COACH: A Consistency-Enhanced and Adaptivity-Aware Framework for Open Heterogeneous Collaborative Perception

Open heterogeneous collaborative perception enables multiple agents equipped with diverse sensors and backbones to cooperatively understand their surroundings. However, heterogeneous perception still faces several fundamental challenges, including inconsistent cross-modal feature spaces, limited adaptability to newly emerging modalities, and high retraining costs for scalable deployment. To address these issues, COACH, a consistency-enhanced and adaptivity-aware framework for open heterogeneous collaborative perception is put forward. In COACH, a foreground-weighted normalized feature alignment (FW-NFA) objective is proposed to enforce fine-grained directional and magnitude consistency in semantically critical regions, achieving robust cross-modal alignment without introducing extra learnable parameters. On top of the consistent feature space, COACH integrates a modality-aware low-rank adaptation (MA-LoRA) module to enable lightweight fine-tuning of collaborative fusion and detection head. This component adaptively modulates the collaboration output based on the active modality configuration, thus improving generalization and scalability across heterogeneous agents. Extensive experiments on the OPV2V-H and DAIR-V2V datasets demonstrate that COACH significantly outperforms existing frameworks under both heterogeneous and incremental perception settings, achieving superior consistency, efficiency, and adaptability. Our code is available at: <uri>https://github.com/supercyt/COACH</uri>