There are a host of obvious reasons why even a "simple" L4 car will beat a human hands down:
* Reaction time.
* 360 awareness / visibility.
* Lack of exaggerated "Human Reflex" to surprising events (i.e. swerving violently to avoid a dog, and hitting other vehicles)
* Keeping perfect space around the vehicle for safe stopping at all times. (don't tailgate, don't get rear-ended)
* Assuming a LIDAR-based system, virtually no difference in day vs. night vision. No "sun in your eyes" or "road glare".
* No fatigue related accidents.
* No DUI and related prescription drug accidents.
* No distraction-based accidents. (kids, cellphone, food)
* No "road rage" based accidents.
I could go on, but this should drive the point home pretty clearly, IMHO.
You make an argument like you have a hammer in your hand and you are looking at a row of nails. I'm suggesting it is short-sighted to look at the problem from only that lens.
For instance, with a more holistic lens you might explore the right place for humans drivers vs autonomous vehicles. You might simulate entire cities from the ground up, specifically optimized for transportation. What if the real breakthrough in transportation systems actually comes from the ability to quickly construct / deconstruct roads? Or dynamic city zoning? Or any number of non-vehicle innovations? What if we were actually optimizing for safety in transportation and not just trying to compare machine ability to human ability in limited contexts?
The worry is this, the cars will not fail like a human will. It may even be dangerous The car may not even be able to move in certain situations (like another car blocking them in front either purposefully or accidentally). Or it may not be able to even see certain conditions a human could.
You're operating on a completely archaic understanding of AI. Modern systems can and do adapt to situations on-the-fly. What's more, they can adapt on aggregate - the experiences gathered by a single car can be shared to all cars. Very quickly, every car in the fleet will have billions of hours of cumulative driving experience.
The overwhelming majority of motor vehicle accidents aren't weird and unpredictable edge cases. They're tragic but mundane events that result from a handful of root causes - inattention, excess speed, poor judgement and unnecessary risk-taking. "Driver/rider failed to look properly" is the key contributory factor in nearly half of road traffic accidents. Computers utterly dominate humans in this respect. For every accident caused by some bizarre and unpredictable set of circumstances, there are thousands caused by someone doing something obviously stupid.
A computer can be programmed to be ultra-cautious in difficult situations. A computer can maintain 100% vigilance 100% of the time. Humans can't. Self-driving cars will undoubtedly fail in new and unexpected ways, but it's abundantly clear that they'll fail much less often than humans.
That's sort of my point. Accidents are rarely caused by edge cases, but by predictable human failures. All the information needed to avoid the typical accident was plainly available, but the driver just didn't see it, process it or react to it appropriately. Driving is generally very predictable, but human attention and perception is hugely fallible. An AI that gets the basics consistently right but occasionally freaks out in an unpredictable situation would be a huge improvement over human drivers.
In the same way, if the world is modelled accurately and succintly, can a car predict the various scenarios and adapt, even if the current sitation was not seen before (extremely high speeds or unusual number of cars, etc.)
From a human's perspective it's a stupid mistake. It seems stupid because computers think differently than people do. On the other hand a computer would probably think a human were stupid if it saw a human make a mistake when calculating the logarithms of every number between one and a million.
It's an improvement as long as those are few percent of the unsurprising and human ways human drivers fail.
It's more like humans can think, why should we trust a computer with a task where unforeseen circumstances can arise?
Further, if I got the current state correctly, no self-driving car system so far works in rainy conditions.
https://www.youtube.com/watch?v=bAxoo6JLmgg
or this
https://www.youtube.com/watch?v=JoGgE55qNBA
and this is just a L2 system...come on brah come harder
* Lack of software bugs due to careless programming.
* Millions of years of brain evolution allows adapting to unforeseen circumstances.
* Vastly decreased likelihood of being tricked by vandals exploiting flaws in recognition algorithms.
* Inability to be hacked en masse and driven into each other/houses/pedestrians.
* Much harder to back-door.
* Can be judged in court if they swerve to avoid a pigeon and hit a crowd of pedestrians.
I could go on, but this should drive the point home pretty clearly, IMHO.
