Hacker News Reader

Comment by msravi | Hacker News Reader

I'd suspect that it's harder to take over and avoid a colision in a self driving car, than a car that you're continuously in control of. You first have to recognize that the system is failing or about to fail. And that has to happen well in advance for you to take appropriate action. Doesn't seem like a reliable failover procedure, even if the person behind the wheel is paying attention.

View on HN

> You first have to recognize that the system is failing or about to fail

For that to work, I'd like that cars that have features like this show you what they're seeing. Tesla does this a bit by showing all the cars it has detected in the dashboard. But it should go further and other car brands should do the same.

Without that, you're right, you can not be sure if the car is going to react to something on its own.

I honestly don't know where Tesla's head is. I'm surprised that when I turn on the turn signal to change lanes, and the car has detected something in my blind spot via sonar sensors, that it doesn't give me a tone or something. Let alone showing me video from one or both of that side's cameras.

Tesla, in putting it's targets on self driving, has omitted a lot of low-hanging fruit that would make significant day-to-day difference to drivers right now.

Interesting. My 2013 BMW vibrates the steering wheel if it thinks I'm going to move sideways into something (and that includes fast moving cars coming from behind) or it beeps loudly if it thinks I'm going to hit something in front of my based on my speed and distance of the object in front. And its highly accurate IMHO. My take on self-driving is someone needs to hook up these signals to controls on the car - the "hard part" does a reliable job.

Similar, my 2015 Audi will rapidly blink "inside" lights on either side as soon as I turn on my indicator if there's something in a blind spot. It's also at least somewhat adaptive to their speed too (i.e. if someone is behind me but at the same speed, it won't warn, but if they're a few car lengths behind at a notably higher speed, it will).

I'm legitimately surprised nobody has a HUD that simply displays all detected objects/potential obstacles. They have this data, just pipe it up to a HUD.

If you are watching a very complex display, why not present it in a format the driver will understand without training, such as a processed video image. Here [1], [2] are some (very optimistic) mock-ups. There was another one released in the last few days but I can't find the link. Showing anything like the raw data with rapidly changing risk annotation would get tiring very quickly, especially as it masks the view the human is actually able to understand, the raw view.

[1] https://futurism.com/self-driving-car-video/

[2] https://next.reality.news/news/augmented-reality-cars-compan...

Nobody does it because it would clearly point out to everyone that the self-driving emperor has no clothes.

This the problem; once you have to pay attention at that level (and I agree that you do to avoid accidents) whats the point of the automation? The taxing part of driving is not the physical act of slightly turning a steering wheel or break pedal but the mental processing power it takes to know when.

> This the problem; once you have to pay attention at that level (and I agree that you do to avoid accidents) whats the point of the automation? The taxing part of driving is not the physical act of slightly turning a steering wheel or break pedal but the mental processing power it takes to know when.

We are in between the technologies. That's litteraly beta testing. Are you saying that when you test your software you expect that everything that get out of it is gospel or you VERIFY and make sure it is?

This is a beta software, this is the testing, this is making sure everything is alright. You have to pay attention at that level BECAUSE it's a test, just like your software you wrote.

Once it's good enough (and we are still pretty far from it), then you won't have to pay that much attention.

The closest you can possibly come is to have your hands on the wheel and feet on the pedals and mime each action you think the car SHOULD be taking... but I suspect that would lead to a shitload more 'intervention events' being recorded for things the car was just doing differently to the particularities of the particular human driver.

Agreed it's too difficult.

Yes. Also, the whole "not paying attention" thing is encouraged by the task. It's completely predictable and these semi-autonomous systems are worse than nothing.

I crash a drone this weekend. I was looking at the controller and not to the drone about to crash. When I finally see it almost crashing, the panic kick in and was unable to stop it.

And this was just a drone...

Reply on news.ycombinator.com