On January 30, 2026, NASA’s Jet Propulsion Laboratory announced that the Perseverance Mars rover had completed the first drives on another world planned entirely by artificial intelligence. The two drives took place on December 8 and December 10, 2025 (sols 1,707 and 1,709 of the mission), covering a combined 456 meters — and the routes were not drawn by human mission planners. They were generated by a vision-language model after reading orbital imagery. JPL led the work in collaboration with Anthropic, using the company’s Claude models.
For developers, the story is not the symbolism of “AI on Mars.” It is a fully qualified AI planning pipeline validated in an environment with punishing communication delays and near-zero tolerance for error: perception data in, structured action plans out, high-fidelity simulation as the pre-flight gate, then autonomous execution.
From Orbital Imagery to Waypoints
The pipeline runs in layers. A vision-language model first analyzed orbital imagery from the HiRISE camera on NASA’s Mars Reconnaissance Orbiter, layered with slope data derived from digital elevation models. It identified terrain features — bedrock, outcrops, boulder fields, sand ripples — and then generated a continuous path from start to goal, annotated with waypoints. The model read the same imagery human planners use; the difference is who proposes first. Engineers used to pick the route from the pictures; now the model drafts the proposal and humans hold veto at the key decisions. ScienceDaily’s coverage adds that these models were tested against JPL’s surface mission dataset — decades of images and measurements gathered during surface operations on Mars and other worlds.
Two Drives and a Digital Twin
Before each drive, commands were verified through JPL’s “digital twin” of the rover — a simulation that checks more than 500,000 telemetry variables — and only then uplinked through the Deep Space Network. On December 8 the rover drove 689 feet (210 meters); on December 10 it covered 807 feet (246 meters). Perseverance’s autonomous navigation software handled execution, with the ground team’s intervention kept to a minimum. NASA Administrator Jared Isaacman called the work a seminal contribution to roverspace, saying the added autonomy is delivering fresh insights from Mars on near-real-time timescales.
Why Hand the Route to AI
The answer is written in physics. Earth and Mars sit roughly 140 million miles (225 million kilometers) apart on average, and communication delays make real-time remote control impossible. The traditional workaround keeps waypoints no more than about 330 feet (100 meters) apart and dispatches drives in short segments, which makes route planning a heavy, repetitive load. JPL space roboticist Vandi Verma put it plainly: as the mission pushes into ever more challenging terrain on Mars, the value of AI-planned routes will only grow. Offloading the repetitive route engineering to a model is what frees human engineers for the science decisions.
The Anthropic Collaboration
The demonstration also marks an unusual pairing: a national space program adopting a commercial frontier model inside an operational milestone. The work was coordinated from JPL’s Rover Operations Center and carried out in collaboration with Anthropic, using Claude models — the same model family that writes code in a developer’s API calls also planned routes on Mars, 140 million miles away. For Anthropic, it is a public validation of Claude under extreme reliability requirements. For NASA, it establishes a workable path for wiring commercial models into mission-critical workflows. The collaboration is also a procurement signal: rather than training a bespoke navigation model from scratch, JPL adapted an existing commercial model and wrapped it in the verification machinery NASA already trusts.
Next: Longer Routes and Edge Deployment
JPL’s roadmap for the capability includes kilometer-scale drives, having the AI flag surface features worth stopping for, and eventually running these models on the edge — aboard future rovers, helicopters, and drones that plan locally instead of waiting on ground uplink. For product and infrastructure teams on Earth, this is a reference architecture qualified in the harshest environment available: exhaust the risks in high-fidelity simulation first, then let the autonomous system run. AI planning has moved from demo to production, and Mars is currently the hardest production environment there is.
Sources
- NASA’s Perseverance Rover Completes First AI-Planned Drive on Mars — NASA JPL
- NASA’s Perseverance rover completes the first AI-planned drive on Mars — ScienceDaily
- NASA’s Perseverance Rover Completes First AI-Planned Drive on Mars — NASA
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.