Slashdot Mirror
AI Monkey Robot
Sircus writes "The BBC is carrying this article regarding a robot 'monkey' that has been taught to swing from rung to rung of a ladder, using only the distance between the rungs and some basic equations about swinging. There's not much technical detail, but this is an interesting approach to the problem of dumb robots, and one which certainly gives the suits something pretty to look at."
IMO, this is not that great. Heres why:
The article states:
"The human instructors have told it some equations for swinging and the distance between the rungs - it must do the rest." - The monkey knows how far it has to go to get to the next rung. What would be amazing is if it could sense where the rung is and calculate that distance itself. It does not solve the problem (to quote the article again) of "bumping into the object" because the monkey knows the object is there.
While it may be great for the mechanics of a robot, (able to swing) it is not that new of a technological developement. I have toured several car making plants. At one plant they had a robot that cleaned the floors and all it had was a map of the building... it simply sensed when objects (people and trash cans) where in the way. The monkey has everything told to it and it also uses trial and error.
What may be new is the fact that they managed to fit all that info and technology into something fairly small (judging from the pictures). Robots that are small and can do what the monkey does are rare (I do not know of any others that exist). I just don't think the swinging (kick legs/hands out until object is caught - then repeat) is a breakthrough.
Remember - this is just my opinion (not flamebait or troll) and I may be wrong... but I have seen far better robots (judging by usefulness, size, and technology) around.
Is it progress if a cannibal uses a fork?
From Reseach Activities 1998, Robotics and Mechatronics Laboratory of Micro System Control, Nagoya University.
Brachiation Robot That Learns By Doing
Study on the Control of a Two-link Brachiating Robbot via "Target Dynamics"
...this approach, with the one taken by Mark Tilden:
The latest Smithosonian had an especially good article about him. Apparently his original demo 'bot was a robust walker that could clear most obstacles, run by all of 12 transistors.
I think somebody needs to look at combining the approaches of AI based robots that make decisions based on modeling their world, and robots such as Tilden's that seem to simply have "body feel" for their environment. Perhaps a Tilden type robot body could carry a reasoning "head" that sets goals for the body to carry out. I think this would be much like ourselves. Imagine if you had to walk or jump with only your eyes to guide you. No sense of balance, or kinesthetic sense. I doubt you could do it. We shouldn't expect robots to do the same.
While I feel that AI is a noble cause, I question the reasons behind an artifical monkey. What's next? Robot monkeys in congress? Robot monkeys for president? Is no one considering the effect this might have on our children? I'm willing to support the advance of technology, but I draw the line at robotic monkeys. If you recall the US attempted to put a monkey into space and frankly, we all know how that ended ... in utter chaos.
Perhaps Jon Katz could write an article on how this will ruin life as we know it unless stopped. Or maybe how I'm oppressing robotic monkeys and forcing them to commit violence in labratories across the nation. I can hardly wait.
Bad Mojo
Bad Mojo
"If you can't win by reason, go for volume." -- Calvin
The same type of calculation of forces is necessary for the function of a walking robot. Now, with walking, you're dealing with a push instead of a pull, and feet that have to balance instead of arms that have to grip, but the calculations are similar.
But most signifigant is that the robot has to try to grab the next rung, and if it doesn't, it has to figure out what adjustments to make on the fly. Simlarily, if a robotic step went wrong, the robot would have to recognise it, and correct itself before it fell
Judge Pag, the Learned, Impartial, and Very Relaxed
But is it truely learning?
Or is it just a simple function? Perhaps all it is, besides a heck of alot of expensive equipment, is the following:
int try_swing(int force, int lastForce, int secondLastForce) // out of reach // We're just swinging between too high and too low and can't ever make the branch; // Try to grab the branch with this amount of force in his "kick" backwards // caught branch // Try a less force // Try a more force // Something happened, we can't do it
{
if (abs(force) > FORCEMAX) return FALSE;
if (force == secondLastForce) return FALSE;
switch (swing_out(force))
{
case GOT_IT : return TRUE;
case TOO_HIGH : return try_swing(--force, force, lastForce);
case TOO_LOW : return try_swing(++force,force, lastForce);
default: return FALSE;
}
}
My point is not to insult the researchers who created the monkey, but instead to point out that just because it can "adapt" enough to swing from branch to branch, that does not mean it thinks! If that is the case, then the thermostat in your house is a thinking robot, because it turns on the heat when it's cold in your house, and turns the heat off when the temperature is too warm.
For this monkey to truely "think", he would need to remember everything he has "learned" in the past and apply it to each new branch, or even to other aspects of its life. For instance, if it ran into a tree, the result should hurt it so much that it would resolve to shy away from trees in the future, rather than run into them again. He should recognize trees that he has already seen before and run them flawlessly, and use skills he learned from one tree on the next.
Notice, I said 'skills'. Anything can be programmed to do one skill. But for it to take that skill and turn it into a new skill, such as taking an unintentional fall and catching the next rung below him, and thus learning how to go from level to level... or to "learn", without ever being programmed, how to catch a branch behind it with its feet...that is learning.
Tepp
This robot is emulating an ape (such as a gibbon) and not a monkey.
Monkeys generally walk *on* the branches and leap from branch to branch. (Sometimes hanging from the branches to grab things, etc.)
Apes are the critters that brachiate. Brachiation allows the animal moving around in the trees to be larger than if it leapt from branch to branch. (Of course, some of the "greater apes" (orangutan, gorilla, chimpanzee and us...) got big enough that they rarely go up in trees anymore -- a full grown silverback male gorilla would break most trees if it tried to climb them)
I guess "robo-ape" doesn't sound as good as "robo-monkey" -- especially since most people think of gorillas and not gibbons when they hear "ape".
Smithsonian magazine (not sure of any link to this) has an article about a guy at Los Alamos named Mark Tilden. He uses advanced circuit design to build a machine that in some ways is much more capable to walk in a random environment than most computer controlled machines. His first 'bot had 4 legs and a total of 12 transistors that allowed it to figure out some complex behavior. Really cool stuff.
Do really dense people warp space more than others?
I will be studying under Prof Fukuda starting next fall. During the winter break, I went to Japan and was invited to his lab to see the various projects there. Besides the monkey robot, the projects that were most interesting to me were a cooperative micro robot project, a set of legs that learn to walk, and a feedback control system for an invasive surgery tool/camera.
The monkey robot is currently most limited by the power/data cables that give it life and can only travel one or two bars before needing to be placed back where it started. It's motions are incredibly lifelike.
For more information about the people and projects at Prof Fukuda's lab, check out this link to the English version of the web page.