[ActorNightly]

I think it can happen, but with current methodology, if has to be a fully mapped data set across known driving roads with car to car communication, with the model having some sort of truth data in relation to gps/compass information for self correction.

Simply mapping sequences of frames to decisions is never going to work without above.

To get fully safe, autonomous driving, we need to have agents that can on the fly adjust to operate on data sets that they have never been exposed to. This is not possible with current training methods.