Imitation Learning from Teleoperation-Based Demonstrations using Gaussian Mixture Regression for a dual-arm Robot

In recent years, significant technological advancement has determined the rising of collaborative robotics. While a couple of decades ago robots were primarily used in industrial applications, today they have found their way into various everyday activities in close co-operation with humans. This era is also characterised by the remarkable progress and development of Artificial Intelligence, which enables robots to exhibit greater flexibility and adaptability in dynamic and highly complex environments. Consequently, modern applications allow collaborative robots to work with humans in a shared environment and interact one with the other. Data-based algorithms using a machine learning approach can be exploited to train robots to perform tasks autonomously. Specifically, robotic manipulators can acquire skills through human demonstration or imitation, using the so-called Programming by Demonstration (PbD) paradigm. This thesis focuses on employing real-time intuitive teleoperation, used as a tool to enforce PbD. Through this approach, a human operator can intuitively perform uni- or bi-manual telemanipulation. In this thesis work, teleoperation was indeed exploited to provide multiple demonstrations of a given task (or skill) to collect a dataset of trajectories from different initial poses. The task was demonstrated to the dual-arm Baxter robot with the specific goal of grasping a bottle on a table with one arm and pouring its content into a glass held by the other arm. Using Gaussian Mixture Regression (GMR) it was possible to embed the skill described by the different demonstration trajectories into a more general probabilistic description, which also exploits the concept of dynamical systems for the generation of online trajectories, called Dynamical Movement Primitive (DMP), which can autonomously adapt to variable conditions. Finally, the GMR-based PbD approach was implemented and tested by means of grasping and pouring experiments with the real Baxter robot.