Our thoughts on NASA’s Perseverance Rover Completes First AI-Planned Drive on Mars

On its space exploration mission, NASA launched a rover named Perseverance to search for ancient life on Mars in 2020. Since its landing in 2021, it has been collecting rocks and Mars dust which it sends back to Earth. This ongoing progress enabled the next step in the mission. In December, NASA engineers used Claude, an advanced multimodal AI model developed by Anthropic, to plan a route for the rover, marking the first AI-planned journey on Mars. How was NASA able to leverage GenAI for this task?

How did the team do this?

To complete the drive, the team used visual imagery obtained from the rover and from satellites. Usually, they will use this to map out the surroundings and find the best path to take to the next point where the rover should perform excavations. Now, they used Claude Code to help out with the analysis and let it plot a trajectory through the Martian landscape. Claude used its vision capabilities to analyse the images received and create a virtual map. From this, it was able to create small segments of a trail, which it then strung together into a full path.

With guidance from the team, who reviewed the generated route and made the necessary adjustments, the rover was then able to complete the journey after receiving its instructions. The 400-meter drive planned by Claude may not sound like much, but if you take into account the rover has driven around 40 kilometres in total the last five years, it becomes much more impressive.  

Driving on Mars is also highly risky, in 2009, NASA lost its rover Spirit when it drove into a patch of loose sand. Claude was able to navigate Mars safely, and it cut the team’s planning time in half! This prompts a larger question: how can a model like Claude handle work that appears so scientifically complex?

Why did this task lend itself well to the use of GenAI

There are three aspects to route-planning for a Mars rover that made Claude successful at it.

1. In- and output of GenAI models

GenAI models operate on machine-readable digital data, so any task they perform requires the inputs -and their outputs - to be in a digital format. The input, visual imagery, plays to a multi-modal model’s strengths as it can use its vision capabilities to map out the rover’s surroundings. The output, XML‑based instructions for the Perseverance rover, is also a natural fit for an LLM to generate. Producing highly structured text is an area where LLMs are very effective. This is because they are trained to predict sequences in pattern-dense data, and structured formats like code have strict, repetitive rules that strongly constrain what the valid next tokens can be. The combination of the in- and output being well-suited for Claude, is a major reason this story was a success, but not the only one. ‍

2. One clear objective

The mission’s objective was clear: create a path for the rover to reach a specific location. LLMs work best when given clear goals, a point supported by research showing that goal‑oriented prompts improve performance. Studies also find that breaking tasks into smaller steps with defined sub‑goals increases an agent’s success. Claude demonstrated this by dividing the route into smaller segments, effectively breaking the problem into sub‑goals.

3. Human expert guidance

The final factor to the success of this story is the expertise put into this project by the engineers involved. The years of experience provided the context that enabled Claude to do its work. While the model could analyse terrain imagery and propose a path, it was the team’s domain knowledge that ensured the plan met NASA’s strict safety thresholds. Without this human review, the rover could have been exposed to the type of dangers that trapped Spirit. If anything, this was a demonstration of a crucial principle for real‑world AI deployment: advanced models are most effective not when they replace human judgement, but when they extend the capability of an expert team.

Reflections

The first AI‑planned drive on Mars marks more than a technical milestone, it shows how human expertise and advanced AI can meaningfully accelerate scientific progress. By reducing route‑planning time, researchers can react faster to new discoveries despite the unavoidable communication delays between Earth and Mars. Just as importantly, this mission demonstrates the types of problems where GenAI delivers the most value: those with digital or easily digitised inputs and outputs, clearly defined objectives, handled by strong expert teams.

For organisations here on Earth, the lesson is not simply to adopt AI, but to understand the conditions under which it performs best. When the data foundation is strong, the goals are well‑defined, and a robust review process is in place, AI can unlock significant gains in speed, precision, and innovation without introducing unnecessary risk. The Perseverance team’s success is a reminder to look at your own operations and ask: Where could AI help you navigate faster, safer, and more intelligently?

‍