Good argument for doubling down on "vision only" approach. Clean dataset can be generated semi-autonomously, An object's velocity and trajectory can be inferred more reliably than noisy radar. And once video frame predictor sees more edge cases from the field. It gets really good at accounting for events like occlusion from puddle spray, dust clouds, etc.
The other neat thing: with the A100 nodes, an exascale GPU cluster can be ready for computing very quickly ;)