NASA’s Perseverance rover has quietly crossed another frontier: in December 2025, it followed a route across Mars that was laid out not by human mission planners, but by Anthropic’s Claude AI – the first time a large language model has helped plan a drive on another world. What sounds like a gimmick – “chatbot plans rover road trip” – is actually a glimpse of how future space missions could lean on AI to make exploration faster, more autonomous, and a lot more scalable.

If you’ve ever tried to remote-control a robot across a room with a bad Wi-Fi connection, you have a tiny sense of what NASA deals with on Mars. Perseverance sits roughly 362 million kilometers away, and every command has to ride a radio link that adds about 20 minutes of one‑way delay. There’s no real‑time joystick driving here. Human “drivers” at NASA’s Jet Propulsion Laboratory (JPL) in California normally spend hours poring over high‑resolution orbital imagery and rover photos to lay out a breadcrumb trail of waypoints – step‑by‑step instructions that tell the rover where to go, how far to drive, and what hazards to avoid. Once they’re satisfied, the plan is vetted, simulated, and finally beamed across space, where the rover executes it mostly on its own.

Perseverance does have some onboard smarts: its AutoNav system can steer around obstacles and adjust its path between those pre‑planned waypoints. But AutoNav only sees what the rover sees – it can’t “look” ahead over the horizon the way humans can from orbit, so the strategic planning still happens on Earth. That’s where Claude comes in. Late last year, JPL engineers asked a very 2025 question: could a vision‑enabled AI model help with the grunt work of route design, without compromising safety?

The test they set up was modest but high‑stakes. Perseverance would attempt two AI‑planned drives on Sol 1707 and Sol 1709 – the mission’s Martian days corresponding to December 8 and 10, 2025 – across a rocky stretch of Jezero crater, the ancient lakebed that is the rover’s home. The target distance wasn’t huge: roughly 400 meters total, about one lap of an athletics track, but over terrain that could easily bog or strand a multi‑billion‑dollar robot if planners got sloppy. JPL is haunted by the memory of Spirit, a predecessor rover that in 2009 drove into a sand trap and never got out. Nobody was about to outsource Perseverance’s fate to a chatbot without guardrails.

To make Claude useful, NASA didn’t just toss it a prompt like “draw me a safe route.” Engineers instead wired up Claude Code, Anthropic’s software‑agent flavor of the model, into their mission tooling and fed it years of detailed rover driving know‑how. That included high‑resolution orbital imagery from HiRISE, the powerful camera aboard the Mars Reconnaissance Orbiter, plus digital elevation models that capture slopes and subtle changes in terrain – the difference between a firm patch of bedrock and a wheel‑eating dune. In effect, Claude got access to the same data that human planners rely on.

From there, the model went to work the way a careful junior engineer might. It generated a continuous path built from short, roughly 10‑meter segments, each with a waypoint in a language only Mars rovers care about: Rover Markup Language, an XML‑style format originally created for the Spirit and Opportunity missions. It iterated on its own output, critiquing and refining those waypoints to hug safe ground, skirt boulder fields, and thread between subtle sand ripples it could spot from above. The idea wasn’t magic autonomy so much as scalable pattern‑matching: let the AI do the first pass over a complex terrain map, then let humans scrutinize and adjust.

Scrutiny is where things got serious. Before any command touches Perseverance, it passes through JPL’s “digital twin” – a virtual replica of the rover that simulates how it will move, down to the behavior of individual wheels and suspension components. For this experiment, Claude’s route was run through that twin, checking more than 500,000 telemetry variables to predict the rover’s pose, wheel loads, and potential collisions for each segment of the drive. Engineers then overlaid the AI‑generated path on their navigation tools, compared it to what an experienced human planner might choose, and combed through ground‑level images that Claude hadn’t seen.

The result: they kept almost all of it. Humans made some tweaks where sand ripples squeezed a “corridor” tighter than orbital images suggested, splitting a section of the route to better thread the needle. But there was no wholesale rewrite. NASA signed off. On December 8, Perseverance drove about 210 meters along the AI‑planned course, then followed up with another 246 meters on December 10, successfully completing the first large language model–planned traverse on another planet.

There’s an important nuance here: Claude didn’t “drive” the rover in the science‑fiction sense. AutoNav still handled the real‑time, wheel‑by‑wheel decisions about where to put its tracks within each waypointed segment. Claude’s role was higher‑level – deciding what those segments should be in the first place, given what the maps and elevation data showed. Think city planner vs. daily commuter: the AI laid out the streets; Perseverance still decided whether to nudge left or right around a loose rock.

