Work

Testing through conflict: Cloud-coordinated closed-field testing for multi-agent continuous interaction scenarios

Year: 2026

Type: article

Abstract: Testing intelligent connected vehicles (ICVs) in closed-field environments requires recreating dynamic and high-risk multi-agent interactions that closely reflect real-world driving conditions. Existing trigger-based or scripted testing methods lack behavioral coupling with the vehicle under test (VUT), often resulting in pseudo-interactions and limited scenario diversity. To address these limitations, this study proposes a cloud-coordinated closed-field testing framework that enables continuous, responsive, and adversarial multi-agent interactions through real-time vehicle-cloud cooperation. The framework integrates a 5G-based cloud control platform, multi-agent cooperative planning, and an adversarial interaction strategy to achieve both realistic behavioral evolution and controllable risk exposure. By leveraging real-time coordination, test agents dynamically transition between adversarial and supportive roles, enabling continuous perturbation of the VUT while avoiding deadlocks or behavioral redundancy. A mixed-integer nonlinear programming formulation jointly considers individual agent dynamics, safety constraints, adversarial intent modulation, and cooperative task allocation, allowing scalable and controllable multi-agent interactions in dense environments. Real-vehicle experiments conducted at the Tongji ICV Evaluation Base demonstrate significant improvements over trigger-based baselines: minimum anticipated post-encroachment time is reduced by 22.5%, average adversarial duration increases by 66.7%, and trajectory entropy improves by 56.7%, indicating stronger temporal conflict, prolonged adversarial engagement, and richer behavioral diversity. These results verify that the proposed cloud-coordinated framework enables realistic, challenging, and diverse multi-agent interaction scenarios, providing a more comprehensive closed-field evaluation environment for autonomous driving systems under high-risk conditions.

Source: Chinese Journal of Mechanical Engineering

DOI: 10.1016/j.cjme.2025.100192

Authors: Xinrui Zhang, Lu Xiong, Peizhi Zhang, Haojie Feng, Junpeng Huang, Mengjie Tian

Institution: Tongji University

Language: English

Topic: Autonomous Vehicle Technology and Safety

Subfield: Automotive Engineering

Field: Engineering

Domain: Physical Sciences

SDG: Industry, innovation and infrastructure

Open Access status: diamond

Funders: National Natural Science Foundation of China, Key Technologies Research and Development Program, National Key Research and Development Program of China