They sell these features as "parking assist" or "smart cruise control" or similar, but don't kid yourself into thinking that there's not a very sophisticated software system that could, if hacked, take over control. Has this ever happened? Have there been catastrophic bugs documented?
More advanced vehicles (i.e. Tesla) have enough hardware to be be nearly fully autonomous today. (i.e. they could rend control from a driver and continue to pilot the vehicle to an alternate destination if hacked)
This is very much not true of humans.
> Vastly decreased likelihood of being tricked by vandals exploiting flaws in recognition algorithms
This is not particularly true of humans. The human eye is very easily tricked. It's true that there are things that would not fool us and would fool a computer, but the inverse is also true.
What in the last few millions of years caused humans to evolve to steer metal objects at 60mph?
For extra credit: describe what the reaction time is for Waymo's driverless cars are during various points of the garbage collection cycle.
People here have a tendency to imagine an idealized version of driverless cars and assume that actual driverless cars are the idealized versions of those cars, and then compare the idealized cars to very non-idealized humans.
I think that the real world is going to be more complicated than that.
It is easy to write code that takes a non-trivial amount of time to process to the point of output on any given hardware. I imagine that all of us have done that at various times. We can certainly imagine that reaction time could be a virtue of driverless cars -- even 99th percentile reaction time. Given sufficient hardware (and not just hardware on the core CPU/ram chips -- this is hardware out to the sensors and the subprocessors associated with each of them).
Similarly, we could imagine putting enough sensors on the vehicle that it truly has 360 degree awareness and visibility. But we do see real cars in the real world that have surprising blindspots.
There are cost tradeoffs to all of these things, and complexity tradeoffs. Is there a happy medium? I think that there very likely is, even without any fundamental advances in the state of the art. Are we within a year or two of that happy medium? Maybe.
>For extra credit: describe what the reaction time is for Waymo's driverless cars are during various points of the garbage collection cycle.
Aight, I'm not familiar with Waymo's cars specifically, but with <10K in hardware (maybe <5K now) you can design and build an autonomous vehicle kit that has a reaction loop that runs at a soft floor of 15 adjustments per second. I'm sure that with better hardware you can increase that to 60+, and 15 per second is already about 5x better than human reaction time, and fast enough that it can react to changes every 6 feet at highway speeds, and every foot at neighborhood speeds. A 60 aps loop means that its recalculating everything every 2 feet your car travels on the highway.
And yes, this isn't incredibly difficult.
For reference, in my experience, running 2 cameras and a low power lidar (along with associated outputs to hardware actuators and such) off of a 2013 (or maybe a 2011) macbook pro, the limiting factor was always the lidar's 15fps framerate.
you can say the same thing for rockets. There is entire field dedicated to this, its called Realtime Systems.
I don't think it's controversial to say that machines can easily beat those human reaction times.
Most or all autonomous vehicle software is written in C++, so garbage collection shouldn't be an issue. Of course, these programs need to be tested carefully and extremely before we trust them.
Not sure I see the point you are making... all cars have blind spots with drivers now. Human drivers are always blind to the spots they are not looking at the moment anyway. I can imagine cars being much safer than they are now.
But when you look at how you imagine driverless cars to be, you might want to temper your imagination with how driverless cars actually behave, such as the Tesla running into a trailer because it was too high off the ground for its ultrasound sensors to detect.
http://www.theverge.com/2016/5/11/11656496/tesla-model-s-aut...
(Note that this is not the famous case of the Tesla being unable to detect a white semi against a bright sky, it's a low speed no-injuries collision).
Obviously, Tesla is not the be-all and end-all of driverless cars. But when you're dealing with sensors combined with AI where each is in many ways less capable than the human eye, with different failure modes, it gets expensive and difficult to design really good fields of vision.
"In the past year our autonomous cars have driven a total of 50 million miles in real-world conditions. Over that period there were 4 collisions, plus an estimated 12 which would have occurred if our trained staff in the drivers seat had not intervened. Insurance analysts estimate that if human drivers had driven under similar conditions there would have been 51 collisions."
