Development of a Physical AI and Infrastructure-Centric AV Control Framework for Mixed-Autonomy Traffic Environments
혼재교통 환경을 위한 Physical AI·인프라 중심 자율주행 제어 프레임워크 개발
Roads will not switch to full autonomy overnight. For a long transitional period, automated vehicles of differing capability will share space with human drivers, cyclists and pedestrians — a heterogeneous mixed-autonomy environment in which no single vehicle has a complete picture. This project asks the infrastructure to supply what the vehicle cannot see, and builds a Physical AI control framework around that idea.
The work centres on a Phygital World: a high-fidelity digital twin in which physical and virtual agents interact in real time, so that high-risk scenarios can be generated, control strategies can be trained, and infrastructure-guided coordination can be validated before it reaches the street.
Research thrusts
- Phygital World — an infrastructure-centred digital twin coupling real vehicles with virtual pedestrians and vehicles for realistic, repeatable validation
- Infrastructure-guided control — roadside intelligence that proposes optimal speeds and manoeuvres, reducing unnecessary stops and conflict risk at intersections
- Heterogeneous mixed traffic — integrated control across automated and human-driven vehicles, including safe handover strategies
- Vulnerable road users — context-aware prediction of pedestrian, cyclist and e-scooter behaviour from multimodal street sensing
- Distributed edge networks — modular edge–cloud architecture for multi-infrastructure network optimisation
Foundations
Five recent studies from the lab supply the source technology the framework is built on.
- A scalable, interoperable C-V2X framework — a modular edge-intelligent mobility operating system (mOS) integrated with a mixed-reality testbed, demonstrating infrastructure-guided AV coordination under realistic latency (Communications in Transportation Research, 2025)
- Information delivery in tunnels — a VR-HMD multi-agent driving simulator showing that advance information cuts driver reaction time by 51%, and that the effect of a head-up display depends on lighting intensity (Tunnelling and Underground Space Technology, 2025)
- E-scooter overtaking perception — replaying a participant's own driving trajectory from the rider's seat, yielding safety thresholds for lateral distance in mixed traffic (Accident Analysis & Prevention, 2024)
- Audio-visual perception and cycling — multimodal street sensing and explainable machine learning for nonlinear behaviour prediction (Transportation Research Part D, 2025)
- School-zone pedestrian crash severity — semantic segmentation of street-view imagery to quantify how occlusion and micro-level street form drive injury risk (Journal of Transport Geography, 2024)