Self-Driving Cars Have a Bicycle Problem
spectrum.ieee.org
spectrum.ieee.org
Summary is that "people riding bikes" (PRB) is a much denser image set than "people in cars" (PIC) and "people walking or jogging" (PWJ), and the PRB objects have a much higher dynamic angular vector capability (they can change direction extremely quickly) combined with a wider dynamic velocity vector to PWJ they strain the ability of the predictive filters to reliably asses the collision threat. As a result you need either faster/better hardware or better algorithms to deal with that particular group.
I'm not particularly surprised by that, I expect the cars to also end up with small animal issues, as they can appear suddenly and aggressive evasive maneuvers to avoid hitting them may injure passengers at the cost of saving a squirrel's life.
And all of that adds up to some of the many things that one has to think about when claiming victory here. It is going to be a long hard engineering slog to get to full autonomy. My question is you can build a computer that can drive a car fully autonomously, what other missions could you create for it? Some of those are kind of scary.
Every example of a NN I've seen is good at one thing, the thing it is trained for. I haven't yet seen a NN that can beat people at chess, recognize faces, drive, and make coffee, or more importantly, decide when to do one or the other. The closest thing I can think of is Watson from IBM.
People can drive (poorly) but they also recognize the value of life in a biker or a squirrel, and in some cases, override their usual learned behavior, for example, instead of continuing to drive, swerving, basically improvising.
Computers and programs again are good for what they have been input and now, what they already have learned. The synthesis part, I'm not quite sure even NN have reached that part yet. I may be wrong, the answer is probably "learn more" or "learn faster" (which is what you suggest), but it's easier to synthesize like this when you have a mind with general knowledge, at which point the mind is less like a NN and more and more like a person.
And, as for their role in autonomous vehicles, I don't think they play a primary role (although I don't design autonomous cars so I wouldn't know for sure.) You can see in this Tesla video that a lot of the computer vision isn't even reliant on neural networks: https://www.youtube.com/watch?v=BLlwm5Dq7Is
Despite NVIDIA's papers, end-to-end neural nets for self-driving cars are dumb, but for the initial perception based on sensor data, neural nets are the only game in town.
Outstanding quote! I'm going to reapply it.
What are you imagining Watson is? Watson is not a singular entity that decides whether to play Jeopardy or analyze medical records or whatever. It's a brand.
The comparison that comes to mind is Colonel Sanders.
Yes, I bike, so I know the worst of the worst. I also remember a few courteous encounters, but probably many fewer due to negativity bias.
These are the sorts of things humans struggle with too. It could be argued that we, also, don't have one neural net that does all these things but rather a variety of specialised brain regions. It would then make sense that a computer might run specialised programs to deal with different tasks.
Anyway, self driving cars and bicycling can be fixed with a simple engineering solution: separate the two with barriers. This is what we should do anyway.
As someone who had their bike destroyed in a high speed car vs. bicycle collision, and was lucky to walk away with a broken thumb and sprained ankle: cycling on roads is insane.
How is it that people in developed countries across the world have learned to do this, and we haven't?
That said, the relative health benefits of cycling outweigh any risks 7-to-1 for an average human. Cycling for an hour, even with American road infrastructure, actually results in a net increase in the average person's lifespan of over an hour. Seems like maybe it isn't so insane after all.
But it is common sense, we know that excerise prolong your life, doing it regularly in modest proportions is best. Using a bicycle for transport to work means you are exercising twice daily, maybe 10-20 minutes for lunch as well. It is obvious that this kind of exercise will do you good.
For me it's one to two hours of exercise per day, it's hard to beat those health benefits.
http://ajph.aphapublications.org/doi/abs/10.2105/AJPH.2015.3... https://www.uu.nl/en/news/dutch-bikers-live-six-months-longe...
I'm okay with the separation if I get the same access as everyone else...and I am not blocked by slow bikers. Of course, I don't need to ride on highways/freeways (nor do I want to).
Nice analysis of why bicycles are so dangerous even now - computational issues like these are just as much an issue for brains as for computers. Not surprising that, in many cities in the US, biking is insane. Nearly everybody I know that bikes to work has a close call or two in their history.
Nearly everybody you know has had a close call or two with driving, I'm sure. And not just an "almost a minor fender-bender where everyone would have been okay" close call either. Cars are one of the leading causes of death for people of all ages. I've had close calls with cars as a pedestrian, as a cyclist, as a motorcyclist, and as an occupant in cars. I've been hit by a car while on a bicycle.
Cars as an entire category are dangerous, as are the humans who operate them.
What about the two who were occupying a whole traffic lane, weaving back and forward to stop cars overtaking, while riding at under 15mph in a 40mph area? They'd start screaming like lunatics at anybody who legally overtook them - the cops wanted a word with them, but they disappeared when they saw me on my phone.
