Domo is a new upper-torso humanoid robot at the MIT
CSAIL Humanoid Robotics Lab.
Photo Donna Coveney/MIT
The biologist who discovered this last fact, Joseph Spagna, currently at the University of Illinois, teamed up with engineers at the University of California at Berkeley to build a robot inspired by nature. The result, named RHex (for its six legs), is a robot that can traverse varied terrain without any central processing at all. At first it had a lot of trouble moving across wire mesh with large, gaping holes. Spagna’s team made some simple, biologically inspired changes to the legs of the robot. Without altering the control algorithms, they simply added some spines and changed the orientation of the robot’s feet, both of which increased physical contact between the robot and the mesh. That was all it took to generate the intelligence required for the device to move ahead. In a related project, Iida and his MIT group are now building legs that operate with as few controlled joints and motors as possible, an engineering technique they call underactuation.
The theory that much of what we call intelligence is generated from the bottom up—that is, by the body—is now winning converts everywhere. (The unofficial motto of Iida’s group is “From Locomotion to Cognition.”) Some extreme adherents to this point of view, called embodiment theory, speculate that even the highest cognitive functions, including thought, do no more than regulate streams of intelligence rising from the body, much as the sound coming from a radio is modulated by turning the knobs. Embodiment theory suggests that much wisdom is indeed “wisdom of the body,” just as those irritating New Age gurus say.
STARFISH TURNS
Bongard and his babies are part of this revolution. Like many roboticists, he is interested in the art of navigating across different terrains, from sand to rock to grass to swamp. Almost by definition, a versatile machine will require this skill; a housecleaning robot might have to climb stairs or ladders, crawl up walls, and wax a floor, one after the other. The old approach would have been to design a central computer that recognized new terrains and geometries, with the characteristics in a database that informed the robot’s gait one hurdle at a time. But this would have failed, because there are too many variables to program in. “Programming a general-purpose machine to anticipate all eventualities is a classic problem in traditional artificial intelligence,” Bongard says. “It has never been solved. Embodiment suggests a different approach.”
Here’s how it works in one of Bongard’s devices, a black, four-legged starfish about the size of a dinner plate. Switched on, the starfish starts to flop about, all the while recording how its body behaves and generating ideas about how its parts might work in the world—how joint a might influence joint b given c units of force, and so on. It then compares the actual behavior of its body with the generated models to see which one made the best predictions, taking the winning model and seeding it into the action in each round. Using something like genetic recombination, the starfish parses and reparses the models, becoming steadily smarter over time. After some specified length of time, the smartest model is chosen as the operating officer—the piece that tells the robot what to do. This event is like the robot’s bar mitzvah, the point at which it is ready to go to work. Now when it receives a task, it hands it to the winning model. The model figures out which joints and motors should do what, distributes those instructions to the parts, and off the robot goes. As it works the robot keeps track of its own performance. Whenever performance falls off, programmers instruct the robot to evolve a new model that might work better.
“Embodiment allows us to avoid building in our human biases,” Bongard says. “It lets the robot figure out itself what is appropriate for its own body.”
MY ROBOT, MY SELF
Embodiment theory speaks to one of the open questions about robots: the nature of our relationships with them. Many versatile machines will work as members of collaborations, including mixed teams of humans and other machines. Perhaps the most obvious such application is soldier robots, but any useful domestic robot would obviously have to interact with its owner.
It is only common sense that the easier it is for these entities to understand each other, the more effective the team will be. If intelligence, including social intelligence, flows up from the body, it follows that making sure all members have approximately the same kind of body will enhance their ability to relate.
Machines smart enough to do anything for us will probably also be able to do anything with us: go to dinner, own property, compete for sexual partners. Will they enrich our lives or, like a new kind of TV, destroy our relationships with real humans
Suppose, for instance, that Robot Bob is struggling to walk on ice. If Robot Alice has learned how to walk on ice herself, her memory of how her body behaved in similar situations will help her recognize the nature of Bob’s problem. Once she understands the situation, she should be able to help Bob by transmitting the right algorithm to him on the spot, or even by offering him a hand. If robots are going to learn by imitation, one of the most powerful learning mechanisms we know, it will clearly be useful for our bodies to be alike.
Even verbal communication will be easier if our machines look like us. “Imagine you use the phrase ‘bend over backward’ with a robot,” Bongard says. “How is the robot supposed to know that this is a metaphor for difficulty? A programmer could write it in ahead of time, but he or she would be working forever. There are just too many phrases like that. But if the robot has a body like ours, if it has a back, then it might be able to understand on its own, automatically.”
Far more is involved in communication than just speech. Candace Sidner, an artificial intelligence expert working for military contractor BAE Systems, is one of many researchers interested in machines capable of understanding and participating in gestural communication. These nonverbal behaviors include eye contact, body movement, shifts in posture, and hand gestures. Sidner’s special interest is engagement, the component of gesturing used to structure the back-and-forth aspect of communication. Her current goal: building a machine that can participate in conversations with two humans at once, figuring out when the humans want to speak to each other so it knows when to fall silent and when to speak up.
