Hacker News Reader

Comment by Animats | Hacker News Reader

Driverless cars are coming along just fine. Waymo is making steady progress. The sensors are getting better. The problem comes from all those "fake it til you make it" startups, Uber and Tesla being the worst. Both have killed people.

This is mostly about sensors and geometry. Machine learning has a role, but only in target identification. That's how Waymo does it. The fake it til you make it crowd had the fantasy that you just hook up some cameras to a machine learning system, train it, and you have self driving. Doesn't work. Machine learning is way too dumb.

You can maybe identify "traffic light", "car", "pedestrian", and "deer" with machine learning. That's just used to guess what they're going to do next. It's not used to decide if they're an obstacle. Obstacle detection and avoidance is all sensors and geometry.

Also, "self driving car", "electric car", and "transportation as a service" are all independent. All will be available, but not from the same companies.

View on HN

> Driverless cars are coming along just fine.

I am curious why you believe this to be the case? Driverless cars have yet to even come close to being workable in adverse conditions, to my knowledge, please correct me if I am wrong.

> This is mostly about sensors and geometry.

And mostly about the limitations of sensors, and the limitation of available algorithms to overcome geometry. A algorithm a driverless car uses to overcome some geographical feature might expect something to be spherical, but in reality, it ends up being elliptical, only it is too late for the car to adjust, for example, or vice versa. There is literally hundreds of thousands of potential corner cases out there. The earth in itself is a good example: although many people try to do distance calculations with lat/long and some standard radius are usually off by n precision points because the earth isn't perfectly spherical, only approximately. Driverless cars have to be precise. A couple points off might spell the difference between safely driving in the lanes, and veering into a ditch, off a cliff, or into a median or obstacle.

Sensors will suffer some of the same problems (or different problems) as our biological sensors do. How will road signs, lines, and likewise be detected in adverse conditions? Snow? How will radar overcome natural geographical corner reflectors?

Driverless cars seem to becoming along just fine in perfect conditions.

Not trying to be defeatist, but I believe if we want true driverless technology, the infrastructure will have to support the vehicles for these edge cases that we cannot overcome. Or we can continue to be idealists. I for one wish we never gave up on trains -- they are the perfect "driverless" technology candidate.

The infrastructure driverless vehicles rely on is their own internal maps, which have lanes and signage marked out. Being able to visually detect these things is important, but it is a measure for added redundancy.

Great strides have been made in using machine learning to filter out obscurants such as snow. Perception for autonomous vehicles is effectively a solved problem.

The biggest technical challenges are related to planning. So say you're approaching an intersection. There is a pedestrian about to cross, a cyclist in front of you and another vehicle waiting to turn left across your path. As a human, you understand that if you behave one way, it will cause the pedestrian, the cyclist and the other car to respond a certain way, but if you respond to the situation another way, it will cause all three to respond differently.

Our ability to game out scenarios like this is intuitive, but for AI to predict how it's behavior as an agent will influence the behaviour of other agents on the road is a daunting undertaking, particularly when taking into account the full scope of scenarios that need to be mastered before an autonomous can reliably safely navigate anywhere.

> Great strides have been made in using machine learning to filter out obscurants such as snow. Perception for autonomous vehicles is effectively a solved problem.

Great strides is not 'perfected' and I don't think perception for autonomous vehicles is a solved problem whatsoever. What is a driverless car to do if it slides out of position on the road and is now facing the wrong direction? Does it have the ability to know whether it needs to call for help, can safely maneuver, it's occupants are in immediate danger and should exit the vehicle (or not exit the vehicle)?

The two key aspects of an autonomous OS are perception and planning. Being able to "see" well enough during a flurry is a perception problem. Knowing whether to maneuver out of it or call for help after spinning out is a planning problem. There are many planning problems left to address and master before autonomous vehicles are ready for widespread adoption. Some planning problems are unique to specific intersections, and dedicated software has to be written just for that intersection, like the intersection of Market and Castro, in SF, for instance.

Can they perceive where on the road they should be located when the road is totally covered in snow and you can't see the pavement, let alone the line markets, possibly only tire tracks, if that?

Yes. An autonomous vehicle with lidar and maps can locate itself using any landmark, or series of landmarks it can see, it isn't limited to lane markings.

And if there are no landmarks, such as you're on a highway or rural area, not in a city?

Not a ditch, or a barbwire fence, or a telephone pole or a tree? I guess that may be tricky.

You can't reliably know the distance from pavement to any of those things unless you plan on indexing every physical object along every roadway.

I think the most likely scenario is self driving cars won't drive in those conditions, hopefully those conditions don't develop while you're on the roadways or you'll be parked somewhere cold.

They do index everything, the maps they rely on are lidar generated 3D point cloud fields that make a model of everything around the road, and are frequently updated.

Just think, the remains of the last self-driving car will help the next self-driving car avoid the trap. But only if Google carefully catalogs it and no one moves the wreckage.

Is "Market and Castro team" going to stay in existence indefinitely as the city of San Francisco modifies this intersection into the next century?

Will similar team exist for weird intersections in the middle of nowhere as well?

Will all these teams solve the problem of all the truck drivers rendered unemployed by "self-driving" vehicles?

Will all America's vehicles just wind-up instead being driven by tech support using joy sticks when the whole system goes south?

> Great strides is not 'perfected'

The goal isn't perfection. The goal is to be as good as humans.

Eg: An autonomous vehicle driving through snow backed by millions of miles of experience driving in snow is better than someone from the tropics driving through snow for the first time.

>> The goal is to be as good as humans.

Which particular humans?

>> An autonomous vehicle driving through snow [..] is better than someone from the tropics driving through snow for the first time

Better than the worst human we can imagine? Better than the average human? Better than the average human who drives on that road at that time of year?

The very best limo driver seems like a good human to emulate.

Autonomous vehicles (rightly so) will be held to a higher standard than humans

> Perception for autonomous vehicles is effectively a solved problem.

The problem is that you can't separate "perception" from "cognition". A human could drive a car reasonably well via a webcam feed, so does a webcam count as "solved perception"?

"Perception" in the context of autonomous agents means a mapping from sensors to a model of the world; how effectively can the agent establish a reliable model given its sensor readings?

What GP means by perception being a solved problem is that existing sensors and algorithms produce high quality and reliable models in a wide range of environment conditions.

A webcam is a sensor device, the human observing the webcam feed most likely cannot construct a good model of the car and its surrounds, as they can't do things like shoulder checking or looking in the rear view mirror, which a good driver does frequently to maintain a model of where other cars are.

But then "reasonably well" for a human driving a car remotely via webcam feed is likely to be a far lower standard than we're holding self-driving cars to.

As a human, you understand that if you behave one way, it will cause the pedestrian, the cyclist and the other car to respond a certain way, but if you respond to the situation another way, it will cause all three to respond differently.

Why does the self driving car have to restrict itself to extremely subtle human nonverbal communication? Just put a loudspeaker on the outside of the car and have it announce its intensions clearly, yielding to pedestrians/bicycles according to the laws of the land.

How about a man carrying a lantern ten paces ahead of it?

I think interesting edge cases include: 'empty cardboard box or log?' 'Deer just crossed road, is there another one close behind?' 'Driving temporarily on the wrong side of the street due to construction or a downed tree limb' 'Police directing traffic' 'Driving on wrong side of road to avoid collision' 'One lane bridges' 'Busy parking garage with blind corners' 'City Bus driver forces themselves recklessly into your lane' 'Other driver in round about about to cut you off' 'Pulling out to cross a busy 4 line highway with no median or traffic signal'

Not to mention, any snow on the road is going to make it supremely hard for a computer to determine where their lane is, or even the road is.

