But so would human drivers. In fact, a self-driving car may actually do better in this scenario because it would have metadata stating that the last 3,000 times a car encountered this intersection, there was a stop sign here.
For another example, there's the nigh ubiquitous question of how to handle ethical dilemmas, like a crash where you have a choice between crushing a nun and crushing a baby carriage. Nobody ever seems to consider the fact that human drivers will almost never consider the ethics of the situation when choosing an outcome, and in fact probably won't deliberately choose any outcome at all, but will merely brake and steer by reflex.
Of course, having automated cars driving at 30km/h through the suburbs is going to piss everybody off, but that's better than dead babies.
In fact, a driverless car may be capable of estimating its own stopping distance given road conditions and other measurements, account for some margin of error, and then use that to calculate how fast it can go down a road, while still adhering to the posted speed limits.
"He won't get there any faster"
(Yes, he will. It's pretty simple to work out exactly how much faster actually)
Sure, a human driver doesn't take any of the relevant factors into account. But a driverless car does and that's the difference.
I think this responsibility falls squarely on the manufacturers of autonomous vehicles.
I think only the people creating the technology really understand the strengths and weaknesses of software drivers compared to human drivers, and are responsible for educating law makers and the general public if they want their products to be accommodated in our laws, road ways, and market places.
Yes, the manufacturers have a much stronger foundation from which to figure this stuff out, but that doesn't mean the rest of us can't stop doing silly things like acting as if human drivers were perfect when talking about potential autonomous vehicle systems.
Any good articles by engineers describing where current autonomous cars are still worse than humans? And which issues are just "bug that hasn't been fixed yet" vs. "we're still not sure how to solve this one in principle".
Agree, though, that a journalist should be interviewing real engineers and trying to summarize, paraphrase, and edit what he hears to the best of his or her ability, instead of just speculating.
I'd hope that a self-driving car has the meta data on board rather than just relying on seeing the sign. I'd also hope that they're designed to treat all junctions with caution, not just those with missing signage.
If you want to draw in a techie audience with clickbait, there's not much better you can do at the moment than writing about how Tesla has some sort of unforeseen flaw whereby both their electric & autonomous tech is doomed to failure.
As for the "best judgement" part - if a human is about to get into an accident, they are going to do what feels best. It's not necessarily the best option, but no court would ever tell you "our simulation tells us that if you reacted 0.1s earlier you could have turned the wheels 30 degrees and hit a car with two occupants instead of four, who had better survival chances". A computer is going to make an algorithmic choice, and that's what the article was about - who makes the choice as to how the car should react? The owner? The manufacturer? The police? The insurer? Should it prioritize its owner life at the potential cost of other lives? Or minimise the damage to protect the interests of its insurers?
We don't think about cars in those categories because we don't(and I would say we can't) make those choices when we drive. But computers will have to be programmed in some way. A lack of stop sign is a specific type of intersection in itself(where anyone going from the right has priority but anyone coming from the left has to yield) - so the car would need to observe and learn. I'm not saying it's impossible, but it's very hard to do.
In terms of edge cases. Lidar gives a 3d image so a floating balloon is going to show up like a bird and not something to be dodged. Same is true of balls and other small fast moving objects, they are not going to ping as kids. An object on the road is going be detected in plenty of time to slow down and stop or just avoid it without hitting any thing. Or, more likely just driven over the wheels are the only contact point and clearance is high enough for many objects in the roadway.
Driver-less cars have vastly shorter stopping distances at all speeds so they can be safe at speeds humans can't. They also rarely need to hard break because they can start slowing down sooner. Dodging might be considered if traction is good and there is open space to move to.
PS: If human drivers where not killing 20,000 Americans a year then we might debate perfection. Until then, self driving cars that kill 1,000 people per month would be a massive improvement end of story.
PPS: Remember google's self driving car from last month is worse than whatever they end up putting into production. The first production cars are also going to be worse than whatever most people end up using.
This is the biggest faulty argument in this entire discussion.
Imagine a radiotherapy machine. Statistically, due to operator errors, it delivers lethal dose of radiation to 1 patient every 10k patients. Then someone brings out a radiotherapy machine that doesn't need an operator, but statistically delivers a lethal dose of radiation every 100k patients. That machine would be banned faster than you can say litigation, even though statistically it saves lives vs. manually operated machines.
