Heterogeneous Collaborative Perception

It studies how agents with different cameras, LiDARs, viewpoints, feature encoders, or tasks can combine observations for detection, tracking, navigation, and inspection. Common approaches align representations with adapters, prompts, codebooks, or online distillation, then fuse BEV features; alternatives exchange sparse evidence or object boxes to reduce bandwidth, including in vehicle-UAV and ground-air teams. Evaluation spans simulated and real-world benchmarks such as OPV2V-H, DAIR-V2X, V2XSet, and V2X-Real, alongside air-ground datasets and embodied tasks. The central challenge is balancing compatibility and accuracy against communication, latency, privacy, localization errors, and adversarial or uncertain inputs; reliable cross-view geometry and operation with unseen agents remain open problems.