Data-driven extraction of drive functions for legged locomotion: A study on Cheetah-cub robot

Abstract

The process of finding working gaits for legged robots always, to different extents, includes manual tuning, systematic search, or optimization of control parameters. This process populates a dataset of control parameter vectors and respective robot behavior factors, like forward speed. The dataset obtained from a tuning process can include many gaits which share a similar performance in one behavior factor, e.g. speed, but differ in the control parameter vectors used. Our question here is, using the tuning dataset, how a continuous drive function can be calculated which takes the desired behavior, e.g. speed, and maps that to a control parameter vector. If this question is answered properly, then the robot operator (or a higher level controller) will have a single control knob to continuously change the desired behavior factor.