Things like snow aren’t necessarily a problem in the beginning because you can start by selling a self driving car that doesn’t work in the snow. There’s still a market for it.

People only have to do real driving in the worst conditions? That sounds like it wouldn't work out too well for them.

No, there are entire countries that don't have snow.

You forgot ‘A woman in an electric wheelchair chasing a duck with a broom’ and ‘people playing frogger on roads’. And Waymo self driving car didn’t seem to have a problem with that.

Waymo can of course cherry pick the scenarios which they present about. Meanwhile a human driver still has to intervene every few thousands miles.

These are tough issues, but they're probably not all relevant in the suburban Phoenix area where Waymo is getting started. Maybe it rolls out one city at a time for many years and that's okay?

"There's a ball in the middle of the road, is a kid going to pop up from behind a parked car and run towards it?"

I've seen this multiple times as a situation that people can deal with instinctively but computers won't. However people need to be trained to handle this situation; it is explicitly mentioned in the driver's handbook.† So I don't see this as a "win" for people over computers.

† CA driver's handbook page 37: http://driveca.org/cms/assets/uploads/2014/09/dl600.pdf

"There's no ball in the middle of the road, a kid pops up from behind a car, you have 0.1 seconds to apply the brakes before you hit the kid."

It's just as easy to come up with any sort of contrived situation for which a machine would outperform a human.

Except us humans drive in the middle of the street in residential areas often below the posted speed limit just for this reason. And if we live on this street, we can better anticipate the conditions.

While I am sceptical about the hype about autonomous vehicles I fully expect them to be equally good at that long before they go mainstream.

Why?

- They aren't as easily distracted as humans,

- you can program them to stick to the rules even if slowing down results in problems for the passengers

- you can program them to care about any neighborhood, not just the ones we live in

'Series of malicious actors dressed as police directing car into a shipping container to kidnap occupants'

Malicious actors posing as police officers pose just as much risk to human-driven cars.

Humans might be used to that situation depending on where they live.

Other edge cases include: Amish buggies, rotaries in Boston, Syracuse's upside down traffic light (green is on top), etc. The possibilities are endless; how can you test them all? Humans are really good about adapting to novel situations on the fly, not so much computers.

I mean, you are right about humans as a species, but oh god, not as drivers. As a cyclist commuter, pass/get passed by a lot of cars and see a lot of irregular/irrational behavior. Everytime there is a change in the traffic rules along my route (detour, construction, closed lane, new signalling) you will get a pretty high rate of people who either ignore the change, panic and drive recklessly, or panic and come to a confused stop.

Maybe we need driverless cars that can show panic faces when they don't know what to do and be coached by a friendly citizen what to do? It would be super cute.

Agree. It's a fun game to think up edge cases that would challenge a self-driving car. On average these cars will far outperform human drivers. I see human drivers blow through stop signs every day completely unaware that they did so.

Didn't Uber car did exactly that? Perfect robot would outperform average human. Sure. Average robot to average human? That's much more interesting...

An uber car also killed someone and the emergency automatic braking system was disabled.

Chernikov tail lights

Zoox has hinted that their vehicles will use sounds (at least) to signal intent and reactions.

One approach would be to record 100-to-1000 trials of how human drivers navigate each particular edge case in vehicles instrumented for autonomous driving.

Sadly, no one seems to have figured out a way to economically motivate human drivers to do this.

The problem are not the expected edge cases, but the unexpected ones. No amount of hard coded reactions are going to do for those.

Isn't this part of what Tesla has been doing though? At least they sell cars that simultaneously has:

- human drivers in demanding situations

- sensor and compute package planned for autonomous driving

From there it is mostly a question of

1. getting permissions to collect the data

2. finding some efficient way to store -> sift for interesting situations -> magic happens here -> lots of regression testing -> self driving car -> profit!

I would think?

I think we need to keep our expectations realistic. Most of those would challenge a significant portion of human drivers (and the scenarios suggested by sibling commenters are even more likely to cause a human to crash), and it's perfectly acceptable for a human driver to say "X crazy thing happened causing me to crash. What the fk?"

How do humans handle edge cases anyways?

I think it comes down to being able to model the world around us, having a model for human behavior, animal behavior, etc. If we see a baby deer, we look for the mother deer.

It feels like we're also really good at creating multiple open-ended stimulations in the back of our mind about how things might play out, where they can easily be paused or fleshed out depending on which became more probable or if we want to learn something. It's kind of a running background process.

> How do humans handle edge cases anyways?

A minimum 14 years of "human learning".

Or, judging by traffic's constant deaths ... Their don't.

Anyone who's seen a Waymo car drive around the valley knows they're much safer than human drivers. They're annoyingly good, because for instance they just never miss a bike and are careful around them, and if you can't tell why it's a bit annoying.

But I've always come away from such incidents with a feeling of "I should have seen that". I've yet to see them screw up even once.

And yet you can't drive from SF to San Jose without seeing a human driver screw up.

> And yet you can't drive from SF to San Jose without seeing a human driver screw up.

I don't think anyone is debating that in perfect conditions the computer cars can sometimes out perform the worst drivers.

If your concern is terrible drivers killing themselves and others, what you should be arguing for is much stricter standards for obtaining a driver's license. Probably removing 'full coverage' insurance and only allowing liability would go a long way as well. Can't keep wrecking your vehicle if nobody's buying you a new one.

>If your concern is terrible drivers killing themselves and others, what you should be arguing for is much stricter standards for obtaining a driver's license.

Or for keeping a drivers license.

>Probably removing 'full coverage' insurance and only allowing liability would go a long way as well. Can't keep wrecking your vehicle if nobody's buying you a new one.

I disagree... a used car is pretty cheap. Cheaper than full coverage, if you have a bad record.

I think this should be attacked by increasing the minimum liability coverage. right now, in California, you can drive with homeopathic levels of liability insurance... $35K isn't going to cover very much hospital time/time off work if you hit someone who has a good job. The minimum liability coverage should be 2-3 orders of magnitude higher.

Of course, for either of these things to be politically possible, we first need a society where it's reasonable to get from A to B without a car. An idea that is fought against daily even in relatively dense places like where I am now.

"Constant deaths" = 1 per ~100 million miles driven

Self-driving cars might be able to beat that, but most of the self-driving car models aren't even close to 100 million miles driven.

https://en.wikipedia.org/wiki/Transportation_safety_in_the_U...

(from page https://en.wikipedia.org/wiki/Transportation_safety_in_the_U... )

Furthermore self-driving cars are at this point ONLY driving through areas with a lot of pedestrians. You'd expect them, therefore, to have much higher accident fatalities than normal vehicles. Why ? They're always around pedestrians. Instead, in the first years of operation, they're safer than humans.

Even with Uber's less than stellar safety practices, and Tesla's ... well ... how do I put this politely ? [1] seems a good link. Let's just say Tesla owners could be more careful, especially since they signed a contract stating that they would be more careful (and youtube has much worse than that video). Given that that's how people use self-driving cars I would argue that's pretty damning for human driving skills.

[1] https://www.youtube.com/watch?v=JT5-Z2MswgU

Not sure what your point is. That graph shows about 10 deaths per billion VMT, which is exactly the number I stated. It is a completely unfounded assertion that they are safer than humans.

> How do humans handle edge cases anyways?

30,000 deaths per year suggest that humans fail at these edge cases all the time. Maybe ate their best a human outperforms a machine but as any person who drives if other drivers are at their best.

Divide by millions. Statistically quite low failure.

Exactly, these statistics don't represent the 'average' driver, let alone a 'good' driver.

I think it's still close to the most likely cause of violent death for an American; Neck and neck with gun deaths, if you count the self-inflicted.

