"Tesla's answer to the height issue is to use its network of cars to build up a database of road signs, bridges and other similar objects using radar to build up a blueprint. The car can then compare that blueprint with the real world as a car travels along the road. If it sees something out of the ordinary, it is far more likely to be a possible obstruction.
"If several cars drive safely past a given radar object, whether Autopilot is turned on or off, then that object is added to the geocoded whitelist," the company notes."
Yeah, good luck with that. With every piece of metal littering the road showing up they'll be going down the road at a snail's pace. Radar data processing is a fiendishly difficult job, and clutter processing especially so. They'll be logging every dropped nail, cats eye, soda can on the road.
Will it see a slab-sided lorry angled across the road as being an obstacle spanning the whole road? Or will it see it as two sets of clutter at either end of the lorry with nothing in between? The F117a is an extremely good example of how flat surfaces don't reflect radio waves back to a monostatic radar if angled just so. The straight metal side of a container lorry (the bit you'll drive into) would have a similar property, but the wheel sets at either end won't.
Crowd Sourced?
As no two Teslas will go down the same road following exactly the same path, they're all going to see the clutter environment differently, and very often that'll be completely differently. That's going to play merry hell with building a meaningful clutter map for every part of every road. The necessary in built pessimism could make the system pretty annoying in real world conditions.
Why not lidar?
Well, its very expensive (in both equipment and processing costs). But at least you get a good 3D map of the environment. In contrast the type of radar Tesla are using is almost certainly not an imaging radar (such as SAR, scanned phased array, etc). At best it'll be getting a set of returns the range to which will be measured quite well, but the angle to which will be pretty vague. In other words, it won't really know whether something is by the side of the road or right in front of the car. It's a poor man's way of attempting 3D scene reconstruction.
Doppler
Doppler processing (if they're doing that) will allow them to tell vehicles from stationary objects. However even that's full of problems. The spokes on a car wheel at the bottom of the wheel are 'stationary' in terms of speed along the road and therefore have the same Doppler shift as something stationary. So they'll look like a set of non-moving objects that magically appear and disappear all the time. That'll make their processing even harder, especially if the spoke return comes and goes (like when the car in front goes round a corner).
All For Nothing?
To me this all sounds a bit, well, desperate. I don't see how this combination of sensors can ever be used to produce a reliable system that can be trusted, with or without a bunch of crowd sourced data and any amount of processing. Sure, it's probably better than nothing, but it's not "perfect".
And thus we return to the fundamental problem; Tesla are selling a car with a fancy cruise control and being honest in saying what it does and how it should be used, but they've given it the name Autopilot. And most people are reading the word "Autopilot" and then not paying any attention to the system's stated limitations.
This new firmware will also have its limitations, and I think Tesla's tightening up how the car monitors the driver's attentiveness is absolutely necessary. But if that means that people may as well not use it, what's the point of having it in the car in the first place? If it goes badly wrong then it's bad publicity for Tesla. It adds nothing to the car's supposed main appeal (battery powered, stonking acceleration, lots of other tech built in).
And with Apple (a company with $200billion in the bank) dropping their self driving car project I think that's a sign that serious companies are beginning to go off the idea of a self driving car. It seems like
* at best it'll be extremely expensive to develop,
* will likely never be allowed to be fully autonomous, not whilst there's bikes, deer, workmen, dogs, children, motorcycles, pedestrians, etc. on the road too
* will be a very difficult sell to all sections of the general public (who, in owning mobile phones that suffer endless software problems, would be wondering why a self driving car from the same company would be any better, and will be disappointed to be told that it won't drive them home from the pub)
* doesn't generate any more useful data on a person that you can't already collect merely by selling them a cheap Android of iOS mobile phone.
In other words, what's the business plan?