Five or six times a day, cyclists run a red light at a busy intersection in town. I see many near-accidents in my fleeting times in that intersection, all of which are caused by the cyclists - and I'm not counting the number of times they've nearly hit pedestrians crossing with the red light. I've seen one car run a red light, three years ago, since I started crossing there in 2012.
Years ago, I was overtaken by a cyclist and his three children, just as I had begun my turn. I nearly ran over the father, and they took their sweet time to move through the intersection a car coming the other way (who legally had the right-of-way) had to stop and wait for them. I could have got out of my car and walked, and I would have got to the other side before they did. They bumbled along as if they'd done nothing wrong.
It's not uncommon for cyclists enter intersections on foot, and so legally become pedestrians for the duration of the crossing, and then jump on their bikes and dodge through traffic while hurling obscenities at vehicles who have right-of-way because they didn't enter as pedestrians. The law here doesn't change your status because it's convenient for you.
Back in 2013, one of my ex-workmates told me quite a tale, about the time she pulled up at a set of lights. When she moved off she signaled to change lanes and only noticed by chance that there was a cyclist holding onto her car to get a free ride. He couldn't have seen her signal, the indicator was near his foot, and if she had turned he would have toppled in high speed traffic, perhaps ending up under her wheels.
Back in the early 1990s, a friend of mine was pulling out of his garage and was hit by a cyclist at some speed - the cyclist threatened to call the police, but stopped short he would have been charged, not my friend.
Don't lump it all on drivers, cyclists are dangerous, too.
However, dangerous on a bike is not even in the same league as dangerous in a car.
While clearly inexcusable, how many of those incidents of bad cyclist behavior do you imagine could have ended with a fatality or catastrophic injury?
> how many of those incidents of bad cyclist behavior do you imagine could have ended with a fatality or catastrophic injury?
Several. My neighbour had around 75 y/o when a cyclist guy passed too close and broke her hip. Result: A long post-operatory, big scar in her leg, and she never could walk properly again. Died a few years later.
Of course, a car would have done much more damage to her in the same situation, and there are many more instances of real injuries from cars to cyclists than cyclists to pedestrians.
Did the cyclist actually make contact with her or just pass too close?
I'm not condoning the cyclist's actions in any way, just asking.
Um, this is true for humans, as well, you know. Every single person I know who decided to start commuting to work on bicycle in the US was in the hospital within a year.
After training, I suspect that computers are going to be MUCH better at avoiding bicycles. They're also going to be much better at avoiding the stupid things that bicyclists do, as well.
Admittedly I've had a few accidents over the years, but none that drew blood (as far as I remember), and certainly not yearly.
Either you live in a dangerous place, or you know bad riders :-)
When I was a teenager growing up in the rural Midwest, I drove a small sports car. I always wanted to see if I could fit under one of these things (a crop sprayer) when I came up behind one on the highway:
https://wapellowarbler.files.wordpress.com/2010/06/edge-of-t...
I wonder if this kind of vehicle would even register on some cars' systems, or if autopilot would just charge full speed ahead underneath to my teenage heart's content.
http://souvenirsfromaprogramme.tumblr.com/post/44133459703/t...
David: The Robo-Cup Federation hope that by 2050 a team of fully autonomous humanoid robot soccer players will win a soccer game, complying with the official FIFA rules against the winner of the most recent World Cup of human soccer. That terrifies me. If they can run around and kick a ball and beat us at football, what else can they do?
John: Bring us breakfast in bed!
David: Yes, and then strangle us.
We are supposed to be terrified by robot soccer players strangling us in bed? Sure it could happen, but there seems to be easier ways to die.
Apparently one of them is riding a bike around self driving cars.
One of the better lines in 'Real Genius' (another comedy) is Lazlo's "All you's need is a tracking system and a, a large spinning mirror and you can vaporize a human target from space." where the consequence of the main character building a multi-GW laser is suddenly clear to that character.
I'm all in favor of self driving cars, and I look forward to the day I can use them to commute and work at the same time, but cars are already used as bombs in the middle east, it isn't a large step from self driving taxi to self driving car bomb is it?
I think his broader point though, as is the point of the person I was replying to originally, that when we design robots for highly adaptable, sophisticated tasks, we design a potentially very dangerous robot.
This is arguably the most expensIve part of the pipeline, but we sHould have atleast two more architectures out by NVDA/AMD before any product comes out.
I haven't rode a bicycle since I was a kid but I do put ~10k miles/year on my Harley. In the last 10 years, I've been hit head-on twice, ran off the road once, and had too-many-to-count "close calls" by drivers who weren't paying attention (not to mention the friends that been killed) -- and that's on a loud, relatively visible (compared to a bicycle), 900-pound piece of steel. I can't imagine the average bicyclist is "seen" better than I am, especially in busy traffic areas/large cities/etc.
