Memory-based Motion Optimization for Unbounded Resolution

Alan L. Jennings and Raúl Ordóñez

Keywords

Optimization, learning algorithms, robotics, motor skills development

Abstract

An algorithm is presented for autonomous motion development with unbounded waveform resolution. Rather than a single optimization in a very large space, a memory is built to support incremental improvements. As an understanding of the motions develops, the freedom to shape the motion is increased. Motions are created as functions of cubic spline interpolation. Nodes can be added to cubic splines so that all previous memory samples can be transferred to the higher dimension space. The transfer is exact, so there is no corruption of the data. The memory-based model, locally weighted regression, predicts the expected outcome for a motion and provides gradient information for optimizing the motion. The motion resolution is scales up with the system's experience so that the search space is never too large nor is final accuracy limited by the initial resolution. Results are compared against bootstrapping a direct optimization. This method shows practical accuracy and scalability.

Important Links:

Go Back