Google says self-driving car hits municipal bus in minor crash
reuters.com
reuters.com
The buses will drive right next to the lane marker, with their mirrors hanging over into your lane. This makes it so everybody has to creep over a little bit into the drivers side lane and hopefully everybody in traffic has a small enough car to deal with it. Otherwise, you have to hang back behind the bus as if it's in your lane and wait for it to get to a stop.
I've had my issues with Google autonomous cars (they drive slow and they used to be exceptionally slow at making right hand turns, causing traffic problems), but in this instance I'm happy to throw VTA under the bus, if you will, and lay blame 100% at their feet.
Here's a link to the DMV Google traffic report: https://www.dmv.ca.gov/portal/wcm/connect/3946fbb8-e04e-4d52...
From reading the above report, it seems it was officially only a one-lane road, but the road was big enough to handle two-streams of traffic. The google car could have just stayed in the middle of the road, but instead, was hugging the right side of the road in preparation for a right-hand turn. Due to sand bags next to a storm drain, the google car had to "merge" back into the one-land road to get around the sandbags. Considering it's still a one-land road, the bus driver should have yielded to any car that was in front of it. I'd place a majority of the blame on the bus.
What I don't understand: Why doesn't the Google car have video of the accident? Or if they do, does anyone know if they will share a video of it?
Right now, lots of people are already blaming the bus, without Google ever actually saying so. Moral high ground for them.
One time I was stopped at a set of traffic lights waiting for them to turn green. It was on a tight bend and I was about 20cm from the line, and 50cm from the inside of the road. A bus came up and decided for whatever reason he had to get in front of me (even though there was a stop directly on the other side of the junction). There was enough space for a car, but definitely not for a bus as it was on a bend.
He slowly kept creeping forward, and the side of the bus kept getting closer and closer. I starting banging on the window as he got next to me, but no he kept going. I was far enough back that I was out of his blind spot. In the end I had to move onto the pavement.
I then walked to the front and banged on the door, gave him a few choice words and the driver just shrugged.
[1] http://sanfrancisco.cbslocal.com/2015/03/02/muni-driver-acqu...
[2] http://hoodline.com/2015/10/muni-bus-hits-kills-cyclist-on-m...
The worst in my experience are trucks, and even more so, pedestrians (almost every day I have to pre-emptively break before someone walks into the road without looking)
I completely avoid ever being next to them now, I'll just stop and wait if there's any chance
f = ma. The bus is considerably more dangerous than a car, even while traveling at a significantly lower speed. As the parent comment noted, there are certainly some aggressive bus drivers out there as well.
The potential damage may be greater, but significantly lower speeds creates a much greater risk of some accident. All things considered, then, it's not clear what's worse at the bottom line.
The chance of an accident is greatly increased when vehicles are traveling at very different speeds. See, for example, http://priceonomics.com/is-every-speed-limit-too-low/
Car drivers are orders of magnitude more of a threat to my safety than buses. Buses are far more predictable and tend to change their vector of travel more gradually when they do change. Also, the drivers tend to be much more alert and aware of their surroundings.
Also, in general and particularly during commuter traffic, they really aren't that much slower aside from the stops.
I've also never, ever had a bus driver intentionally try to harm me in a bus (like trying to 'muscle' me off the road). This has happened on multiple occasions with car drivers.
edit
I'd also like to mention I love the google cars. They're just so predictable and courteous. And in general, they just follow the rules.
For example, at a stop sign that a human driver reaches first. They have the right of way, but they want to be polite so they'll try to wave a cyclist on...then start to go...then stop and wave....then start to go. There's nothing more dangerous to a cyclist then a driver not following the rules and behaving erratically.
Google car? Stops. Waits. Goes. No fuss. I have seen the older models get "stuck" and wait longer than normal. Though usually this is when other drivers refuse to go before the google car.
Taking into account:
1) Number of drivers removed from the roads by average annual bus use
2) Size of the vehicles
3) Random acts of annoyance
4) Dangers posed by each
To confess my bias - I'm a huge fan of mass transit and a bike commuter who touches his car maybe once a week, but curious how else one might model this.
Sometimes (though rarely) the bus stop is such that the bus can pull out of the traffic lane and become less of an obstruction, but even then, right turns are delayed significantly.
I don't mind this when I see a full bus, because I know that many private cars would slow down traffic at least as much. But damn is it annoying when the bus has 3 people on it (as is usually the case in my non-transit-oriented city).
"From now on, our cars will more deeply understand that buses (and other large vehicles) are less likely to yield to us than other types of vehicles"
It will be an endless process for an algorithm driven car. Buses in Dublin will be so much different from buses in London, or trucks in the US different from trucks in Germany. Then there are the bikes, pedestrians. When human drivers/bikers/pedestrians look at each other and know what to expect. These are the things that would be incredibly difficult to formalize, if possible at all. Unless you have something like a Deep Mind on board, but even then I wouldn't be so sure.
This is already an issue humans solve in the context of tourists driving through town, who aren't accustom to local rules like "First driver at an intersection gets to make a left on green instead of yielding to the column of oncoming straight traffic."
The solution is to either run fewer bigger buses, or to inject a whole load of negative feedback into the system, either by having buses wait for a minute at a few stops until their scheduled departure time (which involves making the route slower) or by real time tracking with radioed instructions to the drivers, like they do on some underground train systems.
Buses need bus lanes or they will not be on time and add to traffic.
[1] http://www.bus-bild.de/bilder/hess-trolleybus-bgt-n2c-16700....
Every bus you run requires a qualified employee to drive it, and putting them on for only an hour or two per shift would make it even more expensive (what employee would want to work such minor shifts?)
Humans are actually great drivers when the situation requires thinking. Lot of snow and can't see the lane markers? Millions of people adapt every day to this during winter. Lots of pedestrians, bicycles, motorcycles, etc. doing somewhat unpredictable things? Again, look at any Asian mega city, people can adapt very fast. Basically, if the situation requires being alert, people are very good at driving.
When are people bad a driving? Whenever it's monotonous. Bumper to bumper traffic or a free flowing freeway. Constant repetition of red/green cycles while going down a suburban street. These boring situations make a lot of people basically turn off their alert thinking. Then they do other things like texting, talking on phone, etc.
