Predicting is absolutely key. If you've been around the Waymo cars they horribly manage simple situations. They're VERY safe because they take no risk, but that won't work at scale. "What's that? Stop wait..., what's that? stop... wait wait wait"
They're very jerky at intersections where there's even a mild amount of uncertainty and really only go when it's very clear. It reminds me of lots of first time drivers reacting to new situations where safe is the best way forward. But if we had pure first time drivers, we'd be a gridlock at even a mild amount of traffic that enters and leaves the roadway.
Looking forward to this next set of 5 years. Hopefully us drivers will get replaced in 10? maybe?
I actually prefer the Waymo approach over the Tesla and Uber approach. When there are lives on the line you simply do not cross certain lines and in the longer term Waymo will be the only player left whose reputation is still in one piece. That more than anything will decide whether self driving vehicles gain long term acceptance. Anything else will lead to a 'self driving winter'.
Interestingly enough, Waymo (back than Chauffeur) attempted the Tesla approach - and threw it out once they saw how people behave when the car "drives itself*".
W.r.t."self driving winter", I think Waymo are aiming to protecting themselves from that situation by developing close relationships with regulators, highlighting the "we're not them"-part and sticking to their safety-first approach.
Yes, that could indeed happen. But Waymo will likely be able to comply with such regulation whereas Tesla and Uber never will. The parties hurt hardest are those that would have played nice but that get nipped in the bud by the regulations before they are large enough to comply. So Waymo would benefit.
Tesla's getting enough hours under control at this point that they're likely to win the regulation battle for Uber simply by virtue of statistics, though. Obviously there has been the occasional tragedy (and there no doubt will be more), but the point where they launch the "Autopilot is provably safer than humans!" press release isn't too far out at this point.
Again, the measuring stick is real drivers. And real drivers are really bad at this stuff too.
That press release was sent out more than a year ago and it is about as fallacious as it gets and I'm surprised the whole thing keeps getting airtime.
For the comparison to hold you need to do the following:
- discount all the hours when humans are in adverse conditions and autopilot is not normally engaged
- remove motorcycles and other vehicles from the pool
After you do that Tesla is still much less safe than your average driver. Their whole spiel of taking the easy 80% in miles of driving and claiming victory has been debunked so often now it is getting boring. It's the remaining 20% where the fatalities are and which - coincidentally - are the hard bits.
I’m not sure what brightlines exist in society for when lives are on the line. Lives are already on the line for human driving, so what is the moral story one would extract from the big picture of how people relate to driving today?
The only story I see is politics and complexity, as opposed to moral clarity.
It's quite simple, actually: a self driving vehicle should be a monotonous improvement on non-self driving. In other words, all situations that are currently handled by regular drivers should be handled just as well or better for self driving to be socially acceptable. A statistical argument based on a reduction of overall fatalities including a number of new kinds of accidents will not be accepted.
People (and the media) are going to focus on the accidents that would not have happened if self driving did not exist.
The problem is that we might not ever get to monotonous improvement. We can probably easily get to a state where self-driving would decrease the number of traffic injuries and fatalities in total compared to human drivers.
But machines make different mistakes, mistakes that humans would never do, and that's psychologically not acceptable to most people. All the actual accidents will be "freak" accidents that a human could never cause, and that scares people a lot more than "human" accidents like a drunk driver running someone over at a pedestrian crossing.
Yes, you got it perfectly. That's exactly my fear. And Tesla isn't helping with their marketing because they suggest we are further than we really are.
If I get killed by a drunk driver, my wife will at least have someone to rage at.
If I get killed because my car randomly decides to drive itself into a high up truck and I basically died because of stupid technology. I would hope that it doesn't end up killing my wife too out of rage that she's going up against a major corporation now.
Strange, because there are a LARGE number of driving fatalities caused by this exact behavior every year. The common explanations are: driver dozing or falling asleep, unexpected curve with poor visibility, driver mistaking the shoulder for another lane, and driver striking median to exit ramp by misjudging a lane change. Humans do all of these things, on a daily basis. So by human standards, accelerating and crashing into a barrier is quite commonplace.
The problem with your statement is that some people view handling a situation better as getting more cars through a bottleneck in a certain amount of time, and other people view better as making sure that all humans in the area stay safe.
PR will never be about throughput but it will always be about safety. You'll never read 'due to Tesla's self driving abilities we can now push 3% more cars across the bridge'. What you will read (on page one, complete with 'shocking' picture) is 'Self driving Tesla crashes into bridge support'.
Which might mean that we are going to demand a safety level that's unattainable without decreasing throughput. In other words, human driving might be risky, but it's familiar, which apparently makes it tolerable.
> A statistical argument based on a reduction of overall fatalities including a number of new kinds of accidents will not be accepted.
I think what would be accepted is if the self-driving car would have capabilities that surpass the best human driver (as opposed to the "average" driver). But we're a long way from that.
That makes me wonder: who is verifiably the best human driver? I recall reading about a truck driver who went over a million miles without being in an accident. Probably someone like that?
Yes, difficult to answer, so I wasn't thinking of an actual person (similarly to how the average driver doesn't exist as a person). If you want to imagine someone, then perhaps the person who drives the president?
