Slashdot Mirror
Robot With Broken Leg Learns To Walk Again In Under 2 Minutes
KentuckyFC (1144503) writes When animals lose a limb, they learn to hobble remarkably quickly. And yet when robots damage a leg, they become completely incapacitated. That now looks set to the change thanks to a group of robotics engineers who have worked out how to dramatically accelerate the process of learning to walk again when a limb has become damaged. They've tested it on a hexapod robot which finds an efficient new gait in under two minutes (with video), and often faster, when a leg becomes damaged. The problem for robots is that the parameter space of potential gaits is vast. For a robot with six legs and 18 motors, the task of finding an efficient new gait boils down to a search through 36-dimensional space. That's why it usually takes so long. The new approach gets around this by doing much of this calculation in advance, before the robot gets injured. The solutions are then ordered according to the amount of time each leg remains in contact with the ground. That reduces the dimension of the problem from 36 to 6 and so makes it much easier for the robot to search. When a leg becomes damaged, the robot selects new gaits from those that minimize contact with the ground for the damaged limb. It compares several and then chooses the fastest. Voila! The resulting gaits are often innovative, for example, with the robot moving by springing forward. The new approach even found a solution should all the legs become damaged. In that case, the robot flips onto its back and inches forward on its "shoulders."
Crowbars won't save you now.
No kidding!!! What do you say at this point?
They break the leg's of robots for "educational" purposes? The future has arrived.
I can't wait until Skynet becomes self-aware.
The "new" walking patterns are all pre-programmed. It's not learning, it's just running a few presets and seeing which results in the greatest forward speed. This has use, but I wouldn't throw around "learning" for this experiment. If a novel break comes along that the programmers have not planned for, the machine won't have a working behavior is it's data banks.
Having taken an AI class in College, I'm wondering if this might be a real-world scenario where a self-learning neural net would come in handy....not only for the broken leg scenario, but because you can tell it to learn, the "Gait" can be optimized at run-time under any scenario.
Were it Japanese? Irene
Seems like a missing cost function for picking the new walking mode is to choose something that does not over stress the remaining parts.
So the forward speed with all legs functioning is 0.25m/s, and with one leg broken it is 0.27m/s.
Therefore, if a robot chases you, do NOT break its leg, because that only makes it chase you even faster!
Now strap a 3D printer on each robot and it will 3D print fully functional limbs in 2 minutes!
finally, a roomba with the possibility to give a cat PTSD.
Good people go to bed earlier.
There was a show in the 1980's from the CBC in Canada called "Fast Forward". Every week it focused on different tech innovations and where they'd go. One week they had learning robots. For the most part these were all simplistic things. But they had an "ant" that was about 2 feet long that was autonomous. The MIT crew that had created it realized that a centralized brain was just too big and power draining to build into the robot. So they had a system of sensors with rudimentary data and needs (leg=up, down, forward, backward, touching, not, moving or not, etc). If they shut it down it lost the memory of how to do anything it learned that day. They turned it on for the camera and it was a flailing ball of legs. Within 5 minutes it not only learned how to walk but circumvent objects, falls, danger. It still sticks out as amazing. Watching this video, I wonder what ever happened to that bot from nearly 30 yrs ago and wonder why does this spider seem to have actually gone back in time?
Hey, I can make all kinds of tasks faster by precomputing much of the work and then looking it up in a table. Congratulations, you've (re)discovered another instance of a Space/Time tradeoff.
Now, in particular what they've done is still wicked cool -- it's a great idea to perform may millions of simulations ahead of time so that at runtime (heh) you can quickly draw on that data to adapt. It would be perfectly good research even without the over-the-top claim that they've somehow made the work faster as opposed to cleverly pre-computing much of it.
But that's research -- you do something neat and then you make a ridiculous overstatement to generate buzz ...
Have gnu, will travel.
I recall about a decade ago the Department of Defense or NASA working on this sort of adaptability for controlling damaged aircraft.
An aircraft that suddenly loses use of a part of a wing or a flight control surface may still have sufficient flight control capability to fly home. The problem generally is that the pilot's control inputs won't produce the same motion responses, and the pilot generally has only seconds to map the inputs to control outputs.
The idea was for the computer to do this mapping for the pilot, so the pilot would continue to apply the appropriate inputs (to roll the plane for example). The computer would determine which of the remaining flight surfaces to employ in order to best achieve the desired motion.
One example that I recall was when the aircraft rudder was lost, yaw motion was compensated by dropping the landing gear and speed brakes on only one side of the aircraft to cause more drag on only one side, yawing the aircraft.
Against the Slashdot rules, I read TFA and watched the entire video.
Unless I'm mistaken, all they did was create a giant array of possible motor combinations for movement, and then the robot just randomly tries them until it finds one which lets it more-or-less go in the same direction. It may not be the best one, but one that mostly works (it just stops at the first one that mostly works).
Is that really a super big breakthrough? If the robot dynamically adapted to the broken leg, and figured out how to move using some semi-intelligent algorithm, I would say that is really awesome. But this is literally just trial and error through pre-created movement specs, randomly, then just selecting one that is mostly okay.
Not trying to downplay other's achievements or research or anything, but it just doesn't seem like a big break through, unless "brute force" is something novel.
As far as I understood the article, everything is based on a behavioral repertoire... The only advancement of the study would be the confidence mapping of said repertoire? Wouldn't it be better to work toward the automatic creation of this repertoire by the robot itself?
When animals lose a limb, they learn to hobble remarkably quickly.
Right, I'll do you for that!
It's just a flesh wound.
I have discovered a truly marvelous proof of killer sig, which this margin is too narrow to contain.
I remember when "behaviors" the word was used to describe animal "tricks". As the work tricks was considered not politically correct anymore. When can we call these roboot skills "tricks". Because in no way do I think this is a behavior in anyway.
How could there be a novel break absent the robot growing new limbs? It's a limited search space by the physical nature of reality.
Very Slashdot of you to burn that strawman down so thoroughly. Not that you wanted to shit on the accomplishments of others, but it just feels so good...
How long will it take that robot to figure out how to stop the researchers from breaking its legs in the first place?
Why does it have such difficulty walking on five legs out of six? Four would probably be plenty. It could even lift the one opposite of the broken one just to make things simpler and symmetric.
Stop all A.I. development ! All we need is to simulate the entire universe posibilities in a database. Then focus on designing a database bigger than the universe itself tu put it on every computer.
-no sig today-
Well, I guess screaming does count as a "complaint".
Allow me to make a symantec correction to the story. There are only 3 spacial dimensions that any robot has to contend with. 4 If you include time, but the robot may have to deal with 6 axis on each motor. X, Y, Z, Yaw, Pitch, Roll and those six axis would be multiplied times each limb (6 x 6 = 36).
Axis, not dimensions.
Sort of scary. Maybe the next trick robots will learn is to keep themselves going, if you try to stop them. You pull out its main battery, it (running on reserve battery) puts a new battery in.
Suppose you can't power off a computer. So you unplug it. Running on battery power, the computer reaches out a mechanical arm, and plugs itself back in.
Soon it will be a good idea to put Asimov's Three Laws of Robotics into operating systems.