(video) Applying neural reconstruction in simulation for automated driving

We introduce an automated pipeline designed for training neural reconstruction models by leveraging sensor streams gathered from a data collection vehicle. Subsequently, our simulator, aiSim, is employed to generate a controllable virtual counterpart of the real-world
environment, enabling the replay of scenes in a closed-loop fashion. This video presents a scenario in San Francisco where the static environment is rendered by a Gaussian Splatting model, while dynamic content is populated by aiSim. Additionally, the weather can be altered using aiSim’s built-in features.