[ogreveins]

I don't see why it would. Once you get enough base data you can start simulating the data from what you have, inputting different scenarios without actually encountering them IRL. Faking sensor input and randomizing should get it most of the way there.

[51Cards]

When lives are involved handling edge cases is everything. The person stepping off a curb, the cyclist that falls in front of you, the car that weaves in its own lane and can't be used as a reference, traffic lights that are out of order, stop signs hidden by trees... and on and on. Mess one of these up while autonomous and severely injure someone and you're done.

Human drivers might only see one of these cases a month, or 6 months, but not driving over someone in that case is what is critical. Not saying it's an impossible task, but IMO it will require a lot more training data than humanly possible for one person to generate.

[vonmoltke]

I have significant experience in faking sensor data (specifically radar), and can tell you from it that fake sensor data is terrible. There is way too much going on in the real world to accurately create sensor data without actually recording sensor data. That is, you can manufacture the situation for the sensor to capture much more effectively than you can manufacture the data from a model.

Even pseudo-faking like we were trying to do, wherein as generated signal is injected into actual, recorded background noise, is fraught with problems. Anybody who tries to develop a control system based solely on such data is in for a rude awakening when they try it for real for the first time.

[JustSomeNobody]

Faking input is how most people test their buggy, crappy software. It rarely matches reality.

[cscurmudgeon]

Ah the naivety. Just like 50s AI research :D

History keeps repeating