APR 1 (Thursday): Josh Bongard (Cornell University)
Abstract:
For a robot to function for long periods of time in a hostile, unknown or remote environment, it must be able to deal autonomously with uncertainty: specifically, unanticipated internal damage or external environmental change. The recent difficulties with JPL's Spirit and Opportunity Mars rovers provides a dramatic example. To this end we have recently developed a two-stage algorithm with three functions: it generates an initial gait for a legged robot with arbitrary morphology; it can diagnose unanticipated failure or damage that the robot encounters; and it can generate compensatory gaits for the damaged robot. The algorithm uses two separate evolutionary algorithms to evolve the robot controllers and damage diagnoses, respectively. I will show results and several videos documenting simulated robots that encounter a wide range of unanticipated failures and novel environments, and how they recover. The results will focus on a sample quadrupedal and hexapedal robot. I will also discuss future extensions of this algorithm, including the application to a real-world hexapedal robot, and using it to infer the functional anatomy of the human hand.