Hacker News Reader

Comment by strebler | Hacker News Reader

You've hit the nail on the head - audio recognition is quite poor after (a lot of) concerted effort over decades. Driving a car safely is significantly more difficult (and dangerous).

That being said, Toyota's (and Subaru's) approach is the smart way forward - add sensors and capabilities that augment the human, but leave the human always responsible and in control. In order for a crash to happen, the human AND the machine BOTH have to miss it.

Once the machine is good enough (and we figure out what "good enough" means), then it can take over driving. But not before.

View on HN

That's the approach we know won't work. That was Tesla's first-round autopilot, the self-crashing car. The one that would happily run into obstacles protruding into a lane.[1][2][3] There are two big flaws with that approach. First, if it's good enough that people can tune out, they will tune out. Second, if the driver waits until the self-driving system has clearly made a mistake before taking over, they will be slower at reacting than if they were driving, and may be too late. See especially [3].

Volvo's CEO takes the position that if one of their cars crashes in self-driving mode, it's Volvo's fault. Urmson, while at Google, pointed out the first problem coming up even with their test drivers.

[1] https://www.youtube.com/watch?v=fc0yYJ8-Dyo [2] https://www.youtube.com/watch?v=DvRkSMCDX3o [3] https://www.youtube.com/watch?v=-2ml6sjk_8c

Perhaps a portion of Tesla's problem was with marketing? Autopilot gives very different cues than something like LKAS (lane keep assist).

And it's also called "Autopilot" which sets an extremely different tone and expectation than other driver-assistance/safety technologies.

> And it's also called "Autopilot" which sets an extremely different tone and expectation than other driver-assistance/safety technologies.

Only to those not familiar with autopilot systems in aircraft.

It's like the use of "hacker" - here on this site it still (probably) has most of it's original meaning. To the rest of the world it means a guy in a ski-mask with a russian/chinese accent who's after your bank account.

Unfortunately most of the world thinks pilots in commercial aircraft turn on the autopilot and can then go for a sleep, so when they see "Autopilot" on a Tesla, they think they can do the same.

> most of the world thinks pilots in commercial aircraft turn on the autopilot and can then go for a sleep

They can go to sleep and the plane will continue to fly itself. However, if anyone found out they were sleeping heads would roll.

Autopilot electronics on a plane is trivial. Autopilot for a car is damn hard.

>Unfortunately most of the world thinks pilots in commercial aircraft turn on the autopilot and can then go for a sleep

They can go for a sleep though.

Only if you're happy waking up to find you're about to plow into another large object at a high rate of speed.

The Pilot in Command is expected to sit in their seat and monitor the aircraft, surrounding airspace and the radios, and be ready to take control at a moments notice.

They can't go for a snooze without handing that over to another pilot first.

The same expectation applies to drivers of Tessa's Autopilot.

> Only to those not familiar with autopilot systems in aircraft.

Isn’t that most people?

> That's the approach we know won't work. That was Tesla's first-round autopilot, the self-crashing car.

I think OP was talking about "crash avoidance" systems. Not systems that claim to drive the car but not really.

You're always driving, the car is just silently observing. If the car sees you're about to crash, it steps in and takes over. The opposite of Tesla's approach where you play the role of observer taking over when things go wrong.

The first approach is safe. The second approach isn't safe until full autonomy due to cognitive load and task switching delays.

There are many automotive "auto-brake" systems, usually controlled by a radar. Those are useful. They're usually programmed to act only when a collision is inevitable. They prevent some slow-speed collisions, and reduce the damage in some faster ones.

What they don't do is take over steering. A last-second takeover of steering would make things worse some of the time.

Don’t Volvo and Mercedes do obstacle avoidance? Or is that still in research? I think I remember seeing videos.

Yes they do. They take over steering, automatically drive around obstacles, etc. And this is already deployed in production vehicles today.

Can you elaborate?

Assuming for example the car is fitted with solid state lidar on the four corners and ESR on the four sides or even LIDAR.

What are your thoughts on crash prevention if the car has the same hardware as a a Waymo car?

(edit: to clarify, would you consider autosteering at the last minute safer than a human driver in that scenario?)

The next Volvo model will take over steering for collision avoidance.

http://www.leftlanenews.com/volvo-xc60s-collision-avoidance-...

> What they don't do is take over steering. A last-second takeover of steering would make things worse some of the time.

I have a Honda. While its lane-keeping system won't steer during a crash, it definitely does steer to keep you centered in the lane (when it's able). That could help avoid all kinds of crashes due to inattentive lane-drifting.

