Stop lights, sunsets, junctions are tough work for Google's robo-cars After cruising two million miles of public roads, Google's self-driving cars still find traffic lights, four-way junctions and other aspects of everyday life hard work. To be sure, the hardware and software at the heart of the autonomous vehicles is impressive. But it's just not quite good enough yet to be truly let loose on …
The term "Right of Way" is actually defined as the right to "pass and re-pass" across a piece of land, ie you can do so repeatedly without needing to ask (or be given) permission.
"Priority" means that you are legally entitled to go first, so if a pedestrian steps out into the street in front of a car and gets knocked down, they can't complain that it wasn't their fault because they had Right of Way. There are exceptions, of course, for example at a Zebra Crossing where once the pedestrian has set foot on it, they *do* have Priority and vehicles MUST stop for them.
Similarly, at a roundabout, you MUST give Priority to vehicles coming from the right, if you fail to do so and this results in an accident, you'll be liable.
@AndrewDu
You are quite right. I have visited countries where they paint zebra crossings on the road, but nobody pays any attention, least of all the pedestrians. Americans and Germans usually cross only at a crossing; in Britain we are more flexible. Taking a Google car abroad could be exciting, from keep-left UK to right-on Europe. And those countries where they just drive in the middle anyway.
UK roundabouts have two priority conventions: (1) the Highway Code scheme, (2) the heavy lorry scheme. It will take a smart computer to make the right decisions there.
But the road safety campaigners and the police will want to encourage robotic cars. All speed limits rigidly obeyed, all traffic lights (even the most exasperating roadworks ones) obeyed, cars arrested at the flash of a blue light.
Roads are rarely straight with right-angle intersections. Even roads that are relatively straight can change names several times over the course of a couple miles and signage varies greatly in that time. Five, six and even seven way intersections are common. Staying on 279 through downtown requires diving across multiple lanes. Yinzer lefts(running a red light to make a left turn before opposing traffic begins moving) are the rule. In addition to sun we also have these things called rain, snow, snain and fog.
If a self-driving car can make it here they can make it anywhere.
Just to give you an idea what locals face here is a major intersection: https://www.google.com/maps/place/library+road+and+saw+mill+run/@40.3823127,-79.9954395,3a,90y,222.05h,75.17t/data=!3m6!1e1!3m4!1sBM2MDtZwSJYQWB_U9yrnUg!2e0!7i13312!8i6656!4m2!3m1!1s0x0:0x96fd8fc1f9d775ce!6m1!1e1
Also, that is not a pothole, this is a pothole:
http://3.bp.blogspot.com/_Z5oPaW6LdFY/Rja_-VThYtI/AAAAAAAAAD0/2p3F-8JJuOY/s320/car_in_pothole%5B1%5D.jpg
They'll also need to be aware that if the other car at a 4-way stop (or a roundabout in the UK) is a Mercedes*, they are unlikely to be stopping for anything, regardless of road conditions.
Can they also detect drivers wearing a hat?
* Substitute your own 'favourite' manufacturer as necessary.
Autonomous cars apparently have no difficulty in detecting other cars and measuring their movement vectors. If said whatever-your-pet-hate-is has a vector that puts the autonomous vehicle in danger, I trust that it will be able to avoid a collision by braking before entering the danger zone.
The major issue, apparently, is reacting properly to road conditions, not other drivers.
That has possibilities as it's not so much a bus but a reinvention of the train/bus for in city use. I rather like it.
The only thing I can see being a problem is human drivers trying to beat it to the next corner to turn while the bus is going straight ahead. Also, you're driving down the street and then you see this thing coming right at you. Some serious retraining of drivers will probably be needed.
Some of those don't seem difficult at all? The ball in front of the lights, surely the lidar system or stereo cameras could detect that the red traffic light doesn't usually float on its own. Similarly the lidar or stereo cameras could surely detect the housing of a traffic light unit and therefore place the lights within it, avoiding problems with sunlight, balls, other signs etc.
"...installing computer-friendly traffic signals or embedding wireless guides under the roads or mandating that autonomous cars are painted in one color and human-driven cars in another – is out of the window. It's just not an option."
Sure it's an option. All the corporate-military regime need to is say the magic word: "Security".
That opens the taxpayer till for as many trillions or billions it takes to make us "safe".
The fact that - for the most part - these cars seem to be working is very impressive. Sure, sometimes there will be accidents but I'm confident that the technology will reduce these over time. I'm a lot less confident that human drivers will reduce their accident rates over the same time period.
It's interesting that most of the self-driving cars seem to be in the US which is a very car-concentric country with little public transport - it seems to me that a future with self-driving cars for local A-B trips and a decent public transport system for the A-Z trips might work but I don't see the US trying that path at all.
So far as I'm concerned, the acid test for self-driving cars is not traffic lights, other cars, or street furniture. It's how well they cope with cyclists. (Yes, even the ones who have no concept of the highway code and no concept of self-preservation.) My biggest worry is that those of us who do cycle fully abiding with the law will be kicked off the roads because autonomous vehicles can't cope with the suicidal maniacs who give all the other cyclists a bad reputation. The net result of that is that I'd have to start driving to work...
Changing infrastructure for driverless cars across the *whole* road network will not happen
(a) expensive and (b) rather 19/20th century
Some changes may be made - rather like mobile phones in Africa - build some masts rather than a lot of expensive copper - i.e. limited change.
When there are enough diverless cars on the road, they will talk to each other and solve the problem co-operatively.
Networked cars
Oh no, nothing can go wrong there...
In the UK there seems to be a constant program of digging up the motorway system to make it wider / concrete central barriers / smart ... and I can see something being done in limited circumstances.
However I do see self driving cars being viable on freeways / autobahns / motorways as the problem set is so much more limited.
istr the new merc e class can 'drive' on motorways right now.
