Hacker News Reader

Comment by irjustin | Hacker News Reader

Predicting is absolutely key. If you've been around the Waymo cars they horribly manage simple situations. They're VERY safe because they take no risk, but that won't work at scale. "What's that? Stop wait..., what's that? stop... wait wait wait"

They're very jerky at intersections where there's even a mild amount of uncertainty and really only go when it's very clear. It reminds me of lots of first time drivers reacting to new situations where safe is the best way forward. But if we had pure first time drivers, we'd be a gridlock at even a mild amount of traffic that enters and leaves the roadway.

Looking forward to this next set of 5 years. Hopefully us drivers will get replaced in 10? maybe?

View on HN

I actually prefer the Waymo approach over the Tesla and Uber approach. When there are lives on the line you simply do not cross certain lines and in the longer term Waymo will be the only player left whose reputation is still in one piece. That more than anything will decide whether self driving vehicles gain long term acceptance. Anything else will lead to a 'self driving winter'.

I’m not sure what brightlines exist in society for when lives are on the line. Lives are already on the line for human driving, so what is the moral story one would extract from the big picture of how people relate to driving today?

The only story I see is politics and complexity, as opposed to moral clarity.

It's quite simple, actually: a self driving vehicle should be a monotonous improvement on non-self driving. In other words, all situations that are currently handled by regular drivers should be handled just as well or better for self driving to be socially acceptable. A statistical argument based on a reduction of overall fatalities including a number of new kinds of accidents will not be accepted.

People (and the media) are going to focus on the accidents that would not have happened if self driving did not exist.

The problem with your statement is that some people view handling a situation better as getting more cars through a bottleneck in a certain amount of time, and other people view better as making sure that all humans in the area stay safe.

PR will never be about throughput but it will always be about safety. You'll never read 'due to Tesla's self driving abilities we can now push 3% more cars across the bridge'. What you will read (on page one, complete with 'shocking' picture) is 'Self driving Tesla crashes into bridge support'.

Which might mean that we are going to demand a safety level that's unattainable without decreasing throughput. In other words, human driving might be risky, but it's familiar, which apparently makes it tolerable.

> a self driving vehicle should be a monotonous improvement on non-self driving. In other words, all situations that are currently handled by regular drivers should be handled just as well or better for self driving to be socially acceptable.

I agree with this but I think it implies that people also will not accept safe self-driving cars that take forever to get anywhere in traffic. Especially since a decline in average traffic throughput makes things worse for everyone in a car, self-driving or not.

I think this makes the problem much harder, since a lot of people do demonstrably take risks in their driving now, in the interest of saving time (even knowing intellectually that lives are on the line).

Or, in the more optimistic case, SDV will come to areas where top speed is not a priority - think "bulk cargo".

Yeah, the place where self-driving will make the biggest impact first, is for long-haul truck driving.

If you slowed everyone’s traffic by 25%, how many human lives are wasted (though not ended) per year?

If you assume people drive 10K per year at an average speed of 33 mph, you’re driving ~300 hours per year. Slow that by 25% and it’s an extra 100 hours per year. If you assume 7.5 hours of sleep, you’d spend an extra 1.6% of your waking life in a car because of the slowdown. That’s killing a lot of time, even if not directly killing people.

The problem is that we might not ever get to monotonous improvement. We can probably easily get to a state where self-driving would decrease the number of traffic injuries and fatalities in total compared to human drivers.

But machines make different mistakes, mistakes that humans would never do, and that's psychologically not acceptable to most people. All the actual accidents will be "freak" accidents that a human could never cause, and that scares people a lot more than "human" accidents like a drunk driver running someone over at a pedestrian crossing.

Yes, you got it perfectly. That's exactly my fear. And Tesla isn't helping with their marketing because they suggest we are further than we really are.

Exactly. Accelerating and crashing into a barrier a human would easily avoid is not acceptable. https://www.forbes.com/sites/bradtempleton/2020/02/13/ntsb-r...

It psychologically like being a passenger in a car with a good driver who has a history of epileptic seizures.

Strange, because there are a LARGE number of driving fatalities caused by this exact behavior every year. The common explanations are: driver dozing or falling asleep, unexpected curve with poor visibility, driver mistaking the shoulder for another lane, and driver striking median to exit ramp by misjudging a lane change. Humans do all of these things, on a daily basis. So by human standards, accelerating and crashing into a barrier is quite commonplace.

That's exactly it.

