Rate of autonomous vehicle safety improvement slowing – research Autonomous vehicles in the US are showing a declining rate of improvement in one key safety metric, according to research outfit IDTechEx. Despite this, the industry is making progress and may be nearing early commercialisation of technology such as robotic taxis. IDTechEx has cast its eye over the 2021 autonomous testing and …
You may not have noticed the essence of this report. That the vehicles can't get along without having to hand over to what you call a bag of mostly water but which, in fact, provides parallel processing abilities way in front of whatever amounts of silicon they can contrive to put into the vehicle.
I'm going to try to make a prediction here. I don't claim certainty. Maybe in ten years' time I'll fish it out and either shake my head or nod sagely.
The prediction is: for the general case, we are not going to get a self-driving car that's so reliable that you can stop paying attention. We might get it for special cases - motorways, some kind of special lanes, very slow speeds, things like that. Not for regular urban roads, not any time soon.
Much like so-called "artificial intelligence", current self-driving technology will improve a bit more, and then reach a plateau where further improvement will be very difficult, and it still doesn't work anywhere close to well enough. Getting past that point will require some kind of largely unpredictable major breakthrough, not incremental progress.
I'd still love to have a car that can self-drive on the motorway, though, and that might be feasible, especially with some kind of smart motorway infrastructure.
"I'd still love to have a car that can self-drive on the motorway, though, and that might be feasible, especially with some kind of smart motorway infrastructure."
I think this will be the winner because of all the self driving technology with obvious benefits this is it. I expect it makes sense for haulage to get the truck and goods over long distances with only local driving being needed and it would improve the driving experience for people travelling long distances. It could even increase the speeds on the motorway! Not that the law makers would do that.
And unload itself at the other end.
Not in scope, a truck can simply decouple its trailer and couple to another one and start another journey. The unloading of the container by automated process is easily solved. To some extent there is automated movement of containers in the big shipping ports in places that are not hampered by unions.
One disengagement per 8000 miles in a mixed environment seems quite good performance ... If disengagement simply means human had to take over rather than any accident.
While I agree that general AI suitable for all driving conditions is unattainable, a mixture of better sensors, better processors, improved training sets and specifically designed roads will, I think, bring a 'close to' general case for most roads, not just motorways within 10 years.
It's more a question of the time it takes for infrastructure to be renewed.
I drive around 1000 miles per month. From a technical point of view one disengagement per 8000 miles is fantastic but if I can effectively ignore my car for 8 months while it drives itself then it's very unlikely I'm going to be paying attention the one time in 8000 miles it needs me to do something.
This is the inherent problem with autonomous cars. Them becoming better means the driver pays them less attention which means when they do need attention they're less likely to get it. Autonomous cars are probably already good enough to fall into this awkward middle ground. Moving past it is going to take an awful lot of work. Any disengagements are probably too many and autonomous cars should concentrate on manoeuvering themselves automatically into a safe position if they're unsure of something rather than expecting a distracted driver to help them out.
Forget the fact that the old meat sacks are causing more harm behind the wheel, just keep demanding a higher bar for the automation then flap your arms around and claim because they aren't perfect, they should be banned and we should all walk to work.
We should be focusing on implementing self drive in environments where it is safer than the person behind the wheel, and limiting it in places that it's not. We should probably also acknowledge that the idea that a primarily machine learning driven self drive isn't going to shine on roads that were designed with zero accommodations to driver automation. That means adopting v2v communication, modified signage, high resolution road maps and data, and things like stoplights and signs that tell cars to stop or go. We may also need to have remote pilots monitoring cars running in self drive to avert problems if the passengers miss them(or to allow elderly or disabled people equal access).
Extra sensors and bigger models aren't going to get us there by themselves. But the bottom line is, if the car is killing less people than the drivers, let the car drive. Arguments about perfect safety are like the Trolley Problem, a flawed argument with little bearing on the real world when you don't consider the reaction of actual people in the equation. And even in a robotaxi, there should be an emergency stop any passenger can hit if things are getting out of hand.
