February marks the five-year anniversary of Perseverance‘s time on Mars. But when it comes to trying out new things, NASA’s rover seems to be as active as it was on the day it first arrived on the Martian regolith.

According to NASA JPL, Perseverance completed two driving stints using an itinerary created by Claude, a generative AI model developed by California-based firm Anthropic. To create the rover’s travel plan, the AI program referred to the same datasets used by human engineers in creating waypoints—a “breadcrumb trail” for the rover to follow, Anthropic explained in a statement.

These drives took place on the 1,707th and 1,709th Martian days of the mission, during which Perseverance traveled 689 feet (210 meters) and 807 feet (246 meters), respectively.

“This demonstration shows how far our capabilities have advanced and broadens how we will explore other worlds,” NASA Administrator Jared Isaacman said in JPL’s statement. “It’s a strong example of teams applying new technology carefully and responsibly in real operations.”

Driving a rover from Earth

Suffice to say, operating a rover on Mars is a lot different from driving a car on Earth. On average, Mars lies around 140 million miles (225 million kilometers) from Earth. The distance between Earthbound drivers and the actual rovers creates a sizable communication lag, leaving the rover unprepared for unexpected environmental barriers on the Martian surface.

“This is high-stakes work,” Anthropic noted in its statement. For instance, Perseverance’s predecessor, Spirit, drove into a sand trap and “never moved again,” it added.

To prevent such travesties, researchers will meticulously study images of the Martian terrain taken both from space and by the rover. Based on the images, engineers devise a travel plan that hopefully evades natural obstacles such as bedrock, outcrops, bumpy boulder fields, and sand ripples.

The AI takes the wheel, kind of

This is a process that takes a lot of time and effort. And so, NASA and collaborators wondered if AI could help alleviate the intensity of manually planning rover routes.

Before the demonstration, JPL engineers fed Claude data on how to operate the rover, which the AI used to generate its commands. The commands were then checked for accuracy through simulations verifying over 500,000 variables, NASA said.

The engineers found Claude’s commands to be surprisingly unproblematic, only requiring minor fixes to account for certain images Claude hadn’t seen, Anthropic explained. Once the team was satisfied with the results, the AI-written plan was transmitted to Perseverance and executed in December.

“The fundamental elements of generative AI are showing a lot of promise in streamlining the pillars of autonomous navigation for off-planet driving,” Vandi Verma, an engineer at JPL’s Perseverance team, said in the NASA release.

“It means, in short, that we’ll learn much more about Mars,” Anthropic added, as engineers anticipate the program could help cut the route-planning time by half.

Although Perseverance has only driven a few hundred feet so far, AI could soon start planning mile-long routes across the Martian terrain, Verma said.