Chapter 3: Figure and Ground
53 - Run, Don’t Walk
At Waseda University in Japan, scientists have for some time now been working on humanoids—which is to say, robots that walk like people. Shigeo Hirose at the Tokyo Institute of Technology—where there’s a major lab focusing on this kind of work—has built a number of legged robots, including quadrupeds. He also has a line, called Titan, which they sell to other research labs. The automaker Honda has spent $30 million developing Asimo, a beautifully engineered biped that even looks like a human, albeit a moonwalker encased in white armor, as it trots around booting a soccer ball. Their design strategy involved filming an actual person, then engineering his joint angles and movement characteristics into the robot. Korea’s full-bodied walker, HUBO, takes things a step further, with individually moving fingers. Scientists at Carnegie Mellon University are developing an Anatomically Correct Testbed hand. It has bones and joints that mimic ours, and is controlled by emulations of the neural signals coming from our brains. So far the group has completed one finger. Engineering departments at schools across the U.S. now routinely feature robotics programs, though perhaps none focuses more intensely on this subject than MIT.
There, one finds a science fiction junkyard bounded by vague walls, overworked snack machines, and desks submerged under slurries of paperwork. At first glance it appears to be inhabited by machines that almost resemble living things: tall ostrichlike birds, oversize bugs, humans with their top halves missing. This is the leg lab. Closer inspection reveals a sparse population of real humans bent over obscure tasks. One of these, graduate student Jerry Pratt, sits poking what could be a screwdriver into something that might be a femur but looks more like a spare part for the Terminator.
“One of the nice things about robots now,” Pratt says, “is that they’ve gotten to the point where it’s not painful to watch them move. You know, like fifteen years ago, there were robots but when you watched them you got bored. Now they go on the order of human walking. So they’re a lot more exciting and a lot more practical.” A main goal for designers of legged robots is what Pratt refers to as “biological looks.” By that he means they want to make it look the way a natural organism looks when it’s moving.
As it turns out, walking is a “hard problem”—a term of art that mathematicians use meaning, essentially, that they can’t figure out how it works. But important progress is being made. That progress comes in no small part due to the work of the lab’s founder, Marc Raibert, one of the world’s most prolific roboticists. Raibert has since left MIT to form Boston Dynamics, a leader in the human simulation software used to create lifelike movement in humanoids. But during his tenure here he was interested in balance. An early project was a single-legged hopping machine that behaved like a ten-year-old on a pogo stick. It worked partly by using a gyroscope similar to the sensors in our inner ears.
“From that single-legged machine,” says Hugh Herr, now director of the lab, “he learned how to stabilize the bouncing height and tension of the machine and how to control the speed. Then he applied those principles to machines with two legs, and then with four legs. He eventually developed bipeds that could run and even do flips, really amazing feats.” According to Herr, a Raibert robot holds the land speed record for an artificial biped, thirteen miles per hour—which approaches a four-minute mile.
That’s a fast pace for any biped, though in the end it is the act of walking that presents the bigger challenge.
“One thing we find,” says Pratt, “is that walking is actually harder than running. The difficulty is that you always have to be balanced.” One way of looking at how running is easier than walking, he says, “is if you are walking and you’re about to trip and want to recover, the way you recover is by running a few steps.”
With this principle in mind Herr worked out a system in which a robot’s legs moved through their arcs as if they were spokes on a wheel. “Even though it was a funny wheel,” he says, “with the spokes being elastic and changing length, I still applied that principle. It turns out that when you do that, it just balances and doesn’t require any information on its orientation in space. And it balances a lot better than a lot of the more sophisticated machines.” What that implies, he says, is that “when a horse is trotting or galloping along steadily, it’s not actively using its inner ear to balance. There is no active balance. It is stable by its very nature.”
Roboticists, says Pratt, “are starting to find that when you talk about artificial intelligence or creature intelligence—the intelligence we’re interested in, the intelligence of a cat running around, for instance—you need to think more in terms of dynamical systems than of classical AI stuff like logic. The classical AI approach…and the figuring out and solving of logical equations that way…nature doesn’t operate like that. A living creature is a dynamical system with a bunch of state variables, which are constantly changing.
“In a brain, for instance,” he says, “each neuron might have ten state variables, where the state of the neuron changes with the input from other neurons. So what you’re getting is a network with billions of these, thousands of millions of interconnections between each one…and based on the structure of the system and the physics behind it, they’re all changing over time.”
Email this to a friend | 
Add to del.icio.us | 
Add to digg |
Linkology by Kate (all links are unofficial) | Rights Info
