CooperScene: Multi-Modal Cooperative Autonomy Benchmark with C-V2X Communication Characterization

Autonomous vehicles (AVs) often struggle with incomplete situational awareness due to occlusions and limited sensor range. Cooperative perception, where vehicles and infrastructure share sensor data via Cellular Vehicle-to-Everything (C-V2X) communication, is a promising solution. However, existing benchmarks often rely on unrealistic assumptions about communication bandwidth and agent scalability. The authors introduce CooperScene, a new dataset designed to bridge the gap between research-grade cooperative perception and real-world deployment. The dataset features three connected autonomous vehicles (CAVs) and one infrastructure roadside unit (RSU), all equipped with multi-modal sensors and commercial C-V2X radios. The authors provide 344K 3D labels across 59K frames, with rigorous sensor synchronization, centimeter-level localization, and real-world C-V2X network traces.

By benchmarking state-of-the-art cooperative perception models, the authors demonstrate that while these models perform well under unlimited bandwidth, their accuracy drops by up to 49% when constrained by real-world C-V2X throughput. The study reveals that for many models, the data sharing volume is orders of magnitude higher than what current vehicular networks can support. This mismatch leads to high latency and stale data, which can degrade rather than improve perception performance. The authors also show that even when bandwidth is sufficient, packet loss and latency in real-world channels remain significant hurdles, suggesting that bandwidth optimization alone is insufficient for robust deployment.

CooperScene provides a realistic testbed that forces researchers to confront the trade-offs between perception accuracy and communication efficiency. By integrating real-world network dynamics into a multi-agent, multi-modal dataset, it highlights the urgent need for network-aware cooperative perception algorithms that can operate within the strict constraints of existing C-V2X infrastructure. This work serves as a critical step toward moving cooperative autonomy from controlled lab environments to reliable, field-deployable systems.