DARPA Grand Challenge

So just what were some of the approaches, and what accounts for the big improvement ? Is there real ooh-la-la AI involved here ? Or just sophisticated ad hoc adaptations ?

Congratulations to you and everyone involved, BTW. I can hardly imagine how one would go about doing this, which is why I'm so curious.

I read an account of one team which qualified ( Syracuse ? ) and they just talked about computers catching fire and vehicle problems.

Lew Mammel, Jr.

"Brian Dean" wrote

Cool, so you guys were the Desert Rats! I was cheering for you guys and for Monstermoto. I love to see people with very small budgets beating over engineered limitless budgets. Congrats to your team.

I wonder if magnetic compasses are really worthy for this type of application. If you had to redesign your vehicle, would you still include it in it?

Indeed, one of the guys from the Terramax team was wearing a hat. We were at that back fence waiting for Terramax to be unpaused and the wind took his hat away... I was wondering how funny (or not) and unprobable it would be to terramax to think the hat was an obstacle, take a detour and crash onto the barriers just like alice. 8^)

Kudos!

Exactly my thoughts... One easy thing they could've done is to position a camera on top of that map and broadcast the image of that map, that would be a thousand times better than that one they were using.

Final question to you guys: was it worthy? Would you enter a competition like that again?

It would be very nice if you guys could write a short post-mortem of your project. I know I'd be very interested.

Cheers

Padu

Well, technically it's not my vehicle - I'm just 1 of a 9 person team. Regarding a follow-up race, I doubt it. First off, I'm tired. It was a lot of effort to do the DGC, and I could use a break. Second, a lot of the appeal to me was the challenge of doing something that no one had ever done before. But now that it has been done, a lot of the appeal for a follow-up is now gone as well. But if a follow-up was significantly different and presented new challenges, then yes, maybe.

-Brian

Hey Brian. Thanks for all the info regarding the obstacles. Looks like your vehicles needed to be working well just to get through that. Good job, for going 26 miles too. Oops, my little sumo just ran into the wall - blind as a bat :O).

BTW, what do you think might work for a moose detector? Especially, for a moving moose.

- dan =============

Thanks. We think we won if you measure by miles per dollar :-)

We're not yet sure that was the problem, it's just a theory. We need to analyze the log data to be sure. We just got that back a day or so ago, so we can dig into it and see what was up. But the unit we used is usually quite good, although can be a little finicky at times - it is very sensitive. It's an AHRS with builtin magnetometers so it actually gives us quite a bit of information besides just heading. It gives us pitch, yaw, roll which we can use to determine whether we are on an incline, detect vibration (in order to slow down on rough sections), etc. We are quite happy with the device.

Our Novatel GPS also gives excellent heading when the truck is in motion, but poor heading when sitting still. The AHRS is reliable regardless of motion. Also, the GPS heading flips when we reverse, unlike the AHRS unit which always gives the heading the truck is pointed in. We should probably fuse the two sensors so that the best of both complement each other.

Irony for sure. That guy would be very unpopular on the team after that, I'm sure!

That's a great idea - I wish DARPA would have thought of it.

While I'm tired from the effort, my participation on the team is something I will remember for the rest of my life. I made some good friends and learned a lot of things I thought I already knew. It was a great experience. Those things are priceless, no matter how you slice it.

Would I do it again? Not a competition exactly like that, but maybe something similar in that it is in a field that is of interest to me and the nature of the event is interesting itself. The GC was interesting in part because it was so crazy to think anyone could create a fully autonomous ground vehicle that could successfully navigate 140 miles through an unknown rugged off-road region at useful speeds. That had never been done before, which was part of its appeal to me.

But now that it has been done, doing it again would be sort've anticlimactic, IMHO.

But if another competition came along that had similar characteristics as the DGC, then yes, most likely I would be interested in that as well, just like the original DGC was so interesting to me.

We most certainly will do this.

Likewise.

-Brian

Thanks!

That's a tough one. I thought you were joking when you first posted that, but the more I think about it, I think it would be a great idea. I was driving home at that twilight time in the evening the other day when deer are rustling about and was intensely alert, not that my reaction time still would have been fast enough to avoid one since they come out of nowhere like a bullet. What a great idea to have a sensor to give you some forewarning.

What about an infrared camera - look for its body heat which should be substantial. CMUCam/AVRcam for IR. One should be able to differentiate a moose or other animal (deer, cow, person) from other heat sources like oncoming cars due to temp difference between animals and a car engine. Also, size of the animal could be estimated as well

- do you care about squirrels? If the IR cameras see warm bodies nearby, sound the alert.

-Brian

Moose and dear cause tons of trouble.

I had one miss my truck by inches, almost twilight, into the setting sun, Probably would have totaled the van, as I was doing 70 or so. I'd pay serious money and I think others would as well.

Don't care at all about anything house cat sized on down.

How do you tell a car far away from a deer/cow up close? Do you have range as well? or triangulate from two or more sensors?

Maybe combine IR with motion and/or visual?

I was going on the idea that the IR sensor can actually sense temperature. A car engine should be hotter than an animal's body temp and it can thus be filtered out. Only look for things in the temp range of a large animal.

I'm envisioning one of those infrared cameras that detect body heat. Used by law enforcement to find people at night. Or do those just exist in Hollywood?

A signal image wouldn't provide depth, I think you'd need stereo for that. Given the infrared image(s), standard video processing techniques should work to do the rest - find the moose, sound the alarm.

-Brian

I think this characterization is a little misleading, considering the precisely specified GPS course laid out. GPS was really the sine qua non of the whole venture.

