There definitely exists a sort of pre-cognitive human quality to traffic.

Sometimes you just slow down on a subconscious hunch without any information on that particular situation. Or you might be very conscious not to slow down if you're driving a motorcycle in Asia in heavy traffic trying to make a turn with a huge truck barreling behind you.

There are massive differences in road cultures between countries, and it takes a while to adjust to local habits. Somebody who doesn't know these invisible rules is a very accident prone driver even if they can drive safely in another country.

A couple of personal examples.

1. I've noticed that adaptive cruise control doesn't account for vehicles entering my lane in front me, even with their turn signal on. As a human driver, I see them entering and know to slow down (at least take my foot off the gas), but the system only performs a hard brake AFTER the car has actually moved into my lane. Disclaimer: I don't know if autopilot systems suffer from the same limitation.

2. I was once waiting to make a protected left turn at a busy intersection. There were two protected left turn lanes, and I was in the farther left lane. Light turns green, I take my foot off the brake, and prepare to accelerate. The car on my right, in the other left turn lane, is also barely starting to move forward, when it suddenly rocks to a halt, indicating they applied the brake, hard. Without thinking, I also brake hard. Out of nowhere, a car appears speeding right through where I would have been if I hadn't stopped. Apparently, this car was entering the intersection on the cross street, going right to left. He wanted to beat the red light (he didn't), and also make his left turn, right down my street, in the opposite direction I was headed. I guess he didn't realize there were two left turn lanes. I never saw him. If I hadn't gone with the herd instinct and hit the brakes, I'm sure there would have been a serious impact. I wonder if autopilot would have caught that one, but it probably depends on camera placement.

A single self driving car probably wouldn't (might do the same thing you did at best).

It'll take actual inter-vehicle communication to extend the sensor-net. Much like humans presently do via proxying though the other car's deviations from anticipated behavior. (EG if the freeway traffic keeps going at proper speed around corners)

It's referred to in the motorcycle community as 'Spidey Sense' - the way your subconscious picks up dangerous situations before you consciously become aware of them, giving you a chance to slow down and become more vigilant.

This is a really great point. I wonder if at some point self driving cars will be taught with these regional anomalies in mind. And will they be activated/changed depending on where the car is shipped/activated (Asia vs. America)?

Apart from regional differences, there are a host of other, similar matters. Cruising through your city's bar district at 3am? I'm sure you'll drive more defensively and be prepared for erratic behaviour on the part of others.

However, one thing that I haven't seen discussed anywhere, but that strikes me as a hard problem: by having a non-human driver, you're literally taking out the human element of communication.

Everybody is taught about the importance of eye contact in driver's ed. What will we replace it with? I have no clue, and I doubt anyone else does. Yet it's a crucial technique we all use to negotiate traffic every day, regardless of our mode of transportation.

This is a hard problem, and it's less technical than cultural. It will be a bitch to solve.

Eye contact is another thing that doesn't apply in some Asian countries where by default everyone has dark tinted windows making it virtually impossible to see inside other vehicles.

> some Asian countries where by default everyone has dark tinted windows making it virtually impossible to see inside other vehicles.

Hmm. Perhaps. Personally, I've never noticed this.

Anyway, the concrete example doesn't matter.

I guess you know what I was trying to say: implicit communication between humans. Even without eye contact, there will be a host of minuscule actions (or inactions) that we use to convey intent. There will always be humans on roads, e.g. as pedestrians. Vehicles will be forced to interact with them, and this communication barrier makes it very difficult.

Don't think of highways, think of supermarket parking lots.

Well actually the parking lots are especially difficult due to lack of eye contact.

But I get your point.

People compensate by being very careful when parking and of course you can always roll down your window if the situation requires it.

A robotic vechile simply couldn't manage here. There are way too many dynamic human exceptions. For example, it's common to drive against traffic on the wrong side of the road for short distances since the intersections are so far apart.

Machine learning also has tendency to unlearn things and corrupt it's own models when left unaided.

So do people ;)

Well actually we have not proven that ML works the same way, or proven that our ML has the same input parameters as a human yet have we?

I have had vehicles blow by me on a bicycle at 50 mph, on fairly narrow county roads.

Same here during my road biking years (I switched to Jogging and unicycle). Even though I disliked the risk, I enjoyed the wind as it allowed me to ride faster.

or that at 50mph, a foot away, you're likely to blow the cyclist over ....