Robust Real-Time Coordination of CAVs: A Distributed Optimization Framework under Uncertainty

Haojie Bai, Yan Wang, Cong Guo, Xinhai Zhao, Haiyan ZhuarXiv 2025 · DOI · arXiv

This paper, recently published by researchers from the Harbin Institute of Technology and the Chinese Academy of Military Sciences, tackles one of the "holy grails" of autonomous driving: how to get a swarm of vehicles through a busy intersection safely and quickly when sensors are noisy and the world is unpredictable.

TL;DR

The researchers developed a distributed framework that treats vehicle trajectories as "probability distributions" rather than fixed lines, allowing them to actively "steer" uncertainty to prevent collisions. By combining a parallel negotiation algorithm (ADMM-DTN) with a Transformer-inspired attention mechanism, they achieved a 40% reduction in collision rates while remaining fast enough for real-time hardware.

What Problem Does This Paper Address?

Coordinating Connected and Autonomous Vehicles (CAVs) at unsignalized intersections is a nightmare of "tension." On one hand, you want safety. But in the real world, sensors (like GPS and LiDAR) have noise, and controllers have lag. These small uncertainties accumulate, turning a planned "safe" path into a fuzzy zone of potential collision.

On the other hand, you want real-time performance. Calculating the optimal path for 20 vehicles simultaneously is computationally "heavy." Traditional methods either become too cautious (causing traffic jams) or too computationally expensive to run on a vehicle’s onboard computer. This paper aims to bridge that gap: providing probabilistic safety guarantees without sacrificing the speed needed to make split-second decisions.

Methods

The authors introduce three main innovations:

  1. Covariance Steering & Adaptive Constraints: Instead of planning a single line for a car to follow, the framework plans the mean and the variance (the "spread") of the trajectory. They use a technique called "covariance steering" to actively shrink the vehicle's uncertainty "bubble" during critical interactions. They also use $l_1$-regularization to dynamically adjust safety buffers based on local traffic density.
  2. ADMM-DTN (Distributed Trajectory Negotiation): Using the Alternating Direction Method of Multipliers (ADMM), the team broke the massive "multi-car" optimization problem into smaller pieces that each car can solve in parallel. Crucially, they used "Jacobi updates," which ensure that even if the cars have to stop "negotiating" early due to time limits, the resulting path is still mathematically guaranteed to be feasible and safe.
  3. Interactive Attention Mechanism: Not every car at an intersection matters to you. Borrowing from the "Self-Attention" architecture used in Large Language Models, the framework allows each vehicle to selectively focus on the most "critical" neighbors. This prunes unnecessary data and speeds up the math.

Datasets

The framework was evaluated using:

  • High-Fidelity Simulation: 100 Monte Carlo simulations of a challenging "unprotected left turn" scenario. They used disturbance and noise matrices derived from real-world driving data benchmarks.
  • Real-World Experiments: The team deployed the algorithm on four custom-built mobile robots equipped with Intel NUC computers, ROS (Robot Operating System), and Intel RealSense tracking cameras to validate the "real-time" claims in a physical environment.

Key Insights and Ideas

  • Don't just react to uncertainty—steer it: Most planners treat uncertainty as a fixed "buffer" (like a permanent 2-meter bubble around the car). This paper shows that you can treat uncertainty as a controllable variable. By "steering" the covariance, the car can be "braver" when it has clear data and "tighter" when it's near another vehicle.
  • Negotiation is better than Dictatorship: Centralized planners (one big computer for the whole intersection) fail as you add more cars. This distributed "negotiation" approach scales much better and preserves the privacy of each vehicle's internal model.
  • Attention is Efficiency: By applying an attention mechanism to physical space, the researchers reduced the computational overhead by 14.1%. This proves that "social" awareness in AI isn't just for safety—it's a computational shortcut.

Key Results

  • Safety: In dense traffic scenarios, the proposed method reduced collision rates by 40.79% compared to state-of-the-art "Convex Feasible Set" methods and 29.86% compared to space-time resource searching.
  • Efficiency: Vehicles maintained a higher average speed (9.37 m/s) compared to benchmarks, indicating less "hesitation" at the intersection.
  • Speed: The parallel ADMM-DTN algorithm was an order of magnitude faster than centralized versions. For 16 vehicles, the centralized approach took ~3.7 seconds to solve, while the proposed parallel method took only ~80 milliseconds.
  • Robustness: In real-world tests, the robots successfully dodged an "unexpected" non-connected obstacle (a moving box representing a human-driven car), demonstrating that the probabilistic model handles "surprises" well.

Limitations and Future Work

While the results are impressive, a few areas remain for future exploration:

  • Communication Reliability: The paper assumes a stable V2X (Vehicle-to-Everything) connection. In the real world, "packet loss" or latency in the Wi-Fi/5G link could break the negotiation.
  • Mixed Traffic: While they tested against one "unexpected" obstacle, the framework is primarily designed for a world where most cars are CAVs. Integrating with a high volume of unpredictable human drivers remains a hurdle.
  • Complex Grids: The study focused on a single intersection. Scaling this to a "city-wide" grid where multiple intersections influence each other is the next logical step.

Check out the experiment demo here: https://youtu.be/4PZwBnCsb6Q

Summary updated

Abstract

Achieving both safety guarantees and real-time performance in cooperative vehicle coordination remains a fundamental challenge, particularly in dynamic and uncertain environments. This paper presents a novel coordination framework that resolves this challenge through three key innovations: 1) direct control of vehicles' trajectory distributions during coordination, formulated as a robust cooperative planning problem with adaptive enhanced safety constraints, ensuring a specified level of safety regarding the uncertainty of the interactive trajectory, 2) a fully parallel ADMM-based distributed trajectory negotiation (ADMM-DTN) algorithm that efficiently solves the optimization problem while allowing configurable negotiation rounds to balance solution quality and computational resources, and 3) an interactive attention mechanism that selectively focuses on critical interactive participants to further enhance computational efficiency. Both simulation results and practical experiments demonstrate that our framework achieves significant advantages in safety (reducing collision rates by up to 40.79\% in various scenarios) and real-time performance compared to state-of-the-art methods, while maintaining strong scalability with increasing vehicle numbers. The proposed interactive attention mechanism further reduces the computational demand by 14.1\%. The framework's effectiveness is further validated through real-world experiments with unexpected dynamic obstacles, demonstrating robust coordination in complex environments. The experiment demo could be found at https://youtu.be/4PZwBnCsb6Q.
BibTeX
@misc{bai2025robust,
  author = {Haojie Bai and Yan Wang and Cong Guo and Xinhai Zhao and Haiyan Zhu},
  title = {{Robust Real-Time Coordination of CAVs: A Distributed Optimization Framework under Uncertainty}},
  year = {2025},
  eprint = {2508.21322},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.21322},
  howpublished = {arXiv},
}
Figure