Segmentação automática de demonstrações através da modelagem de séries temporais por processos beta

Abstract

This work addresses the issue of automatic segmentation of manipulation task demonstrations in the context of Learning from Demonstration. Learning from Demonstration is a set of techniques for robot programming and policy learning, based on observation of task demonstrations provided by a human or robot teacher. For a robot programmer to apply Learning from Demonstration, he must divide the demonstrations into activities or actions before using them for actual teaching of the robot. After this segmentation, the organized data is fed into learning algorithms, which are the main point of the LfD. In this work, we focus on the segmentation part of the problem and propose a video-based manipulation task segmentation tool. The proposed data-driven approach does not require any prior knowledge of robot programming or manual segmentation on the part of the human teacher. It uses out-of the-box machine learning tools for automatic image annotation and beta-process autoregressive hidden Markov models that leverage an infinite feature-based representation to create a seamless, easy-to-use tool that achieves relevant semantic segmentation of a completely unsupervised manner. As a result, this method reaches a maximum temporal alignment accuracy of 96% and maximum classification accuracy of 86.8%, without prior knowledge of the actions being segmented, making it clear that this technique can be used to segment manipulation tasks without the need for any manual work beyond of the demonstration itself.