I’m not sure why you believe Google’s is a ‘server run system’ any more than anything else that is out there. As far as I can tell, the only thing ‘server run’ is maps, which would be downloaded.
I’m not sure you know what LIDAR is. LIDAR is what Google says they are using. It is not a server system - its a rangefinding and mapping system. LIDAR sends out tens of thousands of laser beams every minute, and receives them as they bounce off of a target. Its a laser range finder on steroids. This creates an enormous number of points. Each point will have a PNEZ (point number, northing, easting, elevation). Now this is where the software comes in. It has algorithms that detect ‘that’s a car’ from a cloud of points, for instance. We use a version of this algorithm every day, and a LIDAR scan of a busy city street will somewhat magically erase the cars from a point cloud, for example.
But this is where is gets more advanced. Rather than deal with the ten thousand points that make up a ‘car’, the software creates a ‘block’ that represents a car. And the programmers have to spend endless hours programming in what a car ‘block’ should usually do. I hope you can see how very different this is than an accident avoidance system that senses ‘yes, there is a large object in front of me’. Google software has to know speed, direction, 85th percentile driver behavior, etc., so it can know how to act next to this car - not just in an accident avoidance situation, but all the time. As in be smart enough to not ride in somebody’s blind spot, as in ‘read’ what a driver that is merging is doing, to let him into traffic more smoothly, as in calculate whether or not it has time to pass a slow car, before exiting the freeway...and adjust that if the car ‘block’ moves or changes speed.
And Google now has the nearly insurmountable task of anticipating every possible situation a driver may encounter. BTW, LIDAR can usually read paint on the pavement - I honestly don’t even understand how, but it must deal with the paint absorbing more of the laser beams than the surrounding pavement. Reflective signs are the same way. So I imagine it will rely to some extent on OCR to read signs. Every LIDAR scanner I have ever used is also equipped with a high resolution camera.
Nothing I have read has indicated this is anything but ‘on board’. That makes sense, since LIDAR files are so large.
Anyway, this is how Google is doing it. There is no ‘smart road’ or ‘system’, any further than having a good ‘system’ of maps of the area. The really important inputs are the real time world around the car. IOW, you are considering the task of ‘navigation’, vs the much harder task of ‘driving’.
This is quantum leaps above what you are envisioning - and it has to be. Is it possible? To an extent - similar to most other software, there would be constant ‘patches’ to download, as real world use produced situations that Google had not anticipated. LIDAR has revolutionized the survey industry - cutting field time to a fraction of what it once was. But we still have to do ‘data extraction’...as in go through the piles of data points, and assign a ‘D’ to the PNEZ...making it PNEZD. Its very time consuming...now we do have computers that help, with algorithms that erase cars, and interpret the ground line vs vegetation - very powerful software. What Google is attempting to do is similar, but no longer in a static world, and infinitely more complicated, and nearly instantaneously. Possible - yes. But a huge mountain.
Is it worth it? I doubt it. I don’t think the cost will be justified, especially in light of maintenance tasks (just putting fuel in a truck for instance) that can’t be done automatically. And I really do believe this stuff will be too delicate for commercial use....GPS won’t work well in a storm, sensors blocked by ice, RADAR ineffective in dusty conditions, etc.
Its an absolute non-starter for personally owned vehicles. The cost is too high, and the computers probably take up the back seat.
And we haven’t even begun to discuss how people will try to hack these vehicles. Another fear is that more and more components will become ‘drive by wire’, in order to be easily automated - brakes, steering, throttle. This stuff does break, and it can go terribly wrong....and it can be hacked.
Anyway, this is the world Google is operating in. A car driving down the road is a dynamic computer problem, with thousands of computations being done every second, with advanced algorithms ‘thinking’ like a driver would.
The accident avoidance system - its the beam that keeps the elevator door from slamming on my hand.
What LIDAR is is more power than what they need for the problem. That’s Google Segwaying the problem. You simply don’t need that level of detail to drive a car. People don’t have anywhere near that level of spacial analysis and we’ve been mostly successfully driving for a long time.
It’s 2 quantum leaps more intense than they need the system to be. People don’t work that hard at driving. I read an article in a motorcycling magazine in the 90s about how many “driving actions” the average driver takes. I forget the numbers but they were stunningly low with drivers of automatics doing the least, then sticks, then motorcyclists. The simple reality is that we DON’T make thousands of calculations a second while driving, heck we don’t even makes dozens a minute, and a computer won’t have to.
Google is over engineering the problem, and you’re criticisms show exactly why. You’re right in why that system will not work. Which is also why that is not the way self driving cars ARE happening right now.