Accuracy Analysis of a Ground Reaction Predictor-based Feedforward Control

Abstract

Robots with contact interactions can use machine learning-based reaction predictors to build local linear reaction models and use them in feedback and feedforward control. In this paper, a study of effects of noisy and biased reaction predictors on the feedforward control is considered. The paper presents a model of the predictor with exact accuracy, and a simplified model of a noisy and biased predictor. Distributions of the control errors and prediction errors based on the linearized contact model are demonstrated for two robots: a bipedal walking robot and an in-pipe four legged robot. The proposed procedure can be used to find stopping criteria for the training process of the predictors.