Achieving new levels of realism in simulation
Simulation plays a pivotal role in AV development and can be categorized into two main types: open-loop vs. closed-loop. Open-loop simulation uses log replay, where agents follow predetermined trajectories without interacting dynamically with their environment. In contrast, closed-loop simulation allows all agents to interact and react dynamically to each other in real time. Realism in closed-loop simulation is crucial for AV development and can be ensured through precise sensor and reactive agent models. The focus here is on the latter.
DeepScenario has developed a generative AI that can be trained for naturalistic driving behavior in any region – for instance, in the customer’s deployment area. At each simulation step, the model processes the history of all agents and the driving context to predict the agents’ next state in real time. These newly generated agent states are then fed back into the model, ensuring reactive behavior in closed loop. Trained on an extensive database of naturalistic traffic behavior, DeepScenario’s agents provide an unmatched level of realism, enabling highly accurate and reliable AV testing in simulation.
Accessing real-world traffic data at scale with stationary cameras
To train traffic agents end-to-end, massive amounts of data are required. This need can be met with DeepScenario's flexible virtualization pipeline, which enables the rapid, large-scale collection of traffic data. At the core of this solution is the company’s proprietary software that processes video streams from any monocular camera, including dashcams, traffic cameras, or drones, regardless of the perspective view on the scene.
Stationary cameras, such as traffic cameras or hovering drones, are particularly well suited for the acquisition of training data at critical locations like intersections, highway entrances, or roundabouts. A stationary approach allows for continuous monitoring and provides access to all occurring edge cases, which are highly relevant for training and testing. Drones in particular have the benefit of delivering ground-truth data with a 360° occlusion-free surround view. Capturing massive amounts of traffic data makes it possible to derive a comprehensive statistical understanding of real-world driving and provides an exceptional database for the training of traffic agents.
With DeepScenario’s advanced virtualization pipeline and generative AI capabilities, the company can now provide customers with agent models trained for the different traffic behavior in the world. As a result, autonomous vehicles can be tested in a realistic simulation before they are deployed in a specific region, which significantly reduces the development costs and accelerates their deployment.
