Scalable Locomotion for Large Self-Reconfiguring Robots

For large self-reconfiguring robots, any algorithm that requires linear amounts of memory per module (with respect to the number of modules) or linear time for computation or communication per actuation is undesirable. While shape-forming may require linear amounts of memory, locomotion can be performed with simpler shape specifications, and therefore sublinear algorithms are possible. In this paper, we present a locomotion technique that performs both planning and actuation control in sublinear time and memory. The algorithm is inspired by reinforcement learning and uses dynamic programming to plan module paths in parallel. To ensure the physical integrity of the overall robot during motion, we have developed a novel localized cooperation scheme which may also be used with other self-reconfiguration algorithms. Our overall algorithm is able to direct locomotion over arbitrary obstacles, and the formulation of the goal used in the planning encourages dynamic stability