Claude takes the wheel on Mars

Operating a rover on Mars means you are permanently stuck in the past. It takes about 20 minutes for a radio signal to cross the 362-million-kilometer void between Earth and the Red Planet, making real-time steering impossible.

For decades, laying out the safe path has fallen entirely on elite human drivers. If they mess up, a multi-million-dollar rover can end up permanently stuck in a sand trap, which is exactly what killed the Spirit rover back in 2009.

Now, Anthropic's Claude just handled the job.

Navigating Mars in XML

On sol 1707 and sol 1709—Martian days corresponding to December 8 and 10, 2025—NASA's Jet Propulsion Laboratory sent Claude-generated drive commands to the Perseverance rover. The mission was to navigate a roughly 400-meter rocky route across the Jezero crater.

Perseverance has an onboard AutoNav system for dodging immediate obstacles, but it cannot see the big picture or plan far ahead. Human operators normally piece together breadcrumb waypoints using satellite imagery.

JPL engineers set up a workflow inside Claude Code, feeding the AI years of accumulated Mars driving expertise. Claude examined overhead imagery using its vision capabilities, broke the path into 10-meter segments, and plotted out the waypoints.

It didn't just give vague suggestions. Claude wrote executable commands directly in Rover Markup Language, a bespoke XML-based syntax built specifically for Mars exploration. It even reviewed its own trajectory and suggested revisions before handing it over.

Surviving a 500,000-variable simulator

Nobody uploads unvetted LLM output directly to a spacecraft. JPL ran Claude's generated route through its daily simulation pipeline, testing it against more than 500,000 variables to predict machine slip, tilt, and terrain hazards.

When human drivers inspected the results, only minor edits were needed. Ground-level imagery that Claude had not seen revealed narrow corridors flanked by subtle sand ripples, prompting humans to split the line slightly more conservatively. The rest ran as planned.

The route was beamed over NASA's Deep Space Network, and Perseverance executed the drive without issue. JPL estimates that using Claude can cut daily route-planning time in half while making drives more consistent.

Why it matters

Slashing route-planning time by 50% translates directly into more drives, faster sample collection, and more Martian science before hardware components degrade.

We think Anthropic and JPL approached this brilliantly. Instead of pretending an LLM can steer a rover end-to-end, they treated Claude as an expert copilot working inside an existing, battle-tested framework. The model read specialized overhead data, generated obscure domain-specific XML, and let a massive physics simulator verify the results.

The bottleneck, however, is that Claude relied strictly on overhead views and missed subtle ground-level sand ripples. Total autonomy will remain out of reach until models can cross-reference high-altitude satellite data with local hazard cameras simultaneously.

Still, this test sets a massive precedent. As NASA preps for the Artemis lunar missions and looks toward destinations like Europa or Titan—where signal lag stretches to hours—general-purpose models that can plan, code, and self-critique in isolated environments will quickly move from novelty to mission-critical infrastructure.