Jonathan Österberg

Written Works/Publications

Bachelor Thesis

Abstract
With computers being used for more applications where commands can be spoken it is useful to find algorithms which can separate voices from each other so that software can turn spoken words into commands. In this paper our goal is to describe how Independent Component Analysis (ICA) can be used for separation of voices in cases where we have at least the same number of microphones, at different distances from the speakers, as speakers whose voices we wish to separate, the so called "cocktail party problem". This is done by implementing an ICA algorithm on voice recordings containing multiple persons and examining the results. The use of both ICA algorithms result in a clear separation of voices, the advantage of fastICA is that the computations take a fraction of the time needed for the ML-ICA. Both algorithms can also successfully separate voices when recordingsare made by more microphones than speakers. The algorithms were also able to separate some of the voices when there were fewer microphones than speakers which was surprising as the algorithms have no theoretical guarantee for this.

Master Thesis

Abstract
A control system used to control two Panda Franka Emika robots online and simultaneously with two HTC Vive controllers is presented, with the primary purpose of demonstrating tasks for robots. The system is validated by learning from demonstration/imitation learning task via Principle Component Analysis (PCA). The task consists of learning different bimanual movement patterns e.g. for drawing sketches, with latent variables that then can be manipulated by the user to generate new shapes of similar structure. Tasks of various correlations between the arms are tested and compared. The system uses components and adaptations e.g. preexisting modules for sensing, communication, motion planning, etc. to realize the goal of modularity and support for other robots than the one used in this thesis. The most prominent systems used are the Robot Operating System (ROS) for the base framework for handling packages and sending information between different parts of the system, and MoveIt’s planning library (running on ROS) for managing kinematics and collision.