Driving records from NASA's Perseverance rover have been turned into a model that estimates whether Martian ground is safe to cross. The research uses 500 Martian days of operations, 45 kilometres of travel and 43,000 stereo images to learn risks that a geometric obstacle map can miss, such as loose terrain likely to make the wheels slip.
Perseverance already uses its Enhanced Autonomous Navigation system, or ENav, to build a terrain map from stereo cameras and plan routes around rocks and steep surfaces. Geometry cannot fully describe how the six-wheeled rover will interact with the ground. A patch that looks flat may still cause slip, and recovering from a mobility failure on Mars is far harder than stopping an Earth robot for assistance.
The new Mars 2020 dataset links what the rover saw with what happened as it drove. It combines grayscale NavCam stereo pairs, vehicle poses, accelerometer readings, suspension angles, tilt and estimated wheel slip. Images were recorded about once per metre, while motion and body measurements were sampled at roughly eight hertz. The routes came from actual operations directed by human rover planners and ENav rather than a simulated Mars environment.
Learning from the rover's own tracks
The model starts with frozen DINOv2 visual features and adds a five-million-parameter transformer trained to order terrain by driving difficulty. Wheel tracks provide examples of ground the rover crossed, while avoided rocks and sand ripples provide contrasting regions. A second objective aligns the visual representation with acceleration, tilt and the rover's rocker-bogie suspension angles.
A small prediction head then estimates several quantiles of wheel slip for each part of an image. The final hazard score uses whichever is larger: the median predicted slip or the spread between the 10th and 90th percentiles. That construction makes uncertain ground costly even when its median slip estimate is modest. The researchers inserted the score into ENav as an additional hazard layer, so it can influence route ranking alongside height, roughness, tilt and travel time.
On 35 held-out Martian days, the full method reached an area under the receiver operating characteristic curve of 0.874 and an F1 score of 0.758. The strongest reported baseline reached 0.816 and 0.602 on the same measures. Removing the alignment between images and vehicle measurements reduced the area under the curve to 0.723, evidence that appearance alone did not capture the rover-terrain interaction as well.
Those results come from a deliberately skewed dataset. Perseverance normally avoids dangerous ground, so severe slip is rare. The evaluation labels unsafe pixels using the upper tail of measured slip together with rocks protruding in depth data. The comparison therefore tests how models rank hazards within the rover's operational history; it does not measure the probability of becoming stuck on every kind of Martian terrain.
Testing the planner on Earth
The team integrated the learned hazard map into ENav and demonstrated it on a physical rover test bed in a Mars-like rock field on Earth. The model was trained only on Mars driving data and used without further tuning for that test. The published work does not say that the learned planner is running aboard Perseverance, and the project page lists the code and dataset as coming soon.
Sources
- Paper record and abstract, submitted September 21, 2026
- Full paper with dataset, model, comparisons and test-bed integration
- Research project page with dataset examples and deployment video
- Lead image: Figure 1 from Chiu et al., showing the Earth rover analog and the traversability pipeline, reused from the paper under CC BY 4.0.