Would that be sufficient for you to support the introduction of Level 4 automation?
All this would do is make me question the nature of the data collected. For instance, is that 50 million miles of diverse roads and weather conditions or the same mile over and over again on a sunny day? Or were the human-driven cars equipped with contemporary collision-avoidance equipment or a motley representative assortment of cars from the past half-century?
What would sway me is auto insurers offering a discount for relying on self-driving car technology. I can wait patiently for that to happen.
Edit to add... Have the crazy expensive and downright hostile self-driving John Deere tractors eliminated farming deaths?
Obviously, the assumption of all those involved is that the price will plummet as the technology becomes more adopted, so the goal is to get to a price point where the earliest adopters will be willing to pay. For self driving cars, talking about plumbers not buying this at the current price point is a bizarre argument.
Why does everyone assume the cost will decrease with scale? How many people can afford a Tesla right now? Let alone a fully autonomous car?
If flying were more affordable and less strict more people would own their own plane. It isn't and they don't. I see self-driving cars a lot like airplanes and John Deere tractors. That is, expensive proprietary tech that you have no control over. No thank you.
Why does everyone assume that most people want to share a vehicle?
Because you don't need to own the car - you could feasibly rent one on demand (like a taxi), since you can now rent a car without the necessary wage overhead of the taxi driver, or the minimum costs of car-rental (i.e. car rental typically is per-day, for logistical reasons) or the logistical problems of car-rental (i.e. you need to drive to and from the car-rental place, to pick up and drop off the car - and by "drive" I mean "get a lift or use public transport").
Plus, the self-driving stuff shouldn't be particularly expensive, and will be counterbalanced by lower insurance premiums, and likely tax incentives once voters realise that they really do save lives.
"Why does everyone assume the cost will decrease with scale? How many people can afford a Tesla right now? Let alone a fully autonomous car?"
Because just about everything decreases with scale, and it's not inherently hard tech. Tesla's main problem is their lack of scale. As are autonomous cars.
"Why does everyone assume that most people want to share a vehicle?"
Money. It's not "want", it's "can tolerate so as to save money". Your car, when adding together all the costs (purchase, fuel, rego, maintenance, interest payments if you borrow instead of buying outright) is one of the biggest expenses that most people have. Sharing a car between several people should decrease the costs by an order of magnitude, which could be used to buy whatever people spend their money on these days, or saved.
Let's say each ride cost $10... that is $20 to $40 per day just for the kids on most weekdays. Add in birthday parties, saturday games, doctor/dentist, trip to grandma/grandpa that live 7 miles away... and we are talking $1000 per month in "ride-sharing costs" for my children alone. We could add in my wife and I's costs, but it would be silly.
Maybe this service is less than $10 per ride, even at $5 per ride it is wayyyyyy more expensive than I currently spend on a 4 year old Toyota Highlander with low miles with insurance and fuel.
a.) Snacks and supplies in trunk - lockable containers inside. All could be accessible via app/key fob only.
b.) Who says the suburbanite your renting to won't want to use a car seat?
c.) Camera, ID, cleaning deposits, car self-drives to detailer on your lunch break.
$1,000 a month is a heck of a carrot to keep a car clean - even if that's not for you is your worldview really so small that you cannot imagine it being for someone not all that dissimilar from you?
If the total cost of ownership is, say, $1000 per month, you might only pay $200, or about $2.20 per ride.
Why? We literally never see that with top shelf products. This isn't an arduino solder job. This is going to be an integrated software and hardware package that will rival an aircraft autopilot/navigation. Please check you facts on the cost of these things.
I don't want walmart brand, everday low price components on my self-driving vehicle. I want top of the line if my life is in the balance.
These are all things that started high and fell low. The historical prices for each of these systems reflects this fact, as does their gradual movement through the available product line-up. They all started in top of the line vehicles that cost well north of the average person's yearly salary, and are steadily creeping into the lowliest models.
I see nothing in self-driving systems that make them special in this regard. They're already largely an amalgamation of commodity components whose prices are trending downwards.
You can bet on them staying a niche product forever, if you wish. I just think it flies in the face of historical evidence.