Is it less fatiguing?

Yes. You still have to pay attention, but there's not as much pressure and fewer startling surprises.

I really don't want those on in stop and go rush-hour traffic. They're very likely to cause someone /behind/ the automated car to have an accident. (Yeah yeah, following distances; trying building enough capacity for everyone to do the proper following speed and keep distance.)

Tesla took the opposite approach of what I described (the Subaru approach).

Tesla took the approach that the machine is in control and the human has to detect and take over when the machine makes a mistake. That is known to be a very problematic strategy, it's easy for the human to get distracted.

What Subaru has done is to leave the human in control, the machine only steps in when it sees a problem.

It's totally different safety outcomes. In fact, not even the same problem - one is "self driving", the other is "crash prevention".

Yes they are different. They are also WILDLY different value propositions. Crash prevention is nice, just getting safer and safer is nice--though I personally wouldn't stop driving ~$5k cars for even "really good" crash protection.

But a car that can drive me while I sleep and become a Taxi while I work is something I would go into debt for.

But a car that can drive me while I sleep and become a Taxi while I work is something I would go into debt for.

I don't think most regular people will actually own fully autonomous cars to rent them out. Big players will do the capital expenditure to buy a fleet, and rent out transportation service. Renting in uber style will be basically the price of owning, but without the up-front capital expenditure, and it will provide the convenience of choosing the type of car for the specific transport needs of the moment. The margins of these rent-a-fleet services will be low, so you won't have a hope as individual car owner of actually buying a car and renting it out without turning it into a loss compared to just renting.

Certainly if most people owned them, there wouldn't be a market to rent them out.

I would go into debt for that also!

> Tesla took the approach that the machine is in control

That's not really accurate. Tesla says the driver is ultimately responsible. That holds unless a court determines otherwise. The same is true for Toyota/Subaru. They're all SAE level 2 systems.

The opposite approach from this would be Waymo/Volvo who intend to take responsibility for decisions made by their vehicles. Those are SAE level 4 or 5 systems.

Whether or not it "works" it has the overwhelming benefit of being an incremental approach. Incremental (evolutionary) change tends to work while revolutionary is more tricky. e.g. video phones were demonstrated in the 1960s but are still not really widely used. Also no need for Star Trek - level AI, which is a significant advantage.

> but are still not really widely used

Seriously? All smartphones are video phones. I fail to see how they could be more widely used.

You may mean that people choose to make audio-only calls... but just because people choose to do that, it doesn't mean the technology isn't here & fully available in affordable, commercial, "good enough" form.

Interesting. The three of them show the car being hit from the left side when driving on a high-way. It seems that the autopilot is bad at recognizing objects that suddenly appear on that side.

More like, if you are not a car on the left lane, then you are not. Bam, you got an accident.

Edit: Is the last video in the US? Is it normal for other drivers not to stop when they see an accident?

Right. Tesla's round 1 "autopilot" used Mobileye's car recognizer for obstacle detection. If it didn't look like a car, it didn't get recognized as an obstacle.

I wrote a previous YC posting on this, on why you need to use geometry first, then object recognition.

I don't know by what metrics you're making these comparisons. Audio recognition and self-driving vehicles are apples and oranges.

Autonomous driving is difficult primarily because of a changing world and low tolerance for mistakes. The state space of situations you're trying to map is just large. It will take time. But it's not even close to impossible.

Audio recognition is difficult for a different reason: language is difficult to disambiguate without context. So today the limits of our audio recognition bump up against knowing the context of words you're saying. It used to be the case that audio recognition had fundamental difficulties e.g at word boundaries but those challenges are mostly solved. Some of these challenges might not be solved without improvements in areas outside of the raw listening part of ASR.

> That being said, Toyota's (and Subaru's) approach is the smart way forward - add sensors and capabilities that augment the human, but leave the human always responsible and in control. In order for a crash to happen, the human AND the machine BOTH have to miss it.

This is your opinion. Folks at some other companies (e.g waymo) don't think this makes you safer. Human attention might not work the way you think it does.

I think that audio recognition is an order of magnitude less complicated than fully autonomous driving. The world is fully contextless outside the road to an not-even-at-insect-level brain driving two tons of metal, and anything can happen.

Voice recognition is getting just barely tolerable. If you have tried to do something important using voice recognition you'd probably go bat shit crazy after a while. If it had the potential to kill people no sane human would allow it.

Voice recognition also has a lot less market demand than self driving cars. Technology can be invented that is decades ahead of other fields if enough money is invested.

