Abstract
RGB cameras often lose image contrast in smoke, haze, and low-light conditions, which makes robot navigation difficult. In this project, we built a prototype NIR-based navigation system using a Sony Spresense microcontroller. The system captures near-infrared images and classifies them into four navigation commands with a compact neural network. We collected 1,997 NIR images under clear, low-light, light-smoke, and dense-smoke conditions. A student network with 5,988 parameters was trained by knowledge distillation from a MobileNetV2 teacher model with 415,332 parameters. The student model was then quantized to INT8 and deployed on the Spresense. In paired image tests, NIR images preserved object structure when RGB images were strongly degraded. On a held-out validation set of 452 images, the on-device classifier achieved 66.0% top-1 accuracy, retaining about 86% of the teacher model’s accuracy while reducing the deployed model size by about 100-fold (from 1.58 MB to 15.7 KB). These results demonstrate that NIR-based edge inference is feasible on a small prototype system, although the current accuracy is still insufficient for fully unsupervised navigation.