That's a very different skillset from a normal driver; in fact, I think that in driving The Beast aka Stagecoach, road safety is an afterthought to a host of tactical considerations: the goal not being "safe driving", but "keep the President safe, to the exclusion of anything else." With 4 MPG, this is a plainclothes APC, not a normal car ;) https://en.wikipedia.org/wiki/Presidential_state_car_(United...
> a self driving vehicle should be a monotonous improvement on non-self driving. In other words, all situations that are currently handled by regular drivers should be handled just as well or better for self driving to be socially acceptable.
I agree with this but I think it implies that people also will not accept safe self-driving cars that take forever to get anywhere in traffic. Especially since a decline in average traffic throughput makes things worse for everyone in a car, self-driving or not.
I think this makes the problem much harder, since a lot of people do demonstrably take risks in their driving now, in the interest of saving time (even knowing intellectually that lives are on the line).
If you slowed everyone’s traffic by 25%, how many human lives are wasted (though not ended) per year?
If you assume people drive 10K per year at an average speed of 33 mph, you’re driving ~300 hours per year. Slow that by 25% and it’s an extra 100 hours per year. If you assume 7.5 hours of sleep, you’d spend an extra 1.6% of your waking life in a car because of the slowdown. That’s killing a lot of time, even if not directly killing people.
since I let my Tesla 3 drive I have to say, what is out in public is damn shy about what it does and will slow down for curves and the traffic aware cruise control seems to have a second sense. On the open highway if you are navigating on autopilot you can let it change lanes on its own but I still like to control that because I think it doesn't do it as soon as it should; I usually use the must confirm mode. It does react well to people who cut in and my car has dodged someone merging into my lane. Going to be interesting if there is any change to driving when the computer gets upgraded next week.
frankly I don't think any of the teams who have demos are as close as they hint they are. there are just so many exceptions that it becomes maddening.
look, people road rage over the driving habits of others and the computer while it won't road rage has to react to those who infuriate others with their driving either because its reckless or just bad driving
some of my favorite conditions....
people stopping to let you out, how does the car know? people going out of order at an intersection?
* I do not let it drive where I think the situations are complex and obviously it does not have the ability to respond to traffic signals or make turns at intersections.
My belief is that we'll be replaced when vast swathes of roads are replaced with ones designed for automated vehicle transport, which is a fair way out of my lifetime I'd imagine.
Simply because it has been a control theory challenge. Weight on the wheels change, frictions change and drivers are expected to stop the train within inches on average, and a feet at worst. That requires some R&D to implement.
Yurikamome trams in Tokyo runs in L4-L5 and Tsukuba Express runs in L2, in SDC levels term. There was also a L4 PoC in Yamanote line so that’ll probably come by the end of this decade. Elevators had been L4 for like half a century as well, by the way.
The Parisian metro has a L4 line since 1980s...on an isolated track, without significant branching, controlled by a single computer from a command center. Yet emergencies are still escalated to humans, even things as common as "somebody is stuck in the door".
In other words, getting trains to L4 is straightforward, and there's been a rollout for half a century. Making the jump to "doesn't even need the driver to handle emergencies," that's a task comparable in scope to getting cars from L2 to L4.
It's too bad we can't rip the band-aid off and just go straight to full autonomy for every car using Waymo (or similar). The problem you identified becomes a non-issue if 1) it's the norm and 2) the passengers inside aren't paying enough attention to the road to care about the quirks. It's probable in my estimation that roads filled with only automated drivers become safer, more efficient, and more predictable, which acts as a flywheel for the prediction in the cars to get better.
Would be really interesting if Waymo could prove this out somehow.
That boy chasing that ball out onto a crowded street has much better chances with an automated car vs one driven by a human. Heck, any of those situations human drivers perform very poorly in already, the bar is low.
I have first-hand experience with such a case, and I disagree. I avoided hitting a kid that I think any AI would have killed.
I'm driving southbound on Fair Oaks in Sunnyvale, approaching Arques. On the corner is a mini strip mall with a small parking lot.
Out of the corner of my eye, I see a young child, maybe 4 years old, running around that parking lot being chased by his mother or caretaker.
Now, being well off the road and out of its universe of conditions that would be considered by an AI, he would be ignored. But I just knew, somehow, that this was trouble.
The kid, thinking this is a fun, impromptu chase game, then sprints directly into traffic right in front of me. I stop with a foot to spare.
A human could see a group of kids looking in the direction of the street at someone occluded from view and realize the possiblity of a child dashing out onto the street well in advance. The self-driving car would have to detect, classify and react after the child comes into view. It is this kind of common sense AI that we cannot engineer our way out.
A couple days ago there was a viral post about how self-driving cars react to a 35 mph sign vandalized to look like 85 mph. A human would know that the sign is wrong - the self driving car would need this 'use case' programmed in - maybe have location based speed limits programmed in.
> ‘It is worth noting that this is seemingly only possible on the first implementation of TACC when the driver double taps the lever, engaging TACC.’
It isn't a strict self driving situation, but makes for great click bait. I can't imagine that this isn't an easy fix (Tesla's are also easy to update collectively via software update, try doing that to human beings).
There are two bars though, because there's a double standard. If the robot car hits the boy, it's national news. The lower bar of surpassing human safety performance is, in almost all regards, not even relevant to the discussion. Maybe it should be but it's definitely not.
