Learning of Neurobotic Visuomotor Abilities based on Interactions with the Environment

Robotic visuomotor abilities, like grasping, can either be realized through conventional means of independent modules for subtasks like object localization, grasp planning, and inverse kinematics. These modules, however, rely on the availability of accurate robot and environment models. An alternative is to acquire visuomotor abilities through end-toend machine learning. While deep neural networks have proved successful in many areas, they depend on large amounts of annotated training data or long periods of trial-and-error learning. To overcome this issue, developmental robotics leverages principles of incremental learning in biological agents. Increasingly complex visuomotor abilities are learned through mostly autonomous interaction with the environment. Following this paradigm, we present current research on acquiring visuomotor skills with a humanoid robot through self-learning and minimal human assistance. The robot engages in a learning cycle where it repeatedly manipulates an object to gather training samples that link its actions (joint configurations) to states of the environment (images from the robot’s perspective). Human assistance is only requested if errors occur during this phase, e.g., the training object is accidentally dropped out of reach. Based on these training samples, supervised end-to-end learning of visuomotor skills is realized with a deep convolutional neural architecture. The results show that the approach generalizes well to novel objects that were not included in learning. To enable this research, we developed NICO, the Neuro Inspired COmpanion, a humanoid research platform for embodied neurobotic models and human-robot interaction.

@InProceedings{KW17a, 
 	 author =  {Kerzel, Matthias and Wermter, Stefan},  
 	 title = {Learning of Neurobotic Visuomotor Abilities based on Interactions with the Environment}, 
 	 booktitle = {Proceedings of the DGR Days 2017},
 	 editors = {Asfour, Tamim and Beetz, Michael},
 	 number = {},
 	 volume = {},
 	 pages = {14--15},
 	 year = {2017},
 	 month = {Nov},
 	 publisher = {},
 	 doi = {}, 
 }