Researchers at McGill University and Université de Sherbrooke built a vision system that estimates which off-road ground a robot can cross from a single RGB camera. In same-day tests on one mixed forest-and-gravel route using the same planner, its strongest self-supervised version recorded three operator interventions; a model trained on human-labeled terrain recorded seven.
The system, called FLINT (Fast Lightweight Inference for Traversability), tackles a difficult judgment: whether a patch of ground is safe depends both on what the camera sees and on what the particular robot can handle. FLINT uses the RGB camera to assess terrain and wheel encoders for odometry; it does not use lidar or depth sensing.
To avoid manually labeling every image, the researchers trained FLINT on TartanDrive 2.0 off-road driving logs, treating ground the vehicle crossed as evidence of traversable terrain. A pretrained model called DINOv3 turns each frame into visual features, and a small classifier predicts whether each patch is traversable. The team tested three self-supervised labeling methods: the recorded trajectory corridor alone, regions grouped by SAM3 object masks, and visual clusters identified by STEGO.
At runtime, FLINT projects those patch predictions onto a ground-level map that the Nav2 route planner can use. The full vision pipeline ran at 14.7 frames per second on CPU alone during the field tests. The researchers trained the classifier on a desktop GPU, so the CPU result describes deployment inference, not training.
The team compared four versions of the terrain estimator on the same route and day, using the same A* planner; only the estimator changed. The SAM3- and STEGO-based self-supervised versions each recorded three operator interventions; the corridor-only version recorded eight, while the human-labeled baseline recorded seven. The paper defines an intervention as an operator disengaging autonomous mode with the planner’s deadman switch, regardless of the reason. Across all four runs, the robot covered 4.47 kilometers and logged 119.1 minutes of CPU-only inference without dropped frames. The paper also reports autonomy percentages of 99.0% for SAM3 and 96.3% for the human-labeled baseline, but does not state the denominator for that measure.
The result is limited to one route and one day, and the authors say wider terrain testing and online adaptation remain future work. SAFiR Lab lists the paper as under review for the 2027 IEEE International Conference on Robotics and Automation.
Sources
- William Bonilla et al., “FLINT: Fast Lightweight Inference for Traversability,” arXiv:2609.26857
- SAFiR Lab publication record
- Lead image: Figure 3 from the FLINT paper, by William Bonilla et al., CC BY 4.0. Original image.