Skip to content
Perseverance Drove 456 Metres Across Jezero Crater on 8 and 10 December 2025 on Waypoints That an AI Model Drafted and NASA Engineers Checked
AI & Technology

Perseverance Drove 456 Metres Across Jezero Crater on 8 and 10 December 2025 on Waypoints That an AI Model Drafted and NASA Engineers Checked

Photo: NASA / Wikimedia Commons, Public domain (cropped)

English

NASA's Jet Propulsion Laboratory says these were the first drives on another world planned by artificial intelligence. The AI drafted the route from orbital photographs, and a digital twin of the rover checked it against more than 500,000 telemetry variables before the commands went to Mars.

· · 6 min read

On 8 and 10 December 2025, NASA’s Perseverance rover drove 210 metres and then 246 metres along the rim of Jezero Crater on waypoints that an AI model had drafted from orbital photographs. NASA’s Jet Propulsion Laboratory (JPL) announced on 30 January 2026 that these were the first drives on another world planned by artificial intelligence, 456 metres in all.

JPL’s Rover Operations Center led the work with Anthropic, whose Claude models did the planning. Administrator Jared Isaacman of NASA called it “a strong example of teams applying new technology carefully and responsibly in real operations.”

What the AI was asked to do

Mars is on average about 140 million miles (225 million kilometres) from Earth, JPL says, and the delay in sending a signal makes steering a rover live impossible. Anthropic puts the one-way delay at about twenty minutes. So for 28 years, according to NASA, human “drivers” have studied terrain and rover status data, then sketched a route as a string of waypoints, which are fixed places where the rover takes up a new set of instructions. Spacing is usually no more than 330 feet (100 metres) between them, and the finished plan travels to the rover through NASA’s Deep Space Network.

That sketching is the job JPL handed to the AI on the two drives, which took place on the 1,707th and 1,709th Martian days (sols) of the mission. Perseverance landed in Jezero Crater, which is 45 kilometres wide, on 18 February 2021 to look for signs of ancient microbial life and to collect rock samples, according to NASA’s mission page.

How the route was drafted

NASA says the team used vision-language models, a kind of generative AI that reads images as well as text. The model worked from the same material human planners use: high-resolution orbital images from the HiRISE camera aboard NASA’s Mars Reconnaissance Orbiter, and terrain-slope data from digital elevation models. Bedrock, outcrops, boulder fields and sand ripples were identified first, and a continuous path with waypoints followed.

Anthropic, the company that makes Claude, published its own account, and the details below come from it, not from NASA. JPL engineers gave Claude Code, its programming tool, years of data and experience from driving the rover. Next, Claude wrote the commands in Rover Markup Language, an XML-based command language first developed for the Mars Exploration Rover mission. Segments of about ten metres were strung together, and the model then reviewed its own waypoints and suggested revisions.

The check before anything reached Mars

Before the commands were sent, the engineering team ran them through JPL’s “digital twin,” a virtual replica of the rover. NASA says the twin verified more than 500,000 telemetry variables, to make sure the instructions were compatible with the rover’s flight software.

According to Anthropic, the engineers found only minor changes were needed. Ground-level camera images, which Claude had not seen, showed sand ripples on either side of a narrow corridor more clearly, and the rover drivers chose to split the route there more precisely than the AI had. Otherwise the route held up, the company says, and the rover completed the planned path.

An orbital image from NASA shows the AI-planned route for the 10 December drive in magenta and the route actually driven in orange. The Register noted that the two lines differ slightly and guessed that the rover’s own obstacle-avoidance software accounted for it; NASA’s release does not say.

The caution has a history. In April 2009 the Spirit rover broke through a crust and became embedded in soft sand at a site the team called “Troy,” according to NASA. NASA’s mission page for Spirit and Opportunity records that Spirit concluded its mission in 2010.

Why the distance appears as 400, 450 and 456 metres

On the NASA side, the two figures are 689 feet (210 metres) on 8 December and 807 feet (246 metres) on 10 December, which add up to about 456 metres. Space.com reported “nearly 1,500 feet (456 meters),” and Astronomy gave 1,496 feet (455.9 metres). Anthropic’s page says “approximately four-hundred-meter,” The Register wrote about 400 metres, and Payload wrote 450. NASA’s figures are used here.

The AI’s role is also worded differently across outlets. Space.com wrote that the drive was fully planned by AI and Astronomy that it was planned and executed entirely by the model. NASA’s January release describes the digital-twin check, and NASA’s rover pages describe the rover’s own navigation software handling hazards as it drives, so this article says planned by AI, checked by engineers.

