Tutorial on Dynamical System-based Learning from Demonstration


Organizers: Nadia Figueroa, Seyed Sina Mirrazavi Salehian and Aude Billard

Website: https://sinamr66.github.io/TutorialRSS2018.io/

The use of Dynamical Systems (DS) for motion planning problems in robotics has become popular thanks to their ability to generate on-line motion plans inherently robust to changes in dynamic environments. In recent years we have been focusing on formulating DS to model robotic tasks that can be learned from demonstrations (LfD). We have used our DS-based learning techniques in a plethora of robotic applications, from executing simple point-to-point motions, such as pick-and-place and imitating motion patterns to more dynamic scenarios, such as generating golf swings, obstacle avoidance and even catching objects in flight. These techniques have been further extended to learn more complex tasks of repetitive nature, from sequential point-to-point motions to peeling vegetables or rolling pizza dough. In this tutorial, we will introduce various techniques used to learn Dynamical Systems from human demonstrations. We will then provide an overview of different interfaces (haptic device, motion sensors, vision, speech, etc.) available to teach robots and discuss how they serve different purpose. Further, we will introduce a set of techniques to modify a learned behavior locally. We will particularly focus on (i) obstacles avoidance, (ii) non-contact/contact transitions and (iii) locally active dynamics and via-point insertion. Attendees will be able to practice and test the presented techniques through computer-based simulations in MATLAB.