I mean yeah, compared to medieval statistics, it's great, but in comparison to the other non-medical ways to die you face as a 21st century American? automotive fatalities are quite high, and the problem deserves a lot of focus.

Best way to reduce traffic deaths is to reduce traffic. That requires changing of zoning so people can live near where they work.

You could also make it harder to be qualified to drive.

I view autonomous cars as the 'lazy technocratic' solution to a problem caused by total government mismanagement.

You say lazy, I say possible. potato, potato.

(I mean, I agree that reducing miles driven is also a good goal. I'm just saying, making safer cars is a lot easier than talking people into allowing higher density construction and cities built around transit and walking rather than around cars)

No, quite a lot of them die because of stupid stuff like: texting while driving, drink driving, huge ego.

Driverless cars have advantages over humans too. Faster reaction speed, constant 360 degree surveillance, lack of distraction, (theoretically) better control of brakes and steering to handle skids.

Waymo has demonstrated solutions for many of the edge cases you mention. This video is from 3 years ago: https://youtu.be/tiwVMrTLUWg?t=530

Regarding snow, it maybe a difficult problem, but if cars can't see road markings, neither can humans, so how do they do it? Sometimes by following a car in front. Though far from real autonomy, my Tesla doesn't need to see road markings for autopilot to work, it will fall back to following the car in front.

And.. Why does Waymo have to solve this inclement weather problem now? Surely it would be better to demonstrate safety in good weather, then light rain, then heavy rain, then light snow? They can just refuse to operate in bad weather unless they are confident about it.

Weather and trips being not completely predictable, the car would have to be able to pull over and stop if there was weather the car wasn't certified to handle, and if the human was asleep in the back seat or whatever and took a long time to get ready to take over.

I think I've seen at least a prototype Waymo car (it was exhibited in the Computer History Museum) that was designed without the steering wheel and pedals for a human to use. That wouldn't work so well given that the car might pull over and be unable/unwilling to drive anymore. (Maybe you can put an expected upper limit on how long the weather will remain bad ... Nah, storms can last for days, by which point a human who didn't bring food and water could be in bad shape.) I guess you could rely on being able to call AAA. Though a bad blizzard might be exactly the case where AAA would have a hard time reaching you. The human being able to take over eventually seems like an important feature. I guess that is the way all the cars in the field (that I've seen) work.

Does anyone happen to know if self-driving cars are practicing "pull over and stop" maneuvers? For some reason I've never heard of it. I think it is the only potentially reliable fallback mechanism if the car encounters, say, weird terrain it knows it can't handle. (I don't think you can assume the human will be able to respond within n seconds.)

> I think I've seen at least a prototype Waymo car (it was exhibited in the Computer History Museum) that was designed without the steering wheel and pedals for a human to use. That wouldn't work so well given that the car might pull over and be unable/unwilling to drive anymore.

It'd be fine - you just need to have some backup option. A small joystick or two behind some panel, or screen control, or perhaps you control the car with a phone app. It's like having an emergency spare tire - you just need a minimal driving UI that can be used in case of emergencies but users otherwise don't need to see or deal with. It can be inconvenient and speed-limited, since you don't expect to use it much.

Though if there's decent connectivity an obvious intermediate option is remote operation. Some human who is really good at driving might take over for a bit, driving your car from the comfort of their own home or office if you're not comfortable doing so yourself using the backup control pad.

I would add some more edge cases - driving a street in Delhi or Bangkok. Or navigating small Italian town where streets layout had been designed circa 12th century.

> Not to mention, any snow on the road is going to make it supremely hard for a computer to determine where their lane is, or even the road is.

HD mapping allows cars to know precisely where they are on the road even if they cannot see the lane markings below. In fact, georeferenced HD maps can pinpoint the car's location in the world down to less than 10cm of absolute precision (as opposed to relative precision)

> come close to being workable in adverse conditions

Given how people drive in the rain and snow in the Northeast, I would argue human drivers have not even come close to being workable in adverse conditions.

We drive in them anyway, despite often times magnitudes more risk of crashing, because we want to.

I see this fallacious argument made as a counterpoint time and time again and it is getting tiresome.

If you look at all the relative statistics, both IIHS and insurance, the drivers who practice safe driving techniques and do not drive distracted have something crazy like a 3-4 fold lower accident rate. Moreover, those drivers who have had an accident are much more likely to have another one. Have you ever met someone who has gone their entire life without an accident? They exist, I encourage you to speak to them about how they drive -- more often than not it isn't by...accident (sorry)

The other point I am making here -- the drivers who do account for reckless, endangering driving, or who are careless by some measure of frequency (not careless as a whole, but maybe they forgot to change to their winter tires and properly inflate them?) -- they tend to pay the price the most. Their insurance rates are higher, they get tickets and points, and eventually the worst ones aren't allowed to drive at all.

When the "worst" AI's that are responsible for some amount of crashes, who is going to pay the price? Whose license is going to get deducted? Whose insurance is going to go up? In the world of driverless ubiquity, who is going to be the check and balance? What self-selecting mechanisms are going to protect us? You think Waymo will give a shit?

There are other reasons this argument is fallacious, the most obvious one that driverless cars have not yet proven they are safer or can remain statistically significantly safer than safe human drivers. How do we know driverless cars cannot or won't be compromised? What if a foreign nation state deploys an attack that exploits a known weakness? What if a car company is infiltrated by a bad actor and uses bad training data? We were able to break into a country's nuclear reactor. Russians were able to break into our democracy. I imagine this would be a nice juicy target for a country that really doesn't like us. Ect. Ect. It is going to come with a whole host of it's own unique problems, so comparing them to current human driving problems (which have potential solutions that our government won't embark on -- different story for a different day) generally won't be fruitful.

>the drivers who practice safe driving techniques and do not drive distracted have something crazy like a 3-4 fold lower

That's quite the qualifier. Problem is I see at least 3 people with their eyes on their phone during every 30min commute to work ...so...

>When the "worst" AI's that are responsible for some amount of crashes, who is going to pay the price? Whose license is going to get deducted? Whose insurance is going to go up? In the world of driverless ubiquity, who is going to be the check and balance? What self-selecting mechanisms are going to protect us? You think Waymo will give a shit?

These are literally the least difficult issues around self driving cars.

> There are other reasons this argument is fallacious, the most obvious one that driverless cars have not yet proven they are safer or can remain statistically significantly safer than safe human drivers.

so we should stop trying?

> These are literally the least difficult issues around self driving cars.

You are either young, or extremely naive (or both) if you don't think these are going to be challenging problems.

> so we should stop trying?

Not sure why you thought that this begged that question. What I am saying is we should stop making arguments that compare driverless car habits, problems, and safety to the current habits, problems, and safety of human drivers. They are generally bad arguments. They won't bear any fruit. Statistics aren't people. Trying to sell a generation on driverless cars on some unproven projection that "they will be X times safer!" will fall on deaf ears to drivers who have made it a lifetime of safe driving.

> if you don't think these are going to be challenging problems.

are they show stoppers? c'mon assessing risk is old as the hills. There's an entire profession that is devoted to assessing and pricing in risk. With regard to liability, the courts will decide. It'll be messy but it'll get sorted out in time.

>Trying to sell a generation on driverless cars on some unproven projection that "they will be X times safer!" will fall on deaf ears to drivers who have made it a lifetime of safe driving.

people will get over their fear in time. People feared cars when first introduced too.

from https://timeline.com/forget-self-driving-car-anxiety-in-the-...:

"With all the anxiety around driverless cars lately, it’s worth remembering there was a time people worried about cars exactly because they had human drivers. In fact, it was the removal of the horses—the horseless carriage—that gave some people fits.

In the 1890s, the prospect of a person driving without the aid of a second intelligence was a real concern. A horse, or team of horses, acted as a crude form of cruise control and collision aversion.

