Seite - 141 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

Figure3: DecompositionofskillmoveBoxFromLevelToTray intoitsprimitivesandhowexecutivelayer handles the execution. 4.3. BehavioralControlLayer This layer holds all skill primitives. The primitives are platform-dependent and have to be re- programmed for specific robot platforms. The upper layers are immediately portable to other plat- forms. The primitives have a defined interface but their implementation is different for different platforms. Skill primitives can perform perception, manipulation, grasping tasks or any combination of those. Local recoveries are performed in this layer. If a recovery is impossible skill primitives report their error. A skill primitivecanbeusedbydifferent skills. Figure 4 depict for instance the primitive of looking for items once the box is on the tray. For the box detection the point cloud of the top RGBD camera and the PCL implementation [16] of FPFH features[15](initialdetection)andICP(finealignment)areused. Thecontrolof thearmsarerealized using the MoveIt! framework [2]. Items are detected using the RGB cameras in the wrist. Figure 5a shows the inspectionof aboxwhileFigure5adepict the internal representationof the situation. 5.4 Behavioral control Shelf Tray Box Baxter movement Figure5.22: After the box is turned the box’s handle is released and the grasping arm moves away from the handle. Figure5.23: In this figure the different blocks of the detect item skill primitive are shown. First it detects the transport box, then it moves the arm over the box and detects the item using the hand camera. If the skill succeeds the item pose PPG and the box pose BPG ares returned. 67 Figure4: Detect Item primitive.PPG represents the global item pose.BPG represents the global box pose. 5. Results The major result of this work is a working prototype implementation of the proposed order picking system based on the robot Baxter and standard software packages such as ROS or MoveIt!. But we are interested in particular in the dependability of the system. Therefore, we performed a detailed evaluationof the individual skill primitives. For the evaluation we executed individual skill primitives multiple times (around 50 trials each) in 5 141