Our mission is to make robots learn in a developmental fashion - similar to children. Why should that be useful? Already now robots are present in many areas of our life, taking over tasks that are too dangerous, too repetitive, or require too high precision for humans. However, their area of application is bound because preprogrammed robots are not able to successfully interact with our complex and constantly changing world.
In living beings, this problem is solved by an interplay of learning, self-organization and innate
information. Self-organization—ubiquitous in nature—offers promising perspectives for practical applications because it is based on local interactions and typically scales well to large fault-tolerant systems. In our research, we were building the theoretical basis for self-organized robot control. We study how dynamical systems and information theory can be used to generate sensory-motor coordination in robots.
In order to make robots successful in learning new skills, they have to extract as much information as possible from experience. Learning a model of its body and the environment can capture the information needed to master unknown future tasks. Classical regression models are not good at predictions in new situations. We investigate how an algorithm can identify the underlying relationships in data, thereby obtaining a model that can extrapolate well.
For actually learning, improving and controlling goal-oriented behavior, we use reinforcement learning and optimal control methods. Our aim is to make reinforcement learning so data-efficient that it can be applied to real systems easily. We currently take two complementary routes. One route is to combine optimal control methods and planning with reinforcement learning. The other route is to learn many tasks at the same time using hierarchical learners that are intrinsically motivated to explore appropriate goal spaces and understand relational structures in the environment.
Machine Learning methods are the working horses behind our recent developments in model learning, reinforcement learning, and representation learning. While investigating the state of the art methods we realized a potential for speeding up stochastic gradient descent in practical applications. Further, we investigate unsupervised learning of independent representations.