Slashdot Mirror
Inside DARPA's Robot Race
Belfegor writes "The PBS series Nova has a great feature on their website, regarding the coverage of the DARPA-sponsored 'Robot Race' in which driverless vehicles 'competed' in a 130-mile race across the Mojave Desert. The full show is available on the website, and besides that they have plenty more information about the robotics behind the challenge, and also some pretty cool out-takes from the show."
I remember an old nova special about self-navigating robots, and at first it took about a day to cross a room.
But mostly these robots depend on the assumption that everything remains still.
I don't get it.
PBS broadcast that show last night. While I realise that is is a little 2001 to actually watch a program when it is braodcast, I did. And I really enjoyed it. I am hardly current on the status of autonomous robotics and I was pleasantly surprised by how far along the technology is. 130 miles through the dessert using only GPS and local sensors is a pretty amazing feat, and that course was tough. It features mountain switchbacks, tunnels and other hazards. If you even have a passing interest in robotics I recommend watching the show.
Unlike Carnegie's "H1ghlander" and "Sandstorm", Stanford's "Stanley" VW Touareg had no fancy motion compensated sensors and the team didn't flesh out the race course with more GPS data and tell the vehicle how fast it could drive in certain areas. Stanley's software did all that on the fly.
Also, the SuperDAD Toyota pickup looked like it had a tenth of the tech of Stanley but it was doing almost as well. If only the laser sensor hadn't detached itself from the roof.
it is interesting just how involved the contestants are. This contest is their life. They mentioned several times in the show how many months of long workdays they spent to build and program these cars. And, then, who owns the work? Do they at least get patent recognition on some of the innovations? Some of the software they talked about was truly seriously cool stuff.
Sidenote: One hour of Nova or Frontline is like watching 5 days worth of "learning" and "discovery" shows elsewhere. It's amazing how good some of these shows are.
After watching Why We Fight, I'm not so keen on something like this anymore.
A feeling of having made the same mistake before: Deja Foobar
I will say, I was impressed, and surprised that I did not see an article on it at
I will say, that aside from "Stanley" winning the race on completion and time, I also believe that Stanley was the best technology. The H1lander and friend were micromanaged, and there were two vehicles that had different strategies (the tortoise and the hair) and it took almost the whole 2 hours of a team of people to map out the course and program the robots. They then added the fudge factor for human error with the fast and slow strategies.
Stanley was programmed in minutes of receiving the map, and it calculated its speed dynamically on its own. Stanley had "adaptive vision" which overlaid laser, video, and other sensory data to create a dynamic field of view of what was safe to drive through.
Now, what shocked me, was that so many teams finished this year. Nobody got past 7 or 9 miles last year, and many vehicles passed the entire 132 mile trip this year. Watching the vehicles drive was impressive. Most of the time, they appeared to be manned.
The course was not easy, by any stretch of the imagination. With the success of Stanley, I believe that this will increase the adaptive and learning capabilities in current software controlled systems. Currently, software is brute forced into trying to accommodate all possible logical conditions, which is impossible, and often just wrong.
The interesting thing for me is that the method we use (our eyes) was too difficult for machines. That's why all those robots used lasers, and other techniques. We've come far, but we still have a long way to go.
'Coverage' and 'Darpa' in the same paragraph.
Another interesting point is that it seems to me that this is the development arena for the military's new autonomously roving gun platform.
Moderation in All Things... Especially Moderation - gurutc
http://www.mininova.org/tor/266446
I would have entered a giant mechanical penis shaped robot car with "Kill all humans" written on the sides.
Too bad I've been so busy slacking this year.
What do you do in the future when one of these is mass-produced and forgets its turn signal and cuts you off?
Do you scream and give it the finger?
Throw rocks at it?
Run it off the road?
Launch a homing missile at it?
Any way around it, driverless vehicles will have no rights in our future society!
Who will speak up for the robots?
He who knows best knows how little he knows. - Thomas Jefferson
Let PBS know what you thought about the format, show, or anything else.
-Ian
Last Saturday, Digital Village Radio did an interview with Jason Spingarn-Koff, the filmaker of The Great Robot Race, and Sebastian Thrun, the leader of the winning Team Stanford. Here's a link to the mp3.
So DARPA funds this to create autonomous supply vehicles, which might work in a traditional battle with clearly drawn front lines and relatively secure transport routes behind the lines.
It seems to me like 21st century warfare is a whole different animal - how hard would it be for a motivated, talented individual to figure out some simple attacks for the navigation systems on these vehicles, and get loads of sweet US munitions delivered to their doorstep? How effective would one of these vehicles be in an urban setting? How easy would it be to create a series of obstacles that would paralyze one of these vehicles?
It's amazing technology, for sure, and the Stanford and CMU teams deserve kudos. I'm just concerned that with the current rush to technological solutions and shift away from "boots on the ground", this technology will be in battle zones far too quickly.
If anyone is really interested in the technical and mathematical side of this stuff, I definitely recommend Probabilistic Robotics by (among others) Sebastian Thrun, director of the Stanford Artificial Intelligence Lab and leader of the winning team in this race.
The big breakthrough was Stanford's texture vision system. I was very impressed with that. Computer vision in unstructured environments has a terrible track record, yet they made it work. Everything else was basically integration of off the shelf gear.
