3D Perception for Grasping (Stefan Schrandt)

master thesis

[1]

© Robotics

Motivation

Grasping is an essential feature to enable robots to interact with humans or manipulate the physical envi-ronment. For humans grasping is a simple task that does not need much eort. But when a human grasps an object, the hand is oriented and preshaped regarding the volume of the object. Finding this orientation and preconguration of the hand is complex because the number of degrees of freedom of the arm and hand is high and the model is not perceived completely.
Most current grasping approaches need perfect geometrical models of the object and the environmental scene. Using this knowledge the grasp is planned. Those approaches are successful in a simulated environment where the perfect geometrical model of the object is known. Nevertheless, those grasping techniques fail often during real world experiments where the model of the real world is not known precisely.

Expected Outcome

The goal of this work is to classify dierent objects in a given set of grasp categories to nd the pre-grasp that is most suitable for that specic object. In addition to this, part of the work is to nd the position and orientation of the classied object. This allows robots to grasp dierent objects without having knowledge about the exact shape or the function of that object. This approach should conrm the hypotheses that no object recognition is needed and that a coarse model is sufficient to grasp objects successfully. This strategy aims to be more robust and dealing with measurement noise.

Description of Work

Our grasping approach compared with most current ones does not rely on a perfect model. Additionally, the computational cost for the algorithm is signicantly lower as planning is no longer necessary. The compliance of the hand is used, hence a coarse model that does not need to be perfect is sufficient to compute grasps successfully.
In other words, the grasp of similar objects of one category i.e. a tomato, an apple, or a pear of the sphere category is grasped with the same pre-grasp (here: the spherical one). So it is not necessary to dierentiate between above the named objects to grasp them successfully.
The robot perceives the objects to grasp and the environment by the measurement of a 3D sensor. This measurements are searched for a planar surface where the objects are posed on. All measurements above the planar surface are clustered in groups to get for all objects a separate point cloud area. These clusters are used to nd the perceptual primitives that yield the best known pre-grasp for that specic object. Perceptual primitives are characteristics which classify objects to the dierent pre-grasp categories. The position and orientation is also obtained from these clusters.

People:
Clemens Eppner (primary contact)
Oliver Brock [2]



back