I'm all for anything that can makes the roads safer for those of us on two wheels!
Or to put it in other words - if I run a red light by accident, then that's what it is - an accident. I'm still responsible for causing a crash if one happens, but there's nothing more I can do to improve my perception of surroundings(and we already do ban humans who are unfit for driving).
If a self driving car runs a red light, because its sensor is only 99% accurate, then I'm pretty sure the company making it will be sued to oblivion, because they released something that does not work all of the time(and I am talking about normal operating conditions, not sensor malfunctioning).
http://www.sfbike.org/news/a-warning-to-people-who-bike-self...
I guess they've succeeded in acting like most human drivers. I've almost gotten right hooked a few times while running (and wearing bright highlighter-colored high-viz reflective clothing).
</sarcasm>
The sad irony is it's those "pushy" cyclists who stay alive, by making themselves more visible, being a part of typical traffic patterns, and refusing to be pushed off to the side by cars trying to squeeze into the same lane.
While it's completely in one's rights to ride a bike on those streets, it's pretty stupid unless you're hauling ass because the speed differential between the car and bike is too great.
Urban planners can talk all they want about how these streets are bad design and cities shouldn't have them, but it's best for both parties if bikers didn't use those streets. I don't know how someone rides on those streets without thinking it's a bad idea because most of the time, it's someone on a slow mountain bike going 10 mph.
Another comment mentions motorcycles, i'm very curious to see the detection data for them: they should have roughly similar characteristics to a bicycle, but at a similar or higher frequency.
Humans aren't very good at predicting or seeing them either in dense urban areas. Grade separation is the way to go!
I guess they think it's illegal to drive in the bike lane at all, even though the opposite is true: at least in California you're legally required to merge into the bike lane in the last 200 feet before the intersection when making a right turn.
Road rules for bikes and bike lanes are wildly different depending on state, and heck - even city. There are states where you have one set of rules within city limits, and entirely different throughout the rest of the state.
It's a huge mess, and doing things like putting unprotected bike lanes to the right of traffic lanes is almost criminally stupid and dangerous. You cannot undo decades of muscle memory ("things don't pass me in the right lane especially on dense city streets") and then point to "there is a rule that says you have to!" when the expected tragedies start occurring.
I used to think drivers and bikers were just "assholes" to each other in the US - then I spent time in the Netherlands and Belgium and realized these two modes of transportation are purposefully being put at odds against each other in the US for political reasons. It's quite possible to combine these modes of transportation into a network that works for both types of users.
A perspective from a writer who favors separating bikes and cars:
http://www.copenhagenize.com/2010/07/vehicular-cyclists-secr...
To put it another way, I don't think anyone knows this law.
We really need to do something about people parking in bike lanes and in my opinion should completely close market to cars. Allow chariots, public buses etc. Anything with 5+ passangers, but close it to everyone else. There is just too much going on to have that mix.
We also desperately need more grade-seperated bike lanes and bicycle only traffic signals that are aggressively policed.
I too commute a lot via bicyle and also drive, which is why I disputed your comment. I find motorists often see me fine...when they're paying attention and I'm using common sense, ie lights when riding at dawn, not running reds etc.
We're talking about vision / detection here, not people being unattentive, aggressive or inconsiderate, which seems to be the actual problem with motorists and cyclists.
If a self-driving car is 100% dedicated to driving and still has problems detecting bikes 75% (at best) in optimum conditions, that's a different and important issue that I don't believe exists in human drivers today.
I actually hope for both our sakes that if self-driving cars do take over, it's for our benefit as cyclists.
I too am quite optimistic that self-driving cars will be vastly better to deal with than human-driven ones. The constant 360 vision and 100% adherence to turn signaling will be vast improvements. Every day I deal with drivers making unexpected turns, and not checking for the presence of bikers before turning across a path that bikers might be traveling in. Self-driving cars won't have these problems.
My concern is that they will, but policymakers decide a few dead cyclists is a price worth paying. In the UK, we're already ignoring an air pollution crisis (mostly from car fumes) which is killing c. 40k people each year. On top of the KSIs from actual road accidents. Uber have shown all too clearly that some companies will hide or ignore safety issues in this area.
With robot-driven cars, the bicycle problem is an area that engineers can focus on and iteratively improve—essentially forever—even if cycling amongst robot-driven cars right now is as dangerous as cycling amongst human-driven ones.
The solution is likely a device that cyclists can wear that self driving cars can detect - I'm going to bet the AI for bikes is intractable because they are fast, virtually invisible at some angles and travel in between the traffic in strange ways.
