The work, to be described in the Feb. 18 issue of the journal Science, could change the way humanoid robots are designed and controlled and has potential applications for robotic prostheses. It could also aid scientists' understanding of the human motor system.
Developed at MIT, Cornell, and Holland's Delft University of Technology, the three robots are all based on the same principle: they are an extension of several years of research into "passive-dynamic walkers" that walk down a shallow slope without any motors. Passive-dynamic walkers were inspired by walking toys that have been around since the 1800s.
Control programs in the Cornell and Delft robots are extremely simple, because a large portion of the control problem is solved in the mechanical design. The MIT robot uses a learning program that exploits this design, allowing the robot to teach itself to walk in less than 20 minutes, or about 600 steps.
Dubbed "Toddler" because it learns to walk and because it toddles when it does so, the robot "is one of the first walking robots to use a learning program, and it is the first to learn to walk without any prior information built into the controller," said Russ Tedrake, a postdoctoral associate in the Department of Brain and Cognitive Sciences.
Among other things, the learning program allows the robot to navigate efficiently over a variety of walking surfaces, and may eventually allow robots to navigate very rough terrain. That's because the program works so quickly that Toddler is able to continuously adapt to the terrain as it walks.