In 1896 Alfred Sennett warned, “We should not overlook the fact that the driving of a horseless carriage calls for a larger amount of attention for he has not the advantage of the intelligence of the horse in shaping his path, and it is consequently incumbent upon him to be ever watchful of the course his vehicle is taking.”

> yet to even come close to being workable in adverse conditions, to my knowledge, please correct me if I am wrong.

There's some video here for:

>“The Yandex.Taxi autonomous car safely navigated the streets of Moscow after a recent snowstorm managing interactions with traffic, pedestrians, parked vehicles and other road hazards on snowy and icy streets,” https://www.theverge.com/2018/2/16/17020096/self-driving-car...

Presumably it's not perfect but it seems to be coming along. Snow and ice covered Moscow is more adverse than what I usually drive.

Fundamentally, if a human can drive a car with nothing but two relatively poor eyes with a pretty small field of vision set in a single location inside the vehicle, an AI can be trained to drive using the same inputs - any more sensors are a bonus.

The other thing people forget is the bar isn't that high: self driving vehicles don't have to be perfect, they just have to be better than humans are. All of the "whatabout" edge cases people proffer as examples of areas an AI would have trouble with, people have trouble with too. The difference is that once an AI learns to solve that edge case, it doesn't have to relearn going forward.

Aircraft are the perfect driverless technology candidate.

Also trains

Yeah and sailboats it’s and horses too!

anyone have any idea of how frequently horse-drawn carriages crashed in dense urban inner city environments 120 years ago normalized against the incidence of crashes for contemporary cars on the same streets at the same speeds?

Maybe this could form the basis for a kind Turing test for level 1-4 autonomous Vehicles.

I guess rockets are a kind of aircraft. That's working.

Having both flown planes and driven cars, they are definitely not.

Waymo already works in light rain, and they've been testing in snow.

https://www.engadget.com/2018/05/08/waymo-snow-navigation/

Also, the lack of support for harder environment isn't because they're impossible, but rather because they are focusing on getting it perfect in normal environments before expanding. The person said they are "coming along fine", not that they are done.

Also, the lack of support for harder environment isn't because they're impossible, but rather because they are focusing on getting it perfect in normal environments before expanding.

I beg to differ.

I claim that adverse environments are very hard, even pathological problems and placid driving in well laid out suburban road with generally well behaved drivers are merely hard problems.

The thing is that an engineering team working on a problem like self-driving in adverse conditions is always going to be making some progress. The main question is whether they are going to make the kind of progress that justifies the hefty expenses involved.

Improvements have been made in driverless cars, but that doesn’t mean they are perfect.

Much more improvement has yet to come before we get to being safe enough with level 5 autonomy.

But in the meanwhile, we can also still celebrate the intermediate stepswise improvements that have been made.

No one thinks that you just hook up ML to a camera and you're done. Literally everyone working on this only uses ML for identification. When that Model X crashed into a highway divider, it was because the lines on the road were very faint and there was a fork in the line that actually wasn't the _real_ lane line. The car followed that by mistake, but only because of the misidentification of the lane line.

How are more sensors going to tell the difference between a plastic bag and a cat? How does lidar tell you which car is parked on the side of the road and which one is in your path? Relying on a dozen sensors seems like the crutch to me. Nothing about our current world has been designed for lidar and radar. It was designed for binocular vision. Sure, if your only goal is to not hit things at low speed in perfect weather, lidar is great. But you can't get much further than that.

> It was designed for binocular vision

Does binocular vision even matter at such distances (i.e. beyond a few meters)? I thought it was only effective for relatively close objects.

Sure, as long as you have reasonably well-calibrated cameras and optics.

Back-of-the-envelope calculation: assuming 1920x1080 camera sensors with a 90-degree horizontal field of view, each pixel covers a visual angle of approximately 0.001 radians. If you have two cameras separated by 5 feet horizontally, and your measurements are accurate to within a pixel, then you can measure the distance to an object 500 feet away with roughly 10% accuracy. That's a reasonably conservative following distance at typical highway speeds.

And that's just from a single parallax measurement. By using other visual cues (e.g. apparent size of an object whose physical dimensions are known) and averaging measurements across multiple frames, you can get a more accurate measurement.

Ah, I meant for humans. As in, how much is depth-perception actually necessary for driving... Obviously, humans also use higher-level cues, which IMO are much more important.

Whoops, my bad for skimming the comments and overlooking the context of your question.

I found a page [1] that summarizes the level of visual acuity required to obtain a driver's license in each US state. It appears that all states will potentially issue a license to someone with vision in only one eye. Some states impose additional conditions, such as requiring a favorable report from an ophthalmologist, limiting the maximum speed, or restricting the driver to vehicles with outside rear-view mirrors (to compensate for the loss of peripheral vision on one side).

[1]: http://lowvision.preventblindness.org/daily-living-2/state-v...

As evidenced by my username, I can say confidently that you are correct. The depth perception gained from two eyes is only good at very close ranges and doesn’t play a big factor in most driving situations.

Well, my optical sensors(i.e. eyes) provide me with enough information to discriminate a cat from a bag, so I suppose with enough processing we should be able to do the same with a control system. Cars have an added advantage that their vision systems could work in lower light and non-optical wavelengths. I don't see any fundamental issues with getting self-driving cars working. It may not happen in 5 years, but it will happen.

Yes, but your machine learning algos are much much better.

An intelligence animal being able to do something != "with enough processing we should be able to do the same with a control system".

Why not?

One possible answer would be that we aren't intelligent enough to create such systems. I don't think that's true but its a possibility.

> we aren't intelligent enough to create such systems

Evolution suggests that you can, in fact, create systems that are more intelligent than what came before them.

Evolution achieves that in part by a callous disregard for individual lives. The intelligences that evolution has produced are built on an unimaginable pile of corpses. Even Uber doesn't seem quite as willing as mother nature to spend quite so many lives to achieve incremental improvements.

The tough part about this line of reasoning is you can't tease apart the discriminatory power of human knowledge from human vision. How good does our expert system need to be? How good do our sensors need to be? I'm not so sure this is an easy to question to resolve

Exactly. Children have decent hand-eye coordination as toddlers. By 3 or 4, they can drive toy cars; by 5 they're riding bikes. But it takes them another decade or more to have the knowledge and judgment needed to drive a car in traffic. And even then they're not great; teens have a crash rate something like 3x other drivers.

Because of immaturity, not cognitive or physical capability. Kids start racing go karts as young as five - and that's a far more difficult, requiring fast reflexes, precise car control, and situational awareness. It's a far more challenging task than an average commute.

Not really, racing carts is primarily about the ability to control the thing under very well defined conditions (read: much better than your average 3 lane highway with hundreds of other road users) with other road users with the exact same end goal.

Exactly. "Immaturity" includes a big lack of detailed domain knowledge and domain-specific judgment, which is one of the things I think is going to be very hard to teach to computers.

I think domain specific judgement is what computers excel than human.

What human has advantage of is utilizing knowledge in unrelated domain to another domain.

Depends on what you mean by "domain", I suppose. But driving is a very broad domain. One has to understand human custom, human psychology, and human ethics, for example. As well as the behavior of wildlife, some applied meteorology, and a fair bit of understanding of both law and police practice. To a computer, these might be unrelated domains, but they aren't if you put a computer behind the wheel.

If the Model X in your example had LIDAR, it would have been able to build up a decently accurate collision model of the world, which it could have then used to say "there is an object in the path I'm on, maybe I shouldn't plow into it." As opposed to relying on implicit and apparently unreliable cues like lane markers. That particular fatal crash seems like a sensors and geometry problem.

"Nothing about our current world has been designed for lidar and radar. It was designed for binocular vision."

