diplomat33
Average guy who loves autonomous vehicles
I mostly agree with this but to play devils advocate, tesla vision successfully ignored a box while the radar thought the box was a car. It demonstrates the weakness of radar, it has no idea what it is looking at! But also the strength of radar, it can cover for something the vision system, can't classify, just with poor probabilities and insufficient confidence to actually take action other than a warning.
Yes, camera vision and radar have different roles. Radar cannot classify objects. Camera Vision can. But good radar can give you very accurate position and velocity, even in low visibility conditions. So camera vision and radar can be complementary.
It should also be noted that Tesla's radar is poor. I don't think all radar would be that bad at detecting that object. There are higher resolution radars that would do a better job, I think. But to improve radar, Tesla would need to upgrade to better radar which would require either retrofitting all existing cars like they did with the FSD computer (expensive and time consuming and ties up service centers) or adding new radar to new cars and rendering existing cars obsolete. But with camera vision, they don't need to upgrade the cameras and can improve the performance with just better software. So focusing on Tesla Vision makes sense from that angle.
Tesla vision also successfully errors on the side of caution and didn't hit a fake pedestrian, so in an actionable instance no change. I wonder how much general object detection will be in V9? Or if the current vision engine is basically V9, just without the city streets driving policy.
I am guessing V9 will have more object detection since it is designed to be more "FSD". And if V9 is super good at object detection and such, it would make sense to use V9 for basic AP and just remove the city street stuff. After all, it should be the same underlining software.