Slashdot Mirror
DARPA Challenge Prize Money Restored
antispam_ben wrote to mention that, some three months later, DARPA has been able to find the money to offer cash prizes once again. The DARPA Urban Challenge will go forward next November with more than $3 Million on the line. From the article: "The race will see as many as 90 teams 'drive' an unmanned robotic road vehicle through city traffic, competing to finish a 60-mile course within six hours. Set for November 3 of next year, the challenge will call on robots to safely obey traffic laws, negotiate busy intersections, merge into moving traffic, avoid obstacles and navigate traffic circles. DARPA has yet to disclose the race location, but has said it will be in the western United States. The government research group didn't unveil the 2005 Grand Challenge location in the Mojave Desert until weeks before that race, in order to avoid giving any team an advantage."
I sure hope it's a closed course, because I'd hate to be t-boned by an errant robotic Touraeg.
but, while as a software engineer and electromechanical hobbyist I fully appreciate all the challenges involved with these robotic drivers, I'm just not impressed by systems that have courses plotted into them and use GPS and high resolution maps and intimate fore-knowledge of the landscape, etc. As a driver, -I- don't need that fore-knowledge to get from Sacramento to Manhattan - thousands of miles successfully navigated without any more fore-knowledge than that I have to travel generally North East through many states.
I will be impressed when driving automation systems can start with a general idea of where their source and destination locations are and can read the signs to figure out how to get there. They must use perceptive powers to avoid colliding with other drivers or running down pedestrians and following the rules of the road instead of range finders and lasers and GPS-based speed limit adherance and other such nonsense.
Until the system can be boiled down to a pair of eyes and a pwerful set of smarts driving , in my view, it's just an elaborate obstacle course being followed by the likes of this robot. I understand "baby steps", but "they" tend to avoid tackling these big challenges and instead continue to focus on these contraptions that just, plain aren't smart enough.
IMHO, of course.
They didn't need everyone from the last challenge to have finished it. They only need one.
The fact that they got four finishers last time means the cross-country technology works. Now that removing the remaining bugs and improving cross-country technology is just a matter of time and money they can move onto the next step: urban driving.
I live in Chicago and there was a traffic circle by my old apartment - one of the ones they just throw in the middle of the intersection. There were two one-way streets and only one had a stop sign. Each car that came by seemed to go a different way around that thing...and barely any of them obeyed the stop sign. I'm surprised I never saw any accidents there, although one time it was pretty close. The robot drivers can't be any worse at going around them than the human drivers, especially if they are programmed to do it correctly.
I know you are just kidding, but if you think about it, robotic driving is not rocket science, exactly.
If you think about it, all the robotic drivers in computer games such as Grand Theft Auto are pretty damn good, and can follow rules and stick to routes much better than their human opponents. So, driving/navigation algorythms have been developed a decade ago, all they need is a good way to recognize their surroundings.
With this in mind, this whole driving challenge is a problem no different from OCR or voice recognition.
Obama likes poor people so much, he wants to make more of them.
Not to mention, I'm fundamentally opposed to AI drivers because they take the fun out of it :P
When I was 18 and knew everything, I used to think everything was easy also.
Coïncidently I've seen a documentary about the second challenge.
In the first challenge the teams, composed fo bright fellows, all failed exactly because it isn't so straightforward.
The difficulty would exactly be adaptive decisionmaking of the robots; would DARPA (a military instance wanting automated vehicles) put in 2mio USD if it were as easy?
I think we can keep recursing like this until someone returns 1
The algorithm (steer down the middle of the road and avoid obstacles) is not the problem. The problem is machine vision, or interpreting the sensory data about your surroundings. This is a heck of a lot easier to do in a game where your "surroundings" are virtual and created by the same program than it is do to in the real world! In other words, a game already has an accurate model of its surroundings that is updated precisely in real time, and never encounters any object that isn't programmed in. An autonomous vehicle needs to build the model in real time, plus deal with how to interpret objects it has never seen before... sounds a LOT more difficult to me. Plus, in Grand Theft Auto, simply running over any pedestrians encountered is a viable option. One hopes this is not true in the DARPA Challenge!
Video game driving algorithms rely on a discretized version of the world, meaning there is a finite set of possibilities for everything. There are far more possibilities in the real world, and a real-time system like this must take them all into account.
You have a point that this does boil down to the problem of reducing a car's surroundings into meaningful data, much like in OCR or voice recognition, but there is VASTLY more data in the real world than there is in a single image or audio file.
Humans have the benefit of a built-in filter so that we can pay attention to things deemed important, and ignore those that are not. If we had to take the time to process every piece of data that came in through our five senses, we would never be able to keep up, let alone focus on completing any meaningful work.
The mars rovers have the luxury of having nothing but time. They can take a set of pictures of their surroundings, and sit and think and crunch on them for a while before deciding on the best course of action. In a race situation, especially one with moving obstacles, that's just not an option.
AI often has the appearance of seeming easy, because so many of the calculations involved are not even conscious thoughts in human processing. This is a much more difficult problem that it seems on the surface, though.