Design and Development of Ball Catching Robotic Arm

Abstract

The task of tracking and predicting balls was studied as part of robotic ball catching systems ([4], [3], [6], [7], [8]). All of these have a static setup using stereo cameras with rather wide baselines. Detecting the ball is done by pixel-wise segmentation, using color ([6], [7], [8]) or using the difference to a reference image [3]. (Real time trajectory perception paper). A system with a 5 DOF arm on a humanoid robot (only the arm is moving) with a "cooking basket" at the end effector for catching the ball and an active vision system is presented in [9]. In work [3] they used the 7 DOF DLR-LWR-II arm with a small basket at the end effector and a stationary stereo camera and image processing system built from cheap "off-the-shelf" components. III. CONTRIBUTION In this paper we present a ball catching robotic arm setup, which significantly differs in two aspects from previous research. First, we develop a robotic arm with commercially available parts with a low cost. This is challenging for us to build an arm with such a low cost because the hardware and motors of the arm has to be able to reach to the ball fast enough and in addition it has to be light weight, so that the impact of the weight doesn't effects the motors. Moreover, this is also challenging for us to implement the algorithms for the arm movements, because all the work here is to be carried out for the real time results so we need such algorithms which are more efficient than the previous work and also which are able to work on our robotic arm setup. The second new aspect is, that, instead of using dual or stereo vision cameras (300 fps) we use a 60 fps camera to determine the ball and to calculate its future trajectory. Our camera is here responsible for tracking ball, determines velocity of ball and also determines angle at which the ball is getting turned. So using these simple steps we are here able to calculate the trajectory of ball which then allows the arm to move up to the ball. As early mentioned we left the hand grasping work on the future work because of low funds availability. IV. GENERAL SETUP AND ARCHITECTURE Predicting the ball with stereo vision camera is extremely expensive as we need Gige interface cards and more powerful programming language to interface. We therefore use a single camera system mounted on a vertical pole on the beside on the thrower. The viewing angle of camera is 5m horizontally and about 4 m vertically. This is enough for us for tracking of ball from around 5 m. Simple studies of human performance indicates that our system is able to catch the ball if the ball is throwing in virtual 150 cm by 150 cm window. From these basic requirements it is possible to compute flight time and velocities for the trajectory perception, as summarized below.  Throwing distance will be approximately 5-6m.  Ball should be in air or flight time is about 0.9 s.  The ball will travel with an approximate velocity of 5.5 m/s.  The ball should be caught if it arrives within a virtual 150cm × 150cm window. Figure2 Schematic of ball catching experiment The ball is thrown by a human from a distance of about 5 m onto the robot with speed of typically 5.5 m/s, resulting in a flight time of about .9s. The robotic arm is placed on a table about 2.5 m above the ground level. The single camera system (60fps) is placed beside the throwing person at 1m height from ground level. The ball has a diameter of 8.5 cm and a weight of 70 g. Figure 3 System architecture showing the processing modules V. OBJECT TRACKING WITH MEAN SHIFT ALGORITHM Mean Shift is a powerful and versatile non parametric iterative algorithm that can be used for lot of purposes like finding modes, clustering etc.