That is why I think self-driving cars and voice recognition are not comparable problems.

Totally agree ..

> Voice recognition also has a lot less market demand than self driving cars.

Just to expand on your point, today's 'self driving car' technology we may equal to where 'voice recognition technology' was 20 years ago. But 20 years ago, there is no incentive for industry to invest say $100 Billion in 'voice recognition technology', the benefits did not warrant such a huge investment. Where as'self driving car' technology there is potential to reap benefits that is why industry today investing $100 Billion ( all combined investments from all players )

> Technology can be invented that is decades ahead of other fields if enough money is invested.

> That is why I think self-driving cars and voice recognition are not comparable problems.

The huge economic value the can be created with perfected 'self driving car' is in the order of hundreds of Billions of dollars per year world wide combined . Here is an example scenario

In 2025 'self driving Electric Autonomous fleet' vehicles by Uber, Googles of the world offer 'miles plan' ( like our 2005 mobile monthly plan N minutes/month ) that is like 1000 miles per month you can hail any time for a monthly price of $300/month . This is huge economic value for at least 50% of USA drivers and people will embrace it.

Typical USA drivers total cost of ownership of CAR today is in the range of $400 to $600 per month ( CAR price + Repairs costs + car Insurance + gasoline cost) . This is with out counting 1.5 hours/day you have pay attention to drive, the time that can be used for other thing with autonomous fleet

> The world is fully contextless outside the road to an not-even-at-insect-level brain driving two tons of metal, and anything can happen.

See that's where you're wrong. An autonomous vehicle doesn't have to have a human brain. It has to have very accurate sensors and classifiers tuned with generous safety params.

Like any vehicle it also needs redundancies for safety and sane default behavior (mostly "slam on the breaks").

Also, not "anything can happen." There are rules that govern the road that make this problem tractable. There are rules that govern physics that mean if you have adequate sensors + compute you can avoid hitting any objects at all, whether it's a person or otherwise. "Object x is 30 ft away with trajectory y, we're going 15mph, slam on the brakes."

The difficulties mostly lie in how we can "not slam on the brakes all the time and never make progress." e.g if you enter a construction zone you better know what to do and not just sit still, otherwise you'll never get anywhere.

  I think

Wonderful.

Would you please not do this here? We're trying for at least a slightly higher quality level on HN. One thing that helps is not to assume stupidity in the other; we even added a new guideline recently to say that. More diplomatically though.

https://news.ycombinator.com/newsguidelines.html

Audio recognition may be apples and oranges to self-driving vehicles. But as it relates to autonomous driving, audio recognition (and I don't mean speech but general sounds) is something nobody has been talking about (at least publicly) as a sensor input for self-driving cars. Think about how much you rely on your hearing to drive safely and well. Sure, we use sight/depth and so forth substantially but we also rely on our ears listening for different kinds of sounds that can occur from any direction to warn us of changing conditions, unsafe circumstances, and so forth.

We can't even get speech recognition right and meanwhile audio (sounds in general) aren't even being seriously considered for self-driving purposes, AFAICT.

Curious, isn't it? Humans rely on auditory information to enhance situational awareness so you'd think that would factor into the design of a machine that needs as much situational awareness as possible.

For example, emergency response vehicles announce their arrival and direction often long before any visual contact is made. How will deaf autonomous cars recognize incoming emergency responders if the view is obstructed? Will they pull into the intersection just as a Fire Engine is running the red light?

Auditory input was deemed important enough that California even made it illegal to "wear a headset covering, earplugs in, or earphones covering, resting on, or inserted in, both ears" while driving just last year. [0] I suspect other jurisdictions have/will follow(ed) suit.

[0]http://leginfo.legislature.ca.gov/faces/codes_displaySection....

Are deaf people allowed to drive? How do you think they handle these situations?

According to a quick google, noticing that other cars are pulling in/slowing down usually works if they can't see the emergency vehicle. Still, it's not "perfect", but then Hearing drivers are far from perfect either, and there's plenty of failsafes before an accident happens.

Fire trucks and ambulances I've seen when driving against traffic lights/flow of traffic always slow and stop to make sure that there will not be an accident even if they have the right of way - they don't drive like police in pursuit on TV. I'm sure there are exceptions but the general case is the emergency vehicles are planning for other traffic to not yield until it is clear the other traffic is yielding.

I'm sure there are exceptions but the general case is the emergency vehicles are planning for other traffic to not yield until it is clear the other traffic is yielding.