Really? Was the world designed for binocular vision, or is it just three dimensional?

Having more sensors, especially when they have different failure modes seems like the only possible way to create a reliable system. LIDAR isn't super dense, but generally has accurate returns. Binocular vision sucks on untextured objects, like the side of a white truck in the fatal Model S collision. Why wouldn't you want the crutch of both types of measurement?

Was the world designed for binocular vision

Ermm, actually yes, in the sense that anything human made was designed for and by those with binocular vision themselves.

> If the Model X in your example had LIDAR, it would have been able to build up a decently accurate collision model of the world

No it wouldn't. The whole problem with avoiding stationary objects is that they are _everywhere_. Do you really think radar and vision didn't "see" the barrier? Should it have known that it was in it's path? For sure. But that has nothing to do with sensors. Stationary objects are in our paths constantly while we drive, but we tend not to hit them because they are usually only temporarily so. At 80 mph the difference between a parked car being in our path and not, at the stopping distance of over 300 feet, is only a degree or two of wheel turn. We have enough information. Acting on it appropriately every single second of operation is the problem.

> Really? Was the world designed for binocular vision...

Yes. That's why we have road signs, stripes on the road, reflectors, lights on cars, lights hung over the road which light up in different colors to indicate right of way, etc. LIDAR is useless with all of the most important signals on our roads.

> How are more sensors going to tell the difference between a plastic bag and a cat?

Infrared transparency? Radar cross section? There are so many ways. Sensor doesn't just mean "webcam".

Lidar gives you high precision depth, so you can tell whether or not the car is parked or in the lane.

What do you mean "no one"? Press coverage of Waymo is full of references to engineers who are teaching cars and cars that are learning. The AI car is a big part of the hype.

>You can maybe identify "traffic light", "car", "pedestrian", and "deer" with machine learning. That's just used to guess what they're going to do next. It's not used to decide if they're an obstacle. Obstacle detection and avoidance is all sensors and geometry.

Honestly, getting them to work to this basic level is the "easy" part. Real world driving conditions in urban environments are suboptimal at best. You're not only going to need a system that can navigate around and not hit things. You need it to be able to dynamically route around road-construction, do clever things to find parking.

Roads in most places aren't that well standardized. It takes a lot of small judgement calls for a taxi to actually do its job. Uber and Lyft still have trouble dispatching people to pick you up at your actual door!

Roads are currently designed for human drivers. Roads designed for automated vehicles could use different signalling systems for intersections and road works, and even parking (it would help if unattended vehicles could be moved at will for tighter packing).

That would be great but in reality we're going to need self driving cars which are near enough perfect with the existing infrastructure to get some major traction before roads start getting designed for them.

>Roads are currently designed for human drivers. Roads designed for automated vehicles could use different signalling systems for intersections and road works, and even parking (it would help if unattended vehicles could be moved at will for tighter packing).

For the amount of money it would take to completely overhaul the country's entire transportation infrastructure to accommodate self-driving cars in this way, you could have built a phenomenal train and bus system with bike/scooter lanes everywhere.

So what exactly is the value add of cars then?

Coming along fine, sure, but isn't the point about reality versus hype?

The guy who started iRobot, Rodney Brooks, doesn't believe we'll see a true driverless car operating throughout a city on normal roads until 2035 at the earliest: http://rodneybrooks.com/my-dated-predictions/

As software developers, we all know how big the difference between "a demo the boss is excited about" and "reliably working in production for all users" can be. Given how complex the domain is and how many edge cases it has, I can easily believe Brooks is right here.

>until 2035 at the earliest

Could be correct. I think likely it is correct. BUT, everyone thought we still wouldn't be winning games of Go by now. So could be wrong.

I had a look at his predictions re: self-driving cars. He clearly invested quite a lot of thought went into these because he makes a lot of subtle distinctions. However, one of his predictions (probably the one that you're referring to) is this:

> A driverless "taxi" service in a major US city with arbitrary pick and drop off locations, even in a restricted geographical area.

> Not Earlier Than 2032

This is odd because isn't Waymo doing just that in Phoenix with a general release date of 2019? I checked the source because I wanted to verify that it says "a city" and not "any city".

I think his note clarifies it pretty well: "This is what Uber, Lyft, and conventional taxi services can do today."

What I've seen about Waymo (and maybe I've missed something) is very hazy. And it's not out yet. I'll be very surprised if they go from "operating in secret" to "completely competitive with Lyft" in one jump. My guess is that there will be significant limitations for whatever their first release is.

Waymo is available right now, and has been for almost a year, to normal non-employee citizens, in Phoenix, albeit behind a private invite system. And as mentioned above, it'll become an open beta by 2019. Also, most of the limitation at this point has to do with regulation, they'd be in many more places if not for that.

That is much much closer than 2035.

Where did you read 2019? All the sources that I can find say that the initial commercial service launch will be this year.

The question is, does it qualify as 100% driverless? I don't think so. Waymo service is partially teleoperated.

Do you have a proof for that? Waymo is actually the only car that explicitly does not allow teleoperating. They specifically made this decision to bypass the risk of being hacked and taken over.

The best a teleoperator can do is suggest a path to a car that is confused, and the car decides if the path is a good idea. We also don't know how often that's used. If it's rare enough, does it even matter? Does it not make it driverless if once in a while they need a little nudge?

That's like saying your car isn't a real car because 1% of the time you need to take it to the mechanic.

I think you're correct on the first part; "It does have humans in remote operations centers who can communicate with passengers and guide the car in complicated situations—like a cop sending cars the wrong way down a one-way street—but they never take actual control of the driving; it’s the car’s job to stay safe." https://www.wired.com/story/waymo-self-driving-car-service-p...

But yes, if a human has to get involved, I think it's correct to say it's not self-driving. In particular, what Wired describes is either SAE Level 3 or Level 4: https://en.wikipedia.org/wiki/Self-driving_car#Levels_of_dri...

Until it's Level 5, it's not really driverless.

Precisely, not Level 5. They are hiring for their remote operations through Adecco: https://www.adeccousa.com/jobs/self-driving-vehicle-operator... . This is a job where you sit in front of the display(s) with steering wheel in your hand and foot on the pedal. Remote driving - pure and simple. Not very scalable.

Yes, they are operating in secret right now, and are extremely hazy about what "operating" means. And they have said they'll be in open beta by 2019. But it wouldn't be the first time that a software project is late.

Google has promissed "release dates" as early as 2017 in the past. They have managed to keep exactly zero of those. Not only have they pushed back their deadlines again and again, but the scope of their promisses has also been reduced steadily. Let's wait and see what they can actually make work reliably next year.

> You can maybe identify "traffic light", "car", "pedestrian", and "deer" with machine learning.

What you wrote makes a lot of sense to me. However, I would also hope that in a matured self-driving environment, traffic lights would be mapped ahead of time. It's always possible that a traffic light shows up in an unexpected place, but in my mental image of the future, this raises a flag for system management to modify the map of permanent objects.

There is a chance that a powered non-permanent traffic light is found somewhere, a dude walking along the side of the street with a stoplight on his back, and connected batteries. In that case, the police need calling. In machine learning, there are just so many corner cases where best to not try to fully interpret, just call out "dude, look at this thing!"

> There is a chance that a powered non-permanent traffic light is found somewhere, a dude walking along the side of the street with a stoplight on his back

Temporary traffic lights are extremely common, at least here in Australia. They are used in the case of roadworks or other short-term disruptions that requires traffic flowing in both directions to share a single lane, in conditions not suited to manual traffic control (i.e. overnight, weekends, long-term works on rural roads).

http://www.datasigns.com.au/Products/Portable-Traffic-Lights

It's relatively common for road maintenance crews in rural areas to install temporary signal lights during major projects. For example when they close one lane on part of a two lane road so traffic can only move in one direction at a time while the drivers going the other direction have to wait a few minutes.

