After graduation, I became Assistant Professor of Practice at University of Texas at Austin, TX, USA. Currently, I am a Tenure-Track Professor at the Karlsruhe Institute of Technology and an Emmy Noether Group Leader. You can still reach me via email@example.com or my new homepage .
Imitation Learning, Skill Learning, Movement Primitive Representation, Human-Robot Interaction, Movement Segmentation, Grammar Induction Reinforcement Learning, Robot Learning
The Probabilisitic Segmentation (ProbS) Δ method segments initially unlabeled demonstrations are segmented into sequences of movement primitives, while simultaneously learning a library of primitives. ProbS is an Expectation Maximization method, where the E-step computes a new segmentation of the demonstrations given the current model and the M-step learns a new model given the current segmentation. The model is defined as a mixture of primitives. The segmented demonstrations and the learned primitive library are used to induce a Probabilistic Context-Free Grammar for primitive sequencing. The grammar structure is learned using an Markov chain Monte Carlo optimization, where a novel prior based on three Poisson distributions was introduced. The induced grammar serves as an easily comprehensible representation of the capability of a robot with respect to the primitives learned from the demonstrations. The grammar can now be used to generate new sequences of primitives from the library. Finally, the induced grammar is improved through reinforcement learning approaches in order to improve the grammar such that previously unseen tasks can be solved by the grammar and the primitive library.
Before his Ph.D., Rudolf completed his Master Degree in Computer Science at the Technische Universitaet Darmstadt focusing on robotics and machine learning with a minor in bionics. In his masters thesis Rudolf investigated robust and safe policy search methods under the supervision of Gerhard Neumann and Jan Peters. He developed a new policy update approach entitled Information Theoretic Stochastic Optimal Control.
Rudolf is also looks into approaches for anticipatory robot behavior, active learning for human-robot interactions, infusing semantic meaning into learned movement grammars, task and role identification in collaborative human-robot scenarios and robot-assisted learning of human motor skills. Another big interest of Rudolf is the application of machine and robot-learning methods in prosthetic devices. Such intelligent prostheses could improve the adaptation process, the capabilities and the quality of life of the respective person.
Teaching tasks between humans often do not only include the demonstration of a task, but also the correcting and guiding of the executed task. These behaviours can be observed between a parent and a child, a trainer and an athlete and an Instructor and a novice worker. Introducing such additional feedback to a robot can improve it's learning speed and the quality of the resulting policies significantly. This form of interactive learning becomes even more important once the desired task includes a human. Forcing the human to act exactly in the same way and solve a task always in precisely the same way and order is not desirable or even possible. Therefore the robot needs to be able to adapt and interact with the human, and be able to incorporate changes into its behaviour which may vary between each interacting human. However if the robot is able to adapt to the human he can increase the humans productivity, and help him to be more efficient in areas where the creativity and and intelligence of a human can not be replaced by a robot.
The research of semi-autonomous robotics offers many challenges in various areas, of which one is the learning and executing of motor skills. The representation of such skills is still a highly researched topic, which lead to various interesting approaches, such as Dynamic Movement Primitives, Interaction Primitives and Probabilistic Movement Primitives. The application and evaluation of such representations in complex human-robot interactions tasks is an important aspect. At the same time a complex task might require the segmentation of the task into several sub-tasks. These segments allow for more adaptability to changes in the scenario. Instead of manually defining and sequencing such segments it is highly desirable to automatically identify useful, reoccurring motions and store them in a skill library. This library could then be used to automatically find the optimal sequence in order to solve a specified task. The automated segmentation and subsequent compilation of complex tasks are important but unfortunately little researched topics. Further research, new methods and approaches and the evaluation on human-robot interaction tasks in this promising topics is therefore necessary.
- Machine-Learning: Imitation Learning, Reinforcement Learning, Policy Search, Movement Primitive Representation, Skill Acquisition, Movement Segmentation, Structure Learning, Grammar Induction, Skill Composition and Sequencing, Life-Long Learning, Active Learning
- Robotics: Anthropomorphic Robots, Human-Robot Interaction, Semi-Autonomy, Motor Skills, Adaptive Control, Human-In-The-Loop, Robot Learning
- , International Journal of Robotics Research (IJRR), 36, 8, pp.879-894.
- , Proceedings of the International Conference on Robotics and Automation (ICRA).
- , Journal of Aerospace Information Systems, 11, 9, pp.579-595.