Towards Robust Skill Learning With Prediction Guided Autonomy in Unknown Environments

We present first steps towards an architecture that enables autonomous systems to learn and adapt basic behaviors to unknown environments. It can act as layer providing robust behaviors to higher-level layers like e.g. reinforcement learning or planning algorithms, that depend on encapsulated actions. With emphasis on robustness it is able to learn to control its actors without any information about the meaning of its actors. This is made possible by behavior modules that are learned together with their action-effect dependencies and their enabling preconditions by proactively carrying out experiments within their environment