The model generates camera and Lidar data. As if it was a Waymo car that drove through the simulated scenario with its cameras running. This synthetic training data can then be used to train the driving models.
It probably doesn't matter though, "this general blob over there"