Driving world models simulate how a traffic scene unfolds, but they struggle to predict how nearby vehicles will move when the ego car takes a different action. This is because vehicle motion is hard to learn from a camera view alone, objects overlap, and the same pixel movement can mean different speeds.
BirdDreams creates two videos at once: the usual driver's perspective and a simplified top-down schematic view showing roads and vehicles as colored shapes. The top-down view makes vehicle positions and movements clearer, and the two videos influence each other during generation to combine motion understanding with visual detail.
The model can follow a given ego trajectory more accurately than previous methods, and when no trajectory is given, it generates more realistic ego motion. Importantly, surrounding vehicles react plausibly to new ego maneuvers, such as slowing down when the ego car cuts into their lane, without requiring a pre-recorded map or vehicle state.