3mph variance is perfectly normal even on modern cruise control. And even on most modern cards, what you see on the speedometer is often off by ~1 mph at low speeds, and even more at higher speeds. (This is the reason you will often see people talk about 'GPS verified' speeds when it comes to talking about the top speed they have achieved in a car). Federal law in the US only requires they be accurate within 5% when using a standard tire size on normal pavement with the recommended tire air pressure. If you have over or under-inflated tires or have upgraded your car to larger tires for whatever reason, this 5% might become a very different. Even if the Uber was perfectly maintained, that could be 1.75mph variance on the speedometer @ 35mph.
Keeping to exactly the speed limit also just isn't how anyone drives. It wouldn't be, even if speedometers were perfectly accurate. It's impractical. It's hugely fuel inefficient. Everyday driving, be it by person or cruise control, involves accelerating and then letting off the gas. This means if you are targeting a speed limit of 35, you're swinging a few mph in either direction. Add in things like changing road conditions - every highway I've ever driven on will have random patches of road that are made up of different material or different levels of wear after having had potholes, etc, repaired - changes in road incline, weather conditions (winds), or even the behavior of cars in front of you - a vehicle changing lanes is going to result in different air resistance, especially larger vehicles, and plenty of other variables, makes the idea that any vehicle should be kept at exactly the speed limit is pretty unreasonable.
Speedometers aren't perfect. Cruise control or even excellent human driving will never keep you at exactly the speed limit. While saying going '38 MPH in a 35 MPH Zone is speeding' is technically accurate, it's not particularly useful.
Besides, some commenters already pointed out that the google maps of the road in question says that the speed limit is actually 45 northbound (which was the direction the car was going), whereas the 35 number is the speed limit southbound.
I'll ignore ACC for a moment and talk about the modern implementation of regular cruise control, however. You'll have your target speed stored in a control unit, which cares about 3 to 4 things: your set speed, your current speed, the flow of fuel to the engine, and in newer models, access to the brakes. So, how do we determine the current speed? The drive shaft will have a magnet mounted on it, and each time it spins around there is a sensing coil that sends a signal to the control unit. Based on the frequency of these signals, it determines what your speed is, and compares it against your set speed. If it is low, it will send a signal to a servomotor that is connected to the accelerator, sending more fuel into the engine, hopefully increasing your speed. If it is going to fast, it will cut fuel, and on some models, apply breaks if you are more significantly over the set speed.
However, this is a pretty dumb system. It only knows how fast you want to go, how fast you're going, and whatever the formula is that it follows for increasing the flow of fuel to the engine. It doesn't know about millions of other factors that affect how fuel efficient that acceleration is going to be. It is constantly readjusting based on the new speed readings. To my knowledge, all manufacturers also go for something that provides a "smooth" driving experience for their adjustments, rather than how to most quickly achieve the set speed. When you take the limited data available for a cruise control system and combine it with built in smoothing to keep the ride comfortable, you end up in a situation where the cruise control is making frequent adjustments that end up on one side or the other of your target speed. The lack of perfect information means they purposefully do not attempt to perfectly match the target speed - attempting to due so would require so many adjustments that it would be pretty terrible for fuel economy.
Now, hopefully, the Uber vehicle is going to have even more data available than my ACC equipped car. But even with my ACC equipped car, almost all of the work has gone in better determining what the set speed should be, and significantly less on making sure it hits it exactly. I expect this is probably the same on the Uber vehicle. Because quite frankly, it's just not that important. To be quite blunt, the difference between being hit at 35mph and 38mph is probably just how far your corpse flies.
As for saying they should drive 3mph slower than the speed limit, then they might end up going 6mph under the speed limit instead of 3mph under.
I suppose my fundamental argument in the previous post and now this is: What outcome is expected from them being able to stay perfectly at 35mph instead of having enough variance to hit 38mph? What benefits are there? A few mph makes the most difference in damage from an injury standpoint at very low speeds, and these are the speeds that variance is almost always significantly lower - there's no evidence here that Uber cars are going 8mph when they're meant to be going 5mph, for example, and this would not be the type of variance we see in cruise control systems in general - and very little difference to the outcome at high speeds, where the greatest variance is likely to be shown.
It sounds like cruise control is dangerous to use in Germany, based on your description. My 2012 VW GLI certainly did not stay within 3km/h of the set speed when utilizing cruise control.
>If that means going up to 6mph too slow, I guess that's the way it is.
If you want to be in a situation that is more dangerous than going 3mph over the limit, sure. If everyone was going 6mph under the limit this would be fine, but staying closer to the speed of traffic around you is less likely to cause accidents than going significantly slower. (Which, surely, if 3mph is significant, than 6mph is even more so)
>With electric cars, regulating the speed smoothly will be even easier.