Yes, if the driver is well trained. When I was firefighting and instructing drivers, we always emphasized that while the red light and siren gave you the right of way:

a. they don't absolve you of liability for a crash (at least under NC law)

and

b. if you get in a crash on the way to the scene, you're not doing anybody any good, PLUS you've now created another incident requiring another emergency response, PLUS another company (probably coming from further away) has to respond to the original call.

I can't speak for police, but firefighters are actually, in my experience, taught to be pretty conservative when it comes to running red lights, proceeding against traffic on one way streets, and other similar scenarios.

We also always used to emphasize "it does no good to get halfway there, real fast".

There seems to be some limited automotive technology available that would convert sirens to visual signals, akin to doorbell or telephone strobes. But the more interesting adaptive technology seems to be a re-wiring of the brain. [0] Deprived of auditory input and with a heightened reliance on visual cues, the autonomous driving version might be something like Lidar + other driver facial recognition.

[0] https://www.scientificamerican.com/article/superpowers-for-t...

> There seems to be some limited automotive technology available that would convert sirens to visual signals

Yes, I've noticed this too. Sometimes you can almost see blue and red flashing lights on top of a police car.

I applaud your synesthesia! Those of us with more common auditory processing are unable to see sound. :(

si·ren (sīrən)noun 1.a device that makes a loud prolonged sound as a signal or warning.

Per [0] they are allowed to drive in the US, and do not have higher accident rates than hearing drivers[1]. Also, "There are now devices available that can be used in r cars that react to the frequency of sirens and which emit visual warnings of approaching emergency vehicles to assist deaf drivers and those with hearing loss" (which could presumably also "assist" self-driving cars).

[0] https://www.enddd.org/end-distracted-driving/enddd-presentat...

[1] This web site's bona fides aren't clear, but its mission does not seem to specifically include advocating for the deaf, so there's no obvious reason for the information presented to be biased in favor of the deaf.

Microphones cost much h less than LIDAR.

Telling which direction a siren is coming from is really, really easy. You can do it with three electret microphones and a $1 ARM chip. A simple TDOA multilateration algorithm will do a far better job of it than a human, even in complex reverberant environments.

Reliably understanding speech is effectively strong AI. Humans don't speak very clearly, but we're spectacularly good at inferring meaning from context. Accurate transcripts of colloquial speech are often completely incomprehensible, because a vast amount of information is conveyed through context. When we're listening to someone speak, we're decoding the phonemes into symbols, but we're also constructing a model of the speaker's mind and predicting what they mean to say.

"Nobody talking about it" doesn't mean nobody is working on it ;)

> Audio recognition is difficult for a different reason: language is difficult to disambiguate without context.

That's not the only reason audio recognition is hard. If it were, audio recognition would be exactly as hard as interpreting written text.

I get your point but to nitpick, written and spoken language are not equivalent. We don't say things like "quote", "comma" and "question mark". We use things like intonation, gestures etc.

Agreed. I would think audio recognition is not comparable to driving, since the scope of recognizing all voices, accents, languages is vastly different than the set of rules for driving and the complexity of the road.

Realistically speaking, even if Tesla or any other manufacturing company do get to that level of a "super advance Einstein-genius" self-driving car it will still be: "not even close" if it's riding alongside the highway with 18-year old Billy-bob speedster with a penchant for cutting lanes and living dangerously. Most car accidents are due to human error.

If all cars on the highway are self-driving cars, with the ability to learn and communicate with each other through mesh-network or what-not, with predictive capacity -- it will be a lot better than any driving human that can only see what's in front of them. We just need to get to that level where most cars on the highway are smart self-driving capable cars and we can build on top of that. No one should be relying on self-driving cars to do everything at this point in the game because we're not there yet.

> language is difficult to disambiguate without context.

So why don't Siri/Alexa/etc developers give them context? I would LOVE to be able to say "Alexa, play playlist on shuffle mode" or even simple "Alexa, turn on shuffle and go to next song"

But no, the word "and" is completely off limits. Saying "shuffle playlist" instead of "play playlist" also doesn't work. I have to literally use 3 commands to do what I want:

"Alexa, play Taylor Swift" ... "Alexa, shuffle" ... "Alexa, next song"

With context, I could say "Alexa, play Taylor Swift on shuffle"

Apparently Siri is gaining the ability to deal with follow up questions in iOS11. Excited to try it out.

How would you differentiate:

"Alexa, play Taylor Swift on shuffle"

"Alexa, play Taylor Swift on Spotify" (real command)

Spotify is default. She also knows how to do noun recognition. If she can tell the difference between “play” and “stop”, she can differentiate shuffle and Spotify.