"We may also need to have remote pilots monitoring cars running in self drive to avert problems if the passengers miss them"
That will only work if these cars are going to moving in feet per hour. The reaction time at any reasonable speed will be far too slow.
And no one has moved any goal posts. AV are not up to human standards at the moment and will not be any time soon.
“ That will only work if these cars are going to moving in feet per hour. The reaction time at any reasonable speed will be far too slow.”
The robot delivery vehicles that operate in Milton Keynes work this way, because they are relatively low speed. When they encounter an obstacle that prevents progress they call home and a human operator gets a camera view and sorts it out. If the thing is in a pickle then a van is despatched.
So this situation is applicable to full size vehicles, because when they get into stalemate situations (like human drivers do when they don’t cooperate and nobody wants to back down) a remote operator can be used. If the passengers in the car, for some reason, cannot sort it.
"AV are not up to human standards"
I think you overestimate the standards of human drivers.
For appropriate roads (and here motorways are pretty much the ideal case) AV can already do two things:
- Navigate the motorway better than a meatsack
- Hand back to a relatively fresh meatsack at the far end
- Thus improving the safety of roads it's not used on
"Forget the fact that the old meat sacks are causing more harm behind the wheel"
I think you're forgetting that when it gets too hard for the machine it's the old meat sacks that have to engage their greater parallel processing power to dig the machine out of the problem it's got itself into. So exactly who or what is it that's causing more harm?
I'd like to be wrong, but I fear you are pretty much right. Expressway driving is pretty simple most of the time. Except when it isn't. So I think autonomous long distance trucking may be feasible. ... As long as the vehicles reliably pull over and wait for a human to help them out when they come to something -- snow, an accident, construction zone, whatever -- that they can't handle. Maybe a human can shepard them to the expressway on-ramp and to the terminal at the end of the trip? I have no idea whether the economics work. And the potential problems are legion. But just maybe that works. And if that works, you and I will likely be able to put our vehicles into autonomous mode on the expressway with some confidence that it'll wake us up when it needs help rather than assassinating us or innocent third parties.
But the notion that autonomous vehicles can safely drive urban, or suburban, or rural roads now or any time soon seems to me to be crazed. Too bad. I'd feel a lot safer sharing the road with well designed, reliable, autonomous vehicles. If only we knew how to build them.
"And if that works, you and I will likely be able to put our vehicles into autonomous mode on the expressway with some confidence that it'll wake us up when it needs help rather than assassinating us or innocent third parties."
I drive 1000 miles per week and am looking forward to the day. But first, all the autonomous vehicles need a mandated comms system standard so they can talk to each other. Otherwise we'll end up with "road trains" of trucks convoying down lane one and no one else can get off or on at the junctions. It'll need to be "clever" too, otherwise we'll end up with a stationary lane one as everything slows down to create gaps for other vehicles trying to leave or join. It's bad enough as it is now at times, but with AI's trying to optimise truck journeys, the default state will almost certainly lead to long "trains" with almost no gaps.
But first, all the autonomous vehicles need a mandated comms system standard so they can talk to each other.
That'll happen eventually I imagine. But like a lot of other stuff, it's probably harder than it looks. Failure to communicate because of equipment failure or dirty sensors or whatever won't leave anyone the worse off. But negotiating a merge or turn with the wrong vehicle by mistake because of reflections or probably a dozen other unanticipated conditions is likely to be disastrous.
BTW, I imagine that autonomous vehicles not built by sociopaths will follow other vehicles at a safe distance and will probably allow merges according to some set of rational rules. So traffic will still flow. As well as it does now. Probably a bit better. But it can certainly flow even better if vehicles can communicate.
"Maybe a human can shepard them to the expressway on-ramp and to the terminal at the end of the trip? I have no idea whether the economics work."
I suppose they'll work if the human is satisfied with just being paid for the shepherding bit and not for the journey in between. Maybe there's a possibility of a Working From Cab gig to supplement the otherwise constrained income.
We will get there, but the metric is simply that autonomous driving has to reach 1/10th the accident and death rate of human driving to reach the level of public acceptance for regulatory approval.
