The Perseverance rover has made its first Mars drives along a route set by an AI model rather than a human planner. JPL says that on 8 and 10 December 2025 the vehicle covered 210 and 246 metres respectively on a plan produced by Claude models. IEEE Spectrum described on 28 September 2026 how the shift is reshaping autonomy in space missions.
Key takeaways
- Two drives: 210 metres on 8 December 2025 and 246 metres on 10 December 2025
- The route came from Claude models working with the Rover Operations Center at JPL
- The model read orbital imagery and terrain slope data and flagged hazards on its own
- Commands passed through the rover's digital twin, where more than 500,000 telemetry variables were verified
- A compressed IBM model reached the ISS in May 2026, and a maintenance-procedure language model was tested there in July 2026
A route drawn by a model
Waypoints for Mars rovers have always been set by people. Here the model received high-resolution imagery and slope data, then returned a continuous path with waypoints. No human planner supplied a first draft.
| Date | Distance |
|---|---|
| 8 December 2025 | 210 m (689 ft) |
| 10 December 2025 | 246 m (807 ft) |
| Total | 456 m |
The digital twin as a safety catch
Trust in the plan was not unconditional. The engineering team ran the commands through the rover's digital twin?Digital twin: A faithful simulation of a real device, reproducing its mechanics and software. It lets engineers test the effect of a command before it reaches the actual hardware., a virtual copy of the vehicle, checking more than 500,000 telemetry variables?Telemetry: The stream of measurements a vehicle reports: joint positions, motor currents, temperatures, voltages. They reveal whether the machine will behave as planned. before the order went to Mars. Autonomy covered route planning, not the surrender of control.
It's up to the spacecraft to make a decision, and we'll be watching what happened an hour ago.
Robert Ambrose, former chief of NASA's Software, Robotics, and Simulation Division.
Signal delay as the core argument
That one hour is the heart of the matter. On deep space missions real-time control is physically impossible, so either the vehicle decides for itself or no decision happens at all. In Earth orbit the argument is different and concerns bandwidth. That is why the ISS is testing a compressed IBM model alongside a language model supporting maintenance procedures.
Why it matters
Space autonomy was held back for decades not by missing algorithms but by the inability to verify a machine's decisions. The digital twin moves that boundary, because a plan can be checked before execution rather than after. If the pattern holds, agencies will start separating planning autonomy from execution autonomy. The practical risk is that a model behaves unpredictably outside the data distribution it was checked against.
What's next?
- The drives are a capability demonstration, and NASA has not said whether model-led planning will enter routine Perseverance operations
- Icarus Robotics is testing its free-flying Joy system in zero gravity, pointing to where this work heads beyond rovers
