Towards the autonomous robotic gripping and handling of novel objects

摘要

The automatic grasping of objects previously unseen by a robotic system is a difficult task - of which there is no robust solution. In this paper, a 3-phase, learning-based methodology is proposed that generates and scores candidate grasping locations for unknown objects through vision. The point of difference of this approach is mainly in its industrial implementation, definition of a new metric to assess the quality of grasps and use of part-related information and multiple camera perspectives in determining such locations. Through experimentation this methodology has been shown to generate meaningful candidate grasping rectangles for unknown objects in under 2 seconds - given a single image - which demonstrates the potential of the proposed novel object handling research.