Federated learning (FL) emerges as a promising solution to enhance autonomous driving (AD) models against out-of-distribution (OOD) data. However, OOD instances often lack labels, rendering conventional FL approaches less effective in AD. This paper proposes road-supervised FL (RSFL), which leverages road sensors' perception results to annotate vehicle sensors' data, providing a fresh perspective on data annotations for FLAD systems. To get deeper insights into RSFL, the information gain of annotating objects with road sensors is derived by leveraging the expected entropy reduction. Furthermore, a bug-aware sensor placement (BASP) algorithm is developed which strategically reduces (increases) the number of sensors in low (high) complexity scenarios. This is in contrast to traditional sensor placements where sensing coverage or road topology is the only consideration. It is shown that BASP approximately maximizes the information gain brought by road supervision. Experiments confirm the superiority of the proposed RSFL framework and BASP algorithm.