Walking Robot Control with a Machine Learning-based Ground Reaction Force Predictor and Generated Linear Contact Model*

Abstract

This paper aims to present a comprehensive view on a new control method for legged robot: data-driven ground reaction predictor-based control. The idea of the method is to use machine learning tools to build a reliable predictor for ground reaction forces and then exclude them from the control formulation by building local contact interaction models. The advantage of this approach is twofold: first, it allows to avoid making models for every contact scenario, which become less accurate as the robot changes during its life cycle; instead it relies on the data gathered by the robot’s sensors. Second, it allows the use of the wealth of the control methods designed for the serial robots, as is demonstrated in the paper. The paper shows three experiments: with a planar walking robot, with a simplified three dimensional model tailored for optimization-based control, and with AR-601 humanoid robot.