And these boring situations are exactly where AI self driving is better at driving than people. Computers never get bored. The self driving car will be at 100% attention even during the most boring traffic. But when boring suddenly turns to not boring? The current state of AI is very very bad at this.
And unfortunately there's no easy way to use the best of both sides. If the AI is fully driving, then it will do great while the driving is boring. But by the time the AI decides, this is too much, can't handle this, it's too late to alert the human driver to take over. But if the human driver is required to always pay attention, what's the point of the self driving car?
Self driving cars will get there someday, but I think it's much farther away than many assume it will be.
Having driven a corridor with a ton of snow quite a lot (western PA to washington DC), i'm going to strongly disagree with this.
People don't "adapt". They just do it anyway because they have to. The rate of minor and major fender benders in this kind of situation is really high, it's just "acceptable".
In truth, computers that understand the physics of what is happening to the car at a given point in time based on tons of sensor data are going to be much more effective drivers in snow and ice than people (because of even simple data it can use like "this wheel has this much traction currently, so i need to correct this much"). People are really bad at figuring this out in their head and use horrible heuristics, often hitting other cars, guard rails, you name it.
You could also can get data about surrounding conditions (IE the ice the wheel will hit next or whatever) that people can't because they are not close enough/can't process fast enough (IE not just what is going on with the car, but what is the state of the ice on the road 2 inches to the right or left of the car and would it be better for the car to steer in that direction to get more traction, etc)
I remember driving home in a massive rainstorm at night about two years ago with around 2 meter visibility and was going like 20mph because I could not see shit. I had less sensory data than a computer with extra vectors of vision beyond sight would have had, and my reaction time is a ton slower.
Then a year later I'm driving in icy conditions and count five cars that either off roaded or rear ended each other in the span of 15 minutes because of the slick surfaces. I imagine self driving cars would do a much better job slowing down and budgeting for potential slippage than impatient humans.
There is nothing really magical about our eyes that give us better clarity in awful driving conditions, and we don't have radar and GPS fed into our brains to give us a clue where the roads are supposed to be - we just rely on historical context and past driving experience on the same roads.
How many hours would it take a networked fleet of self driving vehicles to outdo my historical context for driving on any of the roads I've driven on for a decade? Two? Maybe three? You just need to get the tunable parameters right though some good old genetic algorithm work like what Google is doing now with its test fleet and while it would be impossible to guarantee perfect performance in every scenario (because inevitably an earthquake is going to make a self driving car crash where nobody could have prevented it) it just has to be better than us, and that has almost certainly already happened.
Maybe not. On March 11, 2011, a magnitude 8.9 earthquake hit Japan. The first seismometer to detect it (out of 92 in the system) was on Kinkazan Island, just off the eastern coast of Japan. Each seismometer has some compute power and is constantly computing a hazard level based on the waveforms of earth movements. The Kinkazan unit did what it was supposed to, and sent an emergency shutdown signal to the Shinkansen high speed rail system. Not through the signaling system; the circuit breakers at power substations were immediately tripped and power to the trains cut. The trains treat a power cut as an emergency stop situation and apply the brakes hard. All trains stopped safely.
[1] http://www.railway-technology.com/features/feature122751/
I'm a relatively new driver (1.5 years or so since taking my license) but this happened to me twice, i.e. catching a rainstorm so big that I could barely see in front of me. In both cases I decided to stop in a refuge on the right-side of the road, waiting for the weather to improve. Sure enough, after about 20 minutes of waiting the rain eased down. What I'm saying is that under some circumstances (like close to no visibility) one should just stop and wait for the external factors to improve.
You do see a spike of people who don't do well to adjust to winter driving. They don't give longer gaps, make sure they have more time before crossing intersections than they do. Computers will do that much better, eventually. I can't wait to go to work or drive around the city and not have to worry about someone texting, doing makeup or making poor driving decisions. Have a nap on long trips, watch a movie, finish up some work.
I can see computers having an amazing reaction time to some kid running into the street to catch a ball. An animal running in front of you at night. It's going to take awhile though. I wouldn't trust it right now to safely take me anywhere. There are just so many variables that the engineers have to account for. Weather conditions, people conditions, car conditions and on and on. It's nuts.
There is a slight problem. In practice, in actual measured statistics, we're utterly fantastic drivers despite all the reasoning.
We need a baseline. How about trains? Trains in the PTC and semi-PTC era don't (can't) speed and are monitored to crazy 1984 levels, usually there are two employees in the cab (not counting subway type things, I mean real Amtrak here). Trains don't have steering wheels, usually can't spin out, etc. Still the death rate is about 1/2 per billion passenger-miles. What can I say, brakes fail, drivers have heart attacks, rails crack. Its apparently impossible to put squishy humans in a sealed box and toss that box around at 90 MPH without killing about one per two billion miles. Note that's to the moon and back 4000 times, safer than even spacecraft.
The death rate for passenger cars is about 7 per billion passenger miles. A little higher for trucks/SUVs, a little lower for cars (cars are safer on the road, more stable).
That's only about 14 times worse than trains.
Anecdotally human drivers do dumb things, because there's 7 billion people and world wide networks every form of idiocy is known to all, yet simultaneously its very rare. Anecdotally a billion passenger miles isn't very far, so every year we stack bodies like cordwood, sadly. But even on the ultimate in safety, a train, we'd still have 10th the number of dead. By reasoning the ratio should be immensely larger than 14, like 1000 or 1000000. But its 14. That has to be answered...
Even crazier to consider is the death rate has dropped by a factor of 3 since the 1970s due to better engineering and regulation. We KNOW how to lower death rates using existing techniques. Yet we also know software written by humans is basically worthless. So given a choice of strategies to further lower death rates, given the track records I'd trust the MechEng and CivEng a lot more than the CS department.
A factor of 14 reduction in death rates is unfortunately far more likely to come from the MechEng/CivEng grads than the CS grads, and I say that as a CS grad.
"Haven't yet attempted" isn't a "cannot."
The main rule for driving in snow is to assume everyone else is actively trying to kill you. :)
It's something I was told when getting a motorcycle at 17, and it's served me pretty well.
