Hacker News Reader

Comment by nlh | Hacker News Reader

Was discussing self-driving cars with a few colleagues on a drive the other day, and a question came up:

Does anyone have a sense how the cars handle situations where traffic is being directed by a human, with hand signals, etc? I.e. a cop standing in the middle of the road, waving through cars through an intersection where there's been an accident, etc.?

We pondered a few scenarios -- are they reading the hand signals? Are they judging other cars' movement? What if it's not a cop but a crazy random person who jumps into traffic and decides to play cop? We humans are remarkably good at determining "soft" things like "that guy looks crazy - let's get out of here" vs "that's a cop, I better pay attention", etc. The rabbit hole gets quite deep here...

View on HN

> What if it's not a cop but a crazy random person who jumps into traffic and decides to play cop?

Most jurisdictions tell you to obey any random person directing traffic.

That's important, because there are good reasons why you'd want random passerby's to be able to direct traffic and be obeyed (eg a big accident that just happened, but it's hard to see). And obeying directions for traffic, even if you desperately want to obey, only really works when you can assume that other people on the road follow them as well.

Of course, if after the fact it turns out that the random guy didn't have a good reason for directing traffic, the law can get him. But you'd still have to obey him in the first place.

(A bit like a commander giving (bad) orders in the military.)

> And obeying directions for traffic, even if you desperately want to obey, only really works when you can assume that other people on the road follow them as well.

This is the key. In the end, the concurrent intelligent-agent-based system that is driving works because all the agents have a mutual goal (to have traffic proceed smoothly so they aren't stuck in it), and so will cooperate to fulfill that goal even without any central authority dictating behavior.

"Traffic" as a whole works as well as it does because every node in the system can intelligently compensate for the failures of the nodes around it, and can decide to cooperate—in a span of seconds—with another driver's compensation strategy.

I would guess that the majority of the Google automatic chauffeur's high-level behavior isn't explicit, but rather "arises" from implicit subroutines about following the flow of traffic. For example, if there's an accident in the middle of a six-lane road with roadside parking, the car could be programmed to notice cones and avoid them, and to notice people directing traffic and follow them—but then it might get stuck, because there's no single "lane" open any more. Or it could just be programmed to notice that up ahead the lanes are merging together and driving half-way between the middle- and right-hand lane—and so, to follow the flow of traffic, it should merge too.

In other words, if all the automatic chauffeur's friends jumped off a bridge, it would likely be a pretty effective optimization for it to jump off too.

(Though I'd hope it was taking the hint mostly from regular cars; a pack of driverless cars following one-another's lead could turn out badly if the one at the head of the pack is malfunctioning.)

> Most jurisdictions tell you to obey any random person directing traffic.

I watched the most amazing form of this in action after Hurricane Andrew. You'd pull up to an intersection and a random person would be directing traffic till (it turns out) the car they were passenger in got through. Then they'd go hop in their car and another random passenger from another car would hop out and direct traffic for a few minutes. Rinse, wash, repeat.

I actually ran into a simple example of a challenge like this. The power went out in mountain view a few weeks ago and a bunch of stop lights stopped working. I pulled up behind the self driving car which had stopped at the disabled light. Instead of treating the light like a stop sign, the car remained stop. After 5 or so seconds stopped at the light (there were no other cars at the intersection), I honked at them. I assume someone took manual control because it started to move right after.

While this situation could probably be addressed, it is a great examples of challenges these cars may come across.

That's why they get tested for years before being sold; so almost every possible situation comes up at least once and can be included in the system. I bet that driver filed a note about what happened and it was fixed soon.

Woah.

What do you think actually went down, there? "Hey, of course, let's just implement the StopHandler::HumanStopSign() subroutine, thanks for the report!"?

This is Google we're talking about. I think they will spend time to generalize this situation as much as possible.

I hope it's not the same people that wrote the VAT handling code on Adwords.

1. Find a way to recognize when this situation occurs. Test that. 2. Decide what the optimal decision should be for the car. 3. Program the car to do 2 when it's in that situation.

I assume someone took manual control because it started to move right after.

This is why I believe self-driving cars will always need a human to take control whenever unusual situations arise, just as planes with autopilot still need a pilot's attention.

But there are autonomous planes known as drones.

Almost nothing "in production" with the military is autonomous for the entire flight, just piloted from the ground. Those that do have some autonomous operation still require manual takeover for critical situations i.e. landing. And we don't trust them with human passengers.

Drones are essentially treated as disposable, because they are. We just lose them sometimes. This is not a state of affairs you can have with a car containing multiple humans navigating pedestrian, cyclist, and construction-laden streets.

There's also a lot less that happens during (most of) flight. Surroundings are relatively static, other planes in the environment actually broadcast what they are and what they're doing, so you don't have to discern it from visual noise. There are no stop signs, no intersections, very little traffic, etc.

> Almost nothing "in production" with the military is autonomous for the entire flight, just piloted from the ground.

We are awfully close:

https://en.wikipedia.org/wiki/Northrop_Grumman_X-47B

http://www.wired.com/2013/07/navy-drone/

There is definitely a point where the drones are going to be better at take off and landing than human pilots, in that case...safety will become an issue (why let humans do it when the drone can do safer?).

The same will happen with cars. It is only a question of whether that happens in 2020 or 2030.

Airplanes on autopilot usually have many minutes and several hundred miles worth of buffer before something bad happens.

Cars that say "I don't know what to do, Jesus take the wheel!" might only give drivers a fraction of a second.

At some point quantitative differences become qualitative ones.

> Instead of treating the light like a stop sign, the car remained stop. After 5 or so seconds stopped at the light (there were no other cars at the intersection), I honked at them.

There are many normal humans who don't know what to do when the traffic lights are in a non-normal situation

The difference is that once a self-driving car is programmed for what to do, it won't get it wrong again.

Self-driving cars will get continually better and will eventually surpass human drivers simply because human drivers will never improve.

It handled it completely correctly. Lights being out is an exceptional situation where the usual rules of the road are suspended.

Sometimes it means "treat it as a stop sign", but other times it means "turn the car around because ignoring the lights on this particular road would be positively suicidal". e.g. a level crossing.

Computers have no conception of danger, and so they don't get to decide what's dangerous and what's not.

Actually, in California, that isn't true. An intersection controlled by a traffic light converts to a 4-way stop when the traffic light is out.

However, even for your example of a level crossing, humans fail regularly. People regularly run into trains in Texas at a level crossing because the train has a long stretch of flatbed cars that someone missed at dusk or at night.

Human drivers follow a bathtub curve. Young drivers (particularly male), are very dangerous in the beginning (so much so, that their insurance rates are much higher). Drivers become more dangerous as they get older, and their reflexes, vision, situational attention/awareness decline. I think one of the best uses of self driving cars initially will be for those populations.

Maybe we're good at differentiating humans with authority vs. humans without. But there's a much simpler social-engineering tactic for rerouting traffic that both humans and autonomous cars will fall for. I'll summarize it like this: road pylons are $8 at Home Depot.

What's more, www.roadtrafficsigns.com will sell MUTCD-compliant traffic signs to... well... anyone.

The car could have a sensor platform that a remote driver could hook into and control over the air. In the 1% of cases the car gives up and can't handle, a human can pop in and guide it out of the situation before returning control to the car.

Reply on news.ycombinator.com