CoSafe: Cooperative Runtime Safety Assurance for Mixed-Traffic Autonomous Driving

Dingji Wang, You Lu, Haowen Jiang, Yanggang Chen, Yuyang Bi, Bihuan Chen, Shenglin Bao, Xin PengarXiv 2026 · arXiv

TL;DR

COSAFE is a modular, ADS-agnostic framework that shifts autonomous vehicle (AV) safety from individual "ego-centric" takeovers to coordinated "scene-level" interventions. By leveraging V2X communication and roadside units (RSUs), it identifies and resolves complex multi-vehicle hazards—like deadlocks at intersections—that single-vehicle safety systems are blind to, reducing safety violations by nearly 88%.

What Problem Does This Paper Address?

Current autonomous driving systems (ADSes) primarily rely on ego-centric safety assurance. Each vehicle monitors its own surroundings and, if a hazard is detected, triggers an emergency maneuver like braking.

However, in mixed-traffic environments (AVs, human drivers, cyclists), safety violations often arise from locally reasonable but jointly incompatible decisions. For example, at an unsignalized intersection, two AVs might both decide to "stop and yield" to be safe. Without a global view to assign priority, they enter a circular wait-for chain, resulting in a deadlock (see Figure 1b). Conversely, they might both decide to proceed, leading to a collision.

Existing solutions either require intrusive modifications to the ADS stack or treat vehicles as independent targets, failing to account for the "cyber-physical system" nature of modern traffic.

Figure 1: Motivation of Cooperative Runtime Safety Assurance. (a) shows a conflict, (b) shows a deadlock from independent takeover, and (c) shows the desired scene-level intervention.

Methods

COSAFE introduces a modular architecture split between Roadside Units (RSUs) and the vehicles themselves. The framework is ADS-agnostic, meaning it works as a "safety layer" on top of existing driving software without needing to modify their internal logic.

  1. Cooperative Perception (RSU-side): The RSU aggregates Bird's-Eye-View (BEV) features from its own sensors and all participating AVs via V2X. It then predicts the future motions of all traffic participants (AVs and non-AVs alike).
  2. Hazard Assessment (RSU-side): COSAFE evaluates four specific "hazard oracles":
  3. Collision Hazard: Standard overlap detection.
  4. Stop-Signal Hazard: Failure to obey traffic lights/signs.
  5. Deadlock Hazard: A novel oracle that detects circular wait-for relations between stalled vehicles.
  6. Stalling Hazard: Identifying vehicles stopped without a legitimate cause.
  7. Joint Hazard Mitigation (RSU-side): If hazards are detected, the RSU generates a pool of candidate trajectories for the "unsafe" AVs using a library of trajectory anchors (Figure 3). It then uses a Branch-and-Bound (BnB) search to select a coordinated set of trajectories that resolve all conflicts while staying as close as possible to the vehicles' original intentions.
  8. Control Switch (AV-side): Each vehicle runs a simple switch. If it receives a "safe reference trajectory" from the RSU, it overrides its native ADS and tracks the new path using a Model Predictive Controller (MPC). Otherwise, it continues with its native ADS plan.
Fig. 2: Approach Overview of COSAFE

Datasets

The authors evaluated COSAFE using the InterDrive benchmark, a closed-loop simulation platform built on CARLA and V2Xverse.

  • Scenarios: 10 typical multi-vehicle interaction types (intersections, lane merges, highway ramps) totaling 46 routes.
  • Traffic: Routes include 2 to 8 AVs interacting with background human-driven vehicles, pedestrians, and cyclists.
  • Target ADSes: COSAFE was integrated and tested with seven state-of-the-art ADSes, including modular systems (Autoware), end-to-end models (TCP, UniAD, VAD), and LLM-driven drivers (LMDrive).

Key Insights and Ideas

  • Interaction-Awareness is Key: The "aha moment" is that safety isn't just about not hitting things; it's about resolving intent conflicts. The inclusion of a specific deadlock oracle allows COSAFE to fix liveness issues that ego-centric systems actually make worse by being overly cautious.
  • Prescriptive Cooperation: The framework utilizes "prescriptive cooperation" (SAE J3216 Class 4), where an infrastructure unit (RSU) can direct specific actions to AVs. This is more powerful than simple intent-sharing because it acts as a centralized "referee."
  • Efficiency via Pruning: To handle real-time constraints, the BnB search for joint trajectories uses aggressive pruning (e.g., removing trajectories that hit static obstacles) so it only has to coordinate the remaining mutual conflicts between AVs.

Key Results

  • Safety Improvement: COSAFE reduced safety violations by 87.80% on average across all tested ADSes.
  • Driving Quality: The average Driving Score (DS) improved by 70.03%. Notably, for systems like VAD and UniAD, which struggled in complex lane-change scenarios, the improvements were even more dramatic.
  • Detection Accuracy: The hazard assessment identified unsafe AVs with an F3 score of 0.930, indicating very high recall (crucial for safety) with few false alarms.
  • Real-Time Performance: The total latency for an active intervention cycle was approximately 80ms, which fits within the typical planning cycles of modern ADSes.

Limitations and Future Work

  • Perception Dependency: While COSAFE uses cooperative perception to reduce occlusions, it still relies on the quality of BEV features. Residual failures in the study were mostly traced back to perception noise.
  • Action Space: The current mitigation module only considers forward-moving trajectories and braking. It does not yet support reverse motion, which might be necessary to resolve the most extreme "stuck" scenarios.
  • V2X Reliability: Although the authors tested robustness against 500ms delays and 30% packet loss, extreme communication failure remains a threat to any cooperative system.
  • Hazard Diversity: The current implementation focuses on four major hazards. Future work could expand this to include map-rule violations (like illegal U-turns) or hardware-level faults.

Summary updated

Abstract

With the rapid development of autonomous driving systems (ADSes), real-world roads are evolving into mixed-traffic environments where autonomous vehicles (AVs) and other traffic participants (e.g., human-driven vehicles (HDVs), cyclists, and pedestrians) coexist and interact with each other. However, current ADSes still incur various safety violations in mixed traffic scenarios, raising a strong demand for runtime safety assurance frameworks that can continuously monitor hazards and prevent potential violations. Existing ego-centric assurance approaches are insufficient for mixed traffic, where violations often arise from locally reasonable yet jointly incompatible decisions among multiple vehicles and cannot be reliably resolved through independent takeover. To bridge this gap, we propose CoSafe, a modular and ADS-agnostic cooperative runtime safety assurance framework for mixed-traffic autonomous driving to extend runtime safety assurance from ego-centric takeover to scene-level cooperative intervention. CoSafe leverages vehicle-to-everything (V2X) communication to enable roadside units (RSUs) to monitor scene-level hazards and jointly mitigate hazards involving multiple vehicles. We integrate CoSafe with seven state-of-the-art ADSes, and our evaluation on the InterDrive benchmark demonstrates that CoSafe improves the driving score of AVs by 70.03% on average, reduces safety violations by 87.80%, and accurately identifies unsafe AVs with an F3 score of 0.930, while satisfying real-time constraints.
BibTeX
@misc{wang2026cosafe,
  author = {Dingji Wang and You Lu and Haowen Jiang and Yanggang Chen and Yuyang Bi and Bihuan Chen and Shenglin Bao and Xin Peng},
  title = {{CoSafe: Cooperative Runtime Safety Assurance for Mixed-Traffic Autonomous Driving}},
  year = {2026},
  eprint = {2611.06204},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.06204},
}
Figure