Ideally you would have a radio transmitter on the traffic light that broadcasts its location, metadata, and state information. You could still have a database, but it would be much more redundant to have live information.

Geometry inferred from range sensors can certainly provide an additional useful source of information, but it's far from clear that the right approach to controlling a self-driving vehicle is mostly to solve a geometry problem. For example, even if a range sensor can provide a perfect estimate of the location and geometry of pedestrian, how do you predict whether the pedestrian will dart out into the road or stop because they see a car coming. Human drivers have learned from years of (mostly) visual experience observing human behaviors to make reliable predictions about such things. Perhaps machines can too.

Since evolution didn't provide us with our personal lidars, human drivers are an existence proof that it is possible to learn reliable driving performance from vision alone. Which isn't to say that machine learning based on vision alone is already close to human-level or competency or that human-level competency is sufficient for a automated driving system or that sophisticated sensors don't make the problem a bit easier. But I don't know how you can be so confident that vision alone (or primarily) won't be route that eventually succeeds.

Hasn’t Tesla autopilot saved more lives than it took? So it’s more net beneficial (so far) than cautious Waymo?

[citation needed]

Quite probably. But that is irrelevant to the lawyers of the people that driverless cars did take. That is the real problem to me.

So did the surgeon who killed that poor girl because he was overconfident and took too many risks.

Is there a specific case you're referring to here?

How could it do that? The safety statistics quoted by Tesla is "fatalities per mile of Autopilot-enabled vehicle", compared to "fatalities per mile for the average car on the road". It's by no means an accurate comparison, in fact they're inflating their relative safety versus other cars by more than an order of magnitude.

Tangential to your point, do you think it's fair to compare AI driven cars to "average" drivers? I think it's more reasonable to compare them to "average [worst_category_of_driver]." For example, how does the Tesla car compare to teen drivers? If it's better than a meaningfully sized category of human drivers, it's probably ready for the road. Being on the real road is where the best data will be collected.

only if you are exclusively giving the Teslas to that smaller, worse, category of driver

The point is that we allow teen drivers on the road with the expectation and understanding that they will improve over several years. Why would we not expect to have to make the same concession for AI?

If a human drives sufficiently poorly they will eventually lose their right to drive on public roads for a time.

Could this/should this also apply for autonomous vehicles? If so, how?

Sure. I'm open to suggestions, but I'll offer financial liability for damages as an opening bid.

I suspect that the corporate entities involved have such deep pockets and/or so many lawyers and lobbyists that won't work.

Instead, how about:

All new autonomous vehicle configurations (let's call that the algos + sensors + vehicle) have to take some kind of actual driving test, just like us humans do.

Maybe the public could even help design a good test? "Not driving at speed into a stationary fire truck which is parked on the highway right in front of you" would be one element I'd want to see tested.

If an autonomous vehicle is involved in an accident, and the algo/sensors/vehicle are found to be (partially) at fault the configuration earns penalty points.

If that configuration earns enough penalty points over a period of time, the entire configuration loses its certification, plus a fine, plus a mandatory re-test.

This method appears to work reasonably well in dealing with us not-always-perfect human drivers, and ought to concentrate the minds of the designers/developers/managers behind autonomous vehicles.

Because they are companies that sell a product to consumers, who expect that this will keep them safe on the road. We need to hold them to a much higher standard than a teenager.

I feel like this position is too risk averse. Fear of a few hundred AI driver deaths will result in hundreds of thousands more human driver deaths.

This seems backwards to me. Shouldn’t we be comparing the performance self-driving cars to the performance of competent human drivers? The variance in human driving ability is quite high, which suggests (to me) that self-driving cars will become better than a significant fraction of human drivers well before they become (what I would consider) safe. I personally don’t want to see any more below-average drivers on the road. It’s true that there’s a constant influx of inexperienced human drivers, many of whom (in my opinion) shouldn’t have been granted licenses, but that’s a separate problem. (I’m talking about the US. I’ve read that licensing requirements are more stringent elsewhere, but here in the states we give out driver’s licenses like candy.)

I think you're confusing the end goal with the path towards that goal. Ultimately, yes, we will want to compare AI driven cars against the best human drivers. But right nwo we're trying to determine if AI driven cars should even be allowed on the road.

The underlying assumption I'm making is that AI driving improvement is accelerated by being on the real road. If that's true, then we want the cars on the road as soon as reasonable possible. Because I would gladly trade hundreds or even thousands of AI driver caused deaths in the short run if I am reasonably convinced that it will prevent the tens of thousands of human driver caused deaths every year. And I am convinced of that. I also acknowledge I'm in the pool of people who might be killed by the AI driver. Just as we let teen drivers on the road with the expectation they will improve over several years, so should we accept a similar risk from AI drivers. The return is far better for the "inexperienced" AI driver because the AIs will continually improve forever, but as you noted we get a new batch of bad human drivers every day.

That’s a fair point. From a utilitarian perspective, I think you’re probably right. Unfortunately, the public will not follow this dispassionate line of reasoning if AI drivers start killing people in “large” numbers, even if there’s strong evidence to support its soundness. It might therefore be best to be just a little less aggressive than would otherwise be optimal, to avoid a public backlash that could—despite being irrational—delay the arrival of competent AI drivers.

How would we get accurate information on the number of lives Tesla's "autopilot" has saved?

Humans die on average every 80 million miles driven ( not sure the metric for highway miles only ), but prove that Tesla can beat that metric and you can assume lives are being saved.

not so fast. youneed to compare tesla death per mile withother high end sedan death per mile. comparing a population of all cars include a lot of less secure cars, like city cars, old cars and cars that haven't enjoyed the same level of maintenance a driver with high income can afford.

iow tesla needs to beat the average 2014+ 3 series death per mile.

You also need to account for the regions where Teslas are most popular, the demographic of people who are purchasing Teslas, cars with similar safety features, and many other factors. Comparing the very tiny slice of the market which drives Teslas against every other car+driver on the road is almost meaningless.

In addition to the confounding variables others have mentioned, I think the most important factor is when autopilot is enabled. If you don't (or can't) use autopilot at night, in inclement weather, or in construction zones, then you'll get into more accidents when not using autopilot -- even if it's a significantly worse driver than you are.

How did you conclude this? There are other similarly-priced but more-numerous models of car where literally nobody has ever died after crashing one. Controlling for price (as a proxy for the wealth of the occupants) Tesla has a poor safety record.

Cite?

https://www.telegraph.co.uk/news/2018/04/15/britains-safest-... for example: "no drivers or passengers killed inside it in the 16 years it has been on sale", with 50k sold in the UK.

> no drivers or passengers killed inside it

To be fair, nobody has been killed inside any driverless car either.

Arguably, but it's hard to demonstrate the negative. At best, we can compare Tesla accident rates to those of other vehicles in the same class (luxury sedans). I'm sure you're right, but proving you're right against the naysayers is a different problem.

You're drinking some strong kool-aid.

I was in a statistics class where the Professor was teaching likelihood, then asked if there are cases when we don't want to maximize the likelihood ratio. The answer is when you want to put a limitation on false positive (or negative) detection.

You wouldn't vet a system that is susceptible to high false negatives and likely to kill people- no matter how good the accuracy is.

You have to compare it to other widely deployed driver assistance, not to the absence of Autopilot.

Sir, you have nailed it. No serious robotics project deployed to date does anything more than Good Sensors + Good Geometry. And I do not think a revolution is required, just good old time.

At some point the right balance of sensor quality and manufacturing processes will converge to be an affordable solution to highway driving. (broad daylight, then dusk, then maybe rain, etc etc).

