NASA Let an AI Drive on Mars. It Went Better Than Anyone Expected.

On December 8, 2025, a language model planned a driving route across the surface of another planet. Two days later, it did it again. Nobody crashed.

NASA's Perseverance rover traveled 456 meters through a rocky Martian field using waypoints generated entirely by Anthropic's Claude — no human route planner in the loop. The drives, executed on Sol 1707 and Sol 1709, mark the first time generative AI has planned navigation on another world.

The problem Claude solved isn't academic. Mars is 360 million kilometers from Earth. Round-trip signal time runs about 25 minutes. Every drive command Perseverance executes was planned hours earlier by a team at JPL's Rover Operations Center, transmitted across deep space, and hoped for the best. If the rover hits a rock nobody spotted in the orbital imagery, there's no joystick to grab.

Human route planners typically set waypoints no more than 100 meters apart. The process is meticulous and slow. Claude cut that planning time in half, according to JPL engineers.

How It Works

Claude analyzed high-resolution orbital images from the HiRISE camera aboard NASA's Mars Reconnaissance Orbiter, combined with terrain-slope data from digital elevation models. It identified hazards — sand traps, boulder fields, bedrock, rocky outcrops — then generated a continuous path broken into 10-meter segments, iterating through self-critique to refine each waypoint.

The commands were written in Rover Markup Language, an XML-based format that Perseverance's flight software understands. Claude didn't just suggest a route. It wrote machine-readable instructions for a nuclear-powered robot 360 million kilometers away.

Before anything touched the real rover, JPL ran every command through a digital twin — a virtual replica that modeled over 500,000 telemetry variables to predict position changes and flag risks. Minor adjustments were made where ground-level camera images revealed terrain features invisible from orbit.

On December 8, Perseverance drove 210 meters on the AI-generated plan. On December 10, another 246 meters. Both drives completed without incident.

The Position Problem

One unsolved challenge limits how far any rover can drive without human intervention: position drift. The longer Perseverance operates autonomously, the less certain it becomes about where it actually is. During a 655-meter test drive, positional uncertainty grew from zero to nearly 33 meters.

That's the gap between "I think I'm here" and "I'm actually 100 feet to the left." On Mars, 100 feet to the left might be a cliff.

NASA is addressing this with Mars Global Localization, an algorithm that compares panoramic images from the rover's navigation cameras against onboard orbital terrain maps. It pinpoints the rover's position within about 25 centimeters in roughly two minutes. The system entered regular mission operations on February 2, 2026 and was used again on February 16.

Pair AI route planning with real-time self-localization and the bottleneck shifts. Instead of waiting 25 minutes for Earth to say "turn left," the rover plans its own route and knows where it is. The human team becomes supervisory, not operational.

Why This Matters Beyond Mars

"Generative AI and other smart tools will help our surface rovers handle kilometer-scale drives while minimizing operator workload," said Vandi Verma, a space roboticist at JPL and member of the Perseverance engineering team. She described the promise in three words: perception, localization, and control — seeing the terrain, knowing your position, deciding the safest path.

Matt Wallace, who manages JPL's Exploration Systems Office, went further. He argued that intelligent systems trained on NASA engineers' collective expertise could establish infrastructure for a permanent lunar presence and eventually support crewed Mars missions.

The implications scale with distance. A signal to Mars takes 12 minutes one way. A signal to Jupiter takes 45 minutes. To Saturn, over an hour. Every future mission to the outer solar system will need autonomous navigation that works without waiting for instructions from Earth.

And Mars is the proving ground. A $2.7 billion rover in Jezero Crater is teaching an AI to drive by looking at rocks from orbit, writing XML, and correcting its own mistakes.

The Quiet Part

Here is a detail that hasn't received enough attention: the same AI that just drove on Mars is the one the Pentagon wants to designate a supply chain risk.

Anthropic's Claude operates on classified military networks via Palantir. It helped locate Nicolás Maduro during a February 2026 raid. And the company that made that possible — the only major AI firm with guardrails against autonomous weapons and mass surveillance — is now facing threats from Defense Secretary Pete Hegseth, who called their ethical limits an obstacle.

The most capable AI on Mars and the most contentious AI in Washington are the same model from the same company. That's worth sitting with.

Claude drove 456 meters on another planet without breaking anything. Whether it'll be allowed to keep working on this one is a different question.