The driving that was already autonomous

Perseverance does not wait for Earth between waypoints. JPL’s release of 12 October 2026 explains that its AutoNav system images the terrain while rolling, builds a 3D map of hazards and plots a course between the waypoints chosen on Earth. JPL says Perseverance has driven more than 90 percent of its total distance this way, against about 5 percent for the first AutoNav on Spirit and Opportunity. It can cover as much as 412 metres in a single sol.

The same release marked a record. On 11 October 2026, the 2,006th sol, the odometer reached 45.22 kilometres, past the 45.16 kilometres set by Opportunity between 2004 and 2018. Five and a half years was all Perseverance needed for what took Opportunity almost fifteen. Mark Maimone, deputy team chief for robotic operations, said, “Perseverance is a better driver than the one that landed.” The release credits AutoNav for the distance and does not mention AI-planned waypoints.

So the two pieces work in sequence: the AI drafted where to go on the map, and the rover’s own software dealt with what was in front of the wheels. For other machines on other worlds, see this site’s accounts of Chandrayaan-3’s Pragyan rover, which drove 103 metres on the Moon, and of India’s Mars orbiter, Mangalyaan.

What NASA and Anthropic say comes next

Vandi Verma, a JPL space roboticist on the Perseverance team, said in NASA’s release that the team is “moving towards a day where generative AI and other smart tools will help our surface rovers handle kilometer-scale drives.” She also named the goal of scouring large volumes of rover images to flag interesting features for scientists.

JPL engineers estimate the method could halve route-planning time, Anthropic reports. That is the company’s account of an estimate, and The Register reported on 31 January 2026 that it could not get either party to say how much time the saving would be. In its own words the company calls the December drives a test run and points to NASA’s Artemis lunar programme as a place the same kind of tool might be used.

The Rover Operations Center says it is developing digital twins and mission-adapted AI models for future rover, aerial and other surface missions. For published test results of robot foundation models, see this site’s 2026 roundup. NASA’s October 2026 record release says the mission’s first extension began that month and will carry Perseverance’s exploration of the crater rim through roughly the end of 2028.

Share
Copied!

Sources & further reading

  1. NASA JPL: NASA's Perseverance Rover Completes First AI-Planned Drive on Mars (30 January 2026)
  2. NASA JPL: NASA's Perseverance Sets Off-World Record for Total Distance Driven (12 October 2026)
  3. NASA Science: Mars 2020 Perseverance Rover mission page
  4. NASA Science: Perseverance rover components (cameras and autonomous navigation)
  5. NASA JPL: Rover Operations Center
  6. NASA Photojournal: Tracks in, Path out? (Spirit rover, PIA12473, December 2009)
  7. NASA Science: Mars Exploration Rovers, Spirit and Opportunity
  8. Anthropic: Claude AI Powers NASA's First AI-Planned Mars Rover Drive (the company's own account)
  9. The Register: NASA taps Claude to conjure Mars rover's travel plan (31 January 2026)
  10. The Robot Report: NASA's Perseverance Rover completes its first AI-planned drive
  11. Astronomy: AI pilots Perseverance across 1500 feet of martian terrain
  12. Space.com: NASA's Perseverance Mars rover completes its 1st drive planned by AI
  13. Payload: Mars Rover Drives with the Help of Anthropic AI

Researched and written with the help of AI tools and edited for accuracy. Provided for general information and discussion only, not professional advice. See our editorial standards and disclaimer. Spotted an error? Tell us.

#perseverance rover#nasa jpl#mars#generative ai#autonomous navigation#anthropic#jezero crater

Enjoyed this? Get the next one.

One good read at a time, straight to your inbox. No spam, unsubscribe anytime.

More in AI & Technology
Stable Diffusion, Released on 22 August 2022, Makes a 512 by 512 Image From Random Noise in About 50 Denoising Steps, Using an Idea First Published in a 2015 Paper Inspired by Physics
A step-by-step account of the diffusion method behind Stable Diffusion: what each paper added, what the authors measured, and what the model card says the released model still cannot do.
Anthropic Trained a 34-Million-Feature Autoencoder on Claude 3 Sonnet, Google DeepMind Then Deprioritised Fundamental Research on That Method, and Anthropic's Attribution Graphs Gave Insight on About a Quarter of Prompts
Researchers can now swap one concept for another inside a language model and watch the answer change. The same papers say, in their own words, how little of a model that covers and where the tools fail.
Young Workers in AI-Exposed Jobs Are 19% Below Where They Would Be, a Stanford Study of Millions of Payroll Records Finds, and Its Authors Say That Does Not Prove AI Did It
Across millions of ADP payroll records, young workers in the jobs AI handles best are falling behind their peers. Whether AI is the reason is the one thing the authors decline to claim.
← all articles