CooperDrive: Enhancing Driving Decisions Through Cooperative Perception
The new framework uses lightweight object sharing to cut reaction time and requires only 90 kbps bandwidth.
A team of researchers including Deyuan Qu, Qi Chen, Takayuki Shimizu, and Onur Altintas has introduced CooperDrive, a novel framework designed to solve a critical flaw in current autonomous vehicles: their blindness in occlusion and non-line-of-sight (NLOS) scenarios. Published in a paper accepted at ICRA 2026, CooperDrive enhances situational awareness by enabling vehicles to share lightweight, object-level perception data. Its key innovation is reusing existing Bird's-Eye View (BEV) features from a vehicle's own detectors to accurately estimate the poses of other cooperative vehicles, eliminating the need for additional, computationally heavy sensor fusion encoders.
This efficient design translates directly to performance and safety gains. On the planning side, the expanded field of view allows the system to anticipate potential conflicts much earlier, transforming reactive stops into predictive, smoother adjustments to speed and trajectory. The team validated CooperDrive through real-world, closed-loop testing at challenging NLOS intersections. The results are impressive: the system operates with an average end-to-end latency of just 89 milliseconds and requires a minimal bandwidth of 90 kbps, all while measurably increasing key safety metrics like reaction lead time and minimum time-to-collision.
- Uses lightweight object sharing, reusing BEV features to estimate poses without extra heavy encoders, keeping native vehicle stacks intact.
- Demonstrated in real-world tests, it cuts end-to-end latency to 89 ms and uses only 90 kbps of bandwidth.
- Increases safety at NLOS intersections by expanding the perceived object set, allowing planners to anticipate conflicts and adjust proactively.
Why It Matters
It provides a practical, bandwidth-efficient path to solving the dangerous 'occlusion problem' that limits current self-driving cars.