If history is any guide (airlines anyone?) This will come along piecemeal, and the CEOs being forced to make ridiculous promises to appease shareholder anxiety will be given nice scapegoat packages along the way.

  Self-driving car: ZF
  Electric car: ZF
  TaaS: ZF
  Self-driving advertising services: ZF

All will be available, all from the same company. EDIT: Waymo laughed at "route-based advertising" when I mentioned, but ZF supports this idea. Route-based advertising or route advertising is an advertising model where your vehicle diverts to make you see a particular place and/or get you to visit a particular shop. For example, it may make you see McDonalds to get you to buy BigMac.

Silly question, but what do you mean when you type "ZF"?

https://en.m.wikipedia.org/wiki/ZF_Friedrichshafen

Zermelo–Fraenkel set theory?

That is extremely dystopian

It sounds like the touts who circle airport arrivals like vultures, looking for tourists to take for a ride...

Also prone to hijacking by terrorists.

My machine learning professor used to say that people tend to sensationalize self-driving car fatalities, without realizing that if every car on the road was self-driving, deaths would go close to zero.

I think the hard part of the problem is the infrastructure. Everyone wants to fix the cars, but they'll leave the hard part to the government- which is the development of autonomous-ready infrastructure.

I think we need to distinguish between assistive technologies, ADAS, self driving, autonomous, and driverless. They don't mean the same thing but are all interchanged.

For your professor, yes assistive technology like lane keeping assist and collision avoidance will save lives, but are no reason to go driverless.

I'm still taken aback by how quickly government (federal, state) allowed testing of driverless cars on public roads. Feels like a cold calculation made on behalf of citizens to say "some people probably need to die in order to bring this to market quickly, and that's ok with us in these areas."

The existence of cars and airplanes (even horses) is a similar trade-off. Some people have to die for us to even have them.

Ultimately the faster we bring driverless cars to fruition, the fewer people will die, simply because it's inevitable that they will quickly exceed the safety of humans.

Then we will have turned automobile safety into an engineering problem. It was partially that before, we could package the victim better to improve their survivability, but now we can modify the driver ... every driver.

> The fake it til you make it crowd had the fantasy that you just hook up some cameras to a machine learning system, train it, and you have self driving.

There are about 7 billion instances of those learning systems though.

Will this be replicated with machines in the next 10 years? I have no clue.

Is this doable at all? Sure.

There's absolutely no equivalence between a CNN created today with deep learning, compared to the Neural Network in a human brain.

The problem with the human brain is one of inattention. CNNs / Artificial Neural Networks can remain at attention 100% of the time due to their artificial / machine nature.

But CNNs, despite being at 100% attention the entire time, still have issues determining if that splotch on the screen is the sky or an 18-wheeler.

https://electrek.co/2016/07/01/understanding-fatal-tesla-acc...

--------------

Artificial Neural Networks / Convolutional Neural Networks have a very long way to go before they reach human equivalence. In contrast, sensor systems or LIDAR brute-forces the problem. LIDAR can see things human's can't see, and advanced sensors can tell you (at least, in clear conditions) the location and velocity of virtually every object around the car.

Fake-it till you make it camera-only driverless cars are clearly hype that relies upon a fundamental misunderstanding. Just because CNNs are kinda-sorta like the visual cortex of the human brain doesn't really mean that it works like one.

CNNs have really cool visual learning properties. But I've yet to see one 100% successfully tell you background vs foreground in pictures like a human brain can do. Even in clear weather conditions, the CNN can confuse a truck for the sky and still run full speed into an 18-wheeler.

I explicitly wrote my comment because @Animats said "machine learning system", not "CNN", which I know definitely not enough about to comment on.

Surely the field of "machine learning" includes things which are not even invented yet, the same way maths comprises of theorems not yet discovered.

I was just saying that I think one could "just hook up some cameras to a machine learning system, train it, and you have self driving", maybe not today though.

The 100% attention thing I why I have tremendous faith that computers will be out-driving humans in short order. Even really good human drivers have a limited perspective and focus. And even the best human drivers are subject to emotional state.

Yeah, humans suck at attention. But until we can program computers to tell me what is a foreground moving object, what is a shadow, and what is the background... you literally can't solve the "Should the car apply the brakes right now" problem.

https://arxiv.org/pdf/1801.02225.pdf

Just go through this recent paper I found on DDG.gg and see the amount of effort it takes to parse foreground / background data that a self-driving car needs.

You gotta figure out if there's a still-object on the road. And whether or not its a shadow (shadows don't move after-all but its safe to drive over). Like, neural networks can't do that stuff 100% reliably yet. And it may never happen.

Or some researcher next year might come out and discover a method to parse background / foreground / shadows out of pictures. But then there are a whole host of OTHER issues involved.

------------

I think self-driving technology has potential. But you need more than just neural networks hooked up to cameras. I really like Waymo's direction with advanced super-human sensors. Avoid the shadow / background / foreground problem entirely and just have LIDAR give you the precise coordinates of all objects within 100-feet of the car.

IMO, if self-driving technology ever happens, it will be because of advances in advanced LIDAR or other kinds of sensors. Stuff that can avoid the research-problems that the "Deep Learning" community hasn't been able to solve for the last 50 years.

That's why I think the future lies with a mix of (possibly layered) sensing and decision-making technologies. LIDAR + cameras + downloaded maps + machine learning + whatever.

The "until we can program computers to..." is a when question, not an if question. There's nothing about driving, in any situation, that doesn't fall in the face of "assuming infinite computing power and infinitely good sensors". Driving isn't a creative act, it's a responsive one.

I think that's reasonable, although our time estimates for when this problem solved may disagree.

My main issue is that a large number of people seem to think that cameras + machine learning are enough to solve this problem. And while I'm not an expert at machine learning, what I know about it makes me a pessimist. There's just too many unsolved problems in the machine-learning community to apply machine learning to the car-driving problem.

Machine learning probably can solve weird cases people don't expect. IIRC, CNNs are better at recognizing blurred or garbled text than humans these days. So CNNs can read speed limit signs, road signs, and other texts and and at least process that.

Even figuring out if its a speed limit sign, an address, or a route-number probably can be solved by CNNs. But higher level reasoning (is that spraypaint messing up the signpost?? Which was common in some of the areas I drove through) seems like an unsolved problem.

----------

Anyway, Machine Learning + Cameras are IMO, at best... a partial solution to some of the problems. Anyone who thinks that cameras + radar + machine learning is sufficient is probably just a TSLA long who wants to believe in the success of their stock. Otherwise, I think most people are reasonable and recognize the importance of experimenting with a ton of different methodologies to solve this problem.

Humans suck at these classification tasks things too, though, if you set your expectations high. Is it a cat or a bag, or a weird shadow? Doesn't matter, I ran it over before I could figure it out and react. Or maybe it turned out to be a boulder? Now I'm one of 35,000 traffic fatalities this year in this country.

Computer vision has been improving rapidly in the last 10 years, I think it's too soon to rule out the viability of a camera-based solution entirely. Though I do hope improved lidar technology can improve on humans.

Expectations for CNNs with Tesla's tech is currently "Does not run into firetrucks". And they're not exactly succeeding right now.

https://www.wired.com/story/tesla-autopilot-why-crash-radar/

Humans are way, way better than current CNNs on this field. We can talk about cats, shadows, and boulders when CNN-based methods stops crashing into concrete barriers, parked fire-trucks, and 18-wheelers making a left turn.

I don't want to dismiss the work of Deep Learning / Machine Learning specialists. I just want to point out that the problem is incredibly difficult. It is very far away from being a solved problem.

