Probabilistic inference methods in robotics — filling the gap between high-level reasoning and low-level motion control

Cognitive robotics is one of the topics of the new PASCAL 2 project. While classi- cal AI approaches can achieve outstanding performances on abstract and symbolic reasoning problems, one of the core problems in robotics is the grounding of high-level planning and rea- soning techniques in the low-level motor and sensor context. Machine Learning techniques can tackle exactly this grounding problem. In this paper we discuss a new approach that uses probabilistic inference techniques to solve low-level motion control as well as high-level planning in Markov Decision Processes. One of the most interesting aspects of this approach is that it does not rely on representing the state of the system as a single high-dimensional state variable. Instead, inference techniques can realise reasoning and planning on structured state representations such as hierarchical and distributed representations. In this way they are a promising candidate for integrating sensor processing, high-level reasoning, and low-level motor control in a single coherent framework.