Online Virtual Repellent Point Adaptation for Biped Walking using Iterative Learning Control

Abstract

We propose an online learning framework to reduce the effect of model inaccuracies and improve the robustness of the Divergent Component of Motion (DCM)-based walking algorithm. This framework uses the iterative learning control (ILC) theory for learning an adjusted Virtual Repellent Point (VRP) reference trajectory based on the current VRP error. The learned VRP reference waypoints are saved in a memory butter and used in the subsequent walking iteration. Based on the availability of force-torque (FT) sensors, we propose two different implementations using different VRP error signals for learning: measurement-error-based and commanded-error-based framework. Both implementations reduce the average VRP errors and demonstrate improved walking robustness. The measurement-error-based framework has better reference trajectory tracking performance for the measured VRP.