Waymo is at the point of being able to deal with those pretty well, right? Either way, I'm suggesting that the impatience OP experienced as a result of being a human driver proximal to a Waymo car would be obviated if all cars were driven by Waymo. If you're not driving the car that's behind the tentative Waymo car, do you really care? You're probably focused on the work you're doing in the car or the TV you're watching.
Sure, they say so, and have been for years! That obviously means they haven't been exaggerating then and aren't now; in that case, I have a very nice bridge here that might be of interest to you!
Not a chance. One thing I found interesting in this article was the graph of successful object identification in images. The author made much of progressing from 50% to 88% in the last ten years (I believe this uncited data comes from [0]), but the great majority of that progress happened in the first five years, and the predictable slowdown of improvement is clearly evidenced by the graph.
We should also be very clear that 88% success rate given for object detection and identification is against a relatively fixed set of static images[1]. And as far as I can tell from the paper[2], there's no indication of how long it takes to come to a conclusion on any single image, and no description of how much hardware was used to reach this rate of success, which to be honest, I'm surprised is so low.
Given all of that, it does not seem reasonable to assume that this 88% success rate has anything to do with how well object detection works in real-time using on-board computing power while the car is in motion. So, don't start planning to throw out your driver's license anytime soon.
Reminds me of when two people walking towards each other struggle with who is passing on the left vs right and go back and forth at the same time.
The best strategy for a resolution isn't to be accommodating. It's to physically shift your entire body to one side and make no eye contact to make it absolutely clear to the other person which side of them you'll be walking on.
That works well when one person does it - but if everyone had that strategy, then it wouldn't work either! (I've spent way too much time thinking about this).
Look into how wifi transmission is solved among the many clients (although I'm sure its quite complex now). You might be able to use a random back off time to settle races like this.
I remember 3-4 years ago when we were told they were 2-3 years away. I think even 10 years is optimistic if we expect them to handle "edge cases" like driving in the rain or snow, dealing with mis-painted lane lines, and differentiating between a plastic shopping bag and a tire in the road.
A friend of mine with a Model 3 says it sometimes works better in rain than he does. Agree there are edge cases and probably 10 years is optimistic. But, of course, plastic shopping bags and tires on the road are a challenge for humans as well.
I live in a rainy area and I drive a Model 3. The other day it got into a high risk situation on autopilot, because it doesn't have any ability to consider rain intensity, big puddles, or rain-filled ruts. It is far more dangerous than an average driver.
And don't even get me started on the "automatic" windshield wipers.
Self driving has potential on finding the lanes (in rain, snow, mis paint) with infrared. The key component to driving in rain or snow is to slow down and increase following distance, something humans, but especially outliers, are relatively poor at.
Slowing down is also the best option for a tire tread or a shopping bag. The bag being less safe than you imply if it is filled with a box of nails or happens to wrap around a driveshaft or exhaust manifold.
I truly believe my children (not born yet) are going to buy their first car and it will be fully autonomous.
Too many people my age and older will be apprehensive about giving up control, despite it likely being safer. So I think it is fair to say that in 20 years there will be many people who learned to "drive" with a self driving car and their parents and grandparents will refuse to give up driving themselves.
Insurance and police enforcement are two reasons why the decision will be removed from law abiding citizens and won't happen linearly.
Once proven safe (or if a manufacturer directly insures the product) the cost to adequately insure a vehicle (as prices of what you hit continue to increase), especially an older vehicle, will increase. With income inequality where it is (and state minimums lagging) it has already become common to have underinsured drivers (and the upper middle class are already paying more through uninsured motorist protection).
When it comes to police enforcement, but especially jurisdictions that use traffic stops to find or create crime to fundraise; the expense of having traffic police in a world where > 50% of cars can do no wrong flips the model.
Finally, there is a large group of people who will not drive a car that's 7 years old or has over 100k miles; combining with less trained ability to do maintenance on more computerized cars. These people claim to be in control, but at one check engine light away from doing whatever a dealer tells them.
Insurance premiums are calculated based on loss rates. In some hypothetical future where level 5 autonomous vehicles are available, there is no reason to expect that the remaining human drivers will have more collisions than they do today. Rather the reverse, considering the increased availability of advanced driver assistance systems.
There's nothing wrong with driving a newer car and relying on the dealer service. Sure it's expensive, but if you can afford it then so what.
If you always drive a newer car and rely on a dealer for service, you no longer control what you drive when those dealers only sell self driving cars. (Or when the service model shifts as electrics have fewer service items and the service department closes)
The myth that electric cars require less service needs to die.
Maintenance needs in any modern car mostly comes back to value engineering. It's literally someone's job to optimize "how cheaply can we make this part, that it breaks just often enough that people won't complain".
The exact same thing is already happening for electric cars, and you'll need to spend exactly the same amount of time and money on service.
Furthermore, almost none of the service on a modern car is related to the combustion engine itself. It's mainly electronics, linkages, wheel beqrings, shocks, suspension, rust damage, tires etc.
If you think about it, a car that's done 200 000 miles at EOL has only been running for around 200 days continuously. A generator, or the engine on a ship, train, airplane etc. runs more than a full order of magnitude longer.
