Navigating off‑road environments is difficult because there is no structured map and no fixed vocabulary for what is traversable. Traversability depends jointly on terrain characteristics and the robot's dynamics, both of which are impractical to label at scale, so the robot must learn from its own experience.
Current platforms typically fuse multiple sensors (RGB‑D, lidar, radar, IMU) and run heavy neural networks on powerful hardware. In contrast, we introduce FLINT, a lightweight traversability estimator that relies solely on an RGB camera. FLINT’s backbone contains 21.6 M parameters, making it 38× smaller than comparable foundation‑model backbones, and it runs at 14.7 FPS on a CPU‑only system. When replaying 24 field logs, FLINT produced a cheaper and more accurate cost map than the deployed foundation‑model system WildOS on 23 of the logs.
We evaluate several self‑supervised learning signals and deploy the resulting models on a real robot in closed‑loop trials. The best self‑supervised head achieved 99 % autonomy over the full route, outperforming a human‑label‑trained baseline run on the same course. These results demonstrate that heavy sensing and computing are not required for reliable traversability estimation.
Review: FLINT proves that high‑precision traversability can be achieved on resource‑constrained robots, opening a path toward low‑cost, robust off‑road navigation.