Seite - 138 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

2. RelatedWork Numerous works exist about high-level planning for robot systems solving complex tasks using sets of simpler system capabilities. This system capabilities are called skills. Dividing a complex task in suchskillshasmultiplebenefits likeflexibility, reuseof skills andgoodsoftwareportability. Pedersonetal. showsin[13] thedivisionofacomplextaskasequenceofmultiplesubtasks(=skills). Skills are described being the fundamental buildingblocksor the systemcapabilities. If a new com- plex task should be executed, the system needs not being reprogrammed. It is sufficient to simply reorder the skills. Similar to our approach, anexecutionmonitor surveys theoutcomeof the skills. In [12] skills are ported to different robotic platforms. Skills get further decomposed into skill prim- itives. With this detailed decomposition the hardware level is abstracted from the skills itself. The advantageof modularity and theabstractionof tasks ispointedout. The authors in [14] introduce a 4-TIER architecture, with the same idea of abstraction for skills and skill primitives as in the previous papers. The lowest layer ensures the hardware abstraction and so the re-usability on different platforms. The next layer contains action and perception primitives. The top layers handles the planning task. As the previous addressed work, this abstraction is used for easing the human robot-interaction. All these papers show a clear distinction between tasks, skills and primitives and focus on portability and easy execution of complex new tasks. But their focus is onhuman-robot interaction. Thehuman in the loopdefinesanewtask through reorderingskills. The next works present a successful task planning utilizing artificial intelligence (AI) planner instead of humans in the looporderingskills. In [6]Huckabydefinedskillswithpreconditionsandeffects in the model spaceof theproblem. The initial stateandgoal are stated in theprocess space. Theyproposed PDDL[7] asplanning language. In [6] the focus lieson thehigh-level. It is assumed that skills and theirprimitivesalwayssucceed. In [11] a system is proposed which transfers the high-level description from the AI planner to a behav- ioral statemachine. Failures in theprimitiveexecutionaredetectedbyavisionsystemandrecoveries areperformed. Finally some works addressing the order picking problem are discussed. The authors in [9] present a mobile bin picking system. Items are picked from a box standing on the ground and placed at a delivery station. Thehigh-levelof this systemisafinite statemachine (FSM). In [3] a software architecture and their implementation for grasping objects is presented. Some of theseconceptsareused inourwork too. Thecