Hacker News Reader

Comment by daeken | Hacker News Reader

Have you seen car accident stats? I'll concede that bugs do occur, but it would take some serious failure to make it any worse than the average driver. Even decent or good drivers are not that great, especially when you factor in distraction, tiredness, poor judgement of road conditions, etc.

View on HN

I understand that you're talking about the interactions between real drivers and autonomous drivers. I still think that the autonomous driver will be better able to react to even a poor human driver than a real one. With all the sensors, they can detect when someone is drifting lanes, not stopping at a stop, etc. and can react accordingly with much more accuracy than a human driver.

You're right, but the software of an autonomous car will be held to much higher standard than just 'has be to better or equal to a human driver'.

The whole thing gets interesting when you look at the judicial side of it: Who is responsible in the case of an accident? The car manufacturer (if it was a software failure)? Then even one accident triggered by it is too much.

For autonomous cars or not, whoever is liable in the case of an accident will acquire insurance to cover the cost of accidents. Currently, that's the driver. With autonomous cars, that may be the manufacturer, in which case 1) The cost to the manufacturer of insurance will be factored into the retail price of the vehicle and 2) The total cost of insurance will be lower because the accident rate will be lower.

It will be interesting to see how this plays out. Hopefully society sees the benefits and gets out of the mindset you mention above (which I understand you are pointing out, not advocating).

The idea that 'even one accident is too many' is a joke when you consider how many accidents there are with human drivers, and yet we accept them willingly. If there are 30,000 deaths per year from human driven accidents, then surely the real answer to how many accidents by computer driven cars is acceptable should have an upper bound of 29,999, right? Why isn't it acceptable if it saves even one human life, rather than being unacceptable if it costs one?

Unfortunately, we as a group don't tend to be very logical, and love to spread blame, so I am afraid we'll take a long time to start saving lives. The data on the self-driven cars on the road is pretty overwhelming that they are safer than human drivers, but unfortunately there will be those (as in this thread) who believe that is true of everyone but themselves.

Its going to be a long road to acceptance.

If you were to replace every human driven car on the road with only autonomous cars, I will agree with you that the accident rate would likely be much lower over all.

But there's two problems here:

1) We won't be replacing all the cars overnight. The problem comes with the interaction between terrible human drivers doing wildly unpredictable, insane things and the inflexible, unadaptable automated cars.

And

2) I'm not comfortable with dying or being injured due to a software error regardless of whether its likelihood is higher or lower. The roads are far more dangerous than the air, and this is why I'm uncomfortable with the whole concept of automated cars on the road and I'm not uncomfortable with autopilot in planes.

> I'm not comfortable with dying or being injured due to a software error regardless of whether its likelihood is higher or lower.

Why on earth is dying due to a software error worse than dying due to a human error if you can chose one or the other, and the chance of a software error is lower?

Individual odds of being in an accident may be lower than the average. Software for all of the cars would be approximately the same, so the chances of being in an accident could increase for extremely cautious drivers. I doubt they would, personally, but it would be an argument against adoption until the cars are demonstrably safer than possible with human drivers.

If the chance of an accident increases then sure, be against that. And I've no idea what the chances are. But to say you'd rather have a higher chance of dying from human error than lower chance of dying from software error... doesn't make any sense.

The truth is our lives are in the hands of more software programs than we may care to admit.

the interaction between terrible human drivers doing wildly unpredictable, insane things and the inflexible, unadaptable automated cars.

There are only so many things bad human drivers can do, the physics is pretty limiting. It isn't like they can teleport in front of another car. Even the craziest things like opposing traffic jumping the median into oncoming traffic isn't all that crazy, it isn't conceptually different from a deer jumping in front of a car.

I'm not comfortable with dying or being injured due to a software error regardless of whether its likelihood is higher or lower.

Hard to argue with that.

> There are only so many things bad human drivers can do, the physics is pretty limiting. It isn't like they can teleport in front of another car. Even the craziest things like opposing traffic jumping the median into oncoming traffic isn't all that crazy, it isn't conceptually different from a deer jumping in front of a car.

I was thinking about your assertion that it's easy to solve, and why exactly I disagree with it so strongly on a base level. What am I basing this belief on?

And I figured it out. Driving games.

The AI in driving games like Gran Turismo has had more work than Google or any other company has done on theirs. And not only that, the AI in driving games has perfect information of the physics involved and the environment around them.

And those AIs are still fucking terrible when interacting with human players. Pack of AI driving around? Just fine. Add a human driver to the mix and it turns into a disaster.

This is why I am very concerned about automated driving in the real world alongside human drivers.

First, I don't think it is easy. I just think it is feasible because it is a reasonably bounded problem.

As for driving games, those have a different set of requirements and incentives. Driving poorly isn't necessarily a bad thing from the position of game play.

The AI in games like Gran Turismo is a terrible example. That AI is not designed to avoid accidents. Accidents are an acceptable part of the game play, and you're racing at high speeds above the norm.

It would be at least as difficult, or less so (but not more so) to program AI to interact with human drivers with the primary goal being safety, not speed and exhilaration.

My experience as a human driver is that most accidents can be avoided by either (A) slowing down in ambiguous situations, or (B) aborting that lane change (usually because two cars are attempting to go into the same space).

Those behaviors are easily improved upon by software. Dealing with lane changes is something software is likely to be vastly better at, because they can point sensors in all directions -- I only have one set of eyes.

Furthermore, a lot of horrific accidents are because of bizarre unexpected things that the human does not react to in a timely fashion precisely they are so far out of the normal scope of traffic flow.

For example, what if a car coming the other direction drifts in front of you on a rainy night? This is an unfortunately common and very lethal kind of accident in rural areas. It will take you 1 to 3 seconds to interpret the scenarios and react appropriately. (You only get 3-4 seconds before your are killed.) A computer can instantly recognize that car 100 yards away is maneuvering in a potentially lethal manner and begin slowing down right now. That buys more time for everyone.

Strangely, this might kill autonomous cars entirely. Any failure to properly respond to something bizarre can be the subject of a lawsuit that could kill the autopilot manufacturer.

Engineering solutions are a tiny part of getting this stuff to market.

Reply on news.ycombinator.com