If I get killed by a drunk driver, my wife will at least have someone to rage at.

If I get killed because my car randomly decides to drive itself into a high up truck and I basically died because of stupid technology. I would hope that it doesn't end up killing my wife too out of rage that she's going up against a major corporation now.

That's kind of maddening.

> A statistical argument based on a reduction of overall fatalities including a number of new kinds of accidents will not be accepted.

I think what would be accepted is if the self-driving car would have capabilities that surpass the best human driver (as opposed to the "average" driver). But we're a long way from that.

That makes me wonder: who is verifiably the best human driver? I recall reading about a truck driver who went over a million miles without being in an accident. Probably someone like that?

Yes, difficult to answer, so I wasn't thinking of an actual person (similarly to how the average driver doesn't exist as a person). If you want to imagine someone, then perhaps the person who drives the president?

That's a very different skillset from a normal driver; in fact, I think that in driving The Beast aka Stagecoach, road safety is an afterthought to a host of tactical considerations: the goal not being "safe driving", but "keep the President safe, to the exclusion of anything else." With 4 MPG, this is a plainclothes APC, not a normal car ;) https://en.wikipedia.org/wiki/Presidential_state_car_(United...

Interestingly enough, Waymo (back than Chauffeur) attempted the Tesla approach - and threw it out once they saw how people behave when the car "drives itself*".

W.r.t."self driving winter", I think Waymo are aiming to protecting themselves from that situation by developing close relationships with regulators, highlighting the "we're not them"-part and sticking to their safety-first approach.

My fear is that Tesla and Uber playing fast and loose with their development will lead to Waymo getting regulated as collateral damage.

Yes, that could indeed happen. But Waymo will likely be able to comply with such regulation whereas Tesla and Uber never will. The parties hurt hardest are those that would have played nice but that get nipped in the bud by the regulations before they are large enough to comply. So Waymo would benefit.

Tesla's getting enough hours under control at this point that they're likely to win the regulation battle for Uber simply by virtue of statistics, though. Obviously there has been the occasional tragedy (and there no doubt will be more), but the point where they launch the "Autopilot is provably safer than humans!" press release isn't too far out at this point.

Again, the measuring stick is real drivers. And real drivers are really bad at this stuff too.

That press release was sent out more than a year ago and it is about as fallacious as it gets and I'm surprised the whole thing keeps getting airtime.

For the comparison to hold you need to do the following:

- discount all the hours when humans are in adverse conditions and autopilot is not normally engaged

- remove motorcycles and other vehicles from the pool

After you do that Tesla is still much less safe than your average driver. Their whole spiel of taking the easy 80% in miles of driving and claiming victory has been debunked so often now it is getting boring. It's the remaining 20% where the fatalities are and which - coincidentally - are the hard bits.

since I let my Tesla 3 drive I have to say, what is out in public is damn shy about what it does and will slow down for curves and the traffic aware cruise control seems to have a second sense. On the open highway if you are navigating on autopilot you can let it change lanes on its own but I still like to control that because I think it doesn't do it as soon as it should; I usually use the must confirm mode. It does react well to people who cut in and my car has dodged someone merging into my lane. Going to be interesting if there is any change to driving when the computer gets upgraded next week.

frankly I don't think any of the teams who have demos are as close as they hint they are. there are just so many exceptions that it becomes maddening.

look, people road rage over the driving habits of others and the computer while it won't road rage has to react to those who infuriate others with their driving either because its reckless or just bad driving

some of my favorite conditions....

people stopping to let you out, how does the car know? people going out of order at an intersection?

* I do not let it drive where I think the situations are complex and obviously it does not have the ability to respond to traffic signals or make turns at intersections.

My belief is that we'll be replaced when vast swathes of roads are replaced with ones designed for automated vehicle transport, which is a fair way out of my lifetime I'd imagine.

As in...railroads?

If I could drive my car onto a railroad track in SF and then not drive again until I drove off the tracks in LA, that would be great.

Hyperloops :P

Speaking of, why are we talking about self-driving cars when trains and trams still have drivers, and those are much more constrained problem spaces?

Simply because it has been a control theory challenge. Weight on the wheels change, frictions change and drivers are expected to stop the train within inches on average, and a feet at worst. That requires some R&D to implement.

Yurikamome trams in Tokyo runs in L4-L5 and Tsukuba Express runs in L2, in SDC levels term. There was also a L4 PoC in Yamanote line so that’ll probably come by the end of this decade. Elevators had been L4 for like half a century as well, by the way.

because the cost of a train/bus driver is negligible compared to the train/bus and its riders.

