[bumby]

>They are absolutely capable of looking at gestures.

But 'looking at gestures' isn't the same thing as having a real-time interpretation of the sensors. This, to the point of another comment above, may be conflating 'perception' with the larger problems of 'self-driving.'

Humans see a gesture and understand the meaning very fast because we process a lot of communication on the subconscious level. If you read the safety report of the Uber incident, it shows that the self-driving system could 'see' the pedestrian, but it struggled to effectively classify them. To be fair, that was years ago and another company, but we (as the public) don't have much insight into how good (or bad) the software is.

So 'seeing gestures,' 'correctly classifying gestures', and 'correctly classifying gestures in a time-sensitive manner' are varying levels of difficulty. The last one is the one that matters for safety-critical operations, and the requirements change with the velocity of the vehicle. IMO, there should be a regulatory framework that addresses a baseline of performance on these systems.