I drive an electric (LEAF) for a little over 5 years and also maintain my wife’s characteristically low maintenance car (Honda CR-V). Despite being very low maintenance, her car is much more maintenance intensive than my car. In five years, I’ve filled the washer fluid several times, changed the wiper blades, and plugged one flat tire. On her car, oil changes alone have taken more time, to say nothing of the brakes, power steering pump, valve adjustment, exhaust gasket change, and a handful of other jobs (all small).
In my case, the “myth” very much matches my experience.
ICE cars these days typically have electric power steering. And BEVs have brakes. I don't think ICE cars necessarily need valve adjustment - doesn't Honda use solid lifters where others use hydraulics?
"there is no reason to expect that the remaining human drivers will have more collisions than they do today"
Seems obvious to me that the people who are economically forced by insurance rates into self-driving cars first would be the worst drivers. That means arithmetically that the remaining drivers will be better than average. So the rates for everybody should go down.
It depends on who chooses automated vs self driving options. If primarily sports oriented people want to drive their cars manually, that is going to be one heck of a risk pool.
Also sometimes you need to move your car to affect the environment, for example if your car is stopped at a parking lot type intersection, and there is a large crowd of people passing, they would keep crossing unless you start to move your car a little indicating, that’s it for you let some cars through
You mean like those drivers who aggressively creep into the intersection to scare pedestrians into letting them through, even when the pedestrian has the right of way? I can't wait for Waymo cars to get them off the road.
OP might be referring situations like a festival, where there's a trickle of cars and a lot of people. And usually cars have to slowly inch forward, some people will stop and let them through, some people won't. The driver of course has to stop, but also not just wait forever to have a clear lane.
Or, another example, a big parking lot with a few buses, people unloading their stuff from the buses (let's at the end of a long trip), relatives/parents/friends pick them up at the parking lot. Relatively dense, many people moving, but still cars come and go. Plus a lot of packs/luggage here and there. Usually in these cases more cars want to enter the parking lot to pick up others, but of course people want to continue unloading, so some kind of balance forms.
This also applies to "living streets": pedestrians are the primary users, but cars are allowed through at lower speeds (20 kph?). "Wait until everyone clears the path" means "forever", a typical livelock; the car needs to signal intent "I wish to move forward" while still yielding to pedestrians.
Years ago I was trying to get through a huge crowd leaving a concert once, because I worked nights in a building on the same block, and while I was edging around the corner, a cop on foot thought I was being too aggressive and started banging on my window hard. I rolled it down and explained to him where I was going and why, but later my power window mechanism failed and it might have been related. I wonder how that should be handled with a hypothetical self-driving car?
The AI takes instructions (orders? requests?) from the user. So if the cop asks the user to go back, the user should tell the car to go back, try a different route, etc.
What happens if the user is drunk? Well, naturally, the cop tells the car that it's a cop and just pick a different route.
And we shall see how well it will work.
Probably it'll take a (few?) decade(s) before a self-driving car should think it can safely wade through a crowd instead of going back and finding a good place to wait for user input :)
I knew it. I suspected the Waymo cars were literally driving on eggshells.
They’ve been peddling that these Level 5 Waymo cars were just around the corner. But they refused to disclose how the cars actually worked, and kept saying that their black box is so mysterious, that their scientists themselves couldn’t understand it, so there was no point in having the rest of us, mere mortals, try to understand it. Lies.
We are a long ways away from being able to trust a robotic car with our lives.
But to their credit, at least they realize that requiring a human to take over vehicle controls in a microsecond, is unrealistic. Ahem! Looking at another car company that thinks this is acceptable.
Of course we're 5 years from L5. Always have been, since the 1970 lane-keeping Mercedes. You just need to understand that this is a codeword for "we have no idea what interesting and hard problem pops up next week, this research is SO EXCITING!" If you are expecting actual L5, in production, five years from any given date...you're gonna have a bad time.
I would argue that driving in traffic is essentially a physically social activity, like a pack of wolves or a flock of birds or a school of fish moving together. There is a tremendous amount of communication happening between the individuals, but not overt language. It's physical communication which requires empathy (i.e. seeing things from someone else's perspective) to accurately interpret and make predictions from.
A simple example is to be in a lane next to a lane that ends. Everyone sees the same signs that indicate that lane is ending; everyone knows the cars in that lane are going to get over. But there are a variety of strategies for doing so: get over early; wait until the last minute; accelerate to merge; slow down to merge; use the opportunity to pass; etc. Some drivers are nervous and cautious; some are aggressive. The way the cars move on the road, the signals they provide like brake lights and turn signals, are a form of communication that you, the potentially yielding driver, can interpret and make decisions about how to deal with the merge. Or prevent it! If that is your inclination.
This seems like a hard problem IMO. Perceiving objects is essentially about physics--detecting and recognizing. Next step harder, making physical predictions e.g. how far a moving object will travel in 200 milliseconds. Maybe that's enough to keep a self-driving car safe, but it won't make it very efficient in heavy traffic (which is the norm in urban areas). To interpret intent as part of a prediction seems way harder than even that. Reminder that when humans start driving they typically have 16-18 years of continuous social learning... and even then it usually takes years to become a comfortably safe driver in traffic.