I’m also ok sayin “Alexa, shuffle Taylor Swift” if that’s easier for their grammar.

I think "shuffle" is the obvious choice to be used as a verb.

As you mentioned, Alexa already understands shuffle (though I don't think it does for Spotify's integration).

Once the machines are better than humans at a macro level, insurance companies will notice, and I think this will cause widespread adoption pretty quickly.

California allows drivers to self-insure. The existing auto insurers are going to lose that business to automated taxi fleets.

This has already started - look up the discount some insurers give to owners of Subarus that are equipped with EyeSight. The numbers show their collision avoidance is making roads safer.

Also, interesting fact: Subaru is achieving these impressive stats using stereo vision, not lidar.

I recently bought a Subaru, and I agree. However, this is in no way at all related to "self driving cars" (which don't actually exist and therefore nobody knows how the insurance companies and general public will react to the liability/safety/skin-in-the-game issues).

Folks are glossing over this quote too easily:

> “Historically human beings have shown zero tolerance for injury or death caused by flaws in a machine,”

It doesn't matter if it's objectively safer; people have no issue doing dangerous things so long as they can maintain the illusion of control.

I don' think that's true in terms of transportation systems.

In terms of cars: Streets where built for people or animals walking on them and cars quickly started killing large numbers of people. The response was not to make cars safer by say physically limiting them to 10 MPH, but to remove people from streets.

A fairly high fraction of early aviators where killed, that did little to slow adoption of aircraft.

Astronaut deaths where sadly common, but also expected and did little to slow progress.

Cost/benefit--if the rewards are judged to outweigh the risks and there is a clear benefit, people will adopt it. If money/time can be saved, it will be very popular.

> The response was not to make cars safer by say physically limiting them to 10 MPH

Umm, reasonably sure that implementing a legal speed limit was an initial response, in the UK at least. It took a number of years (and likely a lot of propaganda, though I haven't looked that closely into it and history is, of course, written by the victors) to allow motorcars to use roads at anything close to a reasonable speed.

Check wikipedia out: https://en.wikipedia.org/wiki/Speed_limit#History

I am referring to a governor which would prevent cars from being capable of 10MPH even on empty roads. Actual in town speed limits would presumably be lower. Now, with our infrastructure that seems like a huge issue, but cites used to be built for walking and even small towns simply had vastly higher density.

Watch this and suddenly those mentioned speed limits are completely reasonable. https://www.youtube.com/watch?v=954L9MpfCEo

Note, 10MPH would allow for 2-3x the speed of a horse over long distances. So, in this alternate history cars would still have been very useful.

Indeed, the speed limits are completely reasonable if you don't want to remove people from the streets - but your suggestion seemed to be that people were near-immediately removed from the streets instead of limiting what cars can do at first. There were 30 years when cars were lawfully limited to 4mph in rural areas, so your "even on empty roads" point is moot.

There where not exactly a lot of cars over most of that 30 years. However, my point was more cars had strong advocates pushing adoption where pedestrians where simply used to how things where and never really organized to protect how things where.

Thus you get jaywalking as a crime, and justified many deaths as people breaking the law instead of the naritive that cars are insanely dangerous and should be banned. In theory cars could have been banned, but in practice their advocates had little trouble changing perceptions.

Note that jaywalking is very much not a crime in the UK. Off of a motorway (which is a relatively recent invention), roads are still for people to some extent. Hopping off the pavement for a couple of seconds to bypass someone who's decided to stop right in front of you is semi-normal, much of rural UK doesn't even have pavements, and we get taught how to Stop, Look, Listen, Live in order to cross a road without a traffic light as kids.

If a pedestrian starts to cross the road, they have right of way. I've been on roads in some city centres where vehicles have to be very, very careful as there's a road through a pedestrianised area and people are rather bold crossing the road. If the vehicle was to cause injury, the driver would likely be liable.

>the illusion of control.

No, it's literal control. They'd rather operate a vehicle in a way that's statistically less effective (where effective is defined by a collection of algorithms, we all know how perfect algorithmic generated routes are) than lose that control for a small benefit. Control itself is worth something to people.

A very large factor in this reasoning is that people look at the statistics and then intuitively decide the statistics will not apply to them because they are much better than the average driver.

The scenario in which insurance companies drive adoption of self driving cars depends upon insurance companies accepting the risk of system failure due to manufacturing defects or limitations that generate the wrong response in particular conditions.

Indeed. You aren't insuring one driver anymore, you're insuring a million instances of the same algo.

Reply on news.ycombinator.com