Visual Grasping Using Passive Joints and Clustered SIFT-Features

This work addresses the problem of visual-based grasping of unknown objects with a five-finger robot hand. The contribution of this work consists of two parts. Firstly, the sensor system of the robot finger tips has been enhanced to increase its sensitivity for reactive grasping of unknown objects. Therefore, two additional passive degrees of freedom are introduced to the finger tips to make them adjust themselves to the object surface. Secondly, a robust approach to visual determination of unknown objects is introduced. The visual part starts with the extraction of interest points which are typical for an object. A cluster of these object features indicate an object candidate. The proportions of this accumulation of 3-dim. points and the distribution of the feature categories within this cluster indicates an object to be grasped. The proposed approach has been tested with the novel tactile sensor system.