Comment by tigen | Hacker News Reader

Comment by tigen | Hacker News Reader

Parent Full thread

Eye contact isn't reliable. I suspect it can't really be something software can safely use as input, even if it wasn't a vision issue.

It's pretty common for motorcyclists or bicyclists to make eye contact with a car driver who proceeds to pull out in front of them anyway.

This is probably in part from car driver brains not interpreting bikes as physical threats. It doesn't have to be deliberate. It can also be the driver scanning for "car" and not seeing "car".

A self driving car is on the other side of that size equation, but eye contact might also be assumed by a pedestrian to mean that you will stop. Especially if a "driver" makes eye contact who isn't driving.

In the end it's best to slow down and not assume. Humans often make dangerous assumptions, for example going around blind turns or over crests too fast to avoid something that might be there. Most of the time it's ok but eventually somebody gets killed. The rational thing to do is make the software safer than humans, not emulate them.

As a pedestrian I often use eye contact for the opposite reason: a multi-way stop where I’m worried if I cross I’ll end up pissing off several cars, and despite that they’re all being too polite and won’t just go, even though it would be fastest and safest for us all.

So what do I do? Get to the corner and stubbornly refuse to look up at any of the cars, and don’t really face my body toward any possible crossing. Not being able to wave me on (I’ve always found being waved in front of a motor vehicle somewhat menacing, tbh), the drivers quickly clear the four-way stop and I can cross without worrying about causing a jam.

Anyways, the AI will have a lot to learn of the crazy habits people build to keep from being lawfully murdered because “but cars”.

Ha. I do this too: as a cyclist who generally follows traffic laws and vehicular cycling behavior, I get annoyed when cars that obviously have the right of way at an intersection want to wait for me to go through first. So, I just stare at the ground and don't move until they do.

I know they're trying to be nice, but it ends up slowing everyone down: suppose a car and a bike are approaching a four-way-stop intersection with the car clearly going to arrive a couple seconds before the bike. As the cyclist, I would assume the car will come to a stop or almost-stop, wait about a half or one second, then start moving. With that assumption in mind, I'll adjust my speed so that I slow down for the stop sign (prepared to stop in an emergency) and roll through right behind the predicted position of the car, without losing too much momentum.

If the car takes their right of way, this is great. But suppose they decide to stop completely and wait for the bike to go through first (against the right of way). Now as the cyclist, I also have to stop completely, since I don't know when they're going to start moving again. If we make eye contact, I'd feel safe to go, but I feel like that's reinforcing bad behavior. Instead I refuse to make eye contact and force them to go first, as they should have initially.

Stop signs are only a suggestion for bikes? I thought that was only true in a few states?

That's why I added the qualifier "generally". When there's little or no intersecting traffic and no ambiguity about right of way, it seems perfectly safe to me to slow down but not completely stop at stop signs (i.e. the "Idaho Stop"). But I admit it's technically not legal where I am. Of course, most cars don't come to a full stop at stop signs in those situations either.

> So what do I do? Get to the corner and stubbornly refuse to look up at any of the cars, and don’t really face my body toward any possible crossing.

I.e., create maximum confusion as to your intentions? Doesn't sound great to me.

One issue with using eye contact for self driving is low reliability (or need for true general intelligence).

But long before that, you'd need to realize a camera suite with an effective resolution between 200 and 500 megapixels, sampling at 60 Hz. We're nowhere close to that today.

Or a pan-tilt camera with a long focal length. I see this as much more viable than a wide angle ultra-high-res one, because it's sort of how the eye works anyway: one really high-res patch, then the rest low-res.

Reply on news.ycombinator.com