08.55pm GMT), which corresponded to December 8 and December 10, 2025, Perseverance executed routes planned by AI instead of humans.
This annotated orbital image depicts the AI-planned (depicted in magenta) and actual (orange) routes the Perseverance Mars rover took during its Dec. 10, 2025, drive at Jezero Crater. Credit: NASA/JPL-Caltech/UofA – Click to enlarge
The rover didn’t follow the set route exactly. NASA’s image of the December 10 path shows that the pre-planned route and the actual route differ slightly, presumably based on decisions made by the AutoNav system. But AI played a role, one many models can be expected to reprise as vision-language-actions models become more capable and get stuffed into robots.
“This demonstration shows how far our capabilities have advanced and broadens how we will explore other worlds,” said NASA Administrator Jared Isaacman in a statement. “Autonomous technologies like this can help missions to operate more efficiently, respond to challenging terrain, and increase science return as distance from Earth grows. It’s a strong example of teams applying new technology carefully and responsibly in real operations.”
Anthropic reports that JPL engineers say Claude can cut the time required for route planning in half. The AI biz however failed specify the amount of time being halved. Representatives of Anthropic and JPL couldn’t immediately be reached to quantify that fraction. ®
