Prof. Jan Peters’ Keynote Talk: Machine Learning of Robot Skills

Machine Learning of Robot Skills   Jan Peters   Computer Science Department – Technische Universitaet Darmstadt, Germany Robot Learning Group – Max-Planck Institute for Intelligent Systems, Germany     Abstract: Autonomous robots that can assist humans in situations of daily life have been a long standing vision of robotics, artificial intelligence, and cognitive sciences. A first step towards this goal is to create robots that can learn tasks triggered by environmental context or higher level instruction. However, learning techniques have yet to live up to this promise as only few methods manage to scale to high-dimensional manipulator or humanoid robots. In this talk, we investigate a general framework suitable for learning motor skills in robotics which is based on the principles behind many analytical robotics approaches. It involves generating a representation of motor skills by parameterized motor primitive policies acting as building blocks of movement generation, and a learned task execution module that transforms these movements into motor commands. We discuss learning on three different levels of abstraction, i.e., learning for accurate control is needed to execute, learning of motor primitives is needed to acquire simple movements, and learning of the task-dependent „hyperparameters“ of these motor primitives allows learning complex tasks. We discuss task-appropriate learning approaches for imitation learning, model learning and reinforcement learning for robots with many degrees of freedom. Empirical evaluations on a several robot systems illustrate the effectiveness and applicability to learning control on an anthropomorphic robot arm. These robot motor skills range from toy examples (e.g., paddling a ball, ball-in-a-cup) to playing robot table tennis against a human being and manipulation of various objects. Bio: Jan Peters is a full professor (W3) for Intelligent Autonomous Systems at the Computer Science Department of the Technische Universitaet Darmstadt and at the same time a senior research scientist and group leader at the Max-Planck Institute for Intelligent Systems, where he heads the interdepartmental Robot Learning Group. Jan Peters has received the Dick Volz Best 2007 US PhD Thesis Runner-Up Award, the Robotics: Science & Systems – Early Career Spotlight, the INNS Young Investigator Award, and the IEEE Robotics & Automation Society’s Early Career Award as well as numerous best paper awards. In 2015, he received an ERC Starting Grant and in 2019, he was appointed as an IEEE Fellow.

Continue reading


Announcement: We expect that the situation will be improved sufficiently in July to continue with the LOD conference as planned

We expect that the situation will be improved sufficiently in July to continue with the LOD conference as planned. We are monitoring the developments and will announce any changes on this website if necessary. The proceedings of LOD 2020 will appear in Springer’s LNCS series as planned, independently of the format of the conference. Ticket […]

Continue reading


The scientific activity of the LOD 2020 Program Committee will be accomplished as usual.

Obviously, the scientific activity of the LOD 2020 Program Committee  has not been suspended or stopped: peer review, accepted papers, camera ready and proceedings will be accomplished as usual. As usual, a Springer LNCS proceedings will be prepared and all accepted papers have to be presented (physically or virtually) at the conference. If the LOD 2020 […]

Continue reading


Call for Virtual Rooms and Socials

Call for Virtual Rooms and Socials LOD 2020 will continue to support the strong community-building role that is so central to the conference. We hope to create opportunities for all participants to meet new people and to share knowledge, best-practices, opportunities, and interests. To enable this, we will support several virtual rooms and socials. A virtual […]

Continue reading


COVID-19 Update

In light of the current pandemic, the organizers are currently considering how best to run LOD2020. Even if it doesn’t go ahead physically, we expect to run a virtual version, so please continue to prepare papers for submission. More details will be posted on the website when they have been decided.

Continue reading