Slashdot Mirror
A Mobile Robot For Modeling The World In 3D
Roland Piquepaille writes "A German team from Fraunhofer AIS has coupled a fast autonomous robot with a 3D laser scanner to digitize the environment. The team reports about their work in this article, one of fifteen on the subject of machine perception published by ERCIM News. "Kurt3D is an autonomous mobile robot equipped with a reliable and precise 3D laser scanner that digitalizes environments. High quality geometric 3D maps with semantic information are automatically generated after the exploration by the robot." This overview tells you more about the four-step method used to generate 3D models with this robot and contains several pictures of Kurt3D and its 3D laser."
How does this thing figure out distances? Does it time the return of the laser reflections?
I also can't help wondering how it models the tops of things - it looks like it's fairly squat.
What's the advantage of a robot like this versus describing every object by hand, as 3d animators do (typically in some kind of interpreted language).
It seems like writing "there's a sphere of radius 3 centered here" would take less time than waiting for the robot to scan it.
Great article (hope it doesn't get /.'d). While they seem to be working on large-scale room features (wall, door, floor, ceiling), I can see the next step being an autonomous robot that can find and identify such basics as a light switch and a power (mains) outlet.
I remember years and years ago, a robot had been developed that could optically recognize a power outlet and plug itself in... but I don't think it did much else. This would have been early 80s, probably, so we're talking Z-80 vs. Pentium.
Future recognition goals:
* Refrigerator door (fetch beer, please)
* Small child (danger! sticky fingers! run away!)
* Other robots for romantic interludes:
(IF Query(Other_Bot, EXCHANGE_CODE) == TRUE Extend_Programming_Probe(Other_Bot))
Stressed? Me? Of course not. Stress is what a rubber band feels before it breaks, silly.
I would think extracting 3D from video footage would be better. This thing can only map places where it can ride. Digital video cameras are pretty decent nowadays. I have seen university projects that say they have gotten pretty decent detection rates from video, but never seen any code nor binary :-/
"I'm waiting for a robot I can fight martial arts with. Any chance of us getting one of those?"
The pre-Qing Dynasty Shaolin Temple (ie prior to destruction) is rumoured to have had a hall of wooden men, basically articulated attack puppets actuated by a mechanism triggered by pressure plates on the floor. Monks had to go through this hallway to "graduate". IIRC there were 18 such dummies, each which had a specific method of attack.
If this legend is true, the engineering boggles the mind.
In the mean time, you can always buy yourself a non-robotic Mook Jong at your nearest kungfu equipment depot.
Seems pretty cool, I think what really sets this apart is the possibility of accuracy that couldn't be easily derived from an on site visit or from video. You can get some very fine grained messurements with this osrt of idea. It seems to me that once they get this thing refined down to a small enough size is method of ploting its currently location is rather condusive the being able to fly around. That would solve some people's concerns about mapping the tops of surfaces.
Just power this thing up, let it cruise around for a bit and you've got an high detail map of an environment. I like it.
Err.
Well, the stated purpose of this thing says nothing about it being used outdoors or to model large-scale terrain features. I mean, that's implicit in its design. This thing is designed to reproduce controlled environments.
And I don't know why you would think that is limiting! Maybe if you're thinking from the standpoint of a modeller/animator. Or maybe you just read the headline, and said 'omg it si small it cannot model WORLD omgomgomg'.
I see a couple of truly kickass uses for this thing. The first is adding texturing ability (you'd probably have to get dozens and dozens of scans, and have some good algorithms, to come up with good and relatively complete texturing, but I gotta' think that would be trivial compared to the sorts of problems they've already solved in making this thing -- and you wouldn't have to recreate the mesh each time, just sync up the coordinates with the one already created.
Ok, the use I see:
Crime scenes.
Bring in, hell, let's say 20 of these. Maybe some of them would be able to raise themselves up (heh, little accordioning platform for the recording mechanism, right out of the cartoons). They would roll around, sense out the room, figure out optimal placements, and then they would all scan the room, creating a near-perfect model of the room, perhaps mere hours of minutes after a crime has taken place. The cops would seal off the room, and the recorders would laboriusly record and texture everything about the room, down to the finest details.
Sure, it wouldn't catch a fingerprint or a peice of hair, and the plane/shape detection that is done actually removes some of the captured information (also removes some 'noise', but the forensic work they'd probably prefer a little noise to averaging out potentially important information) -- but the bottom line is, there wouldn't be a need for crime scene 'reconstruction', from photographs and little sketches and things that come after the fact. This would be absolutely accurate, more accurate than subjective information relayed secondhand from paid expert testimony. "How close would you say they were probably standing, from this photograph of bloodstains?"
So just in forensics alone, I see massive potential.
Funny enough, I came up with this exact idea a few years back, before Columbine and everything made it politically incorrect.
In addition to automatically building Quake maps for the building of your choice, it would help make up for my terrible sense of direction indoors. I get turned around in houses the first time I visit. Larger structures like hospitals and schools are downright labyrynthine. Having a map that builds itself during my travels (a la the self-revealing map in an RPG) would be a boon.
My version would've been a head-mounted stereoscopic camera unit, with software that recognized edges and angles. By watching my motion through the space and computing perspective changes, it could calculate distances and dimensions. Recording samples of textures would be fairly simple.
The system would quickly learn which objects are part of the scenery and which are mobile, by noticing changes in the environment when you visit the same area multiple times.
If miniaturized to the glasses-frame level, it could become always-wearable and answer questions like "where did I put my mug?" and "is the upstairs window still open?" by simply knowing the names of objects and locations.
Apparently the use of laser scanners in the current version indicates that plain-sight image recognition still isn't up to where it should be. Hmmph.