I agree here. That just means solving the speed determination issue, which is not simple. GPS is the best method we have now, but is very problematic if there is no signal, or a degraded signal.
You don't need to stay within 3km/h of the speed limit to follow the law. You need to stay below 3km/h over the speed limit. So if your cruise control has a variance of 5km/h, you should set your cruise control at 2km/h below the speed limit.
> If you want to be in a situation that is more dangerous than going 3mph over the limit, sure. If everyone was going 6mph under the limit this would be fine, but staying closer to the speed of traffic around you is less likely to cause accidents than going significantly slower.
Unless your cruise control has a significantly larger variance than all of the other vehicles, you should be fine because everyone else should be setting their cruise control to roughly the same thing as you.
€15 for an in-town violation, €10 for out-of-town, no points for less than a 20 km/h variation? That's more of a secret tax than an actual deterrent.
Electric cars can't regulate the speed more smoothly than gas cars while on cruise control. It's about the control algorithms, not the source of power, and the gas systems could be on tighter control ranges if it weren't so inefficient and uncomfortable for the passengers. Adaptive cruise control runs on a PID loop control scheme and overshoot is inherently part of the game.
While on cruise control a car will go slightly faster than the speed setpoint for a few seconds, then slower, then faster, until it settles on exactly the correct speed. Then you go down a hill and it takes a little while to slow down, or you go up a hill and it takes a little while to speed up. Electric cars can use regenerative braking when the car is going down a hill but they're still constrained by the nature of control loops and aggressive braking will just lead to wonky acceleration-braking cycles while hunting for the setpoint.
That's also how humans drive. We just don't do it as well. If you're worried about tickets set your cruise for 3 km/h under the posted speed.
The setpoint is the setpoint. Key element of controls engineering. Secretly subtracting from the setpoint behind the scenes to compensate for your local driving laws would be the car lying to you and just leads to more trouble than it's worth.
Every day people tell me that they want a setpoint to be the temperature that the room never exceeds or the temperature it never falls under or five degrees above the highest temperature the boiler hits. It's like setting your clock ahead by five minutes to avoid being late. You can do it if you want to but it's ridiculous functionality to build into the timepiece.
Further, outside of detection errors like the one in this article, the reaction time of a self-driving car is orders of magnitudes faster than a human. You don't need 5m more braking distance per 5km/h because within milliseconds of the computer noticing the issue the car will hit the brakes. Human reaction time is more like 0.7s to 3s. You still need more time to brake as you go faster but the computer doesn't need as much time as our human laws already give us.
In a case like this where the victim enters the path of the car closer than the car's brakes are capable of stopping, even given instant detection, a human would have just hit the victim at a much faster speed because of the reaction time. That's going to happen sometimes. It's just how it works.
Basically everything you're saying is based on an outdated notion of how cars work and in particular your very German desire to follow the rules exactly. The rules are going to change, dude, and in the meantime you're free to set your cruise at 47 km/h to avoid accidentally triggering a photo-radar trap if you like.
That'd be a variance between true ground speed and the speed as estimated by the car. Chances are that the 38mph number came from the car itself, so there should be no "variance", leave alone 3mph over the limit. Modern cruise controls are perfectly capable of maintaining set speed, even at a considerable downhill.
My understanding, which may be incorrect, is that the 38mph number came from GPS recording, and that "cruise control" in Uber's self driving vehicles is not based on GPS speed, due to the impact poor/loss off GPS signal would impact it.
>Modern cruise controls are perfectly capable of maintaining set speed, even at a considerable downhill.
I'm not particularly handy with repairing cars, but I have a pretty decent understanding of how cruise control systems work due to nerding out on it when I bought a car with adaptive cruise control. Everything I have read on the design of cruise control says this simply isn't the case, because it isn't even a design goal. Much more important in the design is keeping close to it while maximizing fuel efficiency and driver comfort, and even in modern cruise control not all systems have access to the brakes, which would be required for keeping it at the set speed going considerably downhill. Anecdotally, all of the cars I have owned fit this and do not keep my vehicle at exactly the set speed, including the most recent one with ACC.
People should not "keep to the speed limit". In England at least the speed limit is the absolute maximum speed at which you drive. You should be driving slower than the speed limit at all times, with no exceptions. Anything else is almost always incompetent driving.
I find it hard to understand why drivers think the laws don't apply to them. Sure, it's annoying when cyclists ignore road laws, but cyclists are far less likely to kill other people.
In theory this may be true but it's absolutely not in practice. It's not uncommon to see people at least 20mph over the limit on motorways. On the M4 that is practically the norm.
It's completely possible to stay below the speed limit in a car, a lot of people just don't bother. Cruise control won't apply the brakes (at least in older cars) if you encounter a significant downhill gradient - that's up to the driver, and while speedometers are often out and GPS is more accurate, speedos universally over-read in my experience. If you're doing GPS 38mph your speedo is probably reading 40, which is not really within an excusable error margin of 35mph IMHO.