Specifically on snow I think we subcontract too much from brain to ABS, traction control, lovely powerful heater and what not. We don't get Scandinavian amounts of snow in the UK, so whenever we do have snow you'll see dozens of ridiculous accidents because people don't take account of stopping distances or speed, or ABS being a hinderance in snow etc.
In any case, you need good tires and working brakes to navigate snow effectively. Anti-lock brakes (in my opinion) are no better than a person that can properly pulse their brakes but ... if you've got even a slight warp in your rotor, you won't brake evenly enough to stop in the snow. The other problem in PA is PennDOT's "new regime". There's been far more ice packed onto the roads when it snows lately - because melting snow that's packed under tires turns to ice. I'd far rather drive on dry, crunchy snow (for traction)
/rant
The hard part of this we're trying to force self driving into our current infrastructure and traffic. If the bus was self driving, they would have communicated properly and merged perfectly.
This could not be less true among people driving in unfamiliar conditions. (http://legacy.wusa9.com/story/news/local/virginia/2016/01/20...)
The difference is that all self-driving cars benefit from the knowledge created by bumping into the edge cases. People can only do this in an ad-hoc manner (i.e. personal experience, taught by mentor, etc)
First snow fall in Calgary looks like what you get in a standard snow storm in a city in the south east. Its even funnier if the first snow fall is early and then there are another two months of sun and no snow and then you get a repeat during the second snow fall. People have to relearn how to drive in snow every year.
People are bad drivers in good conditions, they just happen to be able to be bad drivers in bad conditions as well.
In my experience Saskatoon fares a lot better. They don't salt as heavily because it's often cold enough that salt doesn't work as well, and because judges don't blame the municipality for not salting when there's an accident.
We defined self-driving cars: When I can pull out my phone in any of at least 5 major metropolitan areas around the world, order a driverless car, have it pick me up and deliver me to a specified location within the city in a timely fashion, with as little risk as getting in a taxi.
He is sub 20 years. I'm thinking more like 40-50 mark.
My thesis? We can solve the technical challenges, but so much physical, legal, and regulatory infrastructure needs to change to make this viable that it will be more than 2 decades until we see self-driving cars as a reality across multiple cities. Of course, I would be more than happy to lose the $100 as it would make everyone's lives better that much sooner. But I'm skeptical.
Perspective:
* 60 years ago, we didn't have highways.
* 45 years ago, seat belts weren't required.
* 40 years ago, we expected engines to last 10s of thousands of miles. We didn't really understand crumple zones. Open containers and even drinking while driving.
* 30 years, we added fuel injection, computers, and airbags.
* 20 years ago, the hybrid car, better fuel efficiency, and side airbags
Imagine what we'll know tomorrow.
Vast sums of money have been sunk into developing exceptionally reliable avionics systems since then.
Passenger carrying aircraft still all have a pilot or two.
Not quite. Here's the US highway system in 1955.[1] Construction of the US Interstate Highway System started in 1956.
[1] http://www.mappery.com/map-of/Pre-Interstate-US-Highway-Syst...
What?
> 30 years, we added fuel injection, computers, and airbags.
But we knew about them for a lot longer than that. Fuel injection was used in aircraft during WWII.
It could be a lot more straightforward than you suspect. The three things we have going for us:
1. Car accidents are almost always resolved by deals made between insurance companies (this will be doubly true with unpiloted cars, which will presumably not commit criminal acts)
2. The same few large companies are likely to deal in insurance of piloted and autonomous vehicles, so there's not a huge incentive to go sue-crazy
3. If autonomous vehicles cause significantly fewer accidents and companies can charge the same amount for insurance covering them, that is a huge win for some very large companies with the power to get the ball rolling.
Increased vehicle safety will be a big win for these companies in the short term, although in the long term some price discounts might be factored into the cost of autonomous vehicle insurance.
Self-driving, electric cars will make future taxi-equivalents affordable enough that it's considered for many instances. That makes for a massive market.
It's not as though current taxis have a sparkling reputation as drivers either. Not many will be sad to see the end of them.
Let's go with that. Say that tomorrow Google finished the development of the car. It works very well.
Than it just need to find a single city that would agree to a wide scale deployment. And I imagine there are plenty of reasons why a city might want to do that. Heck if Google cared, they could legally bribe politicians. And it doesn't even have to be a city in the west. It could be a city ruled by a dictator somewhere in africa. Or it could be a self driving car created by a chinese company, and china would be willing to take the risk to lead in this field.
So maybe the first city isn't as hard as it seems.
Now what happens after a car has proven itself in a city for a year or two ?Showing almost zero accidents, and all the other benefits of a self-driving car. How much will people everywhere would want it ? how much political support would it gather ? i imagine a lot.
Would such support and proof be enough for relatively rapid deployment of self-driving cars ?
If you're talking about _everyone_ in major economies having a driverless car, yeah, I think 40 years is realistic given the replacement rate for cars alone, nevermind the legal issues.
¹ Additional car inspection, criminal record check on the driver, etc.
https://en.wikipedia.org/wiki/Taxicabs_of_the_United_Kingdom...
Food for thought.
The problem with driving is you have to consistently do it well. You can make zero mistakes 99.99% of the time and be a terrible driver. One accident per 10,000 minutes would be ~one accident every year. Worse, a lot can happen every minute, so it's closer to not messing up for 600,000+ seconds.
The only reason people can drive is we build in very wide tolerances. ex: Car lanes are far wider than strictly necessary, you can fit ~2 cars in the average lane. Traffic lights have a significant delay where they are all red etc.
And other drivers react to minimize the problem created by another driver.
Coming from a smaller, more suburban area, I learned to recalibrate my driving to be more aggressive in larger cities (primarily NYC & DC), to hold my own against pushy taxis, etc. Most city drivers kind of bluff, and if you show you're not falling for it, they'll stay out of your way.
But city buses don't bluff. If they start pulling into your lane, even if clearly cutting you off, you better slam on those brakes, because that bus is not stopping. It's one of those unwritten rules you learn, at least in my experience driving in major East Coast cities.
I can see how it'd be hard to train AI that you have adjust the driving style depending your locale and for specific types of non-emergency vehicles.
I'm not making an argument here, really just sharing a chuckle about the crash involving a city bus.
