In this paper, we tackle sequence-to-tree transduction for language processing with neural networks implementing several subtasks, namely tokenization, semantic annotation, and tree generation. Our research question is how the individual subtasks influence the overall end-to-end learning performance in case of a convolutional network with dilated perceptive fields. We investigate a benchmark problem for robot command interpretation and conclude that dilation has a strong positive effect for performing character-level transduction and for generating parsing trees.

We present a unified visuomotor neural architecture for the robotic task of identifying, localizing, and grasping a goal object in a cluttered scene. The RetinaNet-based neural architecture enables end-to-end training of visuomotor abilities in a biological-inspired developmental approach. We demonstrate a successful development and evaluation of the method on a humanoid robot platform. The proposed architecture outperforms previous work on single object grasping as well as a modular architecture for object picking. An analysis of grasp errors suggests similarities to infant grasp learning: While the end-to-end architecture successfully learns grasp configurations, sometimes object confusions occur: when multiple objects are presented, salient objects are picked instead of the intended object.