From: 2024-02-02 10:00 to 13:00
Place: E:1406, E-building, John Ericssons väg 2 / Ole Römers väg 3, Lund, Sweden
Contact: matthias [dot] mayr [at] cs [dot] lth [dot] se

Thesis title: Learning with Skill-based Robot Systems: Combining Planning & Knowledge Representation with Reinforcement Learning

Author: Matthias Mayr, Department of Computer Science, Lund University

Faculty opponent: Professor Michael Beetz, Universität Bremen, Germany

Examination Committee:

• Professor Petter Ögren, KTH Royal Institute of Technology
• Associate Professor Todor Stoyanov, Örebro University
• Associate Professor Karinne Ramirez-Amaro, Chalmers University of Technology
• Deputy: Associate Professor Yiannis Karayiannidis, Department of Automatic Control, Lund University

Session chair: Associate Professor Michael Doggett, Lund University

Supervisors:

• Professor Volker Krueger, Lund University
• Professor Jacek Malec, Lund University

Location: E:1406, E-building, John Ericssons väg 2 / Ole Römers väg 3, Lund, Sweden

For download: Follow this link to download the pdf: https://portal.research.lu.se/en/publications/learning-with-skill-based-robot-systems-combining-planning-amp-kn

Abstract

The usage of robots in industry is transforming. Traditionally, robots have been deployed to automate monotonous tasks through manual programming, excelling in speed and precision yet lacking flexibility. Now, as part of Industry 4.0, the paradigm is shifting towards collaborative robotics, where robots are expected to interact dynamically with their environment and handle non-repetitive tasks. This evolution demands a leap towards flexibility and adaptability at both control and task levels. To address these challenges, the concept of “robot skills” — reusable, parameterizable procedures — emerges as a potentially pivotal building block. The skill-based robot control system SkiROS2 is designed to be robot-agnostic and to represent such skills and the necessary knowledge. This knowledge in the world model describes the robot and the environment, facilitating sophisticated reasoning and task planning capabilities.

Despite these advancements, contact-rich tasks remain a complex endeavor, often challenging to fully encapsulate in predefined models. To overcome this, it is possible to allow robot to learn from experience and improve. This thesis presents an approach for robot control and learning based on behavior trees and reinforcement learning (RL). Our integration of robot skills, knowledge and planning with RL does not only enable robots to proficiently learn and execute contact-rich tasks but also allows for the seamless transfer of learned policies to real-world applications. In a comparison with state-of-the-art RL algorithms we show that this combination of planning and learning demonstrates markedly accelerated learning curves. Furthermore, we can demonstrate that the operators can formulate priors for the optimum to guide and speed up the learning process. An extension of this framework further enables robots to adapt to task variations without the need for relearning from scratch, showcasing the system’s robust adaptability and potential for diverse industrial applications.