Hacker News Reader

I have elderly relatives who won't drive at night, or on highways. I have friends who won't drive in various degrees of inclement weather - and I know a LOT of Californians who can't drive in the snow.

Human skill is extremely variable. A car that can drive to any road under any condition will be vastly superior to the average driver.

Reducing such a complex system to a single percentage may be silly, but the idea that a car might have partial coverage under certain conditions is quite reasonable and is reflective of how real human drivers work as well.

I think the GP post also points out that some roads are, frankly, terrible. I've had more than a few times where I wasn't clear if the cones were set in such a way that I could continue down the road or not and I've been wrong at least once (though the cop was very understanding after I explained it).

Some randomly placed cones are enough to confuse humans in many circumstances, as well. I doubt we'll ever have an all-encompassing solution for that, either. Even for things like autonomous-only lane, well it's not like everyone obeys the HOV lanes currently...

There probably isn't a perfect solution, especially when I've seen plenty of ambiguous markings.

Just yesterday I ran into some confusing cones and signage where I made the right merge onto the exit, but the car behind me made the wrong call and drove a kilometre or so down a newly built, unopened highway in parallel to me on my left until they got to some concrete barriers and realized their error.

And there's zero possible way self-driving cars are going to handle country roads that are sometime dirt in the snow during a blizzard.

Humans absolutely can do that. But the very architectural basis via ML that's being used never will. Different architecture? Maybe. Maybe in 50-100 years.

> But the very architectural basis via ML that's being used never will.

Why's that? Can't ML be trained you handle these conditions?

We already have self driving vehicles which can do this.

The average quality isn't the only thing. What about the correlation of errors?

There are situations humans can't handle, and notorious "hot spots" where crashes repeatedly happen, but if AI is more consistent, it might lead to more concentrated failures.

Have you heard about the problems caused by navigation routing, where large numbers of people are directed through roads that aren't suited for heavy traffic?

Lex interviewed Dmitri Dolgov CTO Waymo https://www.youtube.com/watch?v=P6prRXkI5HM I think you are conflating operating a public service within a very tight scope and "only been tested within a very tight scope"

What could "99% complete" even possibly mean? That it goes 99 miles/days/hours/minutes before hitting a tree?

I look at it this way: I haven't had a car accident in over 25 years of driving. Now if there was a 1% chance of having a car accident every time drove, I'd be having multiple accidents a year. This seems bad.

except 99% doesn't necessarily mean "every time you drive"; it's obvious to me that if you drive for 100 miles you are more likely to get into an accident than 10 miles. It's probably closer to "in 99% of situations".

The reality is that long term trips may, in general, be safer since you're on highways. Some statistics show that short term trips are more likely to have accidents.

See https://wtop.com/dc-transit/2020/09/study-25-percent-of-car-...

It doesn't necessarily have to scale linearly like that, but I assume 10 miles on the highway is safer than 100 miles on the highway.

Humans are particularly bad at highway driving long distances because they get highway hypnosis or fall alseep. On local roads on the other hand humans tend to do a good job reasoning with situations.

If this is true then why do 25% of car accidents happen in first 3 minutes of driving

Because manoeuvers are hard. And whatever the task a human takes, they need up to a few minutes to get "in the zone" and fully focused on it. This includes driving.

Highways without intersections are very safe. Only on/off ramps. Mixing traffic directions and intersections are risk factors which mean that non-highways are more risky per kilometer.

A steelman interpretation would for example be "works on 99% of north American roads" in which case preventing it from driving anywhere near known bad patches would get it complete enough.

In that case the problem would be 100% solved, not 99% with "the last 1% [being] the hardest."

> In that case the problem would be 100% solved

Not until you added the geofencing I wrote about :-)

> with "the last 1% [being] the hardest."

I might have misunderstood the context here, but my point was to make an example were the last percentage wasn't that hard.

The original headline of the article was:

Waymo Is 99% of the Way to Self-Driving Cars. The Last 1% Is the Hardest

It's clearly based on the HYPER-idealistic driving conditions of primarily high-angle sun with dreary-yet-predictable Phoenix "stroads".