The Parisian metro has a L4 line since 1980s...on an isolated track, without significant branching, controlled by a single computer from a command center. Yet emergencies are still escalated to humans, even things as common as "somebody is stuck in the door".

In other words, getting trains to L4 is straightforward, and there's been a rollout for half a century. Making the jump to "doesn't even need the driver to handle emergencies," that's a task comparable in scope to getting cars from L2 to L4.

It's too bad we can't rip the band-aid off and just go straight to full autonomy for every car using Waymo (or similar). The problem you identified becomes a non-issue if 1) it's the norm and 2) the passengers inside aren't paying enough attention to the road to care about the quirks. It's probable in my estimation that roads filled with only automated drivers become safer, more efficient, and more predictable, which acts as a flywheel for the prediction in the cars to get better.

Would be really interesting if Waymo could prove this out somehow.

People, bicycles, motorcycles, pets: all unpredictable on the road. What you suggest only works - barely - on motorways.

Waymo is at the point of being able to deal with those pretty well, right? Either way, I'm suggesting that the impatience OP experienced as a result of being a human driver proximal to a Waymo car would be obviated if all cars were driven by Waymo. If you're not driving the car that's behind the tentative Waymo car, do you really care? You're probably focused on the work you're doing in the car or the TV you're watching.

Sure, they say so, and have been for years! That obviously means they haven't been exaggerating then and aren't now; in that case, I have a very nice bridge here that might be of interest to you!

That boy chasing that ball out onto a crowded street has much better chances with an automated car vs one driven by a human. Heck, any of those situations human drivers perform very poorly in already, the bar is low.

There are two bars though, because there's a double standard. If the robot car hits the boy, it's national news. The lower bar of surpassing human safety performance is, in almost all regards, not even relevant to the discussion. Maybe it should be but it's definitely not.

I'm not sure about that. Looking at this system https://www.youtube.com/watch?v=hCWL0XF_f8Y it is disconcerting how people pop in and out of view.

A human could see a group of kids looking in the direction of the street at someone occluded from view and realize the possiblity of a child dashing out onto the street well in advance. The self-driving car would have to detect, classify and react after the child comes into view. It is this kind of common sense AI that we cannot engineer our way out.

A couple days ago there was a viral post about how self-driving cars react to a 35 mph sign vandalized to look like 85 mph. A human would know that the sign is wrong - the self driving car would need this 'use case' programmed in - maybe have location based speed limits programmed in.

I think you have way too much faith in human-based image tracking, and their ability to interpolate situations while listening to the radio. As for https://www.dailymail.co.uk/sciencetech/article-8021567/Tesl...

Reading the article:

> ‘It is worth noting that this is seemingly only possible on the first implementation of TACC when the driver double taps the lever, engaging TACC.’

It isn't a strict self driving situation, but makes for great click bait. I can't imagine that this isn't an easy fix (Tesla's are also easy to update collectively via software update, try doing that to human beings).

I have first-hand experience with such a case, and I disagree. I avoided hitting a kid that I think any AI would have killed.

I'm driving southbound on Fair Oaks in Sunnyvale, approaching Arques. On the corner is a mini strip mall with a small parking lot. Out of the corner of my eye, I see a young child, maybe 4 years old, running around that parking lot being chased by his mother or caretaker.

Now, being well off the road and out of its universe of conditions that would be considered by an AI, he would be ignored. But I just knew, somehow, that this was trouble.

The kid, thinking this is a fun, impromptu chase game, then sprints directly into traffic right in front of me. I stop with a foot to spare.

Edge conditions: it's what's for dinner. (tm)

Not a chance. One thing I found interesting in this article was the graph of successful object identification in images. The author made much of progressing from 50% to 88% in the last ten years (I believe this uncited data comes from [0]), but the great majority of that progress happened in the first five years, and the predictable slowdown of improvement is clearly evidenced by the graph.

We should also be very clear that 88% success rate given for object detection and identification is against a relatively fixed set of static images[1]. And as far as I can tell from the paper[2], there's no indication of how long it takes to come to a conclusion on any single image, and no description of how much hardware was used to reach this rate of success, which to be honest, I'm surprised is so low.

Given all of that, it does not seem reasonable to assume that this 88% success rate has anything to do with how well object detection works in real-time using on-board computing power while the car is in motion. So, don't start planning to throw out your driver's license anytime soon.