Comparing autonomous driving takeovers is flawed as the article states because if you keep testing only in the same area eventually the car figures it out pretty well. Then you give it a new more complex area like San Francisco and it has to learn a bunch of new stuff. New York City and Boston will even worse, and London worse than those two. Don't even ask about third world cities!
Others try to compare accident rates but that's similarly flawed since people let their Tesla etc. drive in occasions for which it is well suited and humans are not (like traffic jam stop and go driving) and less so during poor weather or areas with complex driving scenarios (e.g. major urban centers) so the two numbers are not comparable.
When autonomous cars are able to drive everyone, they will need to reach 1/10th of the human driver accident / fatality rate to gain the level of broad acceptance needed for regulatory approval for "Level 4" or "Level 5" operation. There are two reasons for this. One, most people believe they are better drivers than average, so they will demand that cars be better than average. Two, most accidents are caused by humans who are driving distracted, drunk, sleepy, or other not devoting 100% of their attention to the road.
The human driver accident rate of people driving while paying 100% attention would be far better than the actual rate, but that's the rate you have to beat (plus the "I'm better than average" factor) because people don't want a car that can drive better than they do while they are texting or after having a few beers and should have called a cab. They want it to be capable of driving better than you'd drive if your father in law let you take the wheel of his classic car he loves more than the child you married!
Boston will even worse
No kidding. What those who haven't driven there don't know is that the actual rules of the road in Boston have no relationship at all to the rules you have to learn in order to pass a Massachusetts driver's license test. The actual guiding principle of Boston driving is Vehicle with the least to lose from a collision has the right of way. Once you internalize that, Boston driving becomes not so scary. Actually, in a way it's less scary than Los Angeles where the drivers expect you to know what you're doing and to be paying attention to what's going on around you -- a concept that ends to be beyond the comprehension of many out of towners.
It sounds a bit like Dublin. Belfast drivers always said Dubliners were terrible drivers. On the few occasions I went to Dublin I always felt more comfortable there. I got the impression that drivers knew what was behind them, something Belfast drivers never seemed to manage.
Rural areas are easier, because there is less to hit. Sure, some cows/deer etc. but fewer people in the road, or people driving on the road so fewer concerns about killing people or at fault liability.
Weird stuff like the one lane roads in parts of western Nebraska and rural Ireland will be interesting for autonomous vehicles, but at least reversing 1/4 mile to a place with enough room for two cars to pass is a breeze for a computer!
In the vast majority of cases, I'd be very happy to take even quite a lot longer to get to destination, if it meant I can be distracted or asleep during the trip.
A car can't be an office, but still, 3 hours of doing something useful (including relax activities) while the car drives itself will generally beat 2 hours of driving and 1 hour of full productivity. And, most of the time, sticking to speed limits won't increase trip time that much.
There are two factors that seldom get mentioned. One (and the most important) is that, in the absence of understanding autonomous systems can only be reactive to a limited range of stimuli, so there's inevitably an upper limit to improvement. The other is that testing has almost certainly been conducted in increasingly complex situations as the technologies have developed. Both factors make apparent progress increasingly hard.
However, overridingly, until cars can genuinely think, accumulate complex memories and draw intuitive judgements from them, the technologies will never reach the competence of the highly experienced and attentive driver. The best we'll probably achieve is replacement of the least attentive and most careless.
As an example of what I mean, I was driving one dark morning along a narrow unlit country road with a 50 MPH limit and many concealed tracks on both sides to plutocrat dwellings. Not far ahead off my near side, I briefly saw a moving red light partially concealed by trees. My brain said "fast car emerging suddenly into the road" so I braked. I was dead right - a low sports car shot into the road from an invisible track and turned across in front of me. If I hadn't braked we'd have collided. I'd be surprised if an autonomous system could work that out as the brief glimpse of the red light was about 45 degrees off my near side only lasted a second or so, and the offending car did emerged a few seconds later with no other indications of its presence in between.
You make some great points, and I think you nailed the reason that perceived progress has slowed. I think we need to be ramping up alternate methods like vehicle to vehicle communication and warnings to help address problems like your beyond line of sight sports car example.
