Will There Be A Winning Autonomous Robot in 2005?

An anonymous reader submits "This summer is heating up the DARPA Grand Challenge as multiple top notch schools begin to announce their entry into the competition. The newest organization to announce its entry was the Florida Institute of Technology. Their project is known as Oasis - Autonomous Racing, and they have a team of over 45 students, professors, and advisors that are currently hard at work designing their vehicle and raising funds to pay for it. The DARPA Grand Challenge is a race between vehicles that should be designed to travel up to 300 miles in less than 10 hours through the desert or other harsh medium without any human interaction. The 2005 competition has a $2 million grand prize as authorized by congress. With all of the new entrants does anyone think that the competition will be won the second time around?"

Comments

It might happen...

[sgtsanity]

BTW, for all those interested, Wired ran a list of what went wrong for each team. It reads very comically, but a lot of these things are very "DUH!" after you've gone through the first time. I forsee a lot better results, as teams will have that much more practice. Hopefully some will come up with some more general solutions, rather than brute-force processing the terrain around the known area of the route.

Re:It might happen...

[Anonymous Coward]

None of the teams managed to do much brute-force terrain processing, because in general, their sensors didn't work.

The one that got the farthest just ran off of pre-computed GPS waypoints, and as the GPS accumulated drift error, it started driving to one side of the road, then in the ditch, then off road, until it hit something and stopped.

Against the rules

[dexterpexter]

You can't use a hover-craft, though. I am sure that is the first thing on *everyone's* mind. (I know it was mine)... why not just build a helicopter and make it take the most direct route? There is a reason no one did that:

The rules limit entrants to mechanical-to-ground travel. No hover crafts allowed.

However, there are other non-DARPA competition where flying autonomous bots are preferable. DARPA's competition, however, is limited to road vehicles.

Correction - number of miles

[OblongPlatypus]

I mean 142 miles, not 100. 142 is the number Wired quotes.

The /. blurb mentions 300 miles, but the Q&A on the DARPA page says "will not exceed 300 miles". Apparently the course is randomly selected and only revealed on the race day, to make sure the vehicles aren't trained for the specific race course. I'm assuming the Wired quote means that the course that was picked for this 2004 challenge was 142 miles long.

Re:Doubtful

[renec]

Is that the new math?

You know how to convert to metric, but when you divide 300 by 10, they get 60. I'm assuming you asked google about the metric conversions.

The average speed is 30 mph, or a bit under 50 Km/h

Baja

[crisco]

Depending on the class, the Baja (and other desert races) contestants depend heavily on co-drivers, GPS and proper preparation. They run over the entire course before the race (hence the 'PreRunner' style of trucks) and rely on maps, GPS and the co-driver's experience.

Motorcyles and the trophy trucks averaged nearly 60 MPH on the last Baja 1000, other classes are slower.

I wish Rally driving were more popular over here in the US of A, so much more excitement than big ovals.

Most of them will never work

[Animats]

With the possible exception of CMU, nobody had a system that could avoid a ditch or a pothole. Stereo vision won't do a good enough job on dirt for long range ditch/pothole detection. All the laser rangefinders except CMU's were fixed line scanners, so they couldn't possibly profile the ground ahead reliably from a bouncing vehicle.

CMU's approach is a big hammer. They took a stock line-scanning laser rangefinder and put it in a huge 3-axis gimbal, which they then actively stabilize. That should be able to profile terrain, but it's a huge mechanical kludge. If you miss a spot because you hit a bump, you have a hole in your data. At that point you can either slow down and rescan, or plow ahead blindly. They may eventually complete the course with that rig, but no way is it a commercially viable technology.

The next generation of sensor technology may be ready in time. There are at least three groups with usable sensors in the prototype stage. We're talking to two of them. But that's all I'm going to say for now.

John Nagle / Team Overbot.

(We're recruiting. See our jobs page. No pay, some risk, a fraction of the prize, we cover all expenses. Silicon Valley only. We have our own shop in an industrial park in Redwood City. If you're local, come over and see the thing.)

Re:Most of them will never work

[Animats]

Why not a combination of stereovision, range finding, and a digital horizon to enable real time mapping based off a visual system?

Stereo vision has two fundamental limitations. First, it doesn't work very well unless the scene has clean, sharp edges to match up. Second, the accuracy decreases rapidly with range, beause you're measuring a narrow triangle from angles at the base.