It's true that most deaths and serious accidents occur on highways, where human factors like boredom and sleepiness are probably significant. But humans also have a heck of a lot of fender benders like this incident, and a lot of the ones I have seen weren't exactly situations where "a lot of thinking" was necessary. People are texting and rear-end the person in front of them, or underestimate how much room they have to switch lanes between two cars, or other things like that. I don't see why these situations are uniquely difficult for AI to solve. In fact, they seem like very apt problems for computers to solve.
The problem is: real-life driving does not only consist of situations that even a child could easily manage. Humans are extremely good at letting their driving skills degrade gracefully when conditions get rough. AIs? They drive perfectly until they reach their limit, then they suddenly pull off epic fails - if no human driver is on standby to resolve the situation immediately.
Statistically, Google's Self-Driving Car would've lost it's license by now, if not for human drivers keeping it in check.
Why did the bus move into an occupied lane?!?
It depends on the city, but bus drivers often don't care whatsoever about other vehicles in the way. They merge and let others avoid contact.
Because that's what busses do all the time, whether or not that is safe, legal, or desirable.
Seventy-four per cent of road traffic deaths occur in middle-income
countries, which account for 70% of the world’s population, but only 53%
of the world’s registered vehicles, burdens for 74% of world's road
deaths. In low-income countries it is even worse. Only one percent of the
world's registered cars produce 16% of world's road traffic deaths.
The promise of an AI driver is that there will never be a second accident of the same type. The only question is how fast the learning rate can ramp up.
To me, it makes sense to make use of the brain power provided by the occupant of the vehicle. For example, if one of the main vision sensors in the car breaks, why not let the human occupant drive the car manually?
[0] http://www.amazon.co.uk/Our-Robots-Ourselves-Robotics-Autono...
[1] http://www.econtalk.org/archives/2015/11/david_mindell_o.htm...
If you have years of experience driving a car, maybe it'll work. For some time. But what if your last manual intervention was five years ago? Or twenty years ago? The rarer handoff happens, the more stress it puts on the manual operator, and the more likely things go wrong.
It's much better to have a safe default reaction (stop the car, shut down the reactor, ...) that kicks in when normal, automatic operation can't continue.
Which isn't really any different than any other sort of major mechanical failure.
1) It hands control to an out-of-car human driver, similar to how military drones and driverless trains are operated now.
2) It turns on the hazard lights and pulls to the side of the road. Another vehicle is dispatched to pick you up.
Yes, driving is actually a challenging task. But guess what, humans are BAD AT IT. And are computers going to be better? I'd be pretty damn surprised if they weren't.
Most crashes come down to distraction. Computers never get distracted. So if we can workout errors like the one in this story, then at least we can say the computer isn't going to get lazy or distracted and suddenly forget it's lessons learned.
Yes, its a challenging road ahead for these self driving cars. But I'd feel so much safer in a world of autonomous vehicles than I would with the morons behind the wheel of today.
- not using signals at all, just randomly changing lanes
- not using signals at all and on the top of that do not care about traffic in the other lanes
- driving 45 mph in the left lane because California teaches to new drivers that changing lanes are dangerous, aka "stick to your lane" (I learned that recently and I could not believe)
- driving with the exact same attitude in rain like under dry weather conditions
- do not care about people merging lanes, do not look at them, pretend they dont exist
- being on facebook on the highway driving with 65mph
- do not care about bicycles, pretending they are not legitimate traffic
I am not sure about the rest of US, but several friends told me it is specific to California. I would argue that humans are not as good drivers, at least there is a large part of drivers who aren't. Replacing drivers with software is always going to win, just like it did with airplanes or anything else repetitive task that we automated. I am pretty sure there is going to be a period while AI is not as good as humans but this is why Google is collecting to data at large scale to improve it to beat humans and make driving a safer way to commute.
It doesn't seem too hard to go in that direction. For example you can use satellties or drones or other sensing mechanism to survey a city in real time and tell the driver X minutes in advance before he would need to take control of the wheel - and of course you could adapt this recommendation according the skill of the person, his current state(being drunk) or other data.
(Source: I'm a motorcycle rider, we have a term here "SMIDSY", which stands for "Sorry mate, I didn't see you" - because it's such a common thing for drivers to say when they've just driven into you. There's a great quote from a judge a few years back in a court case over a driver excusing themselves for running into a bike "But the plaintiff was _clearly_ there to be seen.")
Have you seen the accident rate in these places? It's huge, and in many countries underreported.
Imagine how much easier it would be, for example, if roads had embedded markers (like magnets or RFID chips) that designated lanes, speed limits, and other road rules and weather conditions. AI vehicles would simply need to read this information and adjust accordingly.
But, unfortunately, this would be almost unfeasible in today's world vecause a) it depends on every car following the same rules and that's impossible if you have human drivers and b) it would require massive investment and national-level cooperation to change the infrastructure.
Speed limits can be defined as metadata on existing mapping data. Existing vehicles, such as Tesla's with their autopilot system and GPS receivers, can "define" existing lanes by refining existing mapping data using their drivers as expert system trainers [1].
You don't need national-level cooperation to change infrastructure; you just need a massively automated system for collecting data and training based on it. Google and Tesla are doing both, just in different ways.
[1] http://www.teslarati.com/tesla-building-maps-autopilot-drive...
You're talking about 13 million RFID chips that would need to be installed on 164,000 miles of roadway - just to get the interstate highways covered.
Now think about the fact that these chips are going to need to sit static and not be replaced for more than 15 years at a time. So you have to come up with a way to house them such that they will not be affected by > 150 degree heat or by < -50 degree (F) cold.
And then think about the fact that it takes ~6 months for the government wheels to spin in the right direction for a pothole to be filled in.
The project to put RFID, or any new technology on the roadways in the US will have to involve federal cooperation as well as the governments of 50 states and countless local governments. It will be 100 times more expensive than any other venture using the same technology.
Look at the Big Dig project in Boston. It took 25 years, $24 billion and several acts of congress to get it to a state where they could start working out the kinks like "why do the guardrails keep killing people".
The scale of solving this problem "the right way" is incredible. I don't disagree with you entirely. But I think any change like that is a solid 100 years out into the future.
We need visual processing and heuristic logic to handle a kid on the side of the road chasing a ball unlikely to intersect and catch it before he enters traffic. Given that mandatory ability, reading a speed limit sign is pretty trivial.
