FALL DETECTION OF TWO-LEGGED WALKING ROBOTS USING MULTI-WAY PRINCIPAL COMPONENTS ANALYSIS

Abstract

A bipedal robot will fall if it encounters large disturbances. If this fall is detected in advance, an emergency action can be taken to minimize damage or even to prevent the fall. Recently we introduced a new method for the detection of upcoming falls, called the Multi-way Principal Component Analysis (MPCA) detection method, and we showed that it is successful in detecting the fall of the simplest walking model. The aim of this paper is to study the performance of the MPCA detection method on bipedal prototypes. We implemented the detection method on one of our prototypes called 'Meta' and conducted a parameter study for two parameters: sampling time and number of planes used for modeling the nominal behavior. The performance of the detection method was quantified with the average detection time and the amount of false positives. The results of this study show that the MPCA detection method is only able to detect the fall of the prototype after the last stance leg transition before the fall. The detection time is improved for an increase in number of planes or a decrease of the sampling time, but it will never detect a fall before the last heel strike. Thus, the method is useful for triggering bracing actions, but not for preventive foot placement.