From a technical point of view we thought that we need to accumulate a lot of data and thanks to the magic machine learning, it will work out. We used lidars, cameras, or both.
Today the enthusiasm is gone, rumours say Apple canceled the project, Google hasn't released anything, Uber killed a pedestrian, PSA announced that the technology wasn't there, Renault forgot about it while his CEO was busy escaping countries, etc...
Then you have Tesla, with a insanely high stock price, that pretend to be close to full self driving thanks to machine learning. I rented a Tesla model 3 for the first time this weekend and it's insane. The Tesla AI is stupid and dangerous. The AI has issues with simple tasks such as managing the windshield wipers, but will also do dangerous thinks such as smashing the brakes once in a while.
I started to reflect about the limitations of machine learning because of one "phantom breaking" incident I experienced. I was driving on a motorway and I was catching up a car pulling a boat on a trailer and we entered a tunnel. I guess the AI wasn't trained for that situation, seeing a boat in a tunnel in a Norwegian motorway, because it decided to do an emergency stop. A vision based but non machine learning approach would have detected the moving object in front of me, estimated the distance and assumed I was safe.
So, do you think we simply need more data for training and machine learning will stop doing stupid things? We need to accumule data for all kind of objects people can put on a trailer in all light conditions? Is this kind of problem something machine learning can resolve eventually, or is it a major problem that will stay forever?
Tasks that are very difficult for a ML / CV system to solve could become trivial if we could rely on world around us to have the right sensors embedded.
In your example if moving vehicles on the street had beacons on them, you would not had to rely on ML to determine if you are safe. Of course, we do not have access to this kind of data currently and so we have no choice but to rely on ML / CV / AI. But once upon a time there were no highways for cars either. Society as a whole decided that building these highways was worth the investment.
I believe that society will slowly evolve towards building the kind of infrastructure that makes autonomous driving much easier. So ML may only be needed to make the ROI on this kind of infra clearer to more people.