Almost all technical problems of self-driving cars are solved by current algorithms once we get 100x faster HW, i.e. processing at 1000 fps instead of 10, making things like tracking, collision avoidance, vision, etc. trivial; the only problem left is path planning (down to inch resolution with dynamic vehicle models), which is coincidentally NP-hard and there is very little hope to get it to human level in reasonable time (Waymo is still orders of magnitude worse than an average human).