Additional Stability for Single-Unit Pattern Generators

Legged robots have the potential to travel where wheeled robots cannot. While legged robots have many advantages that improve their maneuverability, they are notoriously difficult to control. However, neuroevolution, which combines the nature-inspired fields of neural networks with evolutionary computation, has shown promise in this task. The aim of this paper is to extend prior work that introduced an approach called single-unit pattern generators (SUPGs), which generate oscillatory patterns of activation for controlling the many moving parts of a legged robot. The extended SUPG approach employs a novel adjustment mechanism uniquely suited for SUPGs that allows fine-grained modulation of the SUPG oscillation pattern to potentially react effectively to more challenging conditions such as noise or rough terrain.