As autonomous driving technology and Intelligent Transportation Systems (ITS) rapidly evolve, traditional traffic management and driving models are undergoing significant transformation. Digital Twin (DT) is a digital model of a physical entity that reflects its physical behavior by applying platforms and two ways interactions of data in real-time [1], which is essential for simulation, monitoring, and decisionmaking in both autonomous driving and ITS. It provides a virtual replica of physical systems, enabling testing in controlled environments, reducing risks, and enhancing safety [2]. By predicting outcomes, optimizing performance, and identifying potential failures before they occur, Digital Twin supports predictive maintenance and improves operational efficiency. Go a step further, it also fosters collaboration between autonomous vehicles and ITS, ensuring dynamic traffic management and smoother integration.
DT has been increasingly investigated to integrate vehicles, traffic, and drivers, while imitating real-world changes in real time [3]- [5]. Human cyber-physical system (H-CPS) [6] is proposed and defined as a composite intelligent system comprising humans, cyber systems, and physical systems from driving perspective. Digital twin-based adaptive traffic signal control (DT-based ATSC [5]) models transportation system into roadway network with traffic signals to enhance traffic management. Mobility Digital Twin (MDT [4]) utilizes an artificial intelligence (AI)-based data-driven cloud-edge-device framework for mobility services in the physical space (namely, Human, Vehicle, and Traffic). However, existing works are mainly focused on individual layers, and thus fail to represent real-world information and correlation in different layers of transportation systems.
In this work, we propose a new multi-level digital transportation twin (mDTN), which aims to mimic real-world transportation system from coarse to fine granularity. The main idea is to incorporate multiple levels (each layer corresponds to different granularities), while enabling seamless communication between these layers to share cross-layer information. Combining with real-time information from the real world, mDTN could imitate and track the changes happends in the transportation systems, e.g., crash, construction zone, and traffic jam. In addition, we build a proof-of-concept co-simulation platform to validate mDTN with state-of-theart simulators, including TruckSim, CARLA, and SUMO, in these layers.
II. PROPOSED WORK Fig 1 shows the overview of mDTN architecture, including three layers: dynamic, object, and traffic. We design these layers to represent the real-world transportation systems, from coarse to fine granularities. On the one hand, we leverage state-of-the-art simulators to replicate the real-world scenarios (e.g., vehicles, routes, and roads environments). On the other hand, we incorporate real-time inputs (e.g., traffic signals, video streams, and vehicle controls) from the real world to enhance these simulators, toward more accurate representation of the transportation systems. To share cross-layer information, we exploit the existing messages APIs between these simulators, including the Ackermann steering geometry of vehicles (i.e., SUMO-CARLA) and dynamic controls (i.e., TruckSim-CARLA).
The dynamics layer focuses on vehicle real-time dynamic information based on vehicle control signal. It should be noted that only CAVs are focused for its connection to ITS and Digital Twin Models. Dynamics layer receives the control signal as steering wheel/brake/throttle, and generates realtime vehicle dynamics as acceleration/posture/traction and so on. Then, the object layer receives the real-time vehicle dynamics and applies them to CAVs in DT. Among the SOTA simulators (CarSim/TruckSim Chrono), TruckSim is the core of dynamics layer in mDTN because its realistic vehicle behavior based on empirical data and validated models. In detail, TruckSim, a high-fidelity vehicle dynamics simulator for heavy trucks and buses, used in ADAS development, realtime testing, and automotive engineering is used for vehicle dynamic calculation.
The object layer aims to organize and manage all objects, connect and map real-world objects back to mDTN, simulate scenarios and provide virtual sensors. Object layer receives real-world real-time traffic signals and synchronize them to the virtual traffic signals in the middle DT level. In the same time, object layer receives vehicle dynamics from dynamics layer for CAVs and Ackermann steering geometry from traffic layer for other vehicles. In this way, all vehicles could have more reasonable behaviors by other layers and traffic signals in local layer. There are many game-engine based simulators as CARLA, LGSVL, AirSIm which all provide realistic visualization. CARLA is utilized in mDTN for its convenient Python API, various sensor types and the ecosystem with many other platforms. It is an open-source autonomous driving simulator built on Unreal Engine, offering realistic urban environments, sensor integration, and a Python API for AI research and testing.
The traffic layer provides realistic traffic flow behavior model to object layer. In mDTN, traffic layer takes in the traffic information (vehicles' geometry) comes from realworld traffic sensors (mainly roadside cameras) to mimic real-world traffic flow vehicle by vehicle. Then, the traffic layer could generate vehicles and control their Ackermann geometries and related parameters as speed/acceleration for the object layer. There are many traffic simulators (SUMO, Vissim, MATSim, AIMSUN and so on), and SUMO is used as the core in traffic layer for its large-scale traffic flow simulation ability, easy connection to CARLA by Python APIs. SUMO is an open-source traffic simulation tool for modeling vehicle and pedestrian movement, used in traffic management, smart cities, and autonomous driving research is responsible for traffic flow digital twin in this work. After the real-world traffic flow imitation, SUMO deducts traffic flow in next period based on some predefined driving models which enable short-term predictive ability.
We build a proof-of-concept co-simulation platform based on a workstation with AMD 5955WX CPU, 32GB RAM, and NVIDIA RTX3080 GPU, under Ubuntu 22.04. Specifically, we use CARLA 0.9.15 (built from source), SUMO 1.20.0, and TruckSim 2021.
In particular, we create a simple freeway scenario (with a construction zone) in CARLA, by using RoadRunner. The merging zone is about 5000 feets long and the speed limit changes from 75 mph to 55 mph, and all vehicles follow zipper merge principle. Moreover, we initialize 31 vehicles (including 3 trucks and only 1 truck enables TruckSim and I shows the resource usage during 30 seconds execution.
We introduced a new digital transportation twin (mDTN) to integrate multiple layers of transportation systems, including vehicle dynamics, traffic flow, and road environment. We built a proof-of-concept co-simulation platform by connecting state-of-the-art simulators (i.e., TruckSim, CARLA, and SUMO) and generate multi-layer datasets for advancing autonomous driving research. In future work, we will focus on expanding scalability and refining integration of mDTN.