Self-Modeling Robots

Click To Play

Josh Bongard a roboticist from the University of Vermont has just netted a prestigious 2007 TR35 Young Innovator award for this work on self modeling robots. The work on self-modeling robots is truly brilliant! Consider a humanoid robot like Honda's Asimo. Asimo walks only because it has been programmed by engineers with a bunch of instructions (a model) to walk. Self-modeling robots however, create their own models for performing tasks like locomotion. Professor Bongard and his colleagues Victor Zykov and Hod Lipson at Cornell University developed a robot that is able to create its own set of instructions for moving around. The robot starts off only with knowledge of its motors and parts. It then performs an arbitrary movement (actuation) and senses the result. It creates a number of possible models of its structure to explain what it sensed. These models tell you what is the predicted result if a certain movement is performed. The robot then uses these models to decide what next movement to perform. It does so by comparing the predicted results of all the generated models and finding the movement that results in the most disagreement among these results. This process is repeated 16 times before the final generated model is used to perform the locomotion. At this point, I found it helpful taking a break and looking at some entertaining tongue twisters. Well, this is what is demonstrated in the video. This self-modeling is valuable if, for example, the robot happens to get damaged and lose a limb. It can regenerate a new model to enable it to walk to its destination. Self-modeling seems like a perfect match for USC's SuperBot. See the video page for another video of the Cornell University robot.