Given that its very unusual for an oncoming car to swerve into your lane, and a giant pothole in the oncoming lane, and the car sees three cars swerve into your lane taking turns, what are the odds of the next car hitting you head on? Compared to this puzzler, interpreting a stop sign is pretty easy.
Also the human drivers presumably will not understand the weird behavior of an enhanced self driving car, unless the human infrastructure matched the automatic infrastructure. Of course you could analyze the human data visually to predict the human behavior, but then why bother with the automatic hardware sensors?
I guess a bad automobile analogy is if you must design this pickup truck to safely haul a 8000 pound trailer, then a 300 pound trailer will not overtax it.
On the other hand, making this mandatory would also be the wet-dream of most governments. In the name of safety, they could have access to so much information... Just knowing how many passengers the car is carrying, what's the weight of each passenger, and, of course, a history of the coordinates, would open insane possibilities, not all of them good.
"Boring" versus "not boring" is too vague for this statement to really be meaningful. Examples could be drummed up on both sides. For example computers beat humans hands down in at least some criteria fitting the "not boring" description, like reaction time and seeing in multiple directions at once, and probably some others.
Self driving cars being better than the average driver is absolutely to be expected since they can only get better. They'll be consistent. They will always be focused, not get distracted, not get tired, have more sensors than a human could. It's not just about the environment but also about the mental state of the driver.
Sure it's difficult but your using one case to flip things around doesn't seem to be a very strong argument.
> Self driving cars will get there someday, but I think it's much farther away than many assume it will be.
There's no evidence to support your assertion given the tests that have been going on so far. There's no evidence to support that switching back and forth would be better than having the AI all the time either.
This gets brought up every so often, much to my chagrin. If you asked 10 people what qualifies a good driver, you will get 10 different answers.
Does it mean they have a good safety record?
Does it mean they are skilled at controlling the car?
Does it mean they can navigate traffic efficiently?
Does it mean they communicate with other drivers well enough to not negatively affect traffic?
Does it mean they have good reactions times?
What combination and weighting to you provide the above questions and what questions haven't even been asked?
Of course people respond that they are above average or good drivers since everyone is using a separate metric to define a good driver.
Let's suppose you are given a 2d field in which there are a number of rectangles, each of which have a velocity vector. So you know the location of each rectangle now, and, to a good approximation, in the near future.
How difficult is it to program a path-finding algorithm that will not crash into any of these rectangles, given that lowering speed is always a good option within a city?
I'd say this is really not all that difficult. Yes, there are a lot of special cases to consider, but still, in essence, this seems to me a quite simple problem.
So breaking down the problem is key. If you look at an autonomous vehicle as a black box, then, of course, it will seem dauntingly complicated, and difficult to trust.
hard AI problem, computer vision
> to a good approximation, in the near future.
how near is near?
> lowering speed is always a good option
can't brake if someone faster is behind you
> there are a lot of special cases [..] simple problem
simple problems by definition don't have many special cases, the "simplicity" means you can abstract them into a few cases.
in other words, talk is cheap show me the code
Large swarms on the highway, small swarms (or solitary driving) elsewhere. Best of both!
Turn one out of four one-way streets in business districts to robot only. Make a certain highway on/off-ramp robot only etc...
How about using the "self-driving AI" only as "really advanced automatic assistance". Humans will still drive the cars, but the AI will kick-in in those boring situations where the humans turn off their thinking.
That is an awful result. Humans are awful drivers. They are too easily distracted, don't have enough information, bored, and tired.
What makes you think so?
"A Google Lexus-model autonomous vehicle ("Google AV") was traveling in autonomous mode eastbound on El Camino Real in Mountain View in the far right-hand lane approaching the Castro St. intersection. As the Google AV approached the intersection, it signaled its intent to make a right turn on red onto Castro St. The Google AV then moved to the right-hand sid of the lane to pass traffic in the same lane that was stopped at the intersection and proceeding straight. However, the Google AV had to come to a stop to go around sandbags positioned around a storm drain that were blocking its path. When the light turned green, traffic in the lane continued past the Google AV. After a few cars had passed, the Google AV began to proceed back into the center of the lane to pass the sand bags. A public transit bus was approaching from behind. The Google AV test driver saw the bus approaching in the left side mirror but believed the bus would stop or slow to allow the Google AV to continue. Approximately three seconds later, as the Google AV was reentering the center of the lane, it made contact with the side of the bus. The Google AV was operating in autonomous mode and travelling less than 2 mph, and the bus was travelling at about 15 mph at the time of contact.
The Google AV sustained body damage to the left front fender, the left front wheel and one of its driver's-side sensors. There were no injuries reported at the scene."
[0] - https://www.dmv.ca.gov/portal/wcm/connect/3946fbb8-e04e-4d52...
Interesting to note that the collision was very minor (car moving at 2 MPH). When looking at accident rates we should also consider the severity.
Here's my artists rending. The vertical lane is just the right hand lane, and is "wide", but still one lane. The bus is traveling "up" and the Google AV tries to manenouver around the sandbags blocking it's way. Thinking that it can get around in time, the AV starts moving to the left and crashes into the bus.
I guess the "one lane" aspect is the confounding variable, but since the autonomous vehicle was stopped, then it takes on the aspect of any other stopped vehicle on the side of the road (ie giving up it's right of way). Just like if you have a road where cars are allowed to park on the side of the road, those cars cannot enter traffic unless it is safe to do so.
Edit: Actually, I have no idea. Did a quick read of California's right of way laws, and couldn't find anything that jumped out that would cover this situation.
| |
|
| +----
| SS+----
<-[G]|
| ^ |
| |
|[B] |
[B] |
| |But, if the AV was not stopped, I think the bus is now at fault for attempting to pass a moving vehicle within its own lane, which is illegal (I hope).
https://www.google.com/maps/place/El+Camino+Real+%26+Castro+...
Bus drivers in the Bay Area are notorious for ignoring traffic (and pedestrians). Apparently there is some indemnity or statutes that make suing either the transportation agency or the driver nearly impossible and so pretty quickly people learn that the bus drivers drive with impunity. Plenty of stories posted to the local traffic column in the paper, and shared amongst neighbors and in the department of safety's "blotter" feature.
