Abstract
This research aims at acquiring useful knowledge for robot tasks from many teaching examples by users. We, therefore, develop a system which enables users to easily teach tasks to robots in a virtual world and apply it to acquisition of knowledge for block assembly tasks. Easiness of teaching and a realistic simulation are the keys to a comfortable task teaching in a virtual world. For an easy teaching, the user indicates only blocks and assembly steps by mouse pointing, and the system then automatically calculates the rest of necessary information such as the hand pose for grasping and the robot trajectory. The system also performs a physical simulation for a realistic simulation. Moreover, the interface of the system is implemented using Java for a possible public release in the future. Next, we collect many teaching samples by several users for various block assembly tasks. We then extract useful task knowledge from the samples using machine learning techniques.
