Lets be clear here we are talking some mythical L5 driving nirvana where a driverless, steeringwheel less car is expected to navigate -everywhere-. I do think that musk was over egging it, even by his standards, when talking about wheel less cars. Although I could see it for an urban taxi solution, it is a long way from being most cars (I think, could be wrong). FSD and the next 12 months is a different discussion really.
I have tried hard to take onboard the view that vehicle to vehicle and vehicle to infrastructure communication aren't fundamental requirements for FSD, but I still can't see it.
It is correct that human drivers do (at least need to) use far more than pure vision when driving, and combined inference from these sources is what I believe is essential for safe, effective driving.
Vision + processing (vision+ inference) is what you mean here isn't it? I don't generally use other senses to drive?
There are daily complex scenarios where several people have to reverse to remove a deadlock, it's rarely as simple as who is nearest a passing point, you usually have no way of knowing how close the other car is to a passing point, and what if both cars are close to one and both reverse?
I have to wonder at this point where you are driving? I'd also say that both cars reversing isn't actually a fail? For extreme edge cases this is a pretty OK solution, again, egos aside. I spent the last 2 weeks driving around single track roads and a passing place was always in sight. I can imagine situations in carparks or around roadwork where something might happen, but reversing would be as much a solution there as else where. If traffic was totally jammed up behind you, then the car has the same problem as you would have? No amount of hand waving or getting out of the car improves on everyone reversing? It may not be optimal, but if you are chasing the 999's of avoiding disengagements, it would work?
To look at it from the opposite direction, imagine cars without indicators or brake lights, which are both active communications output, not just passive vision.
and are included in v9 FSD. Brake lights just now, indicators promised (2 weeks, right?) (
Tesla Vision Sees Brake Lights: Turn Signals, Hazards, More Soon)
Signals were not a priority as you can't necessarily trust them - indicators get left on, brake lights can be broken. Hence the nural nets are trained (just now) on what a car does, not what it indicates. But there are obviously improvements to be made by looking at what its indicating, but the set of communications we have just now are adequate for this. I'm not saying that inter car comms wouldn't be awesome (I always fantasise about a brake light that could indicate how hard someone is braking so you can feather your reaction better), but that kind of centralised dictat is not the roads systems we have. It could be, but what can be done with it that can't be done using the visual indicators that exist and work, and what is the implementation plan that makes governments sign up for it? You can't ever ignore the other cars that don't have the system fitted, so you still need all the visual processing too?
True, and it will (potentially, maybe) react quicker.
I do worry a bit about the quality of the vision inputs, though. The recordings don’t seem to be particularly high resolution or high quality and the car seems unable to “drive three cars ahead”.
This I think is a fair accusation to level at the current system. It might be enough, but it definitely isn't optimal. Try going back and playing various version of mario cart and see how you do - in particular I noticed that I am waaaay worse on a Wii 4 player using a quarter of a 480p screen. Extreme example, but games etc are pushing past 960p for a reason. The combo of a wide and narrow camera facing forward helps, but I'd ideally like something like a variable res sensor that get slowly more dense towards the center (or is steerable?) Sounds a little like an eye I guess! Side cameras at 960p is probably fine, but I'd like more pixels forward facing to, as you say, help with understanding things further ahead.
The driving 3 cars ahead is probably a software problem however, separate from the cameras. Or as I call it, driving like a teenager still learning. It needs more planning capacity rather than more input. Big strides are being made in this area of AI too, but not necessarily by Tesla.
Interesting that someone brought up the 10x safer than humans - as that is the stated near term goal really? Car to car coms and other L5 related things are less immediate and it may turn out that L5 does need the re-introduction of radar, other sensors or other options. We have to wait and see really. Does removing the human increase safety 10x? I think it could. Going through idiotsincars on reddit, its the software largely to blame (ie the people) for all the messups on there. Ignoring rules of road, or lights, or just putting getting to their destination in a rush over driving safely is probably most of the problems on there. Accidents on unsighted corners/junctions can be solved by approaching slowly enough to do something about it if there is a surprise - the key is recognising the requirement (I can't see round this corner, ie vision problem) and prioritising that over going fast for fun or being on time.
So I am
confident pure vision could get to 10x safer than human. I am
hopeful both that the hardware in the model 3 is enough to support this, and that it can get most of the way to L4 (driverless in most scenarios).
Bugger, I sound like a right old git for some of that, sorry.