As soon as an automatic car kills the first person(be it due to faulty programming or mechanical breakdown) the manufacturer will recall all of them and the backlash will be huge, even if an average, they are still thousand times safer than manual cars.
Early on flying was incredibly dangerous and that hardly stopped anything. Hell, the general public gets killed at an air show and it's a non story.
People disproportionally care about things we think we should control. Like that fire of a Tesla that happened some time ago - people didn't care that all in all, statistically, electric cars are a LOT less likely to catch fire than regular ones, but media wouldn't shut up about it for weeks, highlighting how dangerous it is to drive around with a huge battery pack(even though it's nonsense). If an automatic car has an accident, it will be a huge deal as well - because people will expect computers to be flawless, and if they make a mistake they will want the head of whoever made the computer.
In tesla's case it got press as tech news, and they kept selling cars just fine. Safty issues can impact sales, but that's rational and it takes a lot more than a one off.
IMO, none of these actually support your viewpoint.
So, what happens if something blows into or falls on a motorway? A car hurtling down the road at 70mph is going to have a hell of a time stopping for an unexpected obstacle, human controlled or otherwise.
But, on a highway where it's safe to do 70MPH a deer running across the road is plenty slow enough to be stopped for. In terms of birds and bunny's just hitting them is the normal human response.
IMO, getting out into the once per year in the US edge cases is kind of missing the point. Ok, if your driving in a hurricane and a roof falls onto your car then yea your going to have a bad day meh.
1. "Driver-less cars have vastly shorter stopping distances at all speeds"
I can understand your expectation of reduced stopping distances due to the near-0 reaction time made possible by sufficient computing power. However, I wouldn't describe the computer's advantage as anything close to a "vast" one. There is no mechanical advantage to braking that is inherent for a driverless car and the difference in stopping distance due to reaction time is less than 1 car length (~10ft) at 30mph assuming the human driver has an average reaction times (0.25s). The advantage is still limited to roughly ~25ft at 60mph. That advantage is simultaneously significant and insignificant in that incredibly common outside factors will more than account for that disparity if not controlled for in the AI's programming. Tire pressure, tread depth, brake pad thickness, rotor wear, alignment, etc are all capable of individually accounting for a ~25ft reduction in stopping distance which leads me to my second point.
2. "Driver-less cars should never drive at a speed where they can't stop in time."
Real world driving conditions present so many uncontrolled variables that a speed considered "safe" by that logic will be pathetically slow. The current state of the art in driverless cars is Google's purely autonomous design which is limited to the pace of a golf cart (25mph). The number of variables introduced by vehicle condition and maintenance alone are so great that any vehicle capable of accurately accounting for each one would be impossibly expensive to buy and crushingly expensive to maintain. While all of our cars are capable of measuring speed (the legal maximum is +/- 5mph @ 50mph in the US, but +/-1-2% is most common) and most cars can measure tire pressure (+/-15% for the most common systems which measure indirectly with the ABS sensor), they don't measure either of those things accurately according to manufacturer specs and things only get worse over time. Personally, I can't see driverless cars ever growing past that hurdle, winding up limited to inner-city commuting/taxi duty.
3. "Same is true of balls and other small fast moving objects, they are not going to ping as kids."
This is actually one of the best examples of why speeds will have to remain slow enough to guarantee a negligible stopping distance.
If a human sees a ball roll into the street from behind an obstruction, they will reasonably deduce that there is a likelihood that a kid is chasing after it and slow down before the danger presents itself. An AI will see the balloon or ball and it will have to treat it as either a complete non-issue or as a mortal danger. If it treats it as a non-issue, it will have to be in a constant state of readiness to stop in the smallest reasonable distance, that means limiting speeds to a relatively glacial pace. If it treats it as a mortal danger, it will constantly produce false positives which leads to frustrated passengers when it brakes hard for every trashbag that blows across a 70mph highway.