[0] https://paperswithcode.com/sota/image-classification-on-imag... [1] https://en.wikipedia.org/wiki/ImageNet [2] https://paperswithcode.com/paper/self-training-with-noisy-st...

Reminds me of when two people walking towards each other struggle with who is passing on the left vs right and go back and forth at the same time.

The best strategy for a resolution isn't to be accommodating. It's to physically shift your entire body to one side and make no eye contact to make it absolutely clear to the other person which side of them you'll be walking on.

That works well when one person does it - but if everyone had that strategy, then it wouldn't work either! (I've spent way too much time thinking about this).

That's basic game theory, and the game of Chicken.

https://en.wikipedia.org/wiki/Chicken_(game)

Look into how wifi transmission is solved among the many clients (although I'm sure its quite complex now). You might be able to use a random back off time to settle races like this.

Based on the actual rate of progress so far (not the misleading demos), we are a lot more than 10 years away from true level 5 autonomous vehicles.

I remember 3-4 years ago when we were told they were 2-3 years away. I think even 10 years is optimistic if we expect them to handle "edge cases" like driving in the rain or snow, dealing with mis-painted lane lines, and differentiating between a plastic shopping bag and a tire in the road.

Self driving has potential on finding the lanes (in rain, snow, mis paint) with infrared. The key component to driving in rain or snow is to slow down and increase following distance, something humans, but especially outliers, are relatively poor at.

Slowing down is also the best option for a tire tread or a shopping bag. The bag being less safe than you imply if it is filled with a box of nails or happens to wrap around a driveshaft or exhaust manifold.

A friend of mine with a Model 3 says it sometimes works better in rain than he does. Agree there are edge cases and probably 10 years is optimistic. But, of course, plastic shopping bags and tires on the road are a challenge for humans as well.

And Tesla recently released a _marketing video_ touting their identification system, that showed their car running a clear stop sign!

I live in a rainy area and I drive a Model 3. The other day it got into a high risk situation on autopilot, because it doesn't have any ability to consider rain intensity, big puddles, or rain-filled ruts. It is far more dangerous than an average driver.

And don't even get me started on the "automatic" windshield wipers.

Then perhaps we will get fusion-powered, self-driving cars in one go.

Flying. Don't forget flying.

I truly believe my children (not born yet) are going to buy their first car and it will be fully autonomous.

Too many people my age and older will be apprehensive about giving up control, despite it likely being safer. So I think it is fair to say that in 20 years there will be many people who learned to "drive" with a self driving car and their parents and grandparents will refuse to give up driving themselves.

Insurance and police enforcement are two reasons why the decision will be removed from law abiding citizens and won't happen linearly.

Once proven safe (or if a manufacturer directly insures the product) the cost to adequately insure a vehicle (as prices of what you hit continue to increase), especially an older vehicle, will increase. With income inequality where it is (and state minimums lagging) it has already become common to have underinsured drivers (and the upper middle class are already paying more through uninsured motorist protection).

When it comes to police enforcement, but especially jurisdictions that use traffic stops to find or create crime to fundraise; the expense of having traffic police in a world where > 50% of cars can do no wrong flips the model.

Finally, there is a large group of people who will not drive a car that's 7 years old or has over 100k miles; combining with less trained ability to do maintenance on more computerized cars. These people claim to be in control, but at one check engine light away from doing whatever a dealer tells them.

Insurance premiums are calculated based on loss rates. In some hypothetical future where level 5 autonomous vehicles are available, there is no reason to expect that the remaining human drivers will have more collisions than they do today. Rather the reverse, considering the increased availability of advanced driver assistance systems.

There's nothing wrong with driving a newer car and relying on the dealer service. Sure it's expensive, but if you can afford it then so what.

If you always drive a newer car and rely on a dealer for service, you no longer control what you drive when those dealers only sell self driving cars. (Or when the service model shifts as electrics have fewer service items and the service department closes)

The myth that electric cars require less service needs to die.

Maintenance needs in any modern car mostly comes back to value engineering. It's literally someone's job to optimize "how cheaply can we make this part, that it breaks just often enough that people won't complain".

The exact same thing is already happening for electric cars, and you'll need to spend exactly the same amount of time and money on service.

Furthermore, almost none of the service on a modern car is related to the combustion engine itself. It's mainly electronics, linkages, wheel beqrings, shocks, suspension, rust damage, tires etc.

