Current experiments and production data from human controlled vehicles have not been encouraging.
What does NP-hardness look like for self-driving car tech? non-deterministic polynomial in: number of objects? number of lanes? time steps in the planning horizon? action/observation branching factor? These things are bounded in practice.
Not saying that the computational problems aren't hard. But ending the conversation at "NP Hard" throws away too many nuances.
2. Who do you blame when a faulty algorithm eventually kills a person? Who do you get mad at? When a drunk driver kills your family member, you can go to court and look at their face and look at the faces of their family members. When that happens with a self-driving car you'll just be looking at corporate lawyers who will shrug their shoulders at you and say "lol sorry our dumb machine killed your daughter".
Slowly with many AI-driven vehicles on the road, we can "optimized" for a better performance on safety and other issues.
People would still get hurt and die directly due to software that cannot be perfected, but the gains to society as a whole might be worth it.
But at this point what makes it better than a train?
Take the train, and you have no vehicle with which to go from the destination station to your endpoint. Also, you could travel at any time instead of being held to the schedule.
Humans can solve these problems due to life experience, not just driving experience. In other words, I think we're gonna need fully-conscious AI to solve self-driving.
The only way self-driving tech will reach production is if the input space is restricted, which is a significant-but-not-groundbreaking iteration on what we've been doing for decades with airplane autopilots and self-driving monorails. Sure, we can have self-driving cars on specifically designed freeways, but nothing more.