The first Darpa Grand Challenge was held last weekend. The challenge was to build an autonomous vehicle that could successfully negotiate a 150 mile off-road course. The most succesful vehicle became stuck in a ditch a mere 7.4 miles from the start line, with many vehicles failing to negotiate the first corner.
My question is why was the general standard of entries so appallingly unsuccesful? Out of fifteen starting entries only six made it past one mile. One team had produced a vehicle based on a two-wheel motorbike - which seems pretty damn stupid as it promptly fell over as soon as the team members released it on its journey.
I can see that completing the entire course is a very difficult technical challenge I was surprised at just how poorly all the entrants faired. Can anyone explain to my why this was?
Simple: it’s hard! The team that went 7 miles showed they were on the right track. They got stuck and their truck dug in. It takes a lot of skill to get yourself unstuck, and when offroading you will get stuck.
Did you see the videos they put up on their web site? Pretty awesome. They need to work on the control system - throw in some fuzzy logic or something, because when the bike is going straight it is constantly correcting a lot. When a human rides, they correct all the time, but just a little.
The March issue of Scientific American has an article (written before the challenge, obviously) about the travails of the Carnegie Mellon team that ending up doing the best.
It’s quite a difficult challenge, since the vehicle has to traverse an unknown course of 200 miles of very rough terrain in less than 10 hours. The vehicles have to be completely autonomous.
But according to the article, they did better in their own testing than they did in the actual challenge: in testing they had a run of over 100 miles, in admittedly less challenging terrain. In the real challenge they got 7.4 miles before putting two wheels over the edge of a bluff and getting stuck.
I think that the main reason is time. The challenge was announced in July 2002. That is just over 18 months to have a finished product. I wonder how well Carnegie Mellon will do with another two years of work.
Another problem is the fact that computers, while able to do many things very quickly, are basically very dumb. They are not able to adapt to things that are not programed. Just about a week before the challenge, the Carnegie Mellon robot over corrected in a turn, and flipped over. A person would probably corrected without even thinking about it.
Or take a look at the wild corrections the motorcycle does in the video I referenced above (the “driving in circles” video). However, the track stand (balancing while not moving) is very cool, and not easily done by a human (although it certainly can be done).
The things that humans think are hard, like finding phrases in millions of documents, playing chess, cracking encrpytion, simulating nuclear explosions, forecasting weather, etc, are all fairly easy for a computer. A lot of the work in these fields has to do with making algorithms more efficient, because using the brute-force approach is often impractical given the current state of computer hardware.
The things we think are easy because any 3 year old can do it: running around, understanding language, having feelings, etc are all very hard for a computer to do.
I don’t think there is universal agreement that this was “a poor showing”, at least not in the robotics field. In fact, here’s an article titled “DARPA Challenge a Stunning Success” at http://www.mikeslist.com/81.htm.
A quote: “But that’s not the story you heard from the mainstream media over the weekend. All the reports I read were laden with disappointed reporters, who snidely focused on the fact that no entry finished the course, or even made it to the eight mile mark in the 142-mile race. This from a bunch of journalists barely technical enough to find the Mojave Desert using Mapquest.”
The best one did about 5% of the course. I’m reminded of films of early attempts at airplanes, which showed various attempts crashing right at take off (very funny film). And I’m wondering how may years it took before any airplane could fly for 5% of the scheduled course. From what I’ve read about this in technical articles (not mainstream media), they mostly are pretty happy with the results.
Thanks for the answers, I guess it was much harder than I initially thought.
That article that t-bonham refewrs to makes the interesting point that teh robots all failed in different ways, so collectively all the challenges have been solved by the various teams so it’s essentially a matter of taking the best from each area and working from there.
Personally, I was impressed the Red Team car even made it 7.4 miles. I think DARPA had unrealistic expectations when they expected these vehicles to be able to traverse offroad switchbacks, bridges, and other obstacles given the constraints of the race.
Remember, these folks only had about a year to design their vehicles, and they had to be completely autonomous, meaning ZERO human intervention. You were only shown the checkpoints you had to hit two hours before the start of the race. That meant you’d need to be super-organized and have a team of guys dividing up the course using the smallest scale satellite/ GPS maps available, which are still pretty crappy because most vehicles can’t traverse a five foot wall, and that’s barely noticable on a topo map.
Could these folks have made better vehicles and done a better job? Sure, given more time. But the prize is only $1 million, so why spend $5 million making a vehicle to accomplish the job?
I think the showing artificially looks bad because the course had many hard obstacles up front which tested the vehicles from the beginning. I’m sure DARPA planned it that way however, so that each of the teams didn’t have to drive 50 miles out into the desert to rescue their vehicles just to find out their vehicle hit a rock too hard and flipped over. Better to find out 7.4 miles in…