They picked a good first experiments location but it's NOT representative of more than 1% of roads on the planet! It's a bizarro-world of simplified challenge and problem space. Drop such a system into any real world driving and it will fall flat on its face very quickly and endanger both passengers, other drivers and pedestrians.

> Specifically, there's an especially tricky road in I-35N in Texas

Why are we ok with streets in such slapdash condition in the first place?

Yes I think self driving cars need to handle it, but I also think we aren't handling it.

Self driving cars have the potential to greatly reduce deaths due to accidents and improve accessibility for often ignored segments of our population. It is sad to see it so close, but held back by reasons like this.

There was always where this was going to end -- demands that we spend (from the public) hundreds of billions to upgrade roads to make self-driving tech "work".

Also,

> and right now, no driverless car from any company can gracefully handle rain, sleet, or snow

Good thing the entire US has the weather patterns of the bay area, and never experiences rain, sleet, or snow!

Lots of people aren't okay with road conditions. But lots more people aren't okay with a myraid of other pet peeves and legitimate grievances they have with just about every single aspect of our society. The process of prioritizing which matters to address is called "politics", a process so notorious I assume I don't need to explain this further.

IH-35 has been under continuous construction since .. the 90's?

Hindsight being what it is .. they should have built a new 8-lane highway a few miles to the West rather than attempting to widen it multiple times (with all the bridge replacements needed) through towns like Temple and Waco. But as the Texas Central Railroad has found, you can't just eminent-domain land in Texas for large infrastructure projects.

I live about an hour west of Boston. There's a sort of half-ring six-lane highway called I-495. It mostly handles the traffic load fairly well. But it's a source of sometimes genuine surprise to me that a highway of this capacity for the area was built in the late 1950s given that must have seemed almost outrageously overbuilt for at least most of its route at the time.

Google Maps navigation can't even reliably call out which lane to use for turning in major metro areas, and that's kinda the basic best case scenario: clear images/footage, unlimited compute time, static road markers, common use case.

I'm not sure how they can do better with far more stringent resources and requirements.

Yes, it’s kind of a silly headline. A percentage compares two numbers, in the numerator and a denominator. If you don’t know what’s being counted or measured, it’s not a real statistic.

However, people do uses percentages metaphorically and we should try not to be so distracted by silly headlines.

That's exactly what this is. It's the classic "The first 90% is 90% of the work. The last 10% is the other 90%." Or whatever the exact quote is. It's not about the number. It's the fact that it's a lot of work to get close. It's as much or more work to get over the finish line.

My current work involves ground-truthing sensor and classifier hits from various automotive OEMs. Take, for example, reading old-fashioned speed limit signs. In principle, doing this correctly and reliably in context is a relatively simple machine learning problem, particularly when restricted to a single country.

I have yet to work with an automotive OEM that doesn't consistently mis-classify some speed limit signs in some common cases that are completely unambiguous. It leads one to wonder about the quality of classifier design and testing such that these end up in production vehicles. The cases that cause a classifier to fail are different across automotive OEMs too, it isn't intrinsic; other automotive OEMs will handle that specific case just fine.

On the other hand, when it works it often works surprisingly well. The good/bad news is that most of the problems I see are ultimately caused by poor model design and testing. These are all surmountable with better quality AI engineering.

There is more AI to self-driving cars than the classifiers that pull features out of their environmental sensors. However, the driving AIs are making decisions based on the output of those classifiers -- garbage in, garbage out.

These are the kinds of situations that will produce relentless incremental improvements over time, and there is still a lot of low-hanging fruit to improve on.

That stretch of road you describe as difficult is actually not difficult at all for Waymo's current technology. Their navigation does not depend on vision.

how does it know where to be in the lane if it doesn't use lane markings?

"stick to the right as far as possible" is not a reliable approach when the emergency lane may or may not be there.

not a gotcha question, I'm actually curious: how does a human driver know where to be in the lane?

Context. The faint wear marks that tires make on the road over years, the seam between lanes, the car in front of them, the distance away from the guardrail or ditch, or there are no marked lanes at all and you just have to share the road with oncoming traffic. Furthermore, it's not uncommon for three lane roads to turn into two lanes during heavy snow, and in those case, the lanes are wherever everyone decided they were, collectively.

I think this is a very good question!

