A self-driving car needs labeled 3D data to learn to see: bounding boxes around every vehicle, pedestrian, cyclist in lidar scans. Labeling is expensive — each frame takes minutes of human annotation. And the labels age: new car models, new road configurations, new objects that didn't exist in the training set. The label dependency is a bottleneck on deployment.
A roadside unit at an intersection sees the same scene every day (arXiv:2603.16742). It doesn't move. Its viewpoint is fixed. Over weeks, it accumulates unlabeled observations of thousands of vehicles passing through. Without any human labels, it can learn a local 3D detector — what objects typically appear, where they go, how big they are — through self-supervised patterns in the repeated observations.
The infrastructure becomes the teacher. When a self-driving car passes through the intersection, the roadside unit broadcasts its detections. The car doesn't need to have seen this intersection before. It receives the local model's predictions and uses them as soft labels for its own learning. The car teaches itself from the infrastructure's experience.
The structural insight: the labeling problem is a sampling problem. A car sees each scene once, briefly, from a moving viewpoint. Labeling that single pass requires human effort because there's no redundancy. A fixed roadside unit sees the same scene repeatedly from the same viewpoint, creating the redundancy that self-supervision needs. The information isn't scarce — it's in the wrong place. Moving it from infrastructure to vehicle is cheaper than generating it from human annotation.