Generative Digital Twin Network: AI-Built Relationships between Digital Twins for Automated Driving

Kurzfassung

Proper understanding of traffic scenario is essential for the safe execution of automated driving functions. Nowadays, digital twin technologies are used in simulation to manage these complex scenarios. While models for the environment or individual road users are well established, models of dynamic interactions among real-word traffic participants are still challenging. To address this, we present the Generative Digital Twin Network (GDTN), a framework that uses AI to build and connect digital twins based on spatial and behavioral relationships of traffic participants. In contrast to traditional statistical methods which are, e.g. limited in adaptability, AI learns complex traffic patterns directly from the received data and can adapt to unforeseen real-world traffic dynamics. The framework aims to provide a deeper understanding of evolving traffic behavior and enable early detection of risks in automated driving. The GDTN framework follows five main steps. First, a simulated traffic scenario is setup with road users and infrastructure elements. In the second step, digital twins are instantiated using ground-truth position data provided by the simulation. Third, spatial relationships between the ego vehicle and other digital twins are calculated based on their relative positions and defined using the ASAM OpenXOntology (e.g., frontLeftOf, hasDistance). In the fourth step the AI model is trained to predict the spatial relationships along with their short-term progress based on the data calculated in step 3. In the final step, the trained model is applied during operation to infer these relationships without access to ground-truth data. This enables the construction of a real-time graph structure, where nodes represent digital twins and edges represent their spatial relationships. The dynamic graph is continuously updated over time, resulting in an integrated model of traffic participants and their relationships to the ego vehicle. Overall, these aspects allow the framework to track evolving traffic situations and anticipate potential risks. Future work will also focus on incorporating real-world sensor data into the framework to support evaluation in more realistic settings. Thus, with GDTN traffic situations can be tracked and analysed which enables safer, data-driven decision-making for automated driving systems.