So why does NASA care so much about shaving a few hours off route planning? In planetary exploration, time really is science. Human rover planners are scarce, highly trained specialists, and every sol they spend hand‑drawing a path is a sol they’re not using to design new experiments, analyze incoming data, or coordinate complex multi‑instrument observations. JPL estimates that delegating waypoint generation to Claude could cut planning time roughly in half for suitable traverses, and make drives more consistent by encoding best practices into a reusable AI “skill.” Over months and years, that translates into more distance covered, more rocks cored, more atmospheric samples taken – more chances to find evidence of ancient life in Jezero’s sediments.

The collaboration also lands at an interesting cultural moment for NASA. This is not the agency’s first brush with AI – JPL has spent decades experimenting with automated planning tools and embedded autonomy – but it is one of the first public demonstrations of a modern, general‑purpose generative model woven into day‑to‑day flight operations. There’s a careful balance to strike: excitement about AI’s potential on one hand, and institutional caution on the other. A bug in a spreadsheet is annoying; a bug in rover commands can end a mission.

That’s why this test is as much about process as performance. Claude was kept behind layers of simulation and human review; it never had direct, unsupervised control over Perseverance. The work flowed more like a teammate than an autopilot: AI proposes, digital twin verifies, humans dispose. For a safety‑first culture like NASA’s, that pattern – “AI in the loop, not over it” – is probably the only acceptable way generative models will find a home in critical systems.

Looking ahead, you can see why JPL is already talking about this as a dress rehearsal for something bigger. The further we travel from Earth, the more painful that speed‑of‑light delay becomes. Even at Mars, round‑trip communications can stretch to 40 minutes; at Jupiter’s moons, they’re in the hours. You can’t wait for a shift of humans in Pasadena to approve every turn of a probe exploring a fissure on Europa or a methane‑drenched plain on Titan. Generative AI systems with strong reasoning and code‑writing skills could help those spacecraft adapt in place – re‑planning routes, tweaking sampling campaigns, or triaging what data to send home when bandwidth is tight.

Closer to home, NASA’s Artemis program – the long‑term push to put humans back on the Moon and establish a base near the south pole – is another obvious target. Lunar missions will juggle power constraints, hazardous terrain, fleets of rovers and landers, and fragile life‑support systems operating days away from Earth‑based help. An AI assistant that can ingest telemetry, update models, write or validate procedures, and even generate code for habitat systems on the fly could act as a force multiplier for small crews living and working off‑world. Claude’s stint as a route planner is a tiny, bounded example of that much broader role.

For Anthropic, meanwhile, this is a marketing milestone with substance behind the hype. A year ago, Claude made headlines for struggling through classic video games; now, a descendant of that model has helped steer a flagship NASA mission across another planet. It’s a neat narrative arc, but it also underlines how quickly these systems can move from fun demos to infrastructure – provided someone does the hard integration work. JPL had to build pipelines, permissions, and safety checks before a single Claude‑generated line of Rover Markup Language was allowed anywhere near Perseverance’s command queue.

There are, of course, open questions. How do you certify an AI model for use in mission‑critical contexts when its internal reasoning is largely opaque? How do you guard against subtle biases in training data bleeding into navigation decisions – for example, consistently preferring one kind of terrain over another in ways that might limit scientific return? And what happens when a future mission leans so heavily on onboard AI that its actions can’t be fully reconstructed after the fact, a key requirement in today’s mishap investigations? Those are problems NASA and its commercial partners will have to solve long before they trust generative models anywhere near astronaut life‑support.

Still, it’s hard not to feel a little sense of sci‑fi made real when you picture the scene: a car‑sized robot, bristling with cameras and coring drills, trundling across the floor of an ancient Martian lakebed on a path suggested by a machine that normally spends its days drafting emails and summarizing PDFs. Back on Earth, a roomful of engineers watches its progress on big screens, the AI’s magenta route overlaid on a monochrome landscape that, not that long ago, was just a smudge through a telescope.

Four hundred meters isn’t much in the grand scheme of Mars exploration; Perseverance has already driven many kilometers since its landing in 2021, collecting rock cores that may one day be returned to Earth. But as a proof of concept for AI‑assisted exploration, those 400 meters might end up being some of the most consequential of the mission. They hint at a future where rovers, landers, and orbiters work less like radio‑controlled toys and more like genuine partners – machines that can understand their environment, write their own playbooks, and push the frontier outward while we’re still waiting for the next downlink.