I agree, merely getting to the point where they can operate entirely autonomously on freeways but hand control back to you when you exit would be huge. While I certainly wouldn't mind being able to drive myself to the bar, have a few drinks, and have my car drive me home, the biggest win for me would be when I'm traveling on the freeway for more than 10-15 miles, and would rather be doing something else like working, surfing the web, talking on the phone, or maybe even catching a nap if the trip was long enough.
If I wanted to go to Times Square in NYC I'd have to catch two flights, and from the time I leave for the airport I have:
30 minutes to airport
60 minutes from arrival to take off
60 minutes first flight
>=90 minutes until next flight
150 minutes second flight
>=90 minutes getting off plane and reaching Times Square
That's at least 8 hours, assuming no delays. It is roughly 1000 miles to drive for me, a self driving car obeying all speed limits could probably do it in 14. If you add in sleep it is basically a wash. If I started out in at say 7 pm Central I could watch movies (or whatever I might do if I was at home on the night before travel) until I get tired, lay the seat back flat and catch some sleep, then wake up refreshed and miss the worst of rush hour with my 10 AM Eastern arrival.
My current car gets about 30 mpg on the highway, for that round trip with gas selling at about $2.20 currently that's only $150 - WAY cheaper than a round trip ticket AND far more enjoyable than a TSA groping and sitting in coach! It would still be a great deal even if gas went back up to $4/gallon. It would be even cheaper with an electric that's both more efficient and can be charged with renewable energy to reduce pollution versus an airplane (assuming enough others did this that there were fewer flights scheduled)
Rubbish.
It's a programming problem.
We also don't really know how people do it and real computer AI isn't invented yet (Only domain specific expert Systems = Databases and badly Simulated Intelligence) as we don't actually even understand any form of biological learning or intelligence.
Why aren't there trains without drivers first? A much more controlled problem. Then trams. Exactly what is Google's motivation on automated cars?
Sadly, no. They are only ATO level 2 (with the exception of the DLR which is ATO 3) which means the driver is still required to operate the doors and tell the train when it is safe to depart.
More info: https://en.wikipedia.org/wiki/Automatic_train_operation
They are only ATO2 due to the unions. The trains don't *need* a driver, they only need an operator to open and close the doors. Due to pay and staff levels this is the 'driver' that does it.
In other news, the unions believes that it is unsafe for a driver to operate the doors and requires a conductor to do it. So if the conductor could operate the doors on the tube then it wouldn't need a driver, surely?
First, the computing power required isn't there yet. All of that kit in the rack would have to be shrunk by several orders of magnitude in size. Even with Moore's law, you will have to have several main processors dedicated to specific tasks and then a central CPU taking that logic in. Most of it will require an RTOS and these systems will require some redundancy.
Second, you will have to deal with obstacles that are also moving and can move erratically or in unpredictable directions. (e.g. pigeons, deer, etc... )
Other drivers....
The point is that while you can constantly improve the driver assist mode, you can't easily make it autonomous.
Lets also not forget the cost.
Google and others have predicted and promised based on hype.
If you cant get xpoint or crossbar memory to work and are already 2 years behind the promised release, then you aren't going to get this working by 2020. And xpoint/crossbar reRam is a critical component in making this work.
There's more, much more in terms of senors and identifying traffic. including modernizing the infrastructure. Not an easy task.
Posted Anon for the obvious reasons.
This is a great article. I'm so tired of reading tons of bollocks about "We're going to be mass producing self-driving cars by 2022" and so forth. This article highlights some of the real world problems that go into creating autonomous vehicles, and the fact that while we've learned a lot, we still have a long way to go before we really have self-driving cars.
While Google may be "projecting" having self-driving cars by 2020, I think it's clear that it will take more than 4 more years to get the bugs out of the project to where the cars actually work.
Every few months I come across a situation where construction or repairs requires the placement of a temporary stop sign, or certainly road workers directing traffic. So how does a Googlemobile tell the difference between a road worker directing you into the oncoming lane (which is supporting alternating shifts of one-way traffic) and some people by the side of the road dancing or waving at cars?
Assuming that most traffic lights are getting their bulbs replaced with LED arrays in the next 10 years, why not agree a standardard that the array also outputs specific wavelenths outside the visible spectrum that a camera can easily detect and isn't drowned out by glare from our star.
could also do something in high-frequency directed sound also to make double sure.
The Sun is only really a problem when it is directly behind the traffic light. It outputs on all frequencies from long radio waves to X rays so just using something outside of the visible range doesn't solve anything. Fortunately the Sun is only behind the traffic light from certain direction/distance combinations, so as long as the hidden signal was strong enough to be visible from a couple hundred feet away you don't have to worry about the Sun because you'll have some views of the signal without the Sun overwhelming it. When it is very low in the horizon it might not be visible to the car until you get close so maybe the car will need to slow when approaching those intersections near both dawn and dusk but that's not a big deal.
Whatever frequency you use it should be 1) able to travel well through fog/rain/snow at least 50 feet or so and 2) directional enough that the car can tell it is coming from the traffic light (to avoid bad guys trying to fake signals and cause trouble - you could encrypt it with a private key, but with millions of traffic lights it would only be a matter of time before it became known)
"Human motorists rely on eye contact to know when it's safe to go or just take the initiative and move first. A driver-less car gets stuck trying to safely nudge its way across the box."
Is everybody in this picture too young to remember (and too dumb to read about) Car Sense Multiple Access with Collision Detect?
If you see someone else's car where you want to be, at the same time you want to be there, you back off, and a semi-deterministic algorithm determines when you have another go (so you don't spend forever saying "you next", "no, you", "no, you").
If you detect a collision, you ... err, OK, more of a problem, that one.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