The algorithms for stereo vision aren't all that forgiving. There are basically two flavors. One finds and matches "features", usually corners. This works nicely for indoor scenes and badly on dirt roads. The other does a straightforward correlation between matching scan lines from two cameras, sliding them back and forth looking for the best match. This has a high false alarm rate on surfaces with high-frequency detail, like gravel roads.

Practical problems include the fact that correlation algorithms are sensitive to high-frequency noise, so any thermal noise from the camera is a major problem. Also, keeping two cameras aligned to within a pixel while jouncing along on an off-road vehicle requires a very rigid mounting with the cameras near the center of gravity along the inter-camera axis. (For an example of a good one, see the Bumblebee from Point Grey. They have the most successful stereo vision products.)

To date, the most successful outdoor stereo vision system used on a mobile robot was on the NASA Hyperion robot. They were able to achieve a range of about 7 meters on rocky terrain with hard edges. This is about a third the range that theory predicts. A DARPA Grand Challenge vehicle needs at least 20 meters of range, and if you want to go fast, 50 meters. You need 1.5 to 2x your stopping distance.

We have a stereo camera setup working on my desk here, and we've had it for over a year. We've tried that.

Stereo from motion, where you work with successive frames from a single camera, has potential. The baseline is the distance you move between frames, which can be much bigger than the distance between two cameras. But people have been trying to make that work for years without much success. If you want to work on vision, that's a good problem. Especially since you can just take data from a camcorder and crunch on it - no special hardware required for development.

Re:Most of them will never work

[rebelcool]

They're getting better, but its still years away from usefulness. This past week I checked out some new work that used temporal data streaming to fill in whats between the discrete frames (much more like how human vision works). This allowed much more detail to be considered, with lower noise errors that plague stereo vision traditionally.

However it still required structured light to work well, meaning outside the lab it wont work well.

One problem that I see with stereo vision research at the moment is that its still focused on turning cameras into fancy rangefinders. Though its nice to be able to determine the entire range of objects in a camera's vision area (as opposed to a singular laserbeam endpoint), it still tells you very little about the nature of those objects which is a great deal more useful than their range.

There is a rule against damaging the terrain

[dexterpexter]

The only problem I could see with this is that driving through things was not seen as an acceptable solution by DARPA. It stipulated that the terrain and obstacles must be left unharmed. I think there are reasonable allowances made, such as running through "weeds" and leaving faint tire tracks.

Sending a bulldozer through something, however, would likely cause harm.

The motive behind this, if I get to guess, is that they are looking for a more covert vehicle. Something that has torn through the terrain and left chaos in its wake is more likely to be tracked/disabled than something that can quickly and nimbly navigate across the terrain.

I think that your idea is a fine idea, though. If they are looking at application for war situations and covert navigation is not an issue, I think that you are onto something.

When I first heard about the competition, that was my first reaction, too. Why not just create a tank and plow through the terrain along the most direct route? A review of the rules showed that they had already taken into consideration this solution and created a rule against it. I can see their reasoning, though.

I'm not so sure

[Einer2]

From what I know of the race course, these vehicles have to average 30 mph going cross-country through the desert. If it's anything like the terrain around the Tucson area, I'm not sure that I could average that without piling straight into a saguaro.

Re:Short answer ...

[pauldy]

Man, why is this modded as troll? Are there slashdoters out there that are so intolerant of the use of the word god.

Re:I think they will

[jbrocklin]

Actually, I've read most of the technical papers that the teams were required to submit, and many of them did use a "track and assign danger levels" as a way of finding a best path (most used this as a way of keeping the vehicle inside the boundaries of the course - assign the off-course sections with infinite danger and the vehicle will never go there).

Overall, the majority of the problems that people were with unplanned problems, such as going up a hill and not switching down gears, stopping to check terrain, and then not being able to start back up again. Or getting a wheel stuck in a small ditch and not being able to get out (no friends to jump out of the car and push ya know).

I think that someone will win next year - or at least make it far enough that the logic part of the system will be proven effective. Sure there will be some little things that will hit just about every team, but I hope someone does a good enough job preparing that it will take a lot of little things to bring them down.

Mod me down

[wonkavader]

But while I'm interested in the beginning of that paragraph, and the end, I'm not gonna read what's in the middle.

The return key is your friend.