Google claims that it dropped the price of lidar by a factor of 10x once it started producing them. That seems like a significant, factual, drop in cost.
If ridesharing was awesome, we would see it now with luxury cars. We don't because it is stupid and will not reduce the price of the car dramatically enough to outweigh the negatives.
A lot of the regulation that drives the cost of planes up won't apply to cars because of the different economies of scale.
The most popular light aircraft in history is the Cessna 172, and since it was introduced in 1955, only 43,000 have been built. Ever.
By contrast, the most popular vehicle on the road today in the US is the Ford F-series pickup truck. Since being introduced in 1977, Ford has sold 26 million of them. So roughly 600x as many as Cessna has sold 172s over a period of time 20 years longer.
The US market alone was responsible for 17.5 million new cars and trucks being sold last year. Globally, it's approaching 100 million per year. Pretty incredible economies of scale available with that kind of market.
I'd quite like the freedom to be able to commute to work, go out for drinks with people after work and come home again at the time of my choosing without relying upon someone else to cart me around, paid driver or otherwise.
THAT seems freedom-enhancing.
Yes, I know that Uber has started to tackle this problem, but as far as I know it only provides 1 car seat, and a car equipped this way may not always be nearby. Maybe this will be a completely solved problem soon, which would be great!
Remember when LCD screens came out and very few people could afford them? We were complaining then too about how could people possibly afford this and who will buy it??!
1. A Tesla with full self-driving hardware can be had today for ~$75k. And the model 3 will cut that in half. And the model after that will be even less. Millions of people can afford this.
2. Who said anything about retrofitting old cars? As far as I know, nobody is planning to do this.
3. This isn't a requirement, and it won't happen all at once. The plumber with an old F150 will continue to drive as he always has until the cost/benefit makes sense for him to upgrade.
4. Self driving tractors move at 3mph, and I don't think farm-field traffic fatalities are the big driver behind automated farming. Different use case.
Additionally maximizing profit (and this may read as condescending but I do not intend for it to be so) does not necessarily mean maximizing price. Obviously if you are not selling at a loss you can potentially make more money with a cheaper price point on volume.
I do agree, however, that first adopters will likely pay a premium.
Eventually I imagine insurance companies will help subsidize these things too - if the price does not become trivial.
Absolutely they will. That's [at least] $60K worth of driver's salary/benefits costs you've just eliminated. $75K is cheap to eliminate the cost of a human employee.
Most people who rely on a vehicle for work are local or long-haul trucking drivers, not local trade workers. The important work a plumber or a/c tech performs is separate from the driving. You're right, those people probably won't move to automated vehicles immediately.
But theoretically, in the future where they could have an automated van meet them at the job site with their tools? And return the tools to a secure storage facility afterwards? I don't see why they wouldn't go for that.
Ever call a taxi (not in city) ? 30+ minutes.
Also, tragedy of the commons ... who was the last one to use the car has to pay for clean up.
How to accomplish that ? ... "No the other guy did it ..."
http://www.thedrive.com/tech/9548/the-biggest-opportunity-ev...
There is a successful precedent in aviation. The latest fly-by-wire flight control systems treat the pilot's inputs as merely suggestions and will modify them as necessary to prevent departure from controlled flight, midair collisions, flight into terrain, and overstressing the airframe.
That seems like the most obvious statement I've read in the last year or so at least. I think the fact that Alphabet spent more than 10 years developing these self-driving cars already is a pretty strong indicator that there are non-trivial engineering challenges involved.
Yes, map services like Google have been receiving some data from government in addition to collecting user's current location and user's generous feedback (e.g. Waze user) to determine best route. But we are no where near the condition we can know what's happening ahead of us. What about weather condition? On intersection who goes first?
A safe driverless vehicles should be able to communicate (check, for as long as the communication is stable), and government and cars will share feedback to other cars. This is a crowd-sourcing effort to make fully autonomous car possible on road. If we just learn as we go on the road, these driverless cars will not work well in very complex road condition.
