Figure 2.

ECM for autonomous robotic playing. For execution a random walk is performed from layer A to layer E. The transition from layer C to layer D is performed by executing the corresponding sensing behavior s, measuring the haptic data and using a time series classifier. All other transitions follow Equation (6). The preparatory behavior b_∅ ≡ (void behavior) is always in the set of preparatory behaviors. The dashed box and lines refer to skills used as preparatory behaviors in order to build skill hierarchies. After preparation, the basic behavior b^σ corresponding to the desired skill σ is executed. The active learning approach presented in this paper enables the robot to perform transitions between clips on layer D. The behavior composition extension helps the robot to add new preparatory behavior compositions in layer E. The ECM shown in this figure is independent of the environment model and will only be influenced (e.g., by adding new clips) in case the environment model becomes more mature.