One accomplishment not oftened mentioned is that, by year two, many of the components that weren't available in year one were available off the shelf. In year one, getting an integrated GPS/INS/compass/odometer system was very tough. Applanix had one that cost $70K, took up a 4U rack, and required air conditioning. (CMU used it.) By year two, you could get something comparable from any of three vendors for about $20-$30K, ruggedized and able to run on 12VDC. All the successful teams had one, usually from Trimble or Novatel. Once you have one of those, just staying on course is straightforward. Then it's all about obstacle avoidance.
At what point do the robots turn on each other and try to smash one another with saws, hammers and spikes? Wait, I think that is a different show...
Do a google search on Sabastian Thrun, he was the team lead for Stanford, and formally at CMU (what a non-coincidence). Most of the software they used on Stanly (Stanford's bot) was either written by Sebastian in his former research or taken from experience gained on CMU's team the previous year. The ladar mapping he used, I know I saw on some former page of his that had all the gory algorithm details. It might just take a little bit of searching. He also has a c library out there somewhere that does a lot of this stuff, but I can't seem to find it now.
u blic_html/papers/thrun.ces-tr.html (sorry, no linky, writing in a hurry)
/ bfl-trunk/
h tml
One paper that's of interest might be here: http://www.cs.cmu.edu/afs/cs.cmu.edu/user/thrun/p
And that paper is mentioned in the readme of the BFL (Bayesian Filtering Library) found here:
http://people.mech.kuleuven.be/~kgadeyne/software
Lastly, at one point all of us competitors were required to give our design documents to DARPA, and they put them up on their webpage here:
http://www.darpa.mil/grandchallenge05/techpapers.
BTW, I wasn't on Stanford's team, but I was on another finalist team.
"We need a fourth law of Robotics: Stop Fingering My Wife"
There were several points made in the program that I hadn't heard elsewhere (and I've been paying attention to the Grand Challenge since the initial press release).
-- The teams get the GPS waypoints a few hours before the race. The waypoints are purposefully vague, so the robots have the choice of driving off a cliff (or into one) while still being within GPS parameters. This is supposed to prevent the race from reducing to "Who can follow GPS the best?" The Red Team had a group of what looked like 20 or 30 people who immediately sat down with the waypoints mapped out on satellite imagery, going through and adding waypoints of their own and adding speed commands for their robots. This seems to me to be a big violation of the spirit of the competition.
-- The Red Team had two entries, which they programmed differently: one more aggressive, the other more conservative (on speed). The faster robot, Highlander, was pulling away from Stanley for the first part of the race, until some unknown issue starting causing problems. Nova didn't say what was wrong, but it looked literally like Highlander was slipping out of gear and rolling back down hills. It _might_ have been doing it on purpose, i.e. a software glitch, but it didn't look that way.
-- One of the Red Team's entries completed the last portion (the hardest portion) of the course with its main sensor non-functional -- it was stuck pointed 90 degrees to the side. This argues even more strongly that the Red Team's vehicles weren't doing much route-finding and were pretty much just following GPS waypoints.
The conclusion I draw from this is that we are still a long way from the DOD's goal of autonomous transport vehicles. In a combat situation, transports need to be able to avoid obstacles put in their way _by the enemy_. The only time during this challenge that the vehicles did anything like this was during the initial trials before the race, and that was very limited. The actual race course was hard -- off-road, dirt, narrow, slippery -- but it didn't have tank traps painted the same color as the dirt they rest on. It didn't have razor-wire barricades, forcing the cars to figure out a route through the bushes around them.
I'm confident that if I had been on the course fifteen minutes before the cars showed up, I could have stalled or disabled all of them. Pile a bunch of bushes across the road and all of them would have stopped. During the trials and race, none of them demonstrated the ability to work around such a very limited obstacle.
All of this is not to minimize what was accomplished. But we're a long way from sitting back sipping champagne while robots do the dirty work of war.
As a student at Carnegie Mellon who has discovered the extent of his school's ties to development (had I known prior... and no, CMU is not unique in this regard, the problem is everywhere) of military products and has since spoken out against them a few times, thank you for realizing that this DARPA stuff isn't all it's cracked up to be.
I'm perhaps one of four people (an exaggeration, I hope) on my campus that isn't gung-ho about helping the DOD build driverless vehicles, and it's lonely at times.
Whatever moderator marked this down as off-topic was clearly just trying to limit the scope of discussion in the same way that DARPA and military contractors are trying to limit the scope of their moral and ethical liability.
-bugg
Honestly, not really. It was so damn dry out there that they water would spray the dust off and dry off in no time. I'd say rarely though did we ever see the water system turn on. Really, only in our mud testing did we ever get major buildup. Those LADAR's were pretty resilient sensors. The sun shining in them was much worse than any dust buildup.
"We need a fourth law of Robotics: Stop Fingering My Wife"
> Now once at Stanford they changed how they did things entirely and wrote a ton of code to make everything play much nicer than CMU's platform.
This sounds a little bit more like that, what I have heard. I've read, that they throw away most of the code and rewrote a large deal. E.g the classification of driveable terrain by the laser scanner was rewritten and learned. AFAIK, most of what has been published (and to what you pointed) is fairly generic stuff.
To the best of my knowledge, it has not been published how they learned the far range vision based on the near range laser scanner, which, to my eyes, is the most interesting part of the project.
> Nice try, I wasn't on CMU either.
Well, the comment on Sebastian Thruns previous affiliation and the code development sounds like something Mr. Whittaker could have said. But from what I've heard, he followed a fully stochastically approach and less reliance on the physical stability of the sensors and GPS, which AFAIK was quite different to the Red Teams approach and resulted in a much smaller code base.
"Between strong and weak, between rich and poor [...], it is freedom which oppresses and the law which sets free"