In reality, the problems encountered by cars are fairly classic and predictably where computers can be very good at: computer vision, 3d projection and modeling, project patterns, control loops (stearing vs. drifting). The beginning was expectedly bad, but is it not a stretch to imagine those solvable, just like Moore’s law and expecting cheaper electronics with mass-market “makes sense”.
Ignoring that, that ransomware wont' even exist for a number of reasons, the biggest being it wouldn't be profitable. Ransomware authors aren't going to get a payout, very few people are going to be able to buy Bitcoin in 30 minutes while unable to leave their car even if they wanted to. Not only that, it isn't remotely difficult to have a manual, physical, kill switch that physically stops the car in a life or death situation. Even if you didn't have a kill switch, you could simply call the police and they can lay down a spike strip.
>who are careful because they fear their own death just as much as I do.
LOL! That's a good one! Have you actually even been on the roads before!?! When I worked drive thru I had people who were so drunk they could hardly form a sentence come through my drive thru. That's just ONE example.
You share the road with these upstanding citizens who are clearly very careful and concerned for their own wellbeing: https://www.youtube.com/watch?v=DcnuIWNv8lw https://www.youtube.com/watch?v=9hddt4bWNns
> Road fatalities are already the number 1 killer of some age groups
You're hiding a weakness in your argument by not specifying the age groups. You're hiding a weakness in your argument by restricting to cross-sections of the population that support your argument. You're misleading your listener by hiding the fact that this statement isn't because traffic accidents are unusually high, but that other causes of death are unusually low.
> Around 40,000 in the US last year.
Careful using absolute numbers, using 40,000 which is a very small number compared to other kinds of deaths like hard disease (800,000 deaths last year) can be misleading because 40,000 sounds like a big number.
> its estimated that over a million people worldwide die on the roads every year
You're hiding the fact that the US has 4% of the world's automobile deaths but 20% of the world's cars. You're using aggregate data to smooth over deaths that are caused by dangerous driving environments, roads, or ineffective traffic laws and misleading your listener into assuming that they're caused by human error.
Total deaths is not a useful statistic. The deaths that could have been prevented by L4 autonomous cars but could have not been prevented by L1-L3 or more conventional safety practices is what you should be presenting.
Give as little surface area as possible to people who would use it to push back.
As for sharing roads with others who fear death as much as I do. That's a wonderful idea, however in my case a pipe dream. I share the roads with people who have no problem speeding, driving under the influence or even refusing to use seatbelts.
Ultimately I'm more optimistic about solving the ransomware problem than I am about putting the fear of death into other humans.
Such malware are only possible because the targeted system has vulnerabilities to begin with. One just has to ensure the absence of such vulnerabilities, possibly using machine checked proofs.
One obvious approach is to properly isolate the driving software and sensors from external input.
Or what if an actor uses something like this:
https://techcrunch.com/2016/06/29/now-you-see-me-now-you-don...
that confounds the car's sensors. Self-driving cars have a whole spectrum of attack vectors that human beings don't have.
Yes, hostile input will have to be taken into account. Fortunately, this is easily detectable (just look at the logs). And if people die as the result, it will count as murder. Finally, one does not simply jam sensors from across the planet. You need a physical presence on site, and that's riskier than a remote hack.
A lot of the current driving issues do suck, but they still will suck with self-driving cars, as any computerized system has its own vulnerabilities. It's just they might be systemic and affect a lot more people in ways they can't compensate for.
Of course even machine checked proofs will have trouble finding vulnerabilities that creep in during the specifications stage or exist in the realm of hardware.
How do you isolate your sensors from external input and expect to do anything?
I did say "very expensive". But yes, if we're serious about correctness, safety, and security, I don't see any other choice. We should scrap the crap and spend the $billions necessary to rebuild it right.
Ideally. We don't really need to go that far. Isolation is possible and not that hard, see Qmail: http://hillside.net/plop/2004/papers/mhafiz1/PLoP2004_mhafiz...
> Of course even machine checked proofs will have trouble finding vulnerabilities that creep in during the specifications stage or exist in the realm of hardware.
I'm no hardware specialist, but you can still prove the correctness of the design of such and such hardware. The actual chips can still fail the specifications, but that's easier to test once you know the design itself is correct.
