Design and Validation of a Digital Twin Test Platform Leveraging Open-Source Autonomous Driving Systems

Abstract

This paper presents an open-source vehicle-in-the-loop simulation (VILS) digital-twin platform integrating CARLA, Autoware Universe, ROS 2 Domain Bridge, and Lanelet2-based high-definition maps. The platform couples a physical Autoware vehicle and a CARLA-based virtual vehicle in both directions: physical-vehicle state and localization information are transferred to the virtual environment, while virtual-object information is supplied to the physical planning stack. The proving-ground environment is reconstructed from LiDAR point-cloud and Lanelet2 map data, and the virtual vehicle is configured with sensor modalities, nominal mounting poses, transform frames, and ROS interfaces compatible with Autoware. Field validation was performed on three geometrically distinct proving-ground layouts under global-path-following and virtual-static-obstacle-avoidance conditions. The Virtual-to-Physical global-path RMSE ranged from 0.188 to 0.322 m, and the mean absolute difference in obstacle-relative avoidance-start distance was 0.08 m across four obstacles. During an approximately 10-min coordinated four-PC live field acquisition, all 4,833 PC4 source virtual-object messages were matched one-to-one in the vehicle ROS domain with identical serialized payloads and preserved order; no application-message omission, duplication, reordering, or payload mismatch was observed. A separate approximately 10-min stationary Chrony acquisition yielded a maximum client 95th-percentile absolute residual offset of $5.649~\mu $ s and a maximum observed absolute residual of $23.044~\mu $ s. These results demonstrate the feasibility of an open-source, bidirectionally coupled VILS platform for quantitative vehicle-level digital-twin testing.