So the claim is there is absolutely no scenario where the other systems can contribute which seems to be false. E.g. Tesla's cameras are blinded pointing straight on into the sun, the car even tells you the cameras are blinded. If the cameras are splashed with mud they'll also see nothing. Tesla's radar was able to see "through" some obstacles which the vision system can not (and you can see that in the traffic visualizations).
Now how far do you want to fuse, do you just want to overlap the unique sensing abilities of each of those systems? How do you handle conflict when you're in the regions where all systems sense "equally" well. Sure, there's questions. But clearly having more data can't make things worse.
Elon's argument that that's the way to go because that's how humans drive is just ridiculous. And I say this as a proud Model 3 owner (great car, will never ever be autonomous, I don't care). It doesn't pass the sniff test.
No. The claim was that pouring resources into those other systems is better spent improving the vision system, within the context that Tesla and Comma.ai is operating.
> But clearly having more data can't make things worse.
There are several examples of cars hitting stuff because one part of the sensor suite thought there was a problem, or that there wasn't a problem.
In general, the more complex you make a system, the more complex the failure modes get.
Doing it with vision only saves on cost, latency & energy consumption.
It's not just about paying more dollars, but also range.
You're assuming that a dual system will be safer, but what if such systems are more prone to perception confusion, or other anomalies?
Even if a dual system were safer, it doesn't make sense to say that you will pay for safety absolutely. For example, you can always add an additional smoke alarm to your house for some marginal improvement in safety, but at some point most people decide they are safe enough.
If you have radar and lidar and vision, then you have at least three different specialist machine learning models running, and then another model running that takes their outputs and decides what the car is going to do. You may have even more than that, some doing specific tasks like localization.
Neural nets and vision only is a more difficult but in the long run straightforward solution. The example he brought up was alphazero vs some other chess engine that has a rook engine, and a knight engine, etc.
Basically he's backing the end to end neural network back approach over some kind of multi-sensor fusion.
https://youtu.be/_L3gNaAVjQ4?t=3801