HydraCollab adaptive framework cuts bandwidth 74% while boosting accuracy

Collaborative perception lets multi-robot systems share sensor data for better situational awareness, but it traditionally requires a trade-off: send more data for higher accuracy at the cost of bandwidth. Real-world networks impose strict bandwidth constraints, making this trade-off untenable for large-scale deployments. HydraCollab, a new framework from UC Irvine researchers, addresses this by adaptively selecting only the most informative sensor features and dynamically choosing between intermediate-level feature fusion and late-stage decision fusion based on spatial confidence maps. This allows the system to allocate communication resources precisely where they improve perception the most, dramatically reducing overhead without sacrificing — and even improving — accuracy.

Evaluated on three datasets (V2X-R for vehicle-to-everything, V2X-Radar, and UAV3D-mini for drone swarms), HydraCollab consistently outperforms state-of-the-art methods. Against Where2comm, it uses only 41% of the bandwidth on V2X-R and just 26% on V2X-Radar, while boosting perception accuracy by 0.78% and 0.75%, respectively. The approach is especially valuable for bandwidth-limited environments such as dense urban traffic or remote drone operations. Accepted for presentation at IROS 2026, the paper includes open-source code, enabling researchers and engineers to adapt the framework for their own multi-agent systems.