Of course this would be way easier if all the cars were self-driving, all at once, but that seems like a really hard way to implement this technology. I think you'd have to have a city with very strong leadership that defines a small area that is self-driving only (working with one or more private companies to provide the vehicles), and then slowly grow the boundaries of that zone.
You may be right, but one of your analogies undermines your point. Realistic bird flocking has been famously replicated in simulations by having birds follow three simple rules, none of them requiring empathy.
So it might be a relatively easy problem for cars, if all the cars could be made to follow a common set of rules. It gets more difficult with humans who may follow different rules, or even actively exploit the rules of the automated vehicles.
I wonder whether game theory would be helpful. If flocking rules could be found which were a Nash equilibrium, then the incentive for human drivers would be to play along. Maybe it would even be a decent predictor of what humans already do.
* Even if it were, prediction requires a theory of mind. What does the other person see, know, and want to do?
* Even if we solved prediction, driving requires social interaction, coordination, and cooperation.
I think self-driving car companies will try solve this with remote internet-connected operators that step in and remotely flash some lights or something to communicate. Of course these people will be overworked and underpaid, will have no skin in the game and not really care (at the end of another long shift) about your particular interaction with their company's vehicles...
Recognizing a car that is going to let you in in front of it is not black magic. Yes, it will take a lot of data. Yes it has to be done safely. Self-driving cars will have to learn and join the dance.
Sure, a fallback remote pilot is probably going to happen too, but that's a very big can of worms. Just think about what happens if the remote pilot wants to take the car through a rail crossing and the connection drops. Whoops.
Hmm, what about letting another car in front of you? I think trying to solve this with pure data will end up like the chat bots that seem real for two sentences and then fail disastrously when you try to actually converse.
I find that most of the problem of merging come from lack of randomness. That is both parties think the other should have already acted, so they think it's not their turn to act. Eg. I already slowed down yet the other party hasn't even started merging, so I'll speed up. Then the other car just starts to turn. Bah, I have to slow down. But they noticed I sped up, so they reverse their course, and we're already approaching the end of their lane, so ... poor other car has to now come to a complete halt, and either wait for a bunch of cars to pass, or this lane will have to come to a stop too to let that car in. And so on.
So some kind of consistency would help. If the self-driving car decides to let the other car in front of it, then slow down, wait a bit, then okay, chance missed, go ahead.
How else would it work? This part is classic reinforcement learning. I don't think this policy part is the hard part. (Eg. similar speed merging, and from-stop merging, or stop and let other car in.)
Probably each brand will have a personality. (Based on their timeouts and other parameters.) And humans will adapt quickly. (And naturally the self-driving cars will change over time too.)
Chat requires an amazing amount of implied context. Merging? Not so much. Driving in general? Yes, that does, because there are strange edge cases. (Like what to do when you are lost, or the car breaks down, etc.)
It’s funny, I see people online in random comments saying “AI” hasn’t materialized in to anything useful. But if you realize that most talk of “AI” is talk of ML, and if you pay attention to robotics, you’ll find it’s making a huge difference. The fact that anyone can credibly say perception is not the major hurdle anymore is a testament to what the current generation of machine learning algorithms have done to advance the state of the art. I’m excited for the time when even prediction feels like a solved problem.
I would think that perception in poor weather remains a real challenge, but certainly we’re leaps and bounds ahead of where we were even 5 years ago.
It's categorically untrue that perception is solved, given that things like this happen: https://twitter.com/greentheonly/status/1228067903666348045. And that's considering that Tesla's perception team is run by one of the world's most respected experts in the field, and that their access to data is essentially unlimited.
Now what bothers me is that leaders in the industry say without batting an eyelid that perception is solved, given that they witness this type of events on a regular basis.
Elon Musk recently claimed „correct vector-space representation“ is the hardest task of autonomous driving. If true or not for Tesla i dont know but i guess its a yet to be solved problem for everyone in the field. Would you agree?
We have cars that can drive themselves at speed under normal conditions. That's pretty huge re: "useful". Extending satisfactory performance under fringe conditions is just a matter of improvement.
In the video (https://reddit.com/link/f4fbfu/video/5yg4oclye5h41/player), skip to 0:38. There is a construction worker walking at a constant speed towards the car's lane. The car could only assume he would keep walking and swerved and stopped.
For a human, we'd probably assume the worker is paying attention and will not walk in front of us. Though at times that also ends up being wrong.
I have no interest in owning a self-driving car myself because I enjoy driving and would never put myself in a situation where I rely on a car, but I'll be very glad if it takes drivers like the one in front off the road. Drivers who can't control their speed without using brakes infuriate me. Not just because it spoils my driving experience but also because I don't like seeing inefficient use of machines.
Back on topic, one thing that is invaluable for human drivers is eye contact. I assume that should be taken into account here, ie. if the pedestrian makes eye contact with the vehicle we can assume he has seen it. Otherwise it will result in a lot of unnecessary swerves and stops especially in towns where pedestrians walk right up the side of the carriageway.
Eye contact isn't reliable. I suspect it can't really be something software can safely use as input, even if it wasn't a vision issue.
It's pretty common for motorcyclists or bicyclists to make eye contact with a car driver who proceeds to pull out in front of them anyway.
