Learning to open new doors

Finding and opening an unknown door autonomously is an unsolved challenge that the most advanced robots in the world have not solved yet. Door handles have diferent locations, shapes, operating mechanisms and are made of different materials. Most approaches to door opening require precise information such as its exact location and a 3D model of the door handle that must be opened. This is one of the barriers that prevents robots from being used outside of controlled environments. In this thesis, we describe an approach to solve the problem of localizing and classifying a door handle with the REEM robot with no human intervention in the process. To do so we use the data obtained from a RGB-Depth sensor to detect the position of the door handle and compute an image of it that is processed by a supervised classi er system to determine the type of door handle. The type of the handle will determine which approach and opening motion is required to open the door. In this thesis we chose to perform stacked generalization with a feed-forward neural network on the prediction of several binary classi ers. The selection of the neural network model and binary classi ers is based on the experimental results of training and evaluating several combinations of supervised classi ers such as K-NN, SVM, Adaboost and Random Tree Forests with the image feature extraction algorithms Histogram of Oriented Gradients, Covariance Matrices and edge detection. In the end we obtain a model able to classify handle images with a performance higher than any of the individual binary classifiers trained