That said, while you are clearly an attentive driver, the accident statistics clearly bear out that both the driver of the other car in your story, and a huge number of other drivers aren't paying that much attention. So I think it's the wrong place to set the bar for a self drive system to be able to beat a fully alert, safe and expert driver.
It also exposes a huge issue that gets close to zero buzz, which is driver training. In the US, for most states the training to get a drivers license is a joke. Human drivers aren't expected to even meet the standards they did 30-40 years ago. Regardless if they are driving themselves or the car is, modern vehicles have so many assists and features, that even safety features like lane following or emergency braking can trip up an active driver if they aren't familiar with how the actually work.
They are basically doing the same thing as the 737-MAX software system did, surprising an operator and either fighting their controls or causing unexpected things whey weren't prepared to react to. The idea you can just drop a driver behind the wheel of one with zero training is one of the big sources of these accidents. People need to understand how these things work, and do the right things at the right times, just like they learned to with the gas, brake, and steering wheel.
Instead, a person can go from driving a 90's geo metro to a "FSD" dual motor Tesla with literally hundreds more horsepower and just drive it off the lot, and probably over the curb across the street.
Both these posts confirm to me that fully autonomous vehicles will only become practical when they are communicating with the other vehicles around them and the road they are on is designed for autonomous vehicles only. I think we will be waiting a long time for that as the car makers wont want to do that.
The comment about modern driving aids is also very true, not had first hand experience myself but have heard much complaining about having to fight the steering wheel on things like lane assist - most people turn things that off.
Driving test ? As my instructor told me, he was teaching me to pass a test, it is afterwards that you learn how to drive !
And the fast car emerging? that is all down to experience, something my daughter has realised after driving for 5-6 years - she now can tell if another driver is going to do something like that. Hence why auto and manual cars wont be allowed to share the same road space !
Or the way we teach people to drive around cones in a parking lot, test them in a quiet suburban street and then let them drive on the freeway. Of course we can't teach them on the freeway because it would be dangerous to let learners drive in freeway traffic
"The last 10% of building anything takes 90% of the effort"
So far, the various autonomous vehicle initiatives have been able to pick off some relatively low-hanging fruit. Motorway driving, where vehicles are generally going at the same speed, lanes are clearly marked and rules are clearly defined. Supermarket car parks, where everything is slow moving and clearly marked. Etc.
But now, we're hitting the complex stuff. Town and city driving, where pedestrians and cyclists can - and do - behave irrationally or unsafely. Road layouts (at least in Europe) inherited from the romans and then mutated through centuries of gradual evolution and hasty modern hacks to deal with population growth and the mass adoption of motorised vehicles. And so on.
And dealing with all these scenarios is exponentially more difficult than what's already been solved. And they all need to be solved to the level where 99.99% of situations can be safely dealt with, because neither the public nor the various government car-safety departments will accept anything less...
If something unexpected comes up, I will never trust autopilot to find and implement a solution.
Instead the vehicle should disengage autopilot and execute its emergency braking. Don't let the meatsack restart autopilot until the vehicle has been stopped for several minutes and then only if airbags have not been deployed.
"Instead the vehicle should disengage autopilot and execute its emergency braking."
So you're barrelling along in the fast lane when somebody changes lane into the gap between you and the car in front. Naturally there's a long train of cars behind. So, because the vehicle is suddenly presented with an object just in front of it it should execute emergency braking and let the cars behind pile into it
Fortunately my wife's car which is equipped with forward-looking cameras for such a situation only throws up a warning and doesn't brake but I've seen it complain twice where emergency braking would have caused an accident and she's complained about other situations when she's been driving. The vehicle doesn't really have situational awareness and doesn't know there's no need to brake let alone that braking would be the worst thing to do.
I find that the problem on busy but fast moving roads is that the actions of other drivers have too much influence over the self driving. It responds in a way that is very frustrating. It’s superb on relatively clear roads, but it isn’t designed to cope with the erratic and selfish behaviour.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