This is probably in part from car driver brains not interpreting bikes as physical threats. It doesn't have to be deliberate. It can also be the driver scanning for "car" and not seeing "car".
A self driving car is on the other side of that size equation, but eye contact might also be assumed by a pedestrian to mean that you will stop. Especially if a "driver" makes eye contact who isn't driving.
In the end it's best to slow down and not assume. Humans often make dangerous assumptions, for example going around blind turns or over crests too fast to avoid something that might be there. Most of the time it's ok but eventually somebody gets killed. The rational thing to do is make the software safer than humans, not emulate them.
As a pedestrian I often use eye contact for the opposite reason: a multi-way stop where I’m worried if I cross I’ll end up pissing off several cars, and despite that they’re all being too polite and won’t just go, even though it would be fastest and safest for us all.
So what do I do? Get to the corner and stubbornly refuse to look up at any of the cars, and don’t really face my body toward any possible crossing. Not being able to wave me on (I’ve always found being waved in front of a motor vehicle somewhat menacing, tbh), the drivers quickly clear the four-way stop and I can cross without worrying about causing a jam.
Anyways, the AI will have a lot to learn of the crazy habits people build to keep from being lawfully murdered because “but cars”.
Ha. I do this too: as a cyclist who generally follows traffic laws and vehicular cycling behavior, I get annoyed when cars that obviously have the right of way at an intersection want to wait for me to go through first. So, I just stare at the ground and don't move until they do.
I know they're trying to be nice, but it ends up slowing everyone down: suppose a car and a bike are approaching a four-way-stop intersection with the car clearly going to arrive a couple seconds before the bike. As the cyclist, I would assume the car will come to a stop or almost-stop, wait about a half or one second, then start moving. With that assumption in mind, I'll adjust my speed so that I slow down for the stop sign (prepared to stop in an emergency) and roll through right behind the predicted position of the car, without losing too much momentum.
If the car takes their right of way, this is great. But suppose they decide to stop completely and wait for the bike to go through first (against the right of way). Now as the cyclist, I also have to stop completely, since I don't know when they're going to start moving again. If we make eye contact, I'd feel safe to go, but I feel like that's reinforcing bad behavior. Instead I refuse to make eye contact and force them to go first, as they should have initially.
That's why I added the qualifier "generally". When there's little or no intersecting traffic and no ambiguity about right of way, it seems perfectly safe to me to slow down but not completely stop at stop signs (i.e. the "Idaho Stop"). But I admit it's technically not legal where I am. Of course, most cars don't come to a full stop at stop signs in those situations either.
One issue with using eye contact for self driving is low reliability (or need for true general intelligence).
But long before that, you'd need to realize a camera suite with an effective resolution between 200 and 500 megapixels, sampling at 60 Hz. We're nowhere close to that today.
Or a pan-tilt camera with a long focal length. I see this as much more viable than a wide angle ultra-high-res one, because it's sort of how the eye works anyway: one really high-res patch, then the rest low-res.
What the hell have we been working on if not that? What good is an autonomous driving system that is not able to consider what happens next?!?!?!?!????
Not crashing into fences. It's still an ongoing area of research in computer vision on how to detect repetitive objects made out of thin parts. Radar cannot reliably detect them due to being thin and optical methods are not yet advanced enough.
Presumably that? I think the author is just saying that it's going to get better.
He is not privy to what his competitors have been working on, in any case, except whatever has been announced publicly. (Unless he is doing industrial espionage.) Maybe it's just that Voyage's cars weren't predicting up until now?
I don't think there's any implication that the technology is perfect in my comment. But the idea that e.g. Waymo's software is not making some attempt at predicting future states, even implicitly, is absurd. This video[1] was released two years ago and references prediction at the provided timestamp.
We've been working on the prediction problem for some time. We should have a blog post coming out about our approach (it's pretty neat and in-line with this post) soon.
That's great to hear! I think upon fully reading the blog post it's clear that you meant the next leap forward it solving prediction, to the degree that perception has supposedly been solved. It will be interesting to see how y'all do. I for one want my self driving car yesterday.
I could be missing something, but this sounds like a horrible idea. I've had two accidents in my life. One of them was caused by inattention--hooray that self driving cars have the potential to wipe this cause out. The other was caused because I was a new driver and I predicted what the car in front of me would do and it didn't do it. It was in my first year of driving and fortunately it ended up being a very minor collision, but I distinctly remember that my conscious takeaway was that I shouldn't try to predict what other cars will do.
If my car is driving, instead of trying to predict the behavior of other objects I want it to do something like a minimax search and do something that will be safe in the presence of the unexpected, not the expected. Prediction sounds like the complete opposite of good driving practices like the rule of following at least 2 seconds behind the car in front of you, which is specifically there to prevent accidents in the rare case where the car in front of you on the freeway does NOT behave in the predicted way of continuing to go 70 mph in the middle of the lane.
> I want it to do something like a minimax search and do something that will be safe in the presence of the unexpected
This is called defensive driving, and you're wrong about it not being based on prediction. Defensive driving predicts that every wrong thing that can happen will happen and positions the vehicle to minimize the harm when they do.