This is not to belittle the achievement, but I think some of the exultations of success ( which attracted my attention ) were overblown. This was being touted as an AI breakthrough, but consider this from the "The Tartan" on the CMU "Red Team(s)" :

While there is an extensive amount of attention to hardware, ?[the] most powerful technology is preplanning the curves in the hours or so before the race when the route is revealed,? Whittaker said. ?We compute the terrain and the details of where we?re going to drive and how fast we?re going to drive, and that is definitely the thinking part of the race.

"definitely the thinking part of the race ..." Compare and contrast the humble bee, plunging into the UNKNOWN, sans GPS, sans LIDAR, with only its vision and its ever so tiny brain to rely upon.

Some say that unto bees a share is given Of the Divine Intelligence ...

- Virgil Lew Mammel, Jr.

They exist, more or less; firefighters use similar equipment to evaluate buildings for residual heat sources. Unfortunately, they tend to be expensive. The low energies involved in thermal radiation require high sensitivity for measurement, and the receiving element must be shielded from background radiation (often by actively cooling the camera)

- Daniel

Actually, I was being semi-serious. As noted in the original post, there are 100s of 1000s of deer collisions every year, so just the cost of the auto repairs, and possible personal injuries, can be high. And I read that in Maine, approx 1 moose is hit every day. How many people even live in Maine? 6, maybe ;-). A 1500# moose will put a dent in the hood.

Down southwest of Denver, they opened up a new ring corridor 10 years or so ago, and there were deer collisions every single day, until they put up about 15 miles of tall fences the deer couldn't jump over on both sides of the road.

The moose/deer sensor would have to:

(a) detect the animal, and maybe note its size (b) track if it's moving (c) compute its speed (d) predict whether it will intercept the path of vehicle

A good engineering problem.

- dan michaels

Hmmm, the obvious place to put sensors is on the front of the car. Right near the screaming hot iron engine, radiators, A/C condensers, etc.'

Much has been said here lately on this subject. I agree that GPS was one of the most utilized sensors in the competition, but so it was last year and results were very different (although I also agree that this year the route was a little bit simpler than last year's). For the sake of robotics field, it would be nice to see vehicles with more intelligent vision systems, and I deeply believe that the real intelligence is in vision.

But look at the two other points of view:

2) Although human vision system still cannot be compared with machine vision, for a course like the GC, we would also need a GPS or at least a map. Although most of navigation was done using GPS, there was collision avoidance present in all vehicles, and we were able to see a demonstration of it on the NQE.

Writing this second view, one question raised on my mind. I believe that according to the course DARPA chose, it was possible to two vehicles to be on the same road on opposite directions. How road sharing was managed when vehicles were about to cross each other? If such thing happened and both vehicles were allowed to be in movement, I take my hat out for them. Such capability is fundamental for real world autonomous navigation.

Padu

Nothing to do with robotics, but two days ago at WalMart I saw this "Deer avoidance" device.

It is a set of two little devices you install on your front bumper and they are "wind activated". From what I saw it is nothing more than a whistle set to a certain frequency that must bother that kind of animal. How effective it is I really don't know.

Padu

"Brian Dean" wrote

Listening to an interview of Montmerlo (spelling?), the team leader of Stanford team, he said it was quite amusing to see Stanley trying to swerve when a bird crossed flew a couple meters in front of Stanley bumper.

Padu

I wonder about the utility of autonomous truck convoys to the military though. For the foreseeable future they're still going to need human guards of some kind. I think the utility for an autonomous device would be more in the nature of a recon or weapons delivery platform--something that goes in harm's way in lieu of a human.

"Padu" wrote in news:6 snipped-for-privacy@iswest.net:

Your forgetting GPS data was a requirement last year and this year. A CD was given out with GPS data and you had to pass over the GPS points, at the required speed for that GPS data point.

You can see in the dark, unless you use ultrasonic, FLIR? SONAR? SAR? Light seeable edge detection is not that easy and how are you going to see a water puddle or pond? that water puddle cause a car to wreck itself last year. What about driving in a sand storm, or fog when you can't see 1 ft. in front of you, heavy rain?

I'm not clear about your definition of intelligent vision.

thanks

Starlight scope.

Which is one of the reasons that vision is the real challenge.

Another reason it's the real challenge.

Then you use alternative sensors. Just remember, though, on the modern battlefield he who radiates is lost.

Not sure I understood what is your point here, but anyway, my point is that I don't see the use of GPS as "cheating" or a bad thing. Although, if you only rely on GPS for navigation, sooner or later you will be doomed.

You are misundertanding vision (as in vision algorithms) with human vision. We can only see in a limited spectrum of light. Cameras can use a much broader spectrum. My point is: it doesn't matter what sensor you are using, the real challenge is making good sense of the data you are receiving. We humans can see a water puddle with our vision, so the computer should be able to do it, it all depends on the intelligence is givem to it. As mentioned above, there is an additional advantage to robots, they are not limited to one type of sensor as we are. How do you drive through a sandstorm or through heavy fog?

That's because I did not define it. What I asserted is that the real intelligence [yet to be found] is in vision. You see, identify that there is a blob in a picture frame is already not a simple thing. Ok, in ideal conditions it is, but the real world is far from mathematical ideal. Light conditions change, lenses get dirt, etc. But assume you got that part nailed and you have the image nicely structured into objects. Now the real problem is to recognize what are those objects and what do they mean. For example, it would be nice for an autonomous tank vision system to identify bushes or soft objects from untraversable ones. I think all falls into the discussion if one day we will be able to have a strong AI, able to implement our horizontal knowledge (or learn its own) and common sense.

Cheers

Padu