Do I subjectively believe in Elon's argument that cameras alone will be sufficient because all humans have is two crude eyes? Yes, MOSTLY, but those cameras are not backed up by the evolved human brain, so the effective visual and execution system starts off way behind in the race. Things like LIDAR, radars, and IR sensors help level the playing field, but they might not be essential. We don't know yet since NOBODY has delivered this stuff.
L5 (and for that matter L4) won't be marketable until they are at least 10x better than humans at the same tasks, so there is also that limitation (which Elon acknowledges). L4 is feasible because it can always just give up and dump it on the human. L5 MUST perform in everything but a hurricane.
You cannot say that LIDAR and IR sensors help level the playing field because you're just adding more sensors to a computer which you feel cannot make rational path planning decisions based on present and perceived future obstacles based on velocity and trajectory.
You can't say L5 and L4 won't be marketable until they are at least 10x better than humans at the same tasks... even 2x better would save lives and that's the initial target. 10x is later on and probably just a matter of software and possibly upgraded processing. L4 does not dump it on a human necessarily, it can do some driving modes and not others. This doesn't mean that it
has to hand off to a human, at L4 it could simply pull over and stop if a human doesn't intervene.
What I see in the nearer term, within the next two years, is a hybrid L2/L4 but not L3. (L3 involves plenty of time for hand offs). At first, I believe we are going to see something more like AP 1 or EAP on AP 2 where the driver still needs to be paying attention but may have to intervene immediately, at any time. The number of driving modes will increase such as the ability to stop at stop lights, drive in neighborhoods, make turns, etc. This is why Elon said you probably won't be able to sleep or read a book while driving at least for the next two years. Once the system gets even better and data shows it's safe, then it will be presented to regulators, and we may see true L4/L5 capability without the need be as vigilant.
L5 cars will not be perfect, it's not true that it MUST perform in everything but a hurricane. There are various other situations where humans are incapable such as bad rain storms where even with full wipers you still can't see, or at super high speed winds which can knock cars and trucks off the road, or very heavy fog with no visibility (at least radar would have an advantage over a human here), really bad snow storm, roads covered in unsafe sheets of ice, if it has a flat tire, if it's been in an accident, sensor failure, etc. All L5 cars should do in such situations is put hazards on, pull over to a safe location, and possibly call for help. There's no good reason an L4/L5 car or a human for that matter should continue driving or else they risk their safety or the safety of others.
You don't need an evolved human brain to drive, you simply need a good set of rules and experience. One rule which has served me pretty well is "don't hit stuff"
. When it's all said and done, I think this system should be capable of that for the most part... I mean
stuff happens.
Next time you travel somewhere, try to follow all traffic laws to the letter, no speeding, no following too close, etc and you can even use my rule of "don't hit stuff" and really see how difficult it is to get from point A to point B. You might be bored to tears driving like a grandma, but you'll realize that the vast majority of driving is super simple if you know what objects are, where they are located, and how fast they are travelling, plus knowing traffic signs/signals helps. Other than that, it's just a matter of handling corner cases which, unfortunately, happens to be the topic of the majority of conversations surrounding autonomous driving.
There are people who say that if a car can't handle situation X then it's not L5/L4 etc. First, it's more about the driving mode then the situation, and as stated there are many situations where humans fail.
Let's look at two tables:
In the last ImageNet competitions have had computers beat humans at image recognition, computers can make decisions faster than a human and based on a more complete picture of their environment. How many times does a human get into an accident because they "didn't see them there"? The evolved human brain isn't helping very much here.
Full self driving, even without being L5 or even L4 can help save lives and prevent many typical human accidents. Will it be perfect? It doesn't even have to be close to perfect, as we all agree, just better than a human.
