Slashdot Mirror
VW Raises the Bar for Self-Driving Vehicles
Old Man Kensey writes "According to the UK Daily Mail, VW has produced a prototype Golf (code-named "53 plus 1" in a reference to Herbie the Love Bug) that successfully steers and accelerates itself at speeds up to 150 MPH on tracks designed on the spot without pre-programming. It sounds almost too good to be true given some of the problems CMU's prototype has had over the years, but perhaps VW has learned from and extended CMU's research (and within-an-inch GPS positioning probably helps too)."
whilst I don't know what the National Inquirer's credibility is like I would highly doubt if the Mail would publish something which is truely fictional. They do exagerate a few things and love to complain about anything, but I've never seen an out and out lie in the paper
*''I can't believe it's not a hyperlink.''
The problem you mention, that you don't know exactly how the system will behave under all possible conditions, is a problem we have with all computer programs, especially those that include learning. On the other hand, this is a problem we have with humans as well. The reason (well, one of them) that we let humans drive is that we have done such extensive testing with humans driving, to see under which conditions they can drive safely and under which conditions the "behave unexpectedly" (icy roads, fog, weird intersection layouts, poorly marked turns etc). In the process, tens of thousands of humans drove themselves to death.
So I wouldn't count out self-driving cars until we have done the same amount of testing with them.
Another way we can try to make self-driving cars safer is to use something like Rodney Brooks' Subsumption Architecture, where the controller is structured into layers. The higher layers are responsible for "high-level" behaviour like navigating and planning fuel consumption, while lower layers do simpler things like avoid driving into walls based on sonar information. If a higher layer falls, the lower layers operate just fine without it. The function of the very lowest layer could then be just to stop the car if all the higher layers fail. Given that the layers are separate circuits, it's very unlikely that all will fail simultaneously.
Much of the design can be automated. I'm myself working on using evolutionary algorithms to design car controllers. So far we do quite simple things and only in simulation, but with a good enough simulation it could then be transferred to a real car. Check out this post, including videos.
The VW project sits on a few different fields of research under the umbrella term "Robotics", such as mobile robotics (path planning, SLAM, range sensor, scan matching), computer vision (getting meaning from video, tracking objects visually) and machine learning (training various software systems based on learning data, like road colour for example). All these fields have plenty of open problems and many problems that can only be solved in "controlled", a.k.a. near-trivial, environments.
Just a few points about your post:
1) AI will not already be available if it was a commerical effort. Not for at least a few more decades. This is coming from what I have read as a PhD student in Robotics and Computer Vision, so I *may* be a bit pessimistic. But many of the researchers I have talked to, both in and our of academia, feel the same
2) Making an intelligent machine is hard, and it is a long term goal. Commercial projects generally require short term results. The academic research setting allows for long term research that may yield useful results 5-10 years from now. To give you an example, try search for SIFT, which came from academic research, and is now used in many commercial software and robotics products.
3) I seriously doubt self-driving cars can make it to our streets anytime soon. Apart from the lack of adaptable machine learning and robotics systems, the *legal* problems will take a long time to overcome. Our legal systems, at lesat, in Western countries, will have major issues dealing with any semi-intelligent systems that makes decisions for us. Decisions which may directly cause injury and death. Liability is a problem when a learning system passes standardized tests, and it makes a mistake. Who is responsible, the system?
Berserk Manga > All
I'm pretty sure I've seen my van do that.
What I wonder about is why the link doesn't go the other way. I live in a hilly area. Some hills are large, some are tiny. The large hills can trip up the automatic transmission on my car. By the time it decides "I ought to downshift" I'm at about the bottom of the hill and and it has to shift back almost immediatly. Now I have learned how to avoid this by when I push the gas on those hills and such, but it got me thinking.
Why not use mapping data to feed the automatic transmission? The data would be a suppliment to tell the car's computer "we're going up a long hill that has a slope of x degrees" or "this hill is short". That additional information would surely be helpful. You would need topographic information in the mapping data (as opposed to pretending the Earth is flat), but it might make up for it. At the very least, you could keep a memory of the 50 miles of road the driver drives on most (their home area) and the elevation for those spots. This would give the same advantage over most of the area the driver drives, without having to have all that topographic data for everywhere else. This data could be gleaned from the GPS and would fix the little "problems" like I described above.
I would think this would all lead to better fuel effiency. At worst, if the system failed, you'd be back to what all automatic transmissions do today.
Of course, CVTs wouldn't need this. And they are simpler. And more efficent. But we can't put those in every new car. Then what would all the transmission shops do?
Comment forecast: Bits of genius surrounded by a sea of mediocrity.
Over 1 million people are killed in automobile accidents each year globally, 43,000 in the USA. Far more are injured or maimed.
Estimates for the costs of crashes range from 10 to 30 cents/mile, factoring in everything -- health, repairs, suffering -- which is more than the cost of gasoline or depreciation.
It's now down to an engineering problem to build self-driving, crash-avoiding cars. It's the largest preventable cause of suffering and death we have.
Has it been over a year since you last donated to the Electronic Frontier Foundation