We're not talking about complicated scenarios with multiple moving actors. Tesla's autopilot cannot even do something as basic as detect stationary obstacles that are directly in front of the car. It will crash into barriers even if the highway is completely devoid of other cars.

You may consider humans as bad drivers but Tesla's autopilot is even worse than that:

It can't even look in one direction!

I'm talking about the pitfalls of human perception, and the low-hanging fruit of ways that self-driving systems can potentially outperform humans.

I'm not claiming Tesla's system is currently better than a human, just that there is plenty of potential for a machine to outperform humans perceptually. As it is, Tesla's system isn't exactly the gold standard.