Tesla Vision is a fascinating project aiming to solve a really difficult AI problem (more precisely, a set of difficult AI problems). There is clearly much more work to be done, but the system’s capabilities today are very impressive.
What is most interesting to me is that the Tesla AI team is so comfortable with their vision-based occupancy neural network, that ultrasonic sensors and radar have been deemed lower fidelity and thus unnecessary.
It’s amazing how they are able to take video frames from a set of cameras and derive a 3D vector space of objects in the scene. The implications of this amazing feat are only now coming to the fore.
A little overly dramatic but a good overview of the dangers of relying on only one type of sensor: Tesla Autopilot Crashes into Motorcycle Riders - Why? - YouTube