Slashdot Mirror
DARPA Grand Challenge 3
Meostro writes "DARPA announced the 3rd "Grand Challenge" today, The DARPA Urban Challenge. "To
succeed, vehicles must autonomously obey traffic laws while merging into moving traffic,
navigating traffic circles, negotiating busy intersections and avoiding obstacles." This year's new twist is two tracks for entry: the first is the same as the previous two challenges (develop on your own without Gov't. funding), but the second involves "submitting a detailed proposal for up to $1 million of technology development funds." Here is the PDF press release ."
Yes, they do research in defense but shouldn't there be a little more than a tiny graphic or blurb about what work they're doing? Couldn't they at least take the time to write an abstract or 1-2 page paper with unclassified information on each project?
Instead, I find the following links in the 'Archives':
- Quantum Computing
- Infrared Focal Plane Array/Uncooled Integrated Sensors
- Advanced Lithography
- Most of the Other Links
My alma mater has produced better papers than this in these fields. I know that a lot of this stuff isn't classified and they list their programs on their sites, why can't they do a better job in showing the American public what they've done with our money?The Grand Challenge Forums are flooded with only vendors. Where are the designs and reports by the teams from older Grand Challenges? Why isn't this structured more like RoboCup where the learning algorithms are released every year so that future contestants can build on this?
The fact that this contest isn't run in a more open way makes it seems like less of a "contest" and more of a "do our research for us!" kind of thing.
My work here is dung.
...how this can be "safe" (as far as it can be anyway) with cars which are automated going on roads for which the system works (I'm assuming most are based on the idea of staying within the white lines) I worry about this quote...
We believe the robotics community is ready to tackle vehicle operation inside city limits. - Dr. Tony Tether, DARPA Director
You can build the safest car in the world but there is always a need to be able to take a very quick decision to avoid some other idiot who might be breaking the rules of the road and not be in an automated car... still, if we all had them...
*''I can't believe it's not a hyperlink.''
Its too early to go urban. They should have spent at least another 2-3 years perfecting autonomous navigation in unstructured environments.
I know last year's challange seemed to be won rediculously easily, but I have seen no proof that that dormain has been fully conquered yet. If they wanted a challenge why not move onto wooded or swampy areas.
In this case it seems they are juat setting themselves up to fail.
Surur
Information is the location of things. Computation is moving things around.
Still trying to think of a clever sig...
but what do the two challenges have to do with each other from a technical standpoint?
When you boil it down, they're the exact same thing -- this is just a couple orders of magnitude more difficult. The previous challenge didn't have them dealing with any dynamic variables -- no passing vehicles, no being passed by vehicles, no boulders rolling off a mountain, etc.
And if you're going to solve those problems, why not do it for real? A boulder falling off the side of the road is reasonably uncommon. A car cutting you off is not (n.b. -- the challenge doesn't actually talk about this as an issue, and it may not be; we'll know more after May 20).
It's still all about road detection, object detection, and avoidance. And you're asking what they have to do with each other technically?
Are they going to give the robots the GPS location of all the stop signs and traffic circles?
Again, we won't know until after the Participant Conference on May 20, but I'd actually suspect they will, along with info on what speed limits apply in different areas (as they did last time). This is not unreasonable -- GPS mapping a city is pretty trivial when it comes down to it, and I doubt that the challenge is geared toward being fully dynamic -- e.g. you'll still follow a predetermined route, there won't be sudden changes in traffic rules (no road crews), and so forth.
That said, even if you have full GPS info on stop signs and so forth the most that's useful for is that you need to be watching out for a sign coming up soon. GPS isn't accurate enough (at least on a moving vehicle) to rely on it for road signs -- coming to a complete stop 3m beyond the stop sign doesn't work so well. So they'll still have to visually recognize a lot of traffic signage.
In some ways this will be easier than the previous challenge -- this is all low speed, so the issue of not being able to process the incoming data in real time will be reduced. On the flip side, you'll have to process a lot more data this time -- as you said, you must be able to recognize the difference between a boulder and a bush for this challenge.
I'll be impressed with no crashing into each other, before they worry about compliance with all traffic laws.
I'll be absolutely stunned if anyone succeeds this year, and moderately surprised if anyone succeeds at the one after.
But once this is complete, on to the next challenge -- mixed mode driving (urban, suburban, highway, maybe offroad). Then you can't tailor your algorithm toward a specific goal.
Think about it. City driving is designed to be easy. In fact it is really really easy. You are told exactly where to go with visible lines, lights, signs, etc which are all designed to be noticed and easily intepreted.
Yeah - there's lots of information, and that's the problem. You're not just concerned about finding the road and avoiding obstacles as in the desert challange, but rather are in the middle of a rapidly changing environment that's presenting an information overload (unpredictably moving cars, people on sidewalks, traffic lights, road signs, other signs) that needs to be correctly sensed, interpreted, prioritized and reacted to (predict path of other dynamic objects, which are operated by intelligent agents vs on predictable trajectories), all in real time while at the same time trying to follow a high level plan such as traversing a traffic circle while staying in lane and trying not to kill anyone. The AI component is at least an order of magnitude more complex, probably much more.
The hardest part of GC1 was finding the road! When it's layed out for you nice and easy.... man thats a cakewalk.
Staying on the road is probably one of the easier tasks, but I don't think it's a cakewalk unless it's so tightly controlled as to avoid all the real-world things that add difficulty such as nightime driving, rain, old road markings, confusing exits/on-ramps, etc, etc. Remember also that all the nice clues such as road markings and curbs may be obscured by traffic, and follow the car in front only works when the car in front isn't switching lane or swerving to avoid hitting something, etc, etc.
You can build the safest car in the world but there is always a need to be able to take a very quick decision to avoid some other idiot who might be breaking the rules of the road and not be in an automated car... still, if we all had them...
This is the Grand Challenge I was really waiting for. I believe that the experience gained in the previous Grand Challenges is practically useless for this new one. This new challenge will involve true AI, that is, AI that has true general learning capabilities and the ability to adapt to new situations. A true autonomous vehicle will need to have common sense understanding and the only way it can have this is by learning through trial and error, imitation/observation, the capability of being trained via communication (with a trainer) and the well known principles of operant and classical conditioning. In addition the AI will need robust and sophisticated perceptual (visual and auditory) system in addition to sound motor control/learning mechanism. This AI will forcibly be based on some sort of neural network or a integrated collection of neurla networks.
What Darpa is asking for is none other than the solution of the AI puzzle. I'm afraid this is worth much more than a few million dollars.