> Everything points to the conclusion that solving the self driving problem for that specific operational domain, congested highway, is almost just as hard as solving self-driving in the general case.
The general self-driving problem is a much harder AI problem than the problem that Mercedes is solving with Drive Pilot.
The operational domain of Drive Pilot is not just any congested highway, it's particular stretches of limited-access freeways in Germany that have been identified and studied by Daimler engineers. These roads have machine-detectable lane markings and no tunnels, toll booths or traffic control devices (stop signs or lights).
The system uses satellite positioning with live corrections applied for transmission errors along these particular roads. The vendor has prepared detailed digital maps of particular signs, fence posts, and guardrails within this geo-fenced environment, and the sensors compare detected objects with the map, combining this positioning data with the satellite method.
This level of preparation reduces the long tail of uncovered cases for this domain greatly, making this solution much less risky for the car and for its neighbors, to the point where the vendor can assume liability and regulators can adopt a framework to support it and evaluate its use.
The general Level 5 self-driving problem throws too many uncovered corner cases at an autonomous car driving on random roadways designed for people. In my opinion, the current approach of throwing more and more data at ML networks cannot by itself bring down the risk to a low enough level to make it practical in my lifetime. Neural networks are not sufficient, and if we choose to enhance the roadways, then neural networks might not be necessary. Conceptual breakthroughs are needed.
I should mention that I build robots for a living, though not self-driving cars, so maybe these other vendors and their investors know something I don't.