Autonomous vehicles are rapidly becoming reality, but their scalability remains challenging due to the complexity of the real world and regional differences in traffic behavior. To overcome these challenges, realistic simulation with precise sensor models and reactive traffic agents is required. In this presentation, we focus on the derivation of realistic agent models using our collected traffic data and generative AI. Our diverse database allows the agents to be trained for different traffic behaviors around the world. This enables autonomous vehicles to be tested in a realistic simulation before deploying them in a specific region, which significantly reduces development costs.

• Our monocular 3D vision pipeline unlocks unprecedented opportunities for capturing large amounts of traffic data
• Data from stationary cameras is particularly suited for training reactive traffic agents
• Realistic agent models are crucial for closed-loop simulation to ensure realism in virtual testing and reduce real-world driving