Also, specifications themselves can be checked. Only the high-level properties must ultimately be decided and reviewed by hand. Not that they won't be complicated or numerous, but that's still much smaller and easier to deal with than the entire implementation.
> How do you isolate your sensors from external input and expect to do anything?
I wasn't talking about the normal sensor input, which of course can be tricked the same way human sensory input can (for instance with a big flash or something). I was thinking about sensor command, such as which way they should be oriented or something. Though I expect most sensors will have no such input, and will only output to the system.
I used to think that was unlikely. Then I had a few discussions with people about the state of C, and things that could be done to make it default to a slightly more deterministic case by changing how undefined behavior is dealt with in regards to optimization. Now I think it's impossible, because nobody is willing to give up even theoretical unknown performance increases for more security. At least for C. You would have to use something that's much more strict about behavior, like Ada with SPARK, or possibly Rust.
I'm willing to bet most the code already written for these projects is C or C++. Good luck getting that changed if so.
Correct software is expensive. But if we make sure incorrect software is even more expensive, we'll get correct software.
Once the industry is forced to get serious about correctness, they will move away from the C/C++ minefield real quick —or at least come up with safe ways of using C and C++.
Correct software can be compromised via the update process, or social engineering. It's necessary, but not sufficient to ensure integrity.
A cheaper, correct option would be a mechanical lever/breaker labelled "manual override" for licensed drivers or "emergency stop" like they have on every industrial robot or heavy machinery since the 60's. Sometimes the answer to a software problem isn't more/better software.
http://cdn.static-economist.com/sites/default/files/images/2...
Since 1970 total airplane casualties have halved, while there are now 7 times the number of passengers. Much of this improvement has been due taking decisions away from pilots and into the hands of 'unfeeling' automated systems.
And then the cops/secret-service would get involved. If there's a death involved (and reason to believe more will occur in the future, which would be necessary for ransom as a prolonged business model), then they will come down on any ransomware (and responsible negligent parties, such as auto companies with shitty security) like a TON OF BRICKS.
Plus, there's still a lucrative malware model that goes like this: Infect the car, then quietly make it steal itself in the middle of the night (or at your moment of choice). Bam, you now have a stolen car that you can't be connected to, and nobody will notice for a while.
People will pay a fair amount of money to avoid a major inconvenience, but "my car won't start" is a lot less likely to have the FBI beating down your door.
Internet infrastructure is decentralized in nature, making it hard to control. Though rogue individuals can have outsized effects: https://www.wired.com/2008/02/pakistans-accid/
One big way in which automated drivers could improve traffic: By not being a jackass. I remember at one point driving down Westheimer, which is four lanes wide and one of the major thoroughfares of Houston, when I had to slow down for someone making a right turn to merge into traffic into a middle lane while their head was buried deep into the passenger side footwell looking through their stuff.
Then, there are the people who feel like they have to tailgate you within 6 feet during rush hour traffic on the highway. Also, the people who won't let you in for some personal justice you can't possibly understand.
It would only take a smallish fraction of cars implementing the "stay between" algorithm that CGP Grey mentions in his video to significantly improve traffic.
https://www.youtube.com/watch?v=iHzzSao6ypE
I've implemented this algorithm manually. (Much easier to do since I have the instant accelerator response of an electric car.) It does seem to improve traffic flow. Also, jackass tailgaters are sometimes confused by this, and decide to pass.
That seems like an odd question to ask. There are huge reams of traffic safety data collected every year by the NTSB and others. Traffic accidents are one of the top two or three treatable public health issues in the modern world, and they get very significant public funding for their study.
Honestly I think the question has to be the reverse: what is is about "driverless" safety data that makes you think it won't be well-measured by the existing "validation" regime?
Obviously not all other things are equal, and potentially we could create greater loss of life by attempting to get self-driving cars on the road too quickly. But that is the nature of problem: how do we avoid moving too quickly while recognizing that the current danger that is human drivers should be removed from the equation as quickly as feasible.
If we are building the safest transportation system, what role do driverless vehicles play? Wouldn't that be the narrative that actually saves the most lives?