I would expect a fully autonomous vehicle to be able to accurately obey the speed limit and I suspect the reason it wasn't/doesn't is because that behaviour has been programmed in or developed. You could make arguments that doing what the other cars around you are doing is safer than being an outlier, but that's not the same argument you're making. I think the 38mph figure is significant.
A speedometer's reading is measured on the gearbox, at the output but there are a myriad of little things after that including variance in the final drive, tire size and tire pressure all of which can make the "calculated" speed and the real speed to be within a few % of each other. All of these components are also affected by wear too.
At the moment you need a GPS to conclusively measure a car's speed with >99% accuracy but GPS signal isn't available/sufficient everywhere. One could argue that we should create some sort of highly accurate local system that measures speed and we certainly can but we'd be trying to solve a 0~10% speed measurement problem: we'd get no real value from it.
In other words: a car measures how fast a specific cog is rotating, adds a few variables and estimates how fast the car is going based on that.
edit: and all of this assumes the car is going straight, calculating speed when the car is curving is more complex.
Oh, certainly, and I apologize if my argument came across as "It's not possible for people to not speed" or "People only speed because of the properties of speedometers or cruise control"
My point is that speed within a few mph is already inaccurate, and that 3mph is pretty much within the current margin of error.
>speedos universally over-read in my experience
I don't know that there's a common leaning one way or the other in actual variance due to speedometer construction, but over-reading is the most common occurrence due to tire pressure being low causing it, and a huge amount of people drive with low tire pressure. Overinflated tires, or moving to larger tires in general, will result in under-reads.
>I would expect a fully autonomous vehicle to be able to accurately obey the speed limit and I suspect the reason it wasn't/doesn't is because that behaviour has been programmed in or developed.
I don't agree. lagadu covers this plenty well in his comment, however.
A very rough calculation: v(t) = -μgt + v_0. To reach 32 kph or 20 mph, t ~= 0.82/μ, guessing a fair μ=0.8 and the chart here https://ec.europa.eu/transport/road_safety/specialist/knowle...
As you go faster, your energy goes up exponentially. This is why there are bitter fights about increasing the speed limit on roads; they become exponentially less safe (big caveats apply, naturally).
But yes, about 1/5th more KE with just a 3 mph difference.
It also fails to address my point about kinetic energy, which is proportional to the square of speed!
Typically it will get you pulled over. Obstructing the flow of traffic is a thing.
One fatality in 4 million miles is safe? Humans are 20 times better than that.
And this is considering the fatality rate for all human drivers, including those that are drunk, playing on their smartphone, sleepy, speeding, street racing, etc. I would not be surprised if those cases account for a very sizable majority of fatal car accidents. Now imagine you would reduce those numbers by relatively simple measures that don't involve AI. Let the car enforce speed limits, mandate smartphone companies to have a 'car mode' that cannot be shut off, mandate attention tracking capabilities for cars, dead-zone indicators, automatic emergency braking, etc. Maybe even mandate an alcohol lock, or at the very least a black box that records driving behavior so it can be analyzed in case of a crash, with sever penalties after the fact if the black box shows signs of drunk driving, etc.
The list of things you can do to improve road safety without requiring AI cars is endless, yet billions are invested in trying to fully automate driving. This just shows that self-driving cars are not about safety, but purely about business opportunities.
As you've basically said, the human stats include all the very worst drivers in the full range of driving conditions, whereas the auto-cars not only have the very best driving conditions, but likely also have the statistical benefit of a likely number of cases where the human actually stepped in and upped the auto-car score even further (since we can't account for the auto-car deaths/fatalities that humans have successfully stepped in and stopped).
Another thing that bothers me is how often the argument 'human drivers are terrible, far worse than computers for situations A, B, C, etc' is used to suggest that AI systems will improve road safety. I will be the first to admit that many people are terrible drivers, they make mistakes, don't pay enough attention, don't obey traffic laws, have bad reaction times, etc, no argument there. But so far I have never seen any statistics about how many of the accidents caused by these kinds of human failures actually lead to fatalities. Taking the human out of the equation could in theory improve all kinds of car accidents, but how many would be just fender-benders as opposed to road fatalities? It's not like it is super-easy to accidentally cause a fatal road accident, and everyone knows one or more people who were involved in one. Who can prove that AI cars will not decrease the total number of accidents of any kind, but increase the kinds of accidents with fatalities, like this one with the Uber car? Maybe human drivers, with all their faults, would have avoided this particular accident by driving more defensively or using social cues to be more prepared for this woman crossing the street?
