I get what you’re saying, and I don’t have the data handy to provide the necessary context, but it’s important to keep in mind that these “what am I looking at?” requests from the cars will represent something like ~0.0001% of all scanned objects on the road. Almost all of the time, the car is confidently identifying everything in its path without any human assistance. Almost every request that the dispatchers would receive to identify was already correctly identified by the car by the time the dispatcher presses their buttons. Most of these requests would be something like “Is this a fire hydrant?”, and that request was automatically sent on the first “frame” where the car couldn’t identify the object; more often than not, the car has figured it out by the very next “frame” and has made whatever necessary adjustments are needed to correct its course before dispatchers even respond.
Of course, I have no way to validate any of this, as any numbers I can provide are just made up based on my own experience doing this job around 7 years ago, and I don’t believe Waymo publishes any actual statistics in this regard, so you’ll just have to take the word of a random guy on the internet for however much its worth. :)
For what it’s worth, I will say that self-driving cars are about the only technology that genuinely excite me these days and some of the things they’re capable of are not talked about nearly enough (you hardly ever hear about the fact that Waymo has autonomous semi-trucks on public roads). With the exception of Tesla, who have been seen as mostly a joke by the rest of the industry, pretty much everybody in this particular tech space are doing some very impressive things.
I don’t suck big tech dick that often, but the self-driving car industry has some good techdick right now.
Why do you need anything beyond “riding a virtual track”, though? The point of a car is to go from A to B, and a track is literally the most effective way to do that (which is why we need more trains in the first place, but that’s an entirely separate rant I could go on about).
IMO, at a certain point you reach a level of autonomy that is sufficient enough. It doesn’t need to know every nook and cranny of obscure roads because realistically, nobody’s going there to begin with. And even then in those situations where you did need to go to a location that wasn’t pre-mapped, it’s still likely to get you 99.9% of the way to your destination, anyway. Waymos are also constantly scanning and remapping their surroundings in real time (to adjust for things like road work, detours, and other blockages), so it stands to reason that if you had to manually engage the car after reaching the end of its mapped area, it would eventually develop a map of the new areas with a high-enough level of confidence to navigate on its own, over time.
I also think you have the wrong takeaway regarding the remote dispatchers being required. I think it’s a good thing that they exist, because that means that Waymo doesn’t trust their own machines to be 100% accurate. Tesla incorrectly does, which is why their self-driving cars have an exponentially higher rate of at-fault collisions than any of their competitors. Waymo at least understands the realistic limitations of their tech, and has a safety net in place for when those limitations are reached. I think that’s the correct way to do it, as you lead up toward an actual finished product.