Hacker News Reader

Comment by ExoticPearTree | Hacker News Reader

Is there a car autopilot system that good on the market that only you know about? And that can replace the driver completely?

View on HN

Waymo is safer than a human driver already.

https://www.nbcbayarea.com/investigations/googles-waymo-safe...

Of course there will be locations where humans are required, but we can already start using computers in those locations where they perform better.

Where I drive daily I constantly have to make decisions which no self-driving car system I know about can possibly do. They don't foresee upcoming gridlocks caused by you driving completely legally but blocking (again, legally) a car from crossing to a side road, that car again blocks the car behind it, which blocks the car in front of you from doing a left turn, which means that you can't move and unlock the gridlock. That's just one on my commute. Then there's driving uphill on snow, seeing a person walking on the other side of the road and a car coming in the other direction. You foresee that this car will have some problems braking for that person (which would be the normal procedure - I'm coming the other way, that car shouldn't try to pass the person). So, what I do is to slow down and give that other driver confidence to pass. And sometimes I have to do the other thing - speeding up a bit in order to create room for another car to let that car avoid problems (if I were a computer I would have no reason to think about that).

This is also something happening all the time where I drive. And tons of other understanding-the-traffic cases.

Self-driving systems only follow the rules. That is not sufficient. When general artificial intelligence eventually arrives - that's when we get self-driving cars better than humans, in any other cases than for trivial cases. And yes, computers could drive cars in trivial places, it's just that those places are pretty much non-existing where I live and drive.

You say all this, yet waymo is demonstrably safer than the average driver.

Uphill in snow maybe you’ll have to take the wheel, I’ll give you that.

Waymo hasn't tried driving outside its comfort zone, so I'll argue that it's not demonstrated.

Humans should also abstain from driving outside of their comfort zones.

Regardless, they don’t and do mostly fine

Have you driven in SF?

No, never been to SF, I have no idea how people drive there or what the conditions are. Except that they're very different from where I drive (except when I'm in the town centre) - roads without sidewalks, mostly no signal lights, snow a large part of the year, and other special conditions.

Good thing that all airports follow similar standards and are all standardized in terms of lighting, signage, etc.

That is not quite accurate. There are certainly standards for airport markings, but there are many non-standard airports... especially when flying at anything smaller than an international airport.

(And yes, even airliners will sometimes land at small regional airports.)

Mate, have you driven in Mumbai? Hell, even in Rome? SF is nothing.

So clearly self driving cars are completely useless because they cannot handle Mumbai traffic.

How has this been demonstrated? They have certainly had their share of crashes already.

Yes https://arxiv.org/ftp/arxiv/papers/2309/2309.01206.pdf

While everything you mention are signs of good, defensive driving and seeing more than "Just following the rules," average human drivers don't do what you do. And that's what the automated systems are compared to.

So again, while I think ultimately the automated systems should also do what you're talking about, right now, if the other article is true in its claims that Waymo is safer, it seems sufficient to follow the rules to achieve a certain level of safety.

[other drivers] But they do - where I live. It's very rare to see people not doing that, so much so that everyone really notices.

Edit: And the point is really that strictly just following rules (i.e. just drive if you're in the right and the signs say so) will cause accidents and worse. For others, not necessarily myself. On my commute and nearby.

What makes you think that a self-driving car cannot have similarly predictive algorithms?

There's nothing so far indicating that they have, at this point. Though I doubt predictive algorithms would work - there are all kinds of situations, all slightly different, and lots of them about understanding what someone else might be thinking. As I mentioned, I think we'll need general artificial intelligence to do this properly.

> Waymo is safer than a human driver already.

Is that data controlled for environment? I.e. it must only compare Waymo driving with human drivers in exactly the same environment. Same road, same conditions, same time, etc. Waymo doesn't take any difficult routes.

Not to mention the fact that when it encounters a situation never seen before it just "freezes" - happened a few times and caused traffic jams.

What I give these cars is their ability to react "instantaneously" to emergency situations like automated emergency braking when there is an obstacle in the front of the car or the car in front suddenly brakes.

IMHO, there was not even enough data to draw this conclusion from the study. 3.8 million miles looks like a lot, but it is laughable for traffic safety conclusions.

And of course Waymo is going to milk as much as possible from their study, we just shouldn't trust it blindly.

We need far more data to say this with certainty. This data is from only 3.8 million miles, in carefully controlled conditions with brand new high quality vehicles.

Besides this comes from Waymo, they didn't release the data used in their study.

Of course Waymo will try to paint their cars in the most favourable light possible. Maybe they are being honest, but that's not the point. The point is that public policy shouldn't trust the allegations of a private company at face value.

Mercedes just released a level 3 car.

Openpilot, while primarily lane keeping, does it way better than any human can.

Driving doesn’t require an advanced degree in physics. It’s mostly a mish mash of fairly mundane tasks.

> Driving doesn’t require an advanced degree in physics. It’s mostly a mish mash of fairly mundane tasks.

Generally not wrong, but when it becomes non-mundane, it can become so in a hurry.

This is why I'm sometimes skeptical about statistics that say "x% of airplane crashes are human error". Well maybe, but by the time the humans get involved the automated systems have often thrown up their proverbial hands and digitally exclaimed "Jesus take the wheel".

I think having computers take over the mundane parts may help collision rates as people can get numbed and bored by the monotony.

One has to be careful about having the right level of automation and control for the given situation, as per the famous "Children of the Magenta Line" presentation:

* https://www.youtube.com/watch?v=5ESJH1NLMLs

> “Children of the magenta line” has become shorthand for pilots who manage and monitor systems but lack the stick-and-rudder skills to fly the airplane with authority. You may hear it wielded as an epithet in a discussion of “real pilots” versus “kids these days.” But in many cockpits today, automation illiteracy may be as dangerous as automation dependency. Vanderburgh’s insights shouldn’t drive us to reject automation altogether, but rather put it in its proper place in our pilot toolkits. Just as you should hand-fly regularly to keep your stick-and-rudder skills sharp, you should also practice using the technology available in the aircraft you fly.

* https://www.aopa.org/news-and-media/all-news/2023/march/flig...

* https://99percentinvisible.org/episode/children-of-the-magen...

* https://skybrary.aero/articles/cockpit-automation-advantages...

* https://www.computer.org/csdl/magazine/sp/2015/05/msp2015050...

See also _The Ironies of Automation_ by Lisanne Bainbridge [1]

[1]: https://ckrybus.com/static/papers/Bainbridge_1983_Automatica...

I definitely get numbed by the monotony even when manually driving. At the very least Openpilot beeps at me loudly if it detects I've fallen asleep.

I'm willing to believe that computers are better at lane keeping than humans are. That's not the only scenario, though: Unfortunately, driving can go from lane keeping to making a life-or-death decision in a second, and that's not enough time for a human hand-off.

A plane's autopilot can beep and bail out to a human with a few seconds for them to get oriented and take over; a car doesn't always have that luxury.

"Is this a person walking into my lane or just a shadow?" isn't a question that comes up in the air. Until a self-driving car can minimize the probability of that detection going wrong to below human thresholds, I wouldn't feel comfortable sitting in the driver (or maybe even passenger) seat of one.

We're already at a point where computer vision models are vastly superior than any human at detecting objects.

> Mercedes just released a level 3 car.

Which only activates on roads mapped extensively by Mercedes.

Reply on news.ycombinator.com