Inside the MIT Leg Lab, M2 is learning to walk. On June 16 the humanoid robot - one of the world's most advanced - successfully put its right foot forward, a single stride that marks a milestone in robotic evolution. M2 is funded by a Darpa program called Tactical Mobile Robotics, which hopes to create reconnaissance bots to replace human soldiers and rescue workers in dangerous situations. The Leg Lab's aim: to mimic the "hardware," elastic construction, and simple reflexes of humans and animals. Indeed, M2 will be the first bot to walk like a person rather than a pricey windup toy.
M2 - so named for no particular reason other than its two legs - builds on 20 years of R&D that began at Carnegie Mellon and continues at MIT. Steadied by safety straps, the nearly 60-pound robot spends much of its time balanced quietly on its feet. M2's anatomy is fully exposed: It has a carbon-fiber and aluminum skeleton, sleek steel and aluminum leg muscles, a nervous system of wires and circuit boards, and a chip of a brain embedded in its armless, truncated torso.
M2's self-steadying motions make it seem eerily alive. Push it lightly, and it shifts back to center. Leave it alone, and it sways in place, balancing against forces in the ground beneath its feet. Humans are constantly engaged in this subtle, unconscious dance, but the behavior is startling in a machine.
This machine hasn't completely mastered walking yet; it's still working on the first step - the one that makes the transition from standing to walking. With a fall demo for Darpa looming, Daniel Paluska, the mechanical engineering grad student leading M2's hardware design and construction, and software developer Allen Parseghian have been madly testing and tweaking and testing again.
It's become an almost daily ritual: Paluska stands behind the bot like a coach spotting a gymnast. He keeps a ready thumb on a remote kill switch, while Parseghian, at a nearby workstation, hits Enter, transmitting the walk command over an Ethernet cable to M2. The robot's 12 series-elastic actuators come silently to life. M2 leans away from Paluska until its center of gravity shifts to its left leg. Then, carefully balancing there, it swings its right thigh up, propelling the right shin forward under its own momentum. As its shin reaches an outstretched position - its body already moving forward - M2 stiffens the leg and brings its weight down onto its right foot.
Except that, until June 16, each graceful attempt ended prematurely with a clumsy crash. The robot would lean too far to the side or misplace its leg and collapse noisily into its safety harness.
"We're confident we'll be able to meet our deadline," Paluska says, "but it was actually kind of horrible that M2 took its first really good step at the end of a Friday. We went home knowing we might have to spend all of next week trying to repeat it."
The long-awaited first step marks a significant leap: While it doesn't mean that the bot knows how to walk yet, it's hard evidence that M2 is getting steadier on its feet. The endless trials help the Leg Labbers understand where they need to refine the bot's controlling software. The software depends on simulations - a cartoonlike M2 walks through 3-D virtual landscapes on a workstation screen - but the transition from the virtual to the real world introduces variables not yet accounted for.
Building machines inspired by nature is a guiding principle of the Leg Lab (www.ai.mit.edu/projects/leglab). Marc Raibert, who founded the lab at Carnegie Mellon in 1980 (and moved it to MIT in 1987), pioneered a method of robotics engineering based on a few physics-inspired control laws - one for height, one for pitch, and one for speed. Raibert also favored simple hardware based on animal limb design. This method has been so successful that the lab is typically filled with odd-looking contraptions that hop and run with an animal sentience, appearing possessed rather than programmed.
Although Raibert left academia in 1995 to focus on a company he'd founded, the M2 project, which launched in 1997, follows his model. Paluska designed the robot's limbs to the specifications of a six-foot man, with joints that mimic human hips, knees, and ankles. M2's hard steel muscle rods attach to its limbs with elastic tendons. Its $90,000 worth of parts are, by robotic standards, surprisingly low tech and relatively low cost. The most expensive component - a $6,900 3-D motion sensor perched in M2's torso - duplicates the balance-keeping function of the human inner ear. Elsewhere, 12 sensors track the angle of each joint, 12 more measure the force applied to each pseudo-muscular actuator, and 8 load detectors measure the ground pressure exerted on M2's feet. The robot's brain is a modest digital signal processing chip, running software developed using the virtual M2.
"Our simulations of the robot are physically realistic," says Leg Lab leader Gill Pratt. "The hardest part is getting the robot to take that first step - to make a change from zero speed and static balance to a dynamic state of movement." Once M2 is moving, it is relatively easy for it to keep moving.
The robot's design hinges on two new ideas, which the Leg Labbers refer to as low impedance control and series-elastic actuators. In plain English, low impedance control means designing limbs so their movements naturally follow the laws of physics. Human knees and ankles, for example, are built to bend and pivot in ways necessary for walking and jumping, rather than every which way.
Series-elastic actuators are robo-muscles that mimic human muscles and tendons, controlling limbs with a flexibility far more forgiving than the rigid actuators of industrial machinery. M2's dozen actuators each consists of a hard, motor-driven shaft attached to its joints by steel cables or carbon-fiber rods. When it comes to walking over uneven surfaces, the elasticity of these pseudo-joints is almost as important as the legs themselves. "In the real world, you're never going to have perfect information," Paluska says. "Your foot is always going to hit the ground a little sooner than you expected, or a little later. M2's springs are like our tendons - a soft interface to the hard world."
By contrast, early humanoid robots mimic walking by moving their limbs through carefully controlled patterns that retrace human movements, but somehow lose their humanity. "Ask a kid to imitate a robot," suggests Pratt. "They'll move their arms and legs in an incredibly stiff fashion that we've even come to describe as 'robotic.' When we walk, we move loosely. Our limbs swing freely, with our muscles adding just a little control at the start and the end of each swing." The swing effect is what separates M2 from the reigning robot-walking champion, Honda's P3. A five-foot-two, 287-pound human-shaped bot in a spacesuit, P3 can be seen in promotional videos climbing stairs, kicking a soccer ball, and balancing on an incline. But compared to M2, P3 is stiff, heavy, and operates in playback mode. "We're behind in terms of flashy demos," Pratt sighs. "But in the end, our stuff is going to work better because it's more flexible. Our approach is based on control laws and reflexes. We don't dictate what the robot does."
It's this approach that attracted Darpa. M2's funding - about a half-million dollars to date - comes from Darpa's multi-year project to create an unarmed reconnaissance bot. Instead of sending a soldier into a building to look for terrorists or snipers, commanders would send in a remote-controlled robot rigged with cameras and other sensors. To be effective in that scenario, M2 needs vision, which will be incorporated sometime next year.
But a few obstacles must be overcome before the researchers tackle sight. "What happens when it falls down? How does it protect itself from being injured during the fall?" Pratt asks. Teaching the robot to recover from a stumble - avoiding a face-plant - is next on the agenda. Even learning to fall properly will take months.
Pratt hopes that success with M2 will inspire more funding for walking research. "With, say, $50 million per year going to a dozen or so universities and companies, a robot capable of running the Boston Marathon could be built in five years. In ten years, its successor could hike the Appalachian Trail, including a climb up Mount Washington." Pratt hopes the walkbot research will also help bring smart prosthetic devices closer to reality. To this end, he's seeking both government and private support for a national center for legged locomotion.
By mid-July, just one month after its first-ever step, M2 was stepping more often than not. But even when the bot does miscalculate, Paluska is reluctant to blame his creation for its shortcomings. "When I say we're teaching M2 to walk," he says, "most of the learning is happening inside of us, not the robot."
