87 
The self-driving car trend would appear to be owned by the global car giants and the infinitely deep pockets of Google, but there’s a new UK upstart on the scene that automates boring, regular trips by learning the route of your tedious commute and taking control. And might be able to do it all for £100.
The system, developed by Oxford University’s Department of Engineering Science, currently uses an entirely electric Nissan LEAF as its test vehicle of choice, with an off-the-shelf computer in the boot linked to various lasers and sensors placed around the car’s exterior — and outputting control inputs to a tablet on the dash.
The concept is to let the car learn its way around routes and environments by monitoring your usual drive. Over time it builds up picture of its surroundings, with the hope being that once it’s sufficiently intelligent you’ll be able to enable auto-drive and have the computer take over a regular drive along a familiar route. You can then get on with staring at your phone while the car drives itself to ASDA, with a tap on the brake pedal bringing back manual control.
Obviously this system isn’t yet passed for driving on proper roads, so while the team badgers the Department of Transport for approval to test it in real-world situations, it’s doing endless laps of simulated town centres on test tracks and running up and down private roads.
Prof Paul Newman from the uni’s Department of Engineering Science said: “Because our cities don’t change very quickly, robotic vehicles will know and look out for familiar structures as they pass by so that they can ask a human driver ‘I know this route, do you want me to drive?’.”
The hardware as it stands currently costs around £5,000 to implement, but the professor dreams of a time when it’ll be a £100 optional extra you’ll stick on the list alongside the USB charger and posh paint when buying a car. [BBC]
This Spooky Google Car Is Driving Itself
A Blind Man Is Driving This Car
How a British Florist Solved the Mystery of the Coming Apocalypse
Cool idea but surely the problem isn’t the route itself but the other road users who will be an almost constant random variable?
However, much respect to Prof Paul Newman just for his name if nothing else!
That’s how the systems work (including Google’s). The system maps out the route without the random variable – i.e. it compares each journey and therefore removes the things that aren’t always there.
Then it can cope with the random variables – it knows the object up ahead isn’t always there and reacts to it, whilst it also knows the other object is a bend in the road, postbox, bollard or whatever because it’s always there and it can just steer around it.
so how would Google’s version deal with a new route then? Surely in that case it’s all random variables unless it’s crowd-sourcing from all the other Google cars so there is no such thing as a ‘new route’.
You’re taking the wrong approach, it doesn’t try to predict any random or unexpected variables in a complicated algorithm before it heads off, it deals with the as it comes across them. The same goes for re-routing, if the car needs to change route it will do so and re-plan a new route, just like your GPS does. The rest of the car’s control systems ensure that it’ll follow that route safely. (see my other reply for a bit more info).
thanks
That’s the idea – all the systems are constantly in touch with main base and updating with current conditions.
That’s how they currently do Google Traffic on the maps app – people using Google Maps update the server with their current speed.
No, that’s not at all how it works. What it does is it uses normal path planning algorithms (same used in your GPS, Google Maps etc) to determine the quickest route based on traffic conditions (crowdsourced), speed limits, road conditions etc (this is what the street view cars were for). It then heads out on the road and uses loads of sensory data (cameras, radars, laser scanners, ultra sonic range finders etc) to determine where on the road it is, what obstacles are around, where other vehicles are, which lane it’s in etc. Add on a ton of rather sophisticated algorithms to predict and respond to what other drivers are doing and a lot of logic to accurately follow to rules of the road (being in the right lane, correctly over taking, letting other drivers in, handling accidents, idiot drivers etc) and voila, you’ve got yourself a self driving car.
It’s extremely sophisticated, complex and very hard to do, especially since you have to demonstrate that it *always* does what it is supposed to do, regardless of the surrounding conditions hence why there are no production cars doing this already. This is also why Google’s self driving cars are so much more impressive, 300 000 km and no accidents so far.
The last article I read about it said that the route is driven manually several times, which builds up a street-view style 3D image, which is compared with the GPS (which isn’t anything like accurate enough) and radar to tell the car exactly where it is on the road. Having a ‘reference’ route means it can also then tell the difference between a litter bin or a pushchair waiting to cross, because one of those objects isn’t normally there.
Without the reference route, it’s extremely difficult to say what is and isn’t normally there on the route. And meant in the demo that this journalist was writing about, the car was able to stop at a crossing to allow someone to cross because it recognised that the stationary figure was a person.
Beware, techy/geeky explanation ahead!
What you’re talking about is a different approach using something known as SLAM (Self-Location And Mapping), and it works really well for mapping out new areas and finding where you are in them. However, it does not help the car deal with unexpected objects, it’s purely a positioning method (and very likely used by Google in their self driving cars).
The challenge is exactly as you say, recognising that there’s a person waiting at a crossing or dealing with a stressed out driver pulling out in front of you. To understand these situations and deal with them correctly you need much more complex object based classification (that’s a person because it has arms and legs, has a certain height and moves at a certain speed vs it’s a person because it wasn’t there last time I went past) and artificial intelligence algorithms to make the right decisions and take appropriate action.
You basically end up with four major systems in the car:
1. vehicle control (low level control of the car such accelerating, braking, turning, changing gears, monitoring fuel etc),
2. positioning (SLAM, GPS, routing and much more, basically where am I and where am I supposed to be?),
3. situational management (lanes, overtakings, what are other drivers doing? is that kid about to run into the road?)
4. vehicle / “mission” management & AI (how do we use all of the information from the other systems to get where we want safely and without killing anyone or breaking any laws?)
I get confused.. is every single aspect of the computer advised on driving? Does it know how to drive defensively or react to being cut off by an asshole in a Transit?
Does it know how to let people out of junctions etc?
Hang about….I smell a lawsuit from Google on this one!!