NASA’s Perseverance rover has just done something no Mars robot has attempted before: follow a route that was planned not by human drivers on Earth, but by an artificial intelligence model. Instead of engineers painstakingly tracing a safe path through satellite images, a large language model generated a traverse across treacherous terrain, and the rover executed it on the ground. It is a quiet but profound shift in how deep space missions might be run, turning AI from a tool that analyzes data into one that helps decide where a billion‑dollar robot goes next.
The experiment unfolded as Perseverance was already hitting a symbolic “marathon” milestone on Mars, proving it is robust enough to keep pushing into more complex terrain. Now, with AI plotting routes that human teams would have struggled to design in the same timeframe, the mission is testing how far it can stretch autonomy without sacrificing safety. I see it as the first real dress rehearsal for a future in which explorers on distant worlds will be guided by software that learned its skills on Earth.
How an AI model became Perseverance’s trail guide
For most of its time in Jezero Crater, Perseverance has driven the way earlier rovers did, with human planners on Earth drawing up detailed paths and sending them as daily instructions. That changed on mission sols 1,707, 1,709 M, when the team handed part of the job to a generative model. But for Perseverance, those drives were different because the rover’s route was drafted by an AI system that ingested maps and terrain data, then proposed a sequence of waypoints that balanced safety and efficiency. Engineers still checked the plan, but they were no longer sketching every twist and turn themselves.
The model at the heart of the test was Claude, a system developed by Anthropic and adapted for mission use. NASA engineers provided Claude with a digital elevation map of the rover’s surroundings and a set of constraints, then asked it to design a safe traverse across the rocky floor of Jezero. According to an official social update, the perseverance used an AI model, Claude, developed by Anthropic to lay out a 400-meter route across the crater’s uneven terrain. The result was not a freewheeling joyride, but a structured set of waypoints that the rover’s onboard navigation system could follow while still using its own sensors to avoid immediate obstacles.
The first AI-planned drive on another world
What made this drive historic was not just that AI was involved, but that it planned the traverse end to end, something no Mars rover had attempted. NASA describes the event as Perseverance completing its first AI‑planned drive on Mars, with the rover following a chain of AI‑generated waypoints across a field of hazards that would have taken human planners far longer to chart. In official mission notes, But for Perseverance, the team deliberately stepped back from its usual hands‑on approach to see whether a language model could shoulder the cognitive load of route design.
The test was not a blind leap of faith. Engineers still validated the AI’s proposal and relied on the rover’s existing autonomous driving software to handle boulders and sand traps in real time. A detailed mission log explains that AI‑generated waypoints guided For the rover across terrain that had previously been considered too complex to plan quickly from Earth. In effect, the language model handled the strategic “where should we go,” while the rover’s onboard systems managed the tactical “how do we get over this rock,” a division of labor that hints at how future missions could be structured.
Inside the AI toolkit: maps, models and Martian hazards
To give Claude a fighting chance at plotting a safe path, NASA fed it more than just a static picture of Mars. The system drew on digital elevation models built from orbital imagery, allowing it to reason about slopes, ridges and potential sand traps in three dimensions. Mission descriptions note that NASA’s planners used data from digital elevation models to help the AI understand where the rover could safely drive and where it might tip or get stuck. One technical overview explains that, to overcome these limits, NASA combined these elevation maps with the rover’s own navigation constraints so the model would not suggest impossible maneuvers.
From orbit, the High Resolution Imaging Science Experiment camera has been tracking Perseverance’s progress and providing the raw material for these elevation maps. An annotated image from NASA’s HiRISE instrument shows how mission teams can trace the rover’s path and overlay it on a detailed topographic model of the crater. In that Description, the High Resolution Imaging Science Experiment is used to visualize the route that the AI helped design and to highlight how the rover’s cameras and sensors filled in the fine details of the surface ahead. By pairing orbital views with ground truth from Perseverance’s own images, the AI could treat Jezero’s floor almost like a city mapped in a navigation app, except that every wrong turn could be fatal.
Why NASA is betting on AI for future exploration
There is a practical reason NASA is experimenting with AI‑planned routes now, while Perseverance is still relatively close to home in planetary terms. As missions push farther from Earth, communication delays will stretch from minutes to hours, making joystick‑style driving impossible. The space agency has said that this milestone could reshape how future missions explore distant worlds as those delays grow, because AI systems will be able to generate and revise plans locally instead of waiting for instructions. One analysis notes that the space agency says the milestone could reshape how future missions explore distant worlds and even hints at lessons for By Stor style automation back on Earth.
Perseverance is an ideal testbed because it has already proven its staying power. Earlier this year, NASA highlighted that Perseverance Just Major Milestone, Not Done Yet, with the rover surpassing the classic marathon distance on Mars and still in good health. Mission updates emphasize that Perseverance Just Major, and that the rover is expected to drive many more miles as it climbs out of Jezero toward ancient river deposits. Testing AI‑driven planning now, while the rover still has years of life ahead, gives NASA time to refine the technique and apply it to more ambitious traverses, including potential sample delivery routes.
From lab experiment to operational tool
What began as a controlled experiment is already being treated as a template for operational use. Engineers at JPL have described how they used Anthropic’s model to generate candidate routes, then integrated those into the same planning tools that mission teams use every day. A mission briefing notes that Image, Anthropic, NASA, JPL and Mars are now linked not just in press releases but in the rover’s actual command chain, with the AI helping Perseverance avoid obstacles in its path. One report on the test explains that Image data and hazard maps were combined with Claude’s suggestions so the rover could thread through boulder fields that would have been tedious for humans to chart manually.
Outside observers have been quick to point out both the promise and the limits of this approach. A detailed account of the experiment stresses that no, the chatbot did not crash Perseverance, and that human engineers remained firmly in the loop. The same report notes that using Claude to plan a Mars drive was not as simple as inputting a single prompt, but required careful tuning and validation. As one summary put it, Igor Bonifacic, a Senior reporter, emphasized that the experiment shows how generative models can augment, rather than replace, the expertise of rover drivers. Another analysis framed the milestone more bluntly, noting that NASA Let AI Drive, Rover, Mars, It Somehow Survived, and quoting Anthropic’s statement on how the model was constrained to keep the rover safe, a point captured in the phrase Let AI Drive.
More from Morning Overview