Continuous Neural Association for Efficient Robot Learning
Event details
| Date | 09.01.2014 |
| Hour | 11:00 |
| Speaker |
Prof. Jochen Steil, Director COR-Lab, Bielefeld University Bio: In 1993, I received the diploma in mathematics from the Bielefeld University, and joined the Neuroinformatics Group headed by Prof. H. Ritter. In 1995/96 I stayed one year at the St. Petersburg Electrotechnical University, Russia, supported by a German Academic Exchange Foundation (DAAD) grant. In 1999, I received the PhD. Degree (Dr. rer. nat) with a Dissertation on "Input-Output Stability of Recurrent Neural Networks". In 2002, I was appointed tenured senior researcher (Akad. Rat) and in 2006 my habilitation was accepted by the Faculty of Technology and I was assigned the venia legendi in Neuroinformatics. From March-July 2006 I was on leave as principal scientist at the Honda Research Institute Europe, Offenbach. In 2008, I was appointed apl. Professor for Neuroinformatics at the Faculty of Technology. Since April 2007, I have been managing director of the Research Institute for Cognition and Robotics, CoR-Lab and since Nov. 2007 a member of the scientific board of the center of excellence in Cognitive Interaction Technology (CITEC). Also in 2007, I started the teutolab-robotics with workshops for school pupils to experience robotics for one afternoon. Since 2010, I am coordinator of the FP7-EU project AMARSi - Adaptive Modular Architectures for Rich Motor Skills. In the academic year 2012-13 I participate in the ZiF research group on "Competition and Priority Control in Mind and Brain". From Sept-Oct. 2013 I have been on leave as a temporary professor to Oxford Brookes University. |
| Location | |
| Category | Conferences - Seminars |
How to represent movement skills in complex behavioral architectures ? This is a persistant research question in cognitive robotics which we tackle through a neuro-robotics approach based on the principle of continous neural association. The latter combines ideas from reservoir computing, input-driven dynamical systems and classical associative memories in a coherent framework. The goal is to learn to bind together sensori-motor data, parametric representations like dynamic movement primitives and low-dimensional embeddings of task-specifying parameters like e.g. via points in a multi-level and multi-scale skill memory. It is further shown that this paramteric skill memory is highly beneficial for application of modern trajectory based learning based on roll-outs and reward-weighted averaging, which we demonstrate in applications to velocity field learning, inverse kinematics control, and skill optimization for humanoid robots like the iCub.
Practical information
- General public
- Free
Organizer
- Mayra Lirot
Contact
- Mayra Lirot