Hacker News Reader

> I want it to do something like a minimax search and do something that will be safe in the presence of the unexpected

This is called defensive driving, and you're wrong about it not being based on prediction. Defensive driving predicts that every wrong thing that can happen will happen and positions the vehicle to minimize the harm when they do.

Fundamentally, if you don't predict that the car up ahead may swerve into your Lane and slam on its brakes, then you're less defended when it does. Defensively choosing what to do and where to be at any moment is based on predicting it happening and acting accordingly.

I think "prediction" in this context means something different. It's not specifically that self-driving cars will optimistically assume they know what other drivers are doing, like a naive driver might (although that is a form of prediction). It's more generally that in order to make any smart decisions at all, the AI has to model the future.

For example, if you are driving behind a car on the highway, and the car _in front_ of that car puts their brake lights on, you may want to pre-emptively start braking now now. That's a type of pro-safety prediction made based on your understanding of the likely future that driving AIs should also have.

As another example, if you're on a roundabout, if there's a car already on the roundabout (to your right in LHS driving countries like the UK, to your left otherwise) then you shouldn't enter it. But an exception is if you can clearly see that the car is going to exit the roundabout before it reaches you: this requires prediction, but not in the sense of guessing what the car might do, or even trusting their indication (ha!) but observing from their current trajectory and whether their wheels have straightened.

The roundabout is a good example of where it would be pretty easy to design a safe approach for self-driving cars (do not go until the roundabout is clear), but in a heavy traffic area, would result in the self-driving car just sitting... and sitting... and sitting... waiting for the roundabout to clear.

A self-driving car that takes forever to get anywhere could easily fail in the marketplace, even if it is clearly safer statistically.

The thing about very safe driving strategies is that humans can already choose them at any time, without buying anything. I, too, can sit and wait for a roundabout to fully clear before going.

The fact is that a lot of people do a sort of risk-cost-benefit analysis in addition to safety analysis when driving. I'm not sure a car optimized heavily for safety, at the cost of efficiency of travel, will meet expectations.

Msybe we will be able to choose a "unsafe" mode from a menu, also by acknowledging the full responsability if anything happens.

Hard to sell indeed.

They'll just call it "sports mode".

Just need to regulate a flagman for self driving cars

https://66.media.tumblr.com/fdd6a0be1711940bd5a9e0f31e9f01ac...

then we can have automation without job loss into the bargain.

Your problem is that your prediction was wrong, not that you shouldn't try to predict what other cars will do. Plenty of times that prediction ability will save your bacon. But you have to work hard at getting it right and making sure that you realize that all predictions should come with error bars that still include all possible futures at different levels of likelihood. Acting like your prediction is a certainty is a mistake in a class all by itself.

I don't understand. 1. The problem is that the GP prediction was wrong 2. You need to make sure predictions are right. 3. All predictions are wrong. 4. So add error bars.

Should the downstream system trust the error bars with certainty?

It's hard to define what a right prediction is. The correct prediction is whatever makes the car safe and effective. If an action leads to a situation with a low probability but very catastrophic outcome, the system should not take that action, even if the prediction correct with its error bars.

Acting on a wrong prediction without a wide enough tolerance for errors in the prediction itself is what caused an issue here. If the OP had not made the prediction they would have had to keep their options open, which was the safer course of action. The fact that they were a beginner compounded the error (and is exactly one of those reasons why beginners tend to pay higher insurance premiums).

I think it's another stream of data feeding the model. For example: you see a child walking along close to the edge of the road, not looking in your direction. You don't slam your brakes on, but as a driver your prediction is that he may step in front of you, so you ease back on the accelerator, move out further into the middle of the road, cover the brake.

The prediction talked about here is if you are at a 4-way stop, and the car to your left has it's right turn signal on, you can predict that they are probably turning right, and so it is ok for you to go through the intersection without waiting for them. Currently, and autonomous car will wait until the other car is no longer in the intersection. With prediction, it will act more like a human. That doesn't mean it 100% believes the car will turn right, but will modify its behavior if it sees the car starting to go straight.

Basically, autonomous vehicles almost entirely assume the worst outcome in all situations, and waits until it is safe. So at a 4-way stop, it will wait for all cars to be out of the picture before going, which results in horrendous performance if the intersection is busy.

When I drive, essentially only use turn signals to restrict my behaviour vs the default assumption. For example, if a car is indicating to turn right, I will assume it is likely to either go right or go straight and wait accordingly.

Reason for this is too many times I've just narrowly avoided a collision because someone forgot to cancel their signal, or they were signalling ridiculously early for a turn they were going to make further up, or even that they just activated the signal accidentally.

Not to suggest prediction is bad, but just that also not treating turn signals with too much authority is safer.

If I am at a 4 way stop and there is a car to my left my prediction is that it's ok for me to go through the intersection regardless of the status of their turn signal because the rules of the road require them to yield to their right. I still pause to observe the trajectory of their vehicle, though, because a lot of people don't seem to know that rule.

I'm not sure if I'm reading your comment correctly, but in case I am...the "yield to the right" rule is only to break a tie if two people arrive at the intersection at the same time. Otherwise, the first to stop is the first to go. I haven't run across many people who don't play by these rules (or maybe my assertive driving style just means everyone else follows my rules anyway).

Well yes, I always yield to vehicles that are already in the intersection, irrespective, again, of their turn signals. Also it is not purely true that "the first to stop is the first to go". Specific minutiae of the law differ depending on region, but generally if I am approaching a 4 way stop with the intent of turning left and another car straight across from me is already stopped and going straight then, while I am waiting for that driver to cross the intersection another vehicle arrives at the stop to my right I am still obligated to yield the intersection to the new car despite having "arrived first". Likewise, when many cars are enqueued at the intersection, turns should continue clockwise, as in most board games, without regard to who is able to zoom up and brake the quickest.

You had the wrong prediction. That doesn't mean prediction is wrong. In fact the 2 second rule is predicated on the prediction of reaction time in the case of something unexpected, which is itself predictable with a probability distribution. Things which can be reasonably expected to occur should be taken account of. That includes e.g. suddenly swerving for an exit, an unusually slow car doing a u-turn on a street, vehicles performing emergency braking when the road ahead has limited visibility, etc. And these predictions take some time to develop.

On my motorbikes, making good progress depends on good predictions, and that means predicting that people don't see me, that they make rash and impatient decisions, that tourists on rental bicycles will do stupid things without looking, etc.

I think prediction in this context means objects actions probability distribution, not fixed action. You need to take care of adversarial distribution (up to realistic threshold).