Robust recovery of piled box-like objects in range images

摘要

This work addresses the vision-guided, robotic bin-picking problem, in the context of which a number of piled objects, should be localized, grasped and transferred by a robotic hand from the position they reside, to a specific place defined by the user. We deal in particular with the depalletizing problem according to which deformable box-like objects piled on a rectangle platform, the {\\it pallet}, should be unloaded. The requirement of a robust system for dealing with this problem stems from almost all industrial sectors, and is expected to substantially reduce the costs associated with product handling and distribution.
<br> 
<br>From the hardware point of view, our system comprises a six degrees-of-freedom industrial robotic arm, on the hand of which a laser sensor is mounted for data acquisition. Besides, a vacuum gripper is mounted on the hand of the robot for object grasping. The object removal process is as follows: Firstly, the top side of the object configuration is scanned by linearly moving the robotic hand along the pallet, and a range image is acquired. Secondly, the image is analyzed and the graspable objects in the pile are localized. Thirdly, the robot grasps the recovered objects from their exposed surfaces and places them at a user defined position. This procedure is executed iteratively, until no objects lie on the pallet.
<br> 
<br>This thesis mainly focuses on the object localization or recovery process, that is, the way in which given a range image the position and dimensions of the objects is determined. Our strategy for object recovery is model based, uses geometric parametric entities for object modeling, and has two aspects, in both of which, in addition to the input range image a boundary image obtained by the former by means of edge detection is employed. Firstly, globally deformable superquadrics are used for modeling our target objects. The object recovery is posed as an optimization problem, in the context of which given the input range image, the posterior probability of the parameters of all graspable objects in the pile is maximized. Our approach extends the recover-and-select paradigm, the most widespread framework for superquadric recovery from range images, since it incorporates object boundary information into the recovery process. This is the main reason why our approach outperforms the recover-and-select framework in terms of both computational efficiency, and robustness.
<br>
<br>Secondly, the boundary of the exposed surfaces of the target objects is modeled as a three dimensional rectangle. The Hough transform is employed to recover the target objects from the boundary image of the object configuration. The seemingly difficult problem of recovering three dimensional rectangles is straightforwardly solved by decomposing the Hough transform into two problems of lower dimensionality: The pose of the objects is recovered, followed by the recovery of their dimensions. This results to a computationally efficient framework.
<br>
<br>The decision on which strategy should be adopted for object recovery depends on the rigidity of the objects. The latter strategy does not account for object deformations, but it is faster than the former. Hence, if we know beforehand that the configuration comprises rigid boxes only, the latter strategy is used and in every other case, the former. Experimental results demonstrate that the resulting robotic system exhibits a variety of advantages such as robustness, flexibility, accuracy, and computational efficiency, the combination of which cannot be found in any existing system up to our knowledge.