Google needs to go back and program their cars to always assume that buses are out to get them and avoid them at all cost, They are an active traffic hazard often operated by a disinterested and distracted driver. The only way to "win" is to not be where ever the bus is.
...I remember once in Zurich when my local bus was two minutes late. The people at the bus stop were really quite cross.
I’ll report what I observed in 2 years of 3 to 4 times of using the bus a day in Germany.
Every time a bus was late, or tried to break the sound barrier™ in the hope of reducing some of the delay, was due to one of the following issues:
(a) A tourist with texan accent trying to get onto the bus, discussing with the bus driver if he can pay with credit card or in dollar (no), then asking the bus driver to wait while he’s going to get money from the nearest ATM (happens at about 5% of stops in the downtown areas where the tourists are)
(b) Rush hour traffic, 200 people squeezing into a single bus, and another few hundred waiting at the bus stop – busses coming every one or two minutes, and it takes quite some time until people stop trying to get into the bus, and leave enough space for the doors to close
(c) some kids with invalid tickets trying to cheat and getting caught
These issues can’t be fixed by automated busses, or trains.
Only by less tight schedules, and more busses and trains.
It could be more easily fixed by accepting foreign credit cards (Gothenburg manages this, London if the card is contactless) or telling drivers not to wait.
(b) is partly solved if the cost of running buses/trains is significantly reduced, as that leaves money for extra vehicles.
(c)... well, who will notice?
Here's an easy solution: every bus stop has a kiosk that dispenses a RFID bus pass if you put in your payment info. Then, when the bus comes, you can enter from any door and the RFID receivers detect your presence. Attempts to board without a bus pass are detected with weight sensors and a camera network (current computer vision techniques are probably reliable enough for this application; they'll be even better in a few years). Unauthorized persons are advised to get off at the next stop by audio recording and identified by displaying a snapshot of their face (taken earlier on the bus). If they don't get off the transit police are summoned.
Put a fare vending machine which accepts cash and credits cards on the bus, which also makes it very obvious how to get a single fare. (This exists for example for trams in Berlin, modulo the accepting credit cards part)
i.e. the passenger would be prevented from boarding the vehicle without a valid ticket and provided there was sufficient capacity for a speedy and comfortable boarding.
This is a case where perhaps the computers have too much imagination. We actually tell human drivers to drive that way, but as humans we all know that we don't really mean that we need to worry about someone driving in front of us suddenly slamming their brakes, drifting 180 degrees, and driving at us full speed. Tell a computer to assume too much malice and the car will refuse to even move, because it's pretty easy to specify the search algorithm that will find that outcome.
We have to specify the exact level of malice the computer can reasonably expect, which is way harder. And it will still, by necessity, at times be an underestimate.
I mean, he's driving a bus and he thinks he has the right of way. This effectively means he has the right of way.
Google needs to increase the settings on "Defensive Driving".
Just because the Google robot drives in a reasonable manner doesn't mean the humans around it will.
Perhaps each autonomous car needs to have a cultural knowledge, based on location.
All the time I'm sitting there at a stop when someone pulls up and stops with no stop sign, then honks and waves at me to go on. No. I have a stop sign. You don't. I am not going. The last thing I need is to get half way through the intersection and you plow into me and claim I ran a stop sign.
No. Follow the damn law. If you stop first, you go first. If I have a stop sign and you don't, you go. If we stop at the same time, the person on the right goes first. It's the fucking law. If you can't figure it out, go back to driving school. Don't wave me on, because the cop and the insurance company don't take much stock in "waves". They expect people to follow the damn law.
It's not defensive driving. It's dangerous.
Of course I don't really like 4 ways in SF where you have to fight tooth and nail for your right of way, but some happy medium where everyone just efficiently goes when its their turn would be nice...
A good percentage of the time two people end up trying to enter the intersection at once because hand signals are not a clear communication method for signaling right of way.
There's no such thing as merge, there is only shove!
Edit: No blame devolves to you. As I said, it's a cultural thing.
Here in South Australia public transport buses are extended unofficial better right of way standing (when departing from the curb), for example. There are also new laws determining the required minimum passing distance between a cyclist and a car to be 1m if travelling 60km/hr or less, and 1.5m if travelling > 60km/hr; and car drivers are allowed to cross double-white lines to avoid cyclists.
It seems to me there are ample opportunity for 'edge case' bugs to occur in the autonomous vehicle software when local conditions are taken into account, and an untested set of requirements are applied specific to a country / city / district.
So true. There are no worse drivers on the roads than MUNI bus drivers. They are terrifying to drive near. I've nearly been crushed by MUNI buses multiple times, while driving in my own lane minding traffic laws.
Regardless, both (and A/C Transit and BART) have extremely strong unions.
My personal fear is that Google and maybe one or two others will get self driving cars right, but then the imitations from other manufacturers will fall short. The liability needs to end up on the manufacturer of the self driving car system, this is not something to be taken lightly at all.
I actually agree that vaccines (and really drugs more broadly) are a rather unusual example of products that can be used as directed and things can still go bad--without the manufacturer necessarily being "at fault." [Edit: Where it's not the fault of the user or other human.]
We'd probably also need some amount of limits on full tort liability though since the manufactures would be bigger targets for lawyers than individuals.
But a car crash will be very easy to figure out who's at fault. Why would such a fund exist for something that can be pin-pointed directly at the party at fault?
People should still be compensated when accidents happen, without the drag caused by trial attorneys attempting to extract as much as possible from self-driving vehicle manufacturers (which is going to slow down progress).
The benefits of self-driving vehicles from reduced fatalities and accidents alone are so great that a process and funding needs to be in place to allow continued innovation (if done with safety put first).
At its current state, self driving cars are not perfect. Roads were designed to be driven by people and it may take many years for road infrastructure and total adoption to properly support self driving cars. The one thing many will be able to agree on is there will be accidents and some people will be hurt at the fault of self driving cars but ultimately removing humans from driving will make roads a much safer place. The reason behind this particular vaccine fund (see link above) appears to align with problems that we will soon see as self driving cars hit the road and make mistakes but for the overall good.
There are lots of scenarios where a person avoids an accident that isn't clearly another driver's fault--skidding on ice, debris on road... Also cases involving pedestrians and bicyclists where it's arguably their fault but drivers still generally have the responsibility to not hit a child running out into the road.
