Automatic task-specific model reduction for humanoid robots

Abstract

The disclosure provides an approach for automatically determining task-specific robot model reductions. In one embodiment, a simplification application determines a smallest order statespace model whose stabilizing controller also stabilizes a full-order robot model. The simplification application may determine such a model via an iterative procedure in which the reduced order is initialized to the number of unstable poles of the open-loop full-order system and, while the closed loop full-order system with the balanced reduced order system's stabilizing controller is unstable, fractional balanced reduction is applied to generate a balanced reduced system. If one or more unstable closed-loop poles exist in the full-order system with the stabilizing controller of the newly-generated balanced reduced system, the reduced order is incremented by one, and fractional balanced reduction repeated, until no unstable closed-loop poles remain. In another embodiment, the model reduction is made task-specific by formulating the full model with task-specific outputs.