BeamNG GmbH cooperated with HDI Deutschland AG on training a deep learning system on simulated data for vehicle damage evaluation based on the simulation software. simulated various types of car crashes ,which were then generated into corresponding ground truth images including detailed metadata about the damage of the individual vehicle parts and semantic annotations. A total of 4.000 scenarios were defined through domain randomization.

Figure 1: Example output of the Impact Data Generator. (left picture) shows a vehicle damaged through a broadside collision, while (right picture) shows the corresponding semantic segmentation.

BeamNG GmbH stands out from among the many other providers of driving simulation software because of its detailed information on vehicle damage. Using a soft-body physics framework allows to easily compute in-depth damage statistics for individual vehicle parts. The subdivision of vehicles into individual parts is also a feature that was advantageous to the project. For example, the model chosen for this pilot project consists of 123 individual parts.

For the data generation BeamNG GmbH developed an Impact Data Generator that is available as an open source project. It is a Python package with a custom extension for

Learn more about the project in the published “Applied Sim-To-Real Transfer for Damage Estimation” paper.