Page - 74 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

AVisual Servoing Approach for a Six Degrees-of-Freedom Industrial Robot by RGB-D Sensing ThomasVarhegyi,MartinMelik-Merkumians,Michael Steinegger, GeorgHalmetschlager-Funek, andGeorg Schitter Abstract—Avisual servoing approach is presented that uses depth images for robot-pose estimation utilizing a marker- less solution. By matching a predefined robot model to a captured depth image for each robot link, utilizing an appro- priate approximation method like the Iterative Closest Point (ICP) algorithm, the robot’s joint pose can be estimated. The a-priori knowledge of the robot configuration, alignment, and its environment enables a joint pose manipulation by a visual servoed system with potential to collision detection and avoidance. By the use of two RGB-D cameras a more accuratematchingof the robot’s links is feasiblewhile avoiding occlusions.Themodeled links are coupled as akinematic chain by theDenavit-Hartenberg convention, andareprevented from divergence during the matching phase by the implementation of an algorithm for joint pose dependency. The required joint orientation of the robot is calculated by the ICP algorithm to perform a pose correction until its point cloud align with themodel again. First tests with two structured light cameras indicated that the recognition of the robot’s joint positions brings good results but currently only for slowmotion tasks. I. INTRODUCTION The fourth industrial revolution involves the use of new robotic technologies for smart and efficient work-flows in an innovative way. Humans will work together with robots side-by-side and integrate them in their every daywork life as a collaborative device. Therefore, a collision detection with humans and the environment has to be established, for instance, with pressure sensitive skins [1, 2] or abnormal force recognition [3, 4] which are two approaches for a collaborative aspect.Another idea is the integrationof visual perception [5, 6]. Robots should see where they are, know and see the environment they move in and know how they can grab and move without disturbing the work-flow. The focus of this paper lies on the application of computer/- machine vision methods for image processing and robot actuation. Vision-based motion control of robots is called visual servoing, where the robot manipulator is operated by the evaluation of visual information from an eye-to- hand (camera fix to workspace position) or an eye-in-hand (camera attached to robot) composition [7]. Figure 1 shows the recording of a robot in an eye-to-hand composition, that is used for the visual servoing approach in this paper. The advantageof visual servoing is that the teach-inprocedureof a robot can be omitted since tool-tip-pose errors caused by low accuracy between the tool-tip-pose and the joint angle All authors are with the Automation and Control Institute (ACIN), Vienna University of Technology, A-1040 Vienna, Austria. Con- tact: melik-merkumians@acin.tuwien.ac.at (correspond- ing author) Base Link1 Link2 Link3 Link4 Link5 Link6 Fig. 1: ABB IRB 120 point cloud model overlaid by the captured point cloud from the IntelR© RealSense R200. can be corrected in addition. These visual information can be exploited as position- or image-based information [8–10]. Position-based detection uses interest-points in the image to detect the object position, while image-based detection uses a template image of the designed object to predict how the camera should be aligned to the object. So far, mainly 2D cameras have been applied for visual servoing applications [11–13]. The accuracy of the interest point estimation in the image as edges or corners determines howprecisely the robotcanbepositionedby2Dcameras.For objects without distinctive characteristics as curved shapes without edges, these kinds of camera systems do not suite perfectly. In this case depth sensing cameras is the better choice. RGB-D imaging systems can be separated into three main groups. First, stereo vision systems [14] which are based on two cameras and feature disparity where the depth information is obtained by the use of triangulation. Second, structured light cameras [15, 16] with the same basic principles as stereo vision cameras but instead of the second camera a projector is used. It emits a patterned light (usually infra-red light) and measures the disparity of the 74