April 20, 2012 12:00 PM
| Matt Bunting|
Years on Job: 3
When Bunting built this bot in 2009, it had no programmed knowledge of its own geometry. It used a camera to teach itself how to walk over four days. Now Bunting uses the device for research on machine learning. "I'm exploring behaviors using only the camera?no fancy sensors," he says. In one experiment, Bunting gave the robot a genetic algorithm that mimicked evolution, allowing it to learn how to walk with "this very eerie, natural motion," he says. "And that only took 10 seconds." For another, the hexapod built a 3D map of the surrounding terrain and stepped over obstacles.
To construct a robotic cheetah for a DARPA project, Bunting and his colleagues built a pair of pneumatically powered legs with simulated feet and toes. Most robots that run have springs in their feet. But the goal here isn't just speed. "If we wanted to build the fastest land robot, we'd make a wheeled machine," Bunting says. "Our constraint is to learn how biology solved a problem. It's hard to make a fully articulated limb that hits the ground all the time. It wears parts out." In the future, this work will help others make more durable robots and prosthetics.