Fundamentally, if you don't predict that the car up ahead may swerve into your Lane and slam on its brakes, then you're less defended when it does. Defensively choosing what to do and where to be at any moment is based on predicting it happening and acting accordingly.
I think "prediction" in this context means something different. It's not specifically that self-driving cars will optimistically assume they know what other drivers are doing, like a naive driver might (although that is a form of prediction). It's more generally that in order to make any smart decisions at all, the AI has to model the future.
For example, if you are driving behind a car on the highway, and the car _in front_ of that car puts their brake lights on, you may want to pre-emptively start braking now now. That's a type of pro-safety prediction made based on your understanding of the likely future that driving AIs should also have.
As another example, if you're on a roundabout, if there's a car already on the roundabout (to your right in LHS driving countries like the UK, to your left otherwise) then you shouldn't enter it. But an exception is if you can clearly see that the car is going to exit the roundabout before it reaches you: this requires prediction, but not in the sense of guessing what the car might do, or even trusting their indication (ha!) but observing from their current trajectory and whether their wheels have straightened.
The roundabout is a good example of where it would be pretty easy to design a safe approach for self-driving cars (do not go until the roundabout is clear), but in a heavy traffic area, would result in the self-driving car just sitting... and sitting... and sitting... waiting for the roundabout to clear.
A self-driving car that takes forever to get anywhere could easily fail in the marketplace, even if it is clearly safer statistically.
The thing about very safe driving strategies is that humans can already choose them at any time, without buying anything. I, too, can sit and wait for a roundabout to fully clear before going.
The fact is that a lot of people do a sort of risk-cost-benefit analysis in addition to safety analysis when driving. I'm not sure a car optimized heavily for safety, at the cost of efficiency of travel, will meet expectations.
Your problem is that your prediction was wrong, not that you shouldn't try to predict what other cars will do. Plenty of times that prediction ability will save your bacon. But you have to work hard at getting it right and making sure that you realize that all predictions should come with error bars that still include all possible futures at different levels of likelihood. Acting like your prediction is a certainty is a mistake in a class all by itself.
I don't understand.
1. The problem is that the GP prediction was wrong
2. You need to make sure predictions are right.
3. All predictions are wrong.
4. So add error bars.
Should the downstream system trust the error bars with certainty?
It's hard to define what a right prediction is. The correct prediction is whatever makes the car safe and effective. If an action leads to a situation with a low probability but very catastrophic outcome, the system should not take that action, even if the prediction correct with its error bars.
Acting on a wrong prediction without a wide enough tolerance for errors in the prediction itself is what caused an issue here. If the OP had not made the prediction they would have had to keep their options open, which was the safer course of action. The fact that they were a beginner compounded the error (and is exactly one of those reasons why beginners tend to pay higher insurance premiums).
I think it's another stream of data feeding the model. For example: you see a child walking along close to the edge of the road, not looking in your direction. You don't slam your brakes on, but as a driver your prediction is that he may step in front of you, so you ease back on the accelerator, move out further into the middle of the road, cover the brake.
The prediction talked about here is if you are at a 4-way stop, and the car to your left has it's right turn signal on, you can predict that they are probably turning right, and so it is ok for you to go through the intersection without waiting for them. Currently, and autonomous car will wait until the other car is no longer in the intersection. With prediction, it will act more like a human. That doesn't mean it 100% believes the car will turn right, but will modify its behavior if it sees the car starting to go straight.
Basically, autonomous vehicles almost entirely assume the worst outcome in all situations, and waits until it is safe. So at a 4-way stop, it will wait for all cars to be out of the picture before going, which results in horrendous performance if the intersection is busy.
When I drive, essentially only use turn signals to restrict my behaviour vs the default assumption. For example, if a car is indicating to turn right, I will assume it is likely to either go right or go straight and wait accordingly.
Reason for this is too many times I've just narrowly avoided a collision because someone forgot to cancel their signal, or they were signalling ridiculously early for a turn they were going to make further up, or even that they just activated the signal accidentally.
Not to suggest prediction is bad, but just that also not treating turn signals with too much authority is safer.
If I am at a 4 way stop and there is a car to my left my prediction is that it's ok for me to go through the intersection regardless of the status of their turn signal because the rules of the road require them to yield to their right. I still pause to observe the trajectory of their vehicle, though, because a lot of people don't seem to know that rule.
I'm not sure if I'm reading your comment correctly, but in case I am...the "yield to the right" rule is only to break a tie if two people arrive at the intersection at the same time. Otherwise, the first to stop is the first to go. I haven't run across many people who don't play by these rules (or maybe my assertive driving style just means everyone else follows my rules anyway).
Well yes, I always yield to vehicles that are already in the intersection, irrespective, again, of their turn signals. Also it is not purely true that "the first to stop is the first to go". Specific minutiae of the law differ depending on region, but generally if I am approaching a 4 way stop with the intent of turning left and another car straight across from me is already stopped and going straight then, while I am waiting for that driver to cross the intersection another vehicle arrives at the stop to my right I am still obligated to yield the intersection to the new car despite having "arrived first". Likewise, when many cars are enqueued at the intersection, turns should continue clockwise, as in most board games, without regard to who is able to zoom up and brake the quickest.