The case I was thinking of is a highway I was in last week, where temporary work has one of the directions going through the other direction, so the 2+emergency lanes in one direction become 2 lanes without emergency in one direction and one lane in the opposite one.

In this case, there are temporary lane markings on the ground indicating you should not follow the old ones, but rather align yourself further to the right then you would normally do.

OTOH, I am familiar with the situation in Rome where often the lane markings are deleted and a road may be 2-3-4 lanes depending on how people queue up, in which case people would generally rely on social/experience cues, which I'm unable to formalize.

I'd be really curious to see how a map/gps based system handles situations where a divided highway is under construction and has temporary lines painted to direct the right side of the road to drive on the left side while left side traffic is held up by a man in a vest holding a sign. This isn't a particularly uncommon scenario, and if the road crews are hard at work, the zone where this is occurring can move by miles a day. To navigate this, a driver needs to rely on road markings, temporary signs, and humans directing traffic. Maps won't cut the mustard in a situation like this.

If a driver assistance system can't handle this sort of situation, that's fine. I'm sure it's still useful most of the time. But it's clearly not "99% of the way to FSD."

I assume that, once you get vehicles driving autonomously on sections of highways, you have some sort of beacon/alerting system that cause the vehicle to tell the driver that it's going to be dropping out of autonomous mode.

The "beacon/alerting system" are the signs, and the man in a vest. If the car can't respond to those, then the words "full self driving" should never be used to describe it.

On a highway, you often don't have a man in a vest and the signs I see are often vague alerts about construction and lane shifts ahead that may or may not be actually happening at the moment.

But I don't really disagree. Some mechanism (perhaps a redundant mechanism) when approaching a possible construction area to give a human a minute or so to take over.

I'd see self driving as not needing a person at the wheel. I don't care if there are changes to the road, it's being driven without a person nat the wheel

Autonomous driving systems do not need to be door-to-door. They also can't require taking over control on a moment's notice. But, systems that can fully autonomous for long stretches of highway driving seem useful.

They just don't address use cases where people aren't able to drive at all. Which are interesting to a lot of people but they may just not happen for a long time.

There are people who just want cheaper Ubers but there are others who are fine with handing off long boring sections of highway driving.

Waymo's system doesn't depend on GPS for navigation either. They use their "realtime sensor suite" (lidar) to localize the vehicle position within the map.

As part of that localization, there will be a metric of how good the local solution is vs. the map. If a divider moves, for example, the data will be inconsistent with the map, and the software will know that something is amiss, without having to intelligently classify anything. The software can drive confidently when in a consistent environment, and drive conservatively when in an inconsistent environment.

Generally, any object that isn't in the map will stand out obviously. That's where having stuff like cameras becomes useful. Asking a computer, "is there a construction worker in this 20MP image of a street?" is hard, but asking a computer, "is this 0.1MP human sized blob a construction worker?" is a lot easier. Waymo has been responding intelligently to bicyclist hand signals for years. "Is this sign shaped blob a construction sign?" is even easier.

The lidar and mapping is a key difference between Waymo and Tesla. Waymo works because they have the big data infrastructure, and the budget to deck out every car with orders of magnitude more sensors and compute.

You should come drive some of the mountain roads in CA. I wouldn't consider roads in TX any trickier than roads elsewhere.

This has always had me wondering about the viability of self driving: what's it going to be like in different countries?

Not only are there different signs, different line marking & colours, & different widths. Cars are different: not all models are sold equally throughout the world. People are different: through their clothing, height and build. But the really big difference is that not everyone drives on the RHS...

The article lists left turns as a problem, whereas on a LHS road system, right hand turns will be the problem.

In NZ (a LHS system), I've heard that turning right across traffic has priority over oncoming traffic.

In Melbourne, VIC Australia (also LHS), right turns are made from the left most lane because of trams lines.

I'm sure there are loads of other issues specific to other geographies.

To me, this never ending self driving experiment screams out that the wrong solution is being pursued. (That doesn't imply that I know what they should be using as a solution)

> In NZ (a LHS system), I've heard that turning right across traffic has priority over oncoming traffic.