> Traffic-Aware Cruise Control cannot detect all objects and may not brake/decelerate for stationary vehicles, especially in situations when you are driving over 50 mph (80 km/h) and a vehicle you are following moves out of your driving path and a stationary vehicle or object is in front of you instead

This is a known issue, a known pattern and has happened multiple times this year. Its repeatable. CNNs today are not working in this case, and fixing it will require a research effort of mammoth proportions.

You're talking about what the tech does. I'm talking about what it can do, with reasonable-sounding non-magical improvements to tech. Just because there are limitations now does not mean those limitations will still exist in ten years.

Umm, Radar?

Comma.ai is doing vision+radar right now and driving cars with it.

The "neural network" terminology is really cringeworthy all things considered. ("Deep learning" is much better.)

I dunno. I think "auto-optimization" best describes the process.

Even calling it "learning" is kinda overselling itself. There really are only two camps of "Learning": auto-optimization (against a trained dataset), and auto-categorization (aka self-training).

Its auto-optimization: the algorithm self-corrects itself to try and look more like the training-set's "ground truth". Or auto-categorization, as the algorithm looks for patterns and tries to draw its own categories.

"Learning" implies model finding. Which... strangely enough... I'd argue that 3-SAT solvers are more "learning" based, at least with colloquial use of the word. Those things really do craft new theories and test them through the process of elimination / resolution / etc. etc. "Neural Networks" explicitly DON'T do this however.

Your standards are even higher than mine. I salute you for that.

Have you looked into Hubert Dreyfus yet?

> Both have killed people.

Are you referring to the midnight jaywalker in Uber's case? I still feel like most of the blame goes to the pedestrian.

Look at other people's videos of driving through that section of road around the same time - the pedestrian would have been clearly visible and the car and safety driver should both have easily seen them and stopped. If you read about the preliminary NTSB report [1], it's clear that the car had six seconds of warning and 1.3s of certainty of impact, yet did nothing to avoid it, not even alerting the safety driver. The pedestrian may have been foolish, but the car could easily have avoided killing them.

[1] https://www.ntsb.gov/news/press-releases/Pages/NR20180524.as...

I'm sorry I can't back up my claims but you will be proved wrong with time.

I would bet in less than 10 years you'll have camera only fully autonomous systems.

From Sterling Anderson, former director of Autopilot who oversaw the implementation of HW2:

>Perception is a game of statistics.Crudely speaking, if we have three independent modalities with epsilon miss-detection-rates and we combine them we can achieve an epsilon³ rate in perception. In practice, relatively orthogonal failure modes won’t achieve that level of benefit, however, an error every million miles can get boosted to an error every billion miles. It is extremely difficult to achieve this level of accuracy with a single modality alone.

>Different sensor modalities have different strengths and weaknesses; thus, incorporating multiple modalities drives orders of magnitude improvements in the reliability of the system. Cameras suffer from difficulty in low-light and high dynamic range scenarios; radars suffer from limited resolution and artifacts due to multi-path and doppler ambiguity; lidars “see” obscurants.

https://medium.com/aurora-blog/auroras-approach-to-developme...

First, thanks for sharing this! Second - based on the language used I can’t tell if he is saying Tesla’s camera and Radar approach will work, or that it won’t. The only thing that I pulled out was that if we increase the types of sensors we get better results because they have different strengths and weakenesses.

Sterling does believe cameras may someday be good enough, but I'm personally of the belief that Lidar will get cheap before vision gets good, and as such there's no good reason to not have the extra redundancy.

He is basically just saying that the errors need to be independent, so you can get that with different capture mechanisms. I think that is wrong. For example, small, fast and poorly reflective will have similar detection performance for lidar and cameras. I think that the real problem is getting consistent statistical estimators and fusing them at a useful update rate. Has anyone seen the lidar from the uber crash? It would be interesting to see if a person could "tell" that the poor pedestrian wasn't an artifact quickly from lidar alone.

The radar, Lidar and cameras on the Uber that struck Elaine Herzberg all detected her, it wasn't a sensor failure that led to her death, it was a software failure, the autonomous OS tuned to not respond to positives below a certain threshold.

Have you ever heard about that thing called fog?

As an existence proof, people drive through fog every day. We prove that it's possible to take binocular visual data and convert that into a series of control signals that will propel a car through fog. Sure sometimes accidents happen, but every day thousands upon thousands of people safely navigate fog in e.g. San Francisco or London. I see no reason machines shouldn't be able to perform at least as well one day.

> I see no reason machines shouldn't be able to perform at least as well one day.

Assuming that day is infinitely far away, then you're right.

Why? What's magical about human beings that a sufficiently advanced sensor and sufficiently advanced AI can't do the same responsive task?

If there's magic in the human mind at all, it's the creative side of the mind. Driving isn't a creative act.

Machines will be (and already are) held to a much higher standard than human drivers— there are a lot of reasons for this, ranging from the emotional to the pragmatic.

Just because thousands of people roll the dice every day driving in unsafe conditions does not mean that we should tolerate machines doing so.

One note is that perhaps people roll the dice because they want to feel some control and forward momentum. But if you're sitting in the car, being autonomously driven somewhere will you be so impatient?

I assume the car will give you an ETA, so you'll check the clock and go back to your book/social media. I'm okay with "slower, but smoother".

I think olivewell's comment in this subthread is what dooms the idea of camera-only self driving for a really long time horizon: you basically need general AI for it.

I work in machine learning in a different area (healthcare) so my perspective may be incomplete, but what I see in ML/AI is that models are really good at memorizing, not so good at understanding. What I mean by that is that a human navigates the world by knowing what a car is, what fog is, what a person is, what a plastic bag is, all sort of things. Object detection can get much of the way there, whether it can get enough of the way there is an open question, but ok we'll grant that. Whether AI/ML can actually make the step from knowing what the object is to knowing how that object interacts with all the other objects in the world is another.

In other words identifying an object is just step one, understanding what that object means in context is the next absolutely required step, and it's way harder. We have a leg up as humans so we can rely on visual alone, but machines almost certainly will need other sensors that report additional information because they don't "get" context in the same way. And I'd say it's an open question whether even with those sensors, they'll be able to get there. I hope so, but the game is far from won.

Just as a point of reference here - I'm not taking sides in this debate, I vividly remember reading about the 1999 Toronto 87 car pile up.

Fog on a straight road caused an 87 car pile up after a trailer jackknifed. The point in the article I remember at the time was a mother recounting watching her 14-year old daughter burn to death. Mom got out, but the daughter had been pinned by her leg.

The quote from the time was ""Mama Sheila, please don't let me die. I'm only 14," Marceya McLamore begged."

I suppose my point is that neither human or machine are good enough yet. and not to treat this as an abstract academic topic, it's real people.

https://en.wikipedia.org/wiki/1999_Highway_401_crash https://windsorstar.com/news/15-years-after-401-crash-trauma... https://www.theglobeandmail.com/news/national/inquest-hears-...

Cars could communicate together though

I do not understand how anyone thinks this will work in any scenario that includes a malicious actor.

Most self-driving efforts do not consider malicious actors and instead aim for sufficient caution to avoid danger. Malicious actors can never be entirely accounted for, as a driver on a narrow road can turn into your path in moments and you can't avoid the impact.

I do not understand how anyone thinks it would be the only thing the car would use

There is no upside to this approach. LIDAR is only getting cheaper, there is no reason not to use it.

I’ll take the “over” on that bet any day of the week. It might not be 20 years, but it’ll be a friggin’ while.

There are scenarios it’s possible to encounter on the road that humans don’t even necessarily know how to handle immediately. In proper, brutal, real-world testing, one very quickly enters the realm of problems that you’d need full-fledged general AI to reason through.

Reply on news.ycombinator.com