Having done this before - it will certainly help in the context of testing, but there are major fundamental issues w/ using this train some big fat deep network

Having done this before - it will certainly help in the context of testing, but there are major fundamental issues w/ using this train some big fat deep network (not sure what they are intending to do with it to be fair) The problem is the same as the real world: safety is dominated by the very very very long tail - things that are so improbable that it requires humans to draw on their vast knowledge outside of driving to address. We know that something light enough to float in the wind is likely safe to drive through, that we can drive over a piece of plywood, but not a pallet. We can predict what other dynamic actors may do (noticing erratic or angry drivers compared to slow, predictable drivers). These are the hard problems of self-driving that haven't been solved yet, and honestly a simulation isn't going to close that gap. It will be an enormous distraction as engineers dive into simulating the chromatic aberration of their cameras, or what happens if cameras get too dusty or whatever ever nonsense is required.