Real-Time LSTM-RNN Classification of Floors With Different Friction Coefficients for a Walking Humanoid Robot Wearing a 3D Force System

Abstract

In the study of biped humanoid robots it is crucial to achieve high precision and robustness in locomotion. Humanoid robots that operate in real world environments need to be able to physically recognize different grounds to best adapt their gait without losing their dynamic stability. This work proposes a technique to classify in real time the type of floor from a set of possibilities learnt off-line. Hence, the paper describes the collection and preparation of a dataset of contact forces, obtained with a wearable instrumented system, mixed with the information of the robot internal inertial sensor to classify the type of underlying surface of a walking humanoid robot. For this classification, the data are acquired for four different slippery floors at a rate of 100 Hz and it is used as input for a long short-term memory (LSTM) recurrent neural network (RNN). After testing different learning models architectures and tuning the models parameters, a good mapping between inputs and targets is achieved with a test classification accuracy greater than 92%. A real time experiment is presented to demonstrate the suitability of the proposed approach for the multi-classification problem addressed.