Decentralized Algorithms for Cyclic Robotic Systems

Eric Klavins, Department of Electrical Engineering and Computer Science, University of Michigan

Control algorithms are difficult to scale up to large decentralized systems because of potentially complex couplings between components. As a result, many such systems function conservatively -- damping out most of their energy and moving slowly or in a start-stop fashion. Without synthesis tools for dynamical behaviors that admit a modular bottom-up approach, the full potential of modern actuators, sensors and computing power likely cannot be realized, dooming robots (and more generally, physically situated computing systems) to a clumsy and inefficient future.

In this talk I describe our efforts to provide a formal basis for designing and verifying decentralized control algorithms for robotic systems, including factories, dynamic manipulators, hoppers and walkers. To avoid the complexity of arbitrary couplings, we consider systems that may be decomposed in very regular ways. I focus on methods for coordinating cyclic behaviors, based on defining ideal, model dynamical systems called reference fields. I show how reference fields, which assume continuous actuation, can be modified for intermittent contact tasks such as "juggling" and hopping -- demonstrating that such behaviors may be "composed" to produce more complex behaviors. Finally, I describe ongoing experiments with a six-legged, scampering robot.