Mobile Manipulation Based on Generic Object Knowledge

Abstract

The control of vision-based mobile manipulators relies on specific information representing the object which has to be manipulated. Conventional methods like position-based or image-based visual servoing use a precise 3D model of the object or image features to move the end effector to the desired position. This a-priori knowledge has to be generated by an expert before the manipulation can be performed and it is limited to a specific object. Similar items of the same category varying in size or color can not be handled with these approaches. We are proposing a new methodology, which uses a generic category description instead of specific object features in order to allow the interaction with a broader range of items. This description includes appearance properties represented through geometrical primitives as well as additional category knowledge like functional properties, mechanical attributes and directions for object handling. Since the online determination of possible grasps for multi-fingered grippers is time-consuming, suitable grasp modes are stored for every object category, including preferred areas of contact on the objects surface