If you think about it, a car that's done 200 000 miles at EOL has only been running for around 200 days continuously. A generator, or the engine on a ship, train, airplane etc. runs more than a full order of magnitude longer.

I drive an electric (LEAF) for a little over 5 years and also maintain my wife’s characteristically low maintenance car (Honda CR-V). Despite being very low maintenance, her car is much more maintenance intensive than my car. In five years, I’ve filled the washer fluid several times, changed the wiper blades, and plugged one flat tire. On her car, oil changes alone have taken more time, to say nothing of the brakes, power steering pump, valve adjustment, exhaust gasket change, and a handful of other jobs (all small).

In my case, the “myth” very much matches my experience.

ICE cars these days typically have electric power steering. And BEVs have brakes. I don't think ICE cars necessarily need valve adjustment - doesn't Honda use solid lifters where others use hydraulics?

It depends on who chooses automated vs self driving options. If primarily sports oriented people want to drive their cars manually, that is going to be one heck of a risk pool.

"there is no reason to expect that the remaining human drivers will have more collisions than they do today"

Seems obvious to me that the people who are economically forced by insurance rates into self-driving cars first would be the worst drivers. That means arithmetically that the remaining drivers will be better than average. So the rates for everybody should go down.

Hopefully they don’t buy any car.

Also sometimes you need to move your car to affect the environment, for example if your car is stopped at a parking lot type intersection, and there is a large crowd of people passing, they would keep crossing unless you start to move your car a little indicating, that’s it for you let some cars through

You mean like those drivers who aggressively creep into the intersection to scare pedestrians into letting them through, even when the pedestrian has the right of way? I can't wait for Waymo cars to get them off the road.

OP might be referring situations like a festival, where there's a trickle of cars and a lot of people. And usually cars have to slowly inch forward, some people will stop and let them through, some people won't. The driver of course has to stop, but also not just wait forever to have a clear lane.

Or, another example, a big parking lot with a few buses, people unloading their stuff from the buses (let's at the end of a long trip), relatives/parents/friends pick them up at the parking lot. Relatively dense, many people moving, but still cars come and go. Plus a lot of packs/luggage here and there. Usually in these cases more cars want to enter the parking lot to pick up others, but of course people want to continue unloading, so some kind of balance forms.

This also applies to "living streets": pedestrians are the primary users, but cars are allowed through at lower speeds (20 kph?). "Wait until everyone clears the path" means "forever", a typical livelock; the car needs to signal intent "I wish to move forward" while still yielding to pedestrians.

Years ago I was trying to get through a huge crowd leaving a concert once, because I worked nights in a building on the same block, and while I was edging around the corner, a cop on foot thought I was being too aggressive and started banging on my window hard. I rolled it down and explained to him where I was going and why, but later my power window mechanism failed and it might have been related. I wonder how that should be handled with a hypothetical self-driving car?

The AI takes instructions (orders? requests?) from the user. So if the cop asks the user to go back, the user should tell the car to go back, try a different route, etc.

What happens if the user is drunk? Well, naturally, the cop tells the car that it's a cop and just pick a different route.

And we shall see how well it will work.

Probably it'll take a (few?) decade(s) before a self-driving car should think it can safely wade through a crowd instead of going back and finding a good place to wait for user input :)

I knew it. I suspected the Waymo cars were literally driving on eggshells.

They’ve been peddling that these Level 5 Waymo cars were just around the corner. But they refused to disclose how the cars actually worked, and kept saying that their black box is so mysterious, that their scientists themselves couldn’t understand it, so there was no point in having the rest of us, mere mortals, try to understand it. Lies.

We are a long ways away from being able to trust a robotic car with our lives.

But to their credit, at least they realize that requiring a human to take over vehicle controls in a microsecond, is unrealistic. Ahem! Looking at another car company that thinks this is acceptable.

Of course we're 5 years from L5. Always have been, since the 1970 lane-keeping Mercedes. You just need to understand that this is a codeword for "we have no idea what interesting and hard problem pops up next week, this research is SO EXCITING!" If you are expecting actual L5, in production, five years from any given date...you're gonna have a bad time.

Whether or not Waymo has been sufficiently forthcoming about how their tech works or not is subjective.

Objectively though they have started performing rides with no drivers in AZ.

https://www.theverge.com/2019/12/9/21000085/waymo-fully-driv...

Waymo has only ever peddled level 4 vehicles.

Reply on news.ycombinator.com