You had the wrong prediction. That doesn't mean prediction is wrong. In fact the 2 second rule is predicated on the prediction of reaction time in the case of something unexpected, which is itself predictable with a probability distribution. Things which can be reasonably expected to occur should be taken account of. That includes e.g. suddenly swerving for an exit, an unusually slow car doing a u-turn on a street, vehicles performing emergency braking when the road ahead has limited visibility, etc. And these predictions take some time to develop.
On my motorbikes, making good progress depends on good predictions, and that means predicting that people don't see me, that they make rash and impatient decisions, that tourists on rental bicycles will do stupid things without looking, etc.
I think prediction in this context means objects actions probability distribution, not fixed action. You need to take care of adversarial distribution (up to realistic threshold).
"Now that we can safely detect the critical objects around us"...except if they're a pedestrian with a bike. Or something equally unlikely; meh, just run it over, probably uninteresting anyway. /s
I completely agree. If prediction algorithms properly implemented, the Uber crash could be avoided.
I think the next step after proper visual prediction is interpreting the noise around. For example, a human can hear the ambulance siren approaching and slow down in a junction without seeing the vehicle. I think sound interpretation is very important especially in city driving.
If Uber had bothered to follow a basic safety protocol that person would still be alive today.
Staying alert on watch is something that nearly every solider, security guard, sailor, will have to do. Where you basically stare into the dark in case someone comes.
There are a number of ways to mitigate the inevitable failure of attention:
* Having a parter;
* Doing short shifts (change with partner);
* Very simple non-distracting tasks - like reading out to your parter a figure at a glance.
You certainly should not:
* Be on your own;
* Have a mobile phone on you (all screens should be locked away while moving).
It's so fucking arrogant that Uber didn't care to implement the most obvious risk mitigation. The cost of an additional person on watch is negligible compared to the cost of their self driving care program. My gut feeling is Uber simply misled regulators with how developed their self driving car program is, so it could continue to mislead investors.
I don’t think accurate prediction is the right long term strategy. It’s edge cases that cause accidents, so a more adversarial approach is safer.
Taken to the extreme this causes problems, but gently slowing down in the case of extreme tailgating is a response to slow reaction times not direct path prediction.
Adversarial approach to driving results in staying at home, because you're likely to be rammed by a bad driver from behind, and you should almost never brake and drive always as fast as possible. Alternatively always stop when you see any pedestrian on the walkway, because they will drive in front of your car.
Driving is cooperative not adversarial, even certain classes of illegal behavior are predictable.
Context matters. For example, at an intersection, both the context of what is happening with others at that intersection as well as what has happened to the vehicle coming up to that intersection matter. An intersection near a soccer field at 4PM has different dynamics than at 4AM. An intersection following a highway exit has different dynamics than an intersection in a neighborhood.
You're right. And that's why I can't imagine a solution without AGI. There are simply too many context-sensitive parameters and we humans are quite good at applying past knowledge to circumstances which are new to us. And we also have access to information, which a computer may not have.
But software in their current state? Is it possible to provide enough logic to handle all cases like humans? Can you do that with pattern matching, feature detection, decision trees, bayesian models and anomaly detection alone? And if we do, didn't we just replicate a human driver with all his flaws?
Looking forward prediction should be superseded by announcement. Each vehicle should be broadcasting it's travel vector and next plan to alter that vector. At some point the SDC traffic will look like a high speed motorcycle team weaving between each other, a pedestrian wanders into the middle is just avoided. High efficiency, high safety.
I think another leap can be achieved with communication between vehicles. Once there are many more self-driving cars on the road, they can feed info to each other as needed. A car experiencing black ice can pass that info to the cars around it. Cars approaching an intersection can adjust their speed so that they all pass through without stopping, etc. But, because there are several manufacturers, this will be hampered without an open standard, and it doesn't solve the randomness of the world that will still be experienced.
"the state-of-the-art in computer vision has moved so significantly that it’s arguably now not the primary blocker to commercial deployment of self-driving cars."
The state of the art in machine perception is still far from human level perception, regardless of how quickly progress has been made. Prediction is a function of perception, so it's also bottlenecked on perception getting better. Let's not confuse recruitment-oriented blogposts with credible scientific writing.
Has anyone observed that the problem of object classification is really just a component of the prediction problem? That the old lady doing donuts in the wheelchair is identifiable as such because her appearance changes in certain (predictable) ways from moment to moment? In other words prediction error minimisation is the basis of the classification, not something made possible by object classification.
No, prediction is not the basis of classification. You can build a classifier that looks at a single out of context frame and learns to label the objects in it. It won't be able to predict where those objects move in the next frame--it was never even trained using subsequent frames.
I've thought in the past, once the compute is available in a low-enough-power form, you can continually increase the resolution (frame rate) that autonomous vehicles view the world.
Essentially giving them 'The Flash' like vision. You could then have all kinds of predictive models based on various patterns.
They're very jerky at intersections where there's even a mild amount of uncertainty and really only go when it's very clear. It reminds me of lots of first time drivers reacting to new situations where safe is the best way forward. But if we had pure first time drivers, we'd be a gridlock at even a mild amount of traffic that enters and leaves the roadway.
Looking forward to this next set of 5 years. Hopefully us drivers will get replaced in 10? maybe?