[TillE]

This is such a well-defined problem with clear constraints, it's not the kind of R&D that can fail in the sense of a drug with unacceptable side effects or a high-tech fighter jet which becomes an overpriced boondoggle. It's simple navigation of paved, mapped roads with an array of sensors. Except perhaps for really harsh weather, even the edgiest edge cases are pretty limited in scope.

The absolute worst case scenario, within a couple years, is that you have limited deployment in favorable conditions, with fail-safes and human operators ready to take over remotely.

[robotresearcher]

You understated the constraints and missed the most important one: almost never, ever kill anyone. Driving around is relatively easy. Driving around and almost never killing anyone is much harder.

[mrguyorama]

You have to do better than not killing anyone. Every single accident, including ones not your fault, will be used against you to argue that your systems are death traps. You can't injure people, or cause property damage, or even be rear ended. You have to produce cars that are able to drive so defensively, that insurance is no longer necessary