Figure 6.

Comparison of the results on a book grasping task between the original approach by Hangl et al. (2016) (A) and the novel extension presented in this work (B). (A) Resulting ECM after training with the original approach by Hangl et al. (2016) on a book grasping task. All required preparatory behaviors have to be available at the beginning of the training process. As the training is only done model-free, training converges slower (c.f. section 11.3) and no preparatory behaviors can be generated when needed. The colored edges denote strong connections with high transition probabilities. (B) Exemplary resulting ECM of our novel approach on a book grasping task. At the beginning of the training process only the behaviors rotate 90, flip and void were available. The behaviors shown in brackets were created using behavior composition. Note that due to the probabilistic nature of Equation (21) other behvaiors could be created like rotate 90, void, rotate 90. For the sake of readability, the weak connections were omitted.