It really shouldn't be very different from, say, a fatal accident from a driver in a corporate-owned car.
Sure...PIP attorneys are going to go after Alphabet and make a big ruckus, but in the end, Alphabet has to have all this covered already as part of the plan.
Here's where it happened.[2] You can see traffic cones around the storm drain.
This is a subtle error. Arguably, part of the problem was that the AV was moving too slowly. It was trying to break into a gap in traffic, but because it was maneuvering around an unusual road hazard (sandbags), was moving very slowly. This situation was misread by the bus driver, who failed to stop or change course, perhaps expecting the AV to accelerate. The AV is probably at fault, because it was doing a lane change while the bus was not.
Fixing this requires that the AV be either less aggressive or more aggressive. Less aggressive would mean sitting there waiting for a big break in traffic. That could take a while at that location. More aggressive would mean accelerating faster into a gap. Google's AVs will accelerate into gaps in ordinary situations such as freeway merges, but when dealing with an unusual road hazard, they may be held down to very slow speeds.
I wonder if Google will publish the playback from their sensor data.
[1] https://www.dmv.ca.gov/portal/wcm/connect/3946fbb8-e04e-4d52... [2] https://goo.gl/maps/QzvVXQGxhX72
Yes, probably true, but not crystal clear. The Google car never left the lane, so it comes down to subtle questions about appearing to be parked or impromptu division of lanes near right-hand turns.
When Cruise (YC 14)'s car hit a parked car at 20 mph in SF last month, there was no press attention.[1] Even though it was across the street from the main police station.
That Cruise crash is an example of the "deadly valley" between manual driving and fully automatic driving, The vehicle made a bad move which prompted the driver to take over, but too late. This is exactly why AVs can't rely on the driver as backup.
[1] https://www.dmv.ca.gov/portal/wcm/connect/bc21ef62-6e7c-4049...
Clearly the algorithm does not take into account the classic attitudes of bus drivers.
If the asshole can't bring himself to drive civilly, eventually he'll lose his license on points and have to ride around in a self driving car!
Here in Europe the attitudes of BMW drivers are not highly regarded.
Most of us are going to expect nothing short of perfection from these machines to really trust them.
I get the feeling the it's the same way with cars just an order of magnitude bigger of a change since they're such a part of our lives.
We absolutely wouldn't be OK with elevators that now and then fail in a dangerous way. Doesn't mean it never happens but it's considered to be someone's fault when it does.
Similarly, any modern passenger jet is doing the same and steering it. It can even land without the pilot's help.
We have already passed over the torch to the machines and self driving cars are just the next logical step I guess.
When you drive, you only look at one thing. If you glance at the rear view mirror, you're no longer processing the front. If you look at a side view mirror, you're no longer looking at the front or back. All this is not counting blind spots. But imagine if you could see all around the car, all the time, and process all of it with the same level of importance.
That's what self-driving cars can do.
People I speak to are usually receptive to that, but there's still the lizard brain aspect, one even I'm victim to. ;)
Imagine that in the world as it stands today that the accidental death rate of auto crashes is 100.000 per year, and we'll call that Earth One. Now imagine a world in which AI reduces the accident rate to 20.000 per year, and we'll call that Earth Two. Given that this is two different worlds, and the types of accidents that human and AI drivers will get into are likely going to be different, then there is likely going to be a large number of people who die in Earth Two that would still be alive in Earth One.
In other words, if AI drivers become the norm, there are some subset of people who are going to die, but would have been alive if AI drivers did not become the norm.
Luckily, we don't live in counter-factual worlds like that, or have knowledge of other timelines, so we're spared from knowing that this would be case.
I hope not. I'd like them as safe as possible, but I already expect that they'll be better than human drivers, and that ought to be sufficient to allow them.
However, there's another issue in play here: in addition to the possibility of holding self-driving cars to a higher standard than humans, people like the feeling of control, and will feel "better" about an accident where they see someone to blame.
It doesn't mean they are bad drivers, but not being in control gives you a very different view of the situation
I guess since you don't know their "internal state" you don't know how recklessly they are driving actually
Robot-driving is just one more layer on top of that. And not one that frankly, seems substantially less safe than autopilot on planes, given how unreliable we ourselves are when it comes to driving. But sure, the emotional impact of hearing a self-driving car malfunction is always going to be emotionally stronger -- i.e. in the man bites dog way -- than the daily fatal accidents that happen to other people that we filter out.
I still think that self-driving stuff should be more enhanced cruise control and less "there is no steering wheel or controls".
Google has been making arguments along these lines -- for instance: http://gizmodo.com/why-self-driving-cars-really-shouldnt-eve...
Now, expecting perfection, and especially giving up 'better than what we have' in exchange for perfection is a bad thing, but but I do think it's completely reasonable to expect the self-driving cars to be way better than an average human, even better than the best humans, just because we lose so many people to auto accidents every year.
I think it's completely reasonable to want to reduce the risks associated with transport, and I think the only politically possible way to do this is with self-driving vehicles, because I don't think it's politically possible to remove the worst half of drivers from the road without offering them similar levels of mobility.
> Most of us are going to expect nothing short of perfection from these machines to really trust them.
I am not sure if this is satire or oblivion.
"Google said in the filing the autonomous vehicle was traveling at less than 2 miles per hour, while the bus was moving at about 15 miles per hour."
Google said in a statement on Monday that "we clearly bear some responsibility, because if our car hadn’t moved, there wouldn’t have been a collision. That said, our test driver believed the bus was going to slow or stop to allow us to merge into the traffic, and that there would be sufficient space to do that."
it was a [double] bus. It isn't the kind of vehicle one stops out of courtesy to let the other car to merge (until such courtesy is the only option for the other car to merge, obviously)
Anyway, it is an obvious software failure as the Google car hit the side of the bus at slow speed. There is no reason to blame it on somebody.
Love this. I'm shamelessly rooting for self-driving cars, and crashes are inevitable. Having the human in the car agree with the computer brings a lot of credibility to the report and follow-up.
Understanding the expected behavior of other drivers is critical to making self driving cars work. And that seems like a pretty hard thing for a computer to figure out.
You're definitely right.