An alternative could be all self driving cars near by sending each other information about cyclists via some means. Or self driving bikes ;-)
If this is true, do you think humans can handle it any better? What you just described would be a nightmare for any human driver as well. Is it reasonable to allow self driving cars that are only marginally safer than human drivers?
This actually raises a good question that may be even more problematic for autonomous cars; people deliberately fucking with them in a way you wouldn't/couldn't before. Maybe this won't happen but who knows.
Obviously this needs some thought. For example, how would you prevent people from messing with the cars - maybe it's very short range, or line-of-sight only?
This is not any different from cars driven by humans, btw. Someone can put up a sign (say a merge left sign and some pylons) that can cause accidents pretty easily, which is why drivers still have to pay attention.
The thing is that wide, let alone 360 degree, vision should provide the computer with more and more useful 'messages' than humans are ever capable of taking in. And probably more than any attack vulnerable information messages could provide too.
Anyway, I'm not suggesting that the car take a remote signal on blind faith, but I still think it might be useful to have some kind of indicator. It can be factored into the other sensor data - "this signaling object does appear to be in close proximity and moving like a cyclist". Just as people don't tend to mess with signage and cause crashes now, I think they're unlikely (on the whole) to do it with these kind of signals. Especially since you'd be in serious trouble if you caused a wreck.
I have yet to see a car capable of reading the body language of a bicyclist or motorcyclist. As one of those, i still want to look a driver in the eye before i trust him not to squish me. Robots arent anywhere near there yet.
If I drive a ferrari, should I hate all those non-turbo car peasants for making me have to drive slower? Of course not, I'm using shared infrastructure that everyone has paid for, even those pesky cyclists and slower vehicles.
I'm a cyclist 80% of the time, a driver the other 20%. I pay my taxes and rego on my car and follow the rules, even while cycling, and you still hate me.
I hate our lack of suitable infrastucture with properly separated cycle paths, not the riders.
When you're driving on a shared road you need to deal with slower vehicles and drive safely. If that's too hard for you to deal with, maybe you shouldn't be on the road.
But seriously, you bring up a good point. Maybe both cyclists and yourself will come up with ways of gaming the system in their favor.
I can't find a link now, but I'm reminded of an experiment where baby chickens would prefer a ruler with three lines painted on it to their real parents.
Perhaps you could have a design on the back of your seat that tricks the self-driving cars into thinking you were a gas tanker truck or similar large, dangerous vehicle, and get yourself a wide berth.
The huge, insurmountable problem is that people will NEVER accept autonomous transport that regularly kills people and causes accidents, even if the computer driver is better than an average human.
As long as self driving cars share roads with humans they will never stop killing people. Humans are unpredictable by nature, and there will regularly be situations a self driving car encounters where death is just unavoidable.
When self driving cars start killing people what is going to happen? You can't sue a machine. What if the machine has to decide who dies? This is an insurmountable problem. Cars may drive automatically but there will always be a qualified human sitting behind the wheel.
Look at planes, it's a lot easier than driving since there's nothing to run into up there. We've had fully automated flying systems of various kinds for decades. Hell, even the space shuttle could land itself... from space. To this day every commerical flight with humans on it has not just one, but two human pilots.
The other big problem with self driving cars is that hardly anyone needs them. Cars are built to transport people, and it turns out people are pretty good at driving. So the vast majority of the time there is zero need for the car to drive itself. In cases where it would be convenient we already have cruise control type inventions and automated lane following.
So what about trucks? For one, if people won't tolerate a 3000 lb piece of automated steel killing people occasionally they sure as hell aren't going to tolerate a 80,000 pound piece of steel doing the same.
So why don't we make special roads for the automated cars and trucks and hitch them together so they can't hit each other? We've had this for over 100 years and it's called a train.
Basically, automated cars with no humans behind the wheel will never happen. This gets rid of most of the advantage of a car driving itself in the first place. Also, if humans need to be able to take back control there goes all the advantages of packing cars close together on highways and stuff too.
For frequent commutes where I don't feel like it, I just take public transport. A lot of time, I bike for the fun of biking, although many times it is also for the sake of transit.
Literally the only thing the self-driving car needs to do is be able to stay on the road, and if the road is blocked, use the brakes. There's no moral dilemmas and no grey areas.
if {obstacle} then {brakes = 1}
Detecting what's an obstacle and what isn't is of course a difficult problem, but I'd think it'll be the legal issues that will actually be the largest barrier to self-driving cars. Someone has to be responsible when it fails. If the human isn't driving, then the car maker has to be responsible. This liability will have to be added to the cost of the vehicle.
https://www.theguardian.com/lifeandstyle/2009/dec/15/cycling...