That's one of my main criticisms of driverless cars in their current state. They are measurably superior in the areas in which human drivers are most flawed, yet their logic leads them to make decisions that no human would ever consider as logical. Even the most state of the art AI isn't even close to being able to account for the nearly infinite number of exceptions that stem from the chaos of the real world and which require an action contradictory to a principle of placing safety first. Just look at the first "unforced error" reported by the Google program where the AI's logic misjudged the actions of a bus due to logic which defined exceptions around the distance needed to merge on to a road. It just went for the merge, turning straight into the side (pretty much in the middle) of a bus moving at 15mph. No human would have made that decision.
2. Google is being conservative with several cars being tested at much higher speeds, don't forget in the computer world being 1/3 as fast is really far less of a jump than you might think. In terms of accuracy GPS can auto correct distance traveled over time to get much higher accuracy. Much like how most computers clocks are far closer to the correct time than their cheep HW clock would suggest.
3. Your thinking like a human. Kids are really really slow at 10mph (which is fast for a kid) 7.5 feet is over 1/2 a second. So, your car can drive like there is always someone about to jump out from every intersection and sprint across the road without being all that slow. Depending on sensor location they can also see under other cars. ED: the Front drivers side headlight is a much better location to see around parked cars than where a driver sits.
Worst case self driving cars might drive slow when stuff is right next to the road. But, if I can be productive and safe it's a non issue as most driving is not in those conditions and I should slow down anyway.
Google has already said that this kind of decision making will not be a part of their cars' system. The car's reaction to situations is based on the rules that we all agreed on when we passed laws surrounding how humans are to behave on the road.
If the car encounter as obstacle, it will brake. It will not calculate death probabilities. It does not need to.
In the absence of any signage, there is always an implicit "Yield." -- ie, no requirement to stop, but you should look for conflicting traffic, and preemptively slow down enough so you could stop if there turns out to be conflicting traffic.
Heck, in some places even a green light is worth defensively treating as a Yield. In Manhattan, when driving, I barrel through green lights when crossing streets while going up avenues (in center lanes), but treat it more like a Yield, slowing to a quickly-stoppable speed, when going crosstown on streets across an avenue intersection. So what if you have a green light? In practice all three road user types, pedestrians, cyclists, and motorists run reds all the time. You're supposed to drive defensively, and a good driverless car should too.
Anyway, I think driverless cars have an implicit advantage here in that their reaction times will be substantially lower. Hopefully they can use statistical evidence to figure out rules like the above ("Treat greens on 3+ lane roads as "No Stopping" sign, but on smaller roads, treat them as yields")
And the only data that needs to be kept updated, is the local environment. The car is in one place at a time. So bandwidth should not be the issue it would seem.
As long as the car has basic route information and knowledge about where it will be able to connect to servers with updated information. Cars could drive anywhere with basic road info, and download updated and more detailed information (intersection metadata, possible closures, etc...) only as needed. It doesn't always need to be in wireless range, it just needs to know when it will be so it can plan accordingly.
The problem with the statement is that it, as the GP post says, ignores the fact that a human driving a car is no less likely to barrel through an intersection with a missing stop sign.
There are a few other low level problems with this scenario as well, mostly hinging on the weird idea that self-driving cars are just going to barrel into an intersection without a stop sign as if there are no other cars going through it. This is why they have radar/lidar/whatever else.
The fact is that GPS navigation is a mature field (I.e. unlikely to get much better very soon), and is still very failure-prone. I just took an Uber this morning and twice the GPS got confused between the street I was going to and the on-ramp to the highway (the two roads were narrow and unusually close together). Twice, the driver followed the GPS instead of looking with his eyes at the road (the on ramp clearly looked like an on ramp). I run into these sorts of things 2-3 times a month.
I expect my driverless car to be able to disobey signs when I indicate that it would be appropriate or necessary to do so.
I may one day enforce this by carrying around two poles that bear "stop", "do not enter + wrong way", "no u-turn", "do not reverse + severe tire damage", and "no stopping any time" signs on them, and creating pop-up paradox traps for unwary driving AIs.
When there is a stop sign, I stop. When there is no stop sign, I don't stop.
Those are the rules of the road. Stopping for a stop sign that doesn't exist is breaking those rules, and thus putting other people in danger.