Online Learning for Foot Contact Detection of Legged Robot Based on Data Stream Clustering

Abstract

Foot contact detection is critical for legged robot running control using state machine, in which the controller uses different control modules in the leg flight phase and landing phase. This paper presents an online learning framework to improve the rapidity of foot contact detection in legged robot running. In this framework, the Gaussian mixture model with three sub-components is adopted to learn the contact data vectors corresponding to running on flat ground, running upstairs, and running downstairs. An online data stream learning algorithm is used to update the model. To deal with the difficulty in obtaining contact data at landing moment online, a “trace back” module is designed to trace back the contact data in the memory stack until the data meet with the probability contact criterion. To test if the foot is in contact with the ground, a projection method is proposed. The acquiring data vector during the leg flight phase is projected onto an independent random vector space, and the contact event is triggered if all projected random variables fall within 1.5σ of the corresponding Gaussian distribution. Experiments on a legged robot show that the presented algorithm can predict the foot contact 16 ms in advance compared with the prediction using only leg force, which will ease the controller design and enhance the stability of legged robot control.