Facilitating heterogeneous sensor information cooperation in multi-agent perception system

Collaboration addresses the fundamental limitations of single-agent perception systems by enabling information exchange. However, most existing approaches rely on simplified assumptions, requiring agents to use either identical multi-modal sensors or distinct yet single modalities. Such assumptions limit real-world applicability, where agents often operate with flexible, heterogeneous sensor suites. Furthermore, in dynamically changing networks where agents may freely join or exit, existing methods frequently exhibit significant performance degradation or even integration failure. To overcome these challenges, we propose DisCo, a novel two-stage distributed paradigm for robust perception under arbitrary intra-agent modality variations and dynamic inter-agent network changes. In DisCo, each agent first learns independently to fuse its own sensor set, and then learns to align these heterogeneous features into a unified collaboration space. We instantiate DisCo with two core modules: a Multi-modal Feature Integration (MFI) module for within-agent fusion, and a Multi-agent Forward Projection (MFP) module, which implements the feature alignment via a lightweight projection pool that dynamically adapts to network changes. Extensive experiments demonstrate that DisCo achieves leading performance across diverse collaborative scenarios, and successfully maintains robustness to agent joining and exiting, as well as modality changes. Our code is available at https://github.com/byzhaoAI/DisCo.