With regard to this particular instance, based on my experiences in many cities in different countries, buses pulling out into traffic when they don't technically have the right of way is an expected behavior. :)
Yeah, I was more trying to make an observation about general behavior of buses. When I read the article there was no mention of fault or reference to a report, so I didn't know. But I do know that buses disregarding right of way is fairly common. "With regard to this instance" was probably too strong/literal of a phrase to use to tie-in my observation.
Looks like the far right lane has double the normal lane width to accommodate on-street parking. The Google car, looking to turn right on red, attempted to use the right-most part of the lane to pass cars that were waiting to go straight, but there were sandbags covering the storm drain near the corner, so it had to stop. When the light changed, the cars it had passed continued on while the Google car waited for a gap, which ended up being in front of a bus, likely caused by the bus accelerating more slowly than the cars in front of it. The Google car tried to use the gap to get around the sandbags, assuming that the bus wouldn't just plow into it, but the bus plowed into it. Perhaps the bus driver assumed the Google car was parked and wasn't paying much attention to it.
I'm wondering if the bus was cruising along and saw the AV inching out into its lane and just thought "Oh, once they see this big-ass bus coming along, they'll stop trying to merge." And the AV thought "Once that bus sees me inching out, it'll slow to let me merge." And neither one was right.
I would group them into a category along with police cars and ambulances due to the prioritization of speed over safety. (Although, in my experience ambulances, are usually very good a prioritizing safety.)
I don't know if all professional bus drivers' conditions are the same as Toronto, but I can imagine this could be extremely distressing for the bus driver involved. At least nobody was injured.
That's for insurance reasons. You're never supposed to admit fault at the collision in the US. You let the police decide and that's that. It's like the first thing listed on the guidelines for collisions on your insurance card.
[0] http://www.reuters.com/article/us-google-selfdrivingcar-idUS...
I guess its simialr to some roundabouts in London.
I have no idea how an AV would handle that(unless they were all AVs at the junction)
The article make it sound like the AI blindly changed lanes into the bus. It seems most likely that the AI knew about the impending collision and decided colliding was safer than any other option. It'd be great to know what the other options were, but I imagine we'll probably never get more detail.
I get your point, though, and I imagine it could have gone something like this: Car predicts bus will slow down (as reported), moves back to center of lane. Bus doesn't, and the car's sensors inform of it. During the seconds (or the fraction of a second) before the collision, the car came up with the safest avoidance maneuver, and started executing it. Crash.
And play with touch football rules to keep the costs down. Cover vehicles with touch sensors and cameras that record each crash.
> "But three seconds later, as the Google car reentered the center of the lane it struck the side of the bus"
Presumably the car's software predicted the crash before it occurred but was unable to completely avoid it. I'd love to know what the programmed behavior is in these 'unavoidable' crash scenarios. edit: formatting
If the test driver stated that s/he thought the car should yield to the bus, would the test driver still have a job? I would have been shocked if the test driver said the software was at fault.
This sounds like a very strong assumption. When was the last time you saw a bus slow down for a car or anything much? There's a reason "Fuck you I'm a bus" is a meme.
Deleted comment
A Google Lexus-model autonomous vehicle ("Google AV") was traveling in autonomous mode eastbound on El Camino Real in Mountain View in the far right-hand lane approaching the Castro St. intersection. As the Google AV approached the intersection, it signaled its intent to make a right turn on red onto Castro St. The Google AV then moved to the right-hand side of the lane to pass traffic in the same lane that was stopped at the intersection and proceeding straight. However, the Google AV had to come to a stop and go around sandbags positioned around a storm drain that were blocking its path. When the light turned green, traffic in the lane continued past the Google AV. After a few cars had passed, the Google AV began to proceed back into the center of the lane to pass the sand bags. A public transit bus was approaching from behind. The Google AV test driver saw the bus approaching in the left side mirror but believed the bus would stop or slow to allow the Google AV to continue. Approximately three seconds later. as the Google AV was reentering the center of the lane it made contact with the side of the bus. The Google AV was operating in autonomous mode and traveling at less than 2 mph, and the bus was travelling at about 15 mph at the time of contact.
The Google AV sustained body damage to the left front fender, the left from wheal and one of its driver's-sidde sensors. There were no injuries reported at the scene.
Serious question I'm curious about. Is this societal, is it a technology problem (do the way buses work somehow encourage this behavior towards other motorists), is it likely a combination of both?
However an important take away for all bay area drivers from the article is
You are supposed to HUG the right shoulder when making a right turn. I am yet to see a single driver who embraces that principle. It is not just for your safety and convenience of the traffic behind you but also because of the safety of bike riders, if you make a sudden right turn they might hit you and get injured.
From here, more accidents are able to happen without being huge news stories. Undoubtedly, the first time someone dies in an accident involving a self-driving car, there will still be lots of questioning of the technology, but it won't come as a complete surprise, now that smaller accidents have occurred.
http://vignette2.wikia.nocookie.net/headhuntersholosuite/ima...
It is even more amazing that there have been apparently 0 injuries or fatalities. I wonder how these autonomous cars compare in terms of hours on the road to # of incidents (not accidents mind you, I hate that term).
This was a situation where there were about 40 cars merging onto the freeway and they mostly just do a fairly standard zipper merge.
There are hundreds of millions of cars operating in all sorts of conditions with little incident. There are literally hundreds of thousands of preemptive actions, foresight, experience and instinctive actions at play for every possibility on roads that millions of people adopt easily.
To say that's not good enough one must articulate a system that is clearly better or has the potential while avoiding a broad brush with 0.1% incidents or diminishing human drivers. That's an argument of convenience and could reflect a lack of understanding of the scope and scale of the problem. This is just sandbags, there are literally millions of obstacles and scenarios negotiated without incident everyday.
Crashes in snow is not always about bad decisions. It could also be extremely poor conditions that cars should not be in, inadequate vehicle or tyres and computing will not help if hardware is deficient. Presuming the worst of others or jumping to conclusions about their intelligence levels is unsavory and dangerous if used to push something.
An AI vehicle in any crowded Asian city is going to be literally stranded by indecision. On relatively empty and organized roads the computing power needed will barely scratch the surface of what a proper self driving AI system needs to be unless you redesign the roads and place constraints, which then becomes a different discussion to approach carefully.