Abstract:

For robots to become truly autonomous and be able to overcome changes in either the working environment or its own embodiment, it needs to ability to self-learn tasks and gradually adapt to changes. In our work we have tried to move the human designer as far as possible from the learning process and allow the robot to create its own model of the world and with minimal prior information learn to exploit its embodiment through exploration, often in collaboration with a human partner. We will present some work in this direction from simple reactive behaviors learned from ground up to more complex predictive behaviors learned in stages that consist of both simulations and real robot experiments. Using combinations of Gaussian process models, deep neural networks and reinforcement learning, emphasis is placed on data efficiency and adaptivity, allowing the robot to learn in a data-driven manner with a minimal number of trials.

Brief bio:

Mårten Björkman is an Associate Professor at the school of Computer Science and Communication at the Royal Institute of Technology, KTH. He received a MSc degree in computer science and engineering from Lund University in 1994 and a PhD degree in computer vision and robotics from KTH in 2002. He has been actively contributing with research to the EC funded projects CogVis, PACO­-PLUS, eSMCs and socSMCs. His research interests are human-robot collaborative system, real­-time object detection and segmentation, and data ­driven mobile manipulation.