I'm not trying to make an argument about the viability of AI cars that are safer than humans here by the way. I personally don't believe AI cars will ever be safer overall than humans in all possible situations that people drive in, but that's a different discussion. What bugs me is that suddenly it seems as if the whole world is willing to bend over backwards to interpret statistics about the presumed safety of AI cars in the most favorable way possible, and at the same time ignore all these relatively simple things that could actually improve safety right now, at much lower costs than developing fully autonomous cars.
There's not enough emphasis on the strategy and skill of driving, and that's for the same reason, I think. We've tried so hard to make driving as basic as walking, someone anyone should easily be able to do, and that doesn't fit well with something that takes a long time to master, that you can always learn a little more about.
So instead there's a narrative that all a driver has to do is obey some cut and dry rules, stop at all stop signs, never drink and drive, always do exactly 5mph over the speed limit (or is it 10? or is it 4? 7? 0?), and if they do all that they can zone out and not worry about anything but what's right in front of them in the windshield.
No immediate source, but pretty sure I read somewhere the safety stats between the best and worst drivers is one or two orders of magnitude. And the above is why. The best drivers are playing chess, the worst drivers are playing checkers. The worst drivers look at the same playing board but ignore almost all the information and just stay zombie-like in their lane till they need to brake to avoid something or their GPS says time to turn.
I'm really curious how self-driving software approaches all this. I have a sad suspicion they're programmed from a naive rule-following non-defensive point of view but would love to be proven wrong.
To me, these kinds of observations only solidify the idea that AI that is 'better than the average driver' does not necessarily mean 'less road fatalities', and that to improve driving safety, it would be more effective to start by taking the most common factors in accidents out of the equation first.
1) people who, like, you know, just drive, man, and like, as long as everyone else does the right thing, then like, i guess things will uh, like, work out dude
2) defensive drivers.
Obviously, self driving cars need to be more like #2 if they plan to actually be as safe as an average human, and have a shot at actually beating humans on this... I suspect Waymo gets this and Uber, well, is uh, ''moving fast and breaking things''
Also of note:
> Arizona has experienced a surge of pedestrian fatalities recently, with more than 10 in a single week of March in Phoenix alone. The state has the highest rate of pedestrian fatalities in the United States.
It seems like you're working pretty hard to justify this woman's death.
She was apparently walking her bicycle through an intersection. Did she grab her bike and jump directly in front of the car? Maybe, but let's not jump to assumptions.
> it seems the algorithm is pretty safe
It ran at least six red lights in San Francisco and would apparently swerve into the bike lane when making turns, now it's killed a pedestrian. It just doesn't seem that good. How many similar incidents have you heard about with Waymo? They've driven more than twice as many miles in city streets.
https://www.theverge.com/2016/12/20/14020720/uber-self-drivi...
Defensive driving means slowing down and paying more attention in areas of low visibility, especially if there are likely pedestrian crossings. It means increasing your following distance. It means increasing the odds that even if something unlikely happens you, the driver, will still be in a better position to react, or you will be less likely to cause harm. It's not about prevention, it's about risk management and harm reduction. This is definitely something that is likely to be difficult to teach to autonomous systems.
Why? What you listed seem like conditions that should be easy enough to recognize even with simple algorithms.
That is, unless they use neural network magic black boxes to implement the self-driving part.
Is this actually done? I thought neural networks were more or less used for the vision system to detect parts of the scene (pedestrians, stop-signs, etc). This is the only area where to me it makes sense to take the blackbox tadeoff right now.
But if you looked at the data (just numbers I've seen repeated a lot in this thread, heh), you'd find Uber's achieved 1 fatality after only around a couple million miles of driving. The average human being goes a hundred million miles between fatalities. Not a lot of data but so far Uber looks 50 times more dangerous than an average human driver. They have a lot of catching up to do
edit: https://www.reddit.com/r/SelfDrivingCars/comments/85ozqr/exc...
(https://arstechnica.com/cars/2018/03/police-chief-uber-self-...)
https://www.google.com/maps/@33.4350531,-111.941492,3a,75y,3...
The hangover from internal combustion is real.
And that's why I go by tram. :)
This sounds like straight out of the "hell" scenario for driverless cars: "we cannot accept responsibility for what happens to pedestrians and cyclists foolish enough to stray out of their designated safe zone".
I'm referencing Robin Chase (the Zipcar founder)'s "heaven or hell" framing - https://www.citylab.com/transportation/2014/04/will-world-dr...
For a recent, thoughtful talk about the damage self-driving cars might do to the city if we are not careful, see this Congress for the New Urbanism talk about "Autonomous Vehicles & The Good City" at https://www.youtube.com/watch?v=utnPEbDNbrE
It's also worth mentioning that there's a bill (AV START, or S. 1885) in the current Senate that would actually preempt local and state regulation of autonomous car safety, but it has been put on hold.