This is incorrect and would be utterly chaotic. You may be thinking of an outdated rule in NZ where right-turning traffic takes precedence over left-turning coming in the opposite direction to the same side road. This was the case for a long time but was reversed[0] in 2012, overnight, almost without a hitch.

[0] https://www.nzta.govt.nz/driver-licences/getting-a-licence/r...

It sounds like a pretty good idea to me, and will ensure low speed, low traffic roads

I think responding to the human element is huge for safe driving. You would not want to safely drive the same way in LA as Boston, or Seattle. Other drivers make very different assumptions on your response to their (and others) actions, and that can create danger irrespective of navigation.

I e.g. think that many self driving car systems are good enough to drive on German highways (and kraftfahrstraßen ~ slower smaller highways).

But only IFF there is no external accident and no construction site.

If both would be handles truckers could sleep/nap while driving on the highways, which could have interesting effects tbh.

Handling construction sites could be made feasible.

But I have no idea how to handle accidents happening around/in-front of the truck in a good way.

Similar I would be worried about improvised marking of fresh accident sites before emergency responders arrive...

That awful stretch of I95 can be special cased. It's a great example of what a computer can be better at than humans who each have to individually figure it out.

An unprotected left turn is probably a harder problem. They'll be expected to do it aggressively like humans, and also expected to do it safely. Those are incompatible directives.

The (IMO) "right" way to do an unprotected left is also somewhat situational. If there isn't a lot of traffic generally on a country road but a somewhat bunched up group of vehicles is coming, I'll be patient until the bunch has cleared (unless someone flashes their lights at me). If on the other hand, I'm in a busier urban area and I can see there aren't a lot of openings, I'll be more aggressive including at traffic light changes.

In general, my expectation--which may be completely wrong--is that we'll see autonomous driving on sections of highways in good weather long before we have door to door almost everywhere. Which is useful by itself for long highway drives.

I almost feel that relatively short range (meaning they can make the hops between charging stations) self-flying passenger drones/planes would be better than cars. Leet air-traffic control systems with decades of anti collision experience interact with and track all flights and inform all flying systems of the telemtry data for all slights so they have spatial awareness that way. Set flight paths that are low, but avoid densly populated areas/neighborhoods/infrastructure etc.

Create small emergency landing areas all over...

The only ground based vehicles I care about being 99% auto would be long-haul trucking. With a station to pickup a local last-mile-human-driver for actual drop-offf etc.

> Specifically, there's an especially tricky road in I-35N in Texas

Seems like you could just get some data on that before declaring that "we're not even close". How do shipping Teslas do?

FWIW: my impression is that the things people think are hard aren't the problems that are actually hard. People here and elsewhere on the internet have spent years screaming about lidar vs. radar vs. vision. And... it turns out vision basically won. Something like half the Tesla fleet in ths US are vision only in the US now, and... the cars see stuff just fine. We're not having detection failures. It works.

If you watch all the FSD beta video being spread around youtube, much of it by drivers specifically looking for edge cases like this, most of the remaining problems are in pathing and decisionmaking, not detection. The cars see a busy intersection and then get paralyzed (there was an amusing clip from one guy of a car trying to make a left turn, giving up, and repathing a right turn that took it back in a loop only to fail at the same left turn repeatedly), or creep too slowly and annoy other drivers. Or they choose the wrong lane, or misread a multiple traffic light environment, or misread the need for detection (the one really dangerous video I remember had the car deciding to take a left turn across a broad street with a complicated blind corner instead of creeping out farther).

Those are bugs to fix, for sure. But they're not the bugs we thought we had to fix. Those bugs... got fixed.

So you can look at this as a pessimist and say "we'll never get there", or you can judge from previous experience that the hard problems won't turn out to be that hard in the end.

Personally, having been watching the progress in this space, my guess is that Tesla and Waymo (and maybe Mobileye -- no one else is close) get this basically fixed within a year or two.

At which point the debate will shift from "we'll never get there" to "we'll never be able to regulate this appropriately" or somesuch.

> Something like half the Tesla fleet in ths US are vision only in the US now, and... the cars see stuff just fine. We're not having detection failures. It works.

You and I have extremely different experiences with Teslas. About the only place I could conceive arguing it works is highways and even then I wouldn't trust riding one down the highway without a steering wheel. Hell, my coworker can't even call his out of the driveway because it has been thinking his mailbox is a car for the last 6 months.

I think we'll get there and I don't think it'll matter if it's with LIDAR or pure vision but we're certainly not close to being able to remove the steering wheel (regulation aside).

Mine drove me to Yellowstone and back with like two disengagements which were my paranoia about other traffic.

But yes: the example was an interstate in Texas. And I happen to know it actually works quite well on interstates. So a very reasonable experiment about whether or not automation "will never get there" on edge cases like the referenced interstate is to take a commonly deployed, already available automation technology there and test it.

But also recognize that shipping autopilot is now about a year old. All the FSD work hasn't made it to released cars yet. It's very much evolved, but visible only on youtube right now.

Summon uses the old, pre surround video NN. The FSD stack is a whole different beast.

  We're not having detection failures.

Then what do you call them hitting stationary things regularly?

Radar

A lie. Or at least a giant [citation needed]. What evidence exists seems to show (not really surprisingly) that the cars are safer with AP engaged than when not.

https://www.necn.com/news/national-international/us-agency-o...

https://www.foxnews.com/auto/tesla-smashes-overturned-truck-...

How often do people hit a giant stationary object in the middle of a nearly-empty freeway?

> What evidence exists seems to show (not really surprisingly) that the cars are safer with AP engaged than when not.

Good way to change goal posts. There are many, many instances of Teslas hitting large stationary objects on roads, and who knows more instances of near misses due to driver intervention.

When the AP can be engaged, which is usually 'better' conditions. Unfortunately humans don't always have the option of driving in 'better' conditions, so this is an entirely cherry picked subset of data.

Cmon. There's a video of a Tesla FSD driving on the wrong side of the road, cutting off a car it did not see at an intersection, then accelerating straight into a fence, all in the space of five minutes in Oakland. Tesla has by no means solved the vision problem. Their software is laughably bad.

https://youtu.be/antLneVlxcs?t=633

Ancient footage, they fixed this bug

It's sort of on you to back up the extraordinary claim that these defects have disappeared in the last five months.

Edit: FSD 9 accelerates across double lines toward bridge support in San Francisco. https://youtu.be/GlIdu7prsAw?t=155

Like it's just a forgotten semicolon which explains all this behaviour. This isn't 'a bug', it's a big, complex, systemic reliability problem.

For anyone wondering what "ancient" means - it was 5 months ago. Maybe the commenter has a different definition of ancient than me.

It's incredible to me that they let this thing on the road. It could not pass a driver's test!

Teslas are not self driving cars regardless of what Musk claims. They aren't even close. Even Tesla admits that their "full self driving" mode is only level 2 automation.

https://techcrunch.com/2021/05/07/tesla-refutes-elon-musks-t...

What do you call the video of a Tesla mistaking the moon for a yellow traffic light[1], if not a detection failure? That wouldn't happen with a lidar-only or lidar-assisted system.

[1] https://www.youtube.com/watch?v=7UF-S2czdCk

they said most, not all

I'm not sure how lidar would help in detecting yellow lights correctly. Seems to me like that has to be a vision system, at least until we get all traffic lights to report their state via radio.

I was thinking distance—if the object emitting the yellow light is 240K miles away from your car, it can be assumed that it isn’t a traffic signal, right?

"how far away is the light?"

This is a strange bug. As if they weren't making predictions about where the yellow light would have to be one second later, by using at least speed and accelerometer data, and compare the expected result with the new reality.

This is not necessarily a job for lidar, they are just omitting some checks.

At second 18 [1] a street light comes into play which looks similar to the moon. That relative movement is what the system should have been expecting from the moon, but they apparently just re-computed the assumption that the moon is a traffic light which is moving along with the car.

It's somewhat revealing that they are not doing these kind of checks. I actually can't believe that they aren't doing it, but I can't see a case where the traffic light would always be at the same spot, if you have velocity and acceleration data at hand.

A traffic light is a very important thing.

[1] https://youtu.be/7UF-S2czdCk?t=18

The traffic light detection runs on an older NN which works over individual image frames instead of the new video-based system. They are migrating towards the latter. In any case, they can ask the fleet for examples of this and train the NN against it.