In the development of automated driving functions, typical methods for generating radar models quickly reach their limits. Both, the complex modeling of relevant physical effects and the required computing time often prevent the practical use of such models.

Our method learns from real sensor data using a variational autoencoder. This enables the fast prediction and simulation of a real radar scan of the environment, captured by cameras and laser scans.

However, instead of just a regular radar scan, our method provides a distribution of realistic radar scans that can represent possible physical clutter effects for any situation.