| The problem of an aging society is real and will affect everyone. There will be too few young people that can ensure adequate living conditions for the elderly. Personal robots have the potential to assist in day-to-day tasks whenever there are too few humans to cope with societal needs. However, for personal robots to become useful they need to be able to skillfully manipulate objects in their environment. Unfortunately, the problem of autonomous manipulation is very complex and progress towards creating autonomous behaviors seems to have reached a plateau.;In this thesis, we will present a data-driven approach to movement generation. We argue that movement generation (motor output) and perceptual processing (sensor input) are inseparably intertwined and that the ability to predict sensor information is essential for skillful manipulation. Movement generation without sensor expectations defaults to open-loop execution which is prone to failure in dynamic and unstructured environments. However, predicting sensor information for an increasing number of sensor modalities including force/torque and tactile feedback through physics based modelling is challenging given the variety of objects, the diversity of possible manipulation behaviors, and the uncertainty in the real world. Instead, our approach leverages a key insight: Movement generation can dictate expected sensor feedback. Similar manipulation movements will give rise to sensory events that are similar to previous ones. Thus, stereotypical movements facilitate to associate and accumulate sensor information from past trials and use these sensor experiences to predict sensor feedback in future trials.;We will call such movements augmented with associated sensor information Associative Skill Memories (ASMs). We will present a coherent data-driven framework for manipulation that implements this paradigm. First, we will introduce a modular movement representation suitable to encode movements along with associated sensor experiences. Second, we will show how stereotypical movements can be learned from demonstrations and refined using trial-and-error learning. Third, we will show how ASMs can be used to monitor task progress, to realize contact reactive manipulation, and to purposefully choose subsequent movements. Finally, we will present a method that can learn forward models for these stereotypical movements. |
