CoSafe coordinates safety interventions across self-driving cars

A safety system that watches one autonomous car may miss a problem created by several cars together. At an intersection, two vehicles might each choose a reasonable maneuver that conflicts with the other’s plan. Independently overriding both vehicles to stop can prevent an immediate crash while creating a deadlock.

In the CoSafe paper, submitted to arXiv on November 4, 2026, Dingji Wang and colleagues propose moving part of runtime safety assurance to roadside infrastructure. The versioned arXiv landing page states that the paper was accepted by the 49th IEEE/ACM International Conference on Software Engineering (ICSE 2027). Their framework checks the wider traffic scene and coordinates interventions across vehicles. It leaves each vehicle’s normal driving software in charge unless the safety layer identifies a need to intervene.

A shared view, with selective control

CoSafe relies on vehicle-to-everything communication, or V2X. Participating autonomous vehicles send a roadside unit their position, heading, speed, intended trajectory and bird’s-eye-view scene features. The roadside unit combines these features with its own observations.

The distinction between intentions and predictions matters. For connected autonomous vehicles, CoSafe uses their uploaded plans. For other detected road users, such as human-driven cars, cyclists and pedestrians, it predicts future movement. Its default assessment horizon is three seconds.

The system checks four kinds of hazard: potential collisions, failures to stop for active stop signals, multi-vehicle deadlocks and unexplained persistent stalling. Deadlock detection builds a graph of who appears to be waiting for whom, then looks for circular dependencies. A short retention window keeps recently flagged vehicles under intervention, reducing rapid switching caused by noisy observations.

CoSafe cannot take over human-driven vehicles or pedestrians. It also treats autonomous vehicles not selected for intervention as outside its immediate control. Alternative plans must therefore accommodate those participants’ expected movements.

Choosing maneuvers together

For each vehicle requiring intervention, CoSafe generates possible trajectories from representative driving patterns learned from expert demonstrations. A neural refinement component adjusts those patterns using the shared scene representation, and the system adds a deterministic braking option.

It filters moving candidates against background traffic and the drivable road region, then searches for a compatible combination across the affected vehicles. The branch-and-bound search avoids exploring combinations that cannot improve the current solution. Persistent yield priorities help prevent vehicles from repeatedly reversing who should go first.

A vehicle-side control switch executes the assigned reference trajectory when one arrives; otherwise, it passes through the native driving system’s plan. The prototype uses model predictive control to track intervention trajectories.

The authors describe CoSafe as independent of the underlying driving architecture because it does not modify or retrain native planners. Integration still requires access to the necessary plans, state information and compatible scene features. Their experiments used a shared auxiliary perception encoder rather than demonstrating arbitrary vendor-specific feature compatibility.

What the simulation showed

The evaluation used InterDrive, a benchmark running in the CARLA simulator. It covered 46 routes across three simulated towns, with two to eight autonomous vehicles per route alongside background traffic. Scenarios included intersections, merging and lane changes. The authors tested seven driving systems and averaged results over five repetitions.

They report an average 70.03% relative improvement in driving score over the unprotected systems and an 87.80% reduction in the counted violations. Driving score combines route completion with penalties for infractions; these figures are not measurements of real-world crash reduction.

For TCP, for example, driving score rose from 71.35 to 92.42, while reported collisions fell from 21 to two. CoSafe also outperformed the ego-centric ARGUS safety framework on the three systems where that comparison was applicable.

A component-removal experiment supports the importance of coordination: independently choosing each vehicle’s best candidate produced 82 collisions on TCP, compared with 21 without assurance and two with full CoSafe. In a separate mixed-deployment experiment, vehicles were assigned TCP, LMDrive and UniAD in rotation across the benchmark. CoSafe increased driving score from 55.16 to 83.00.

Hazard assessment achieved reported average precision of 0.836, recall of 0.943 and an F3 score of 0.930. F3 emphasizes recall; it should not be read as overall accuracy. Necessary interventions were defined using a three-second window before violations in unprotected baseline runs.

Promising results, conditional protection

Mean processing time was approximately 80 milliseconds per active intervention cycle on the evaluation hardware. TCP experiments with communication delays or packet loss retained benefits, but performance weakened.

The evidence remains simulation-based. Protection depends on perception, prediction, communication and trajectory tracking. Reverse maneuvers are absent, and braking fallbacks can leave vehicles deadlocked when forward motion cannot resolve a conflict. Sensor hardware faults and actuator failures are outside the evaluated scope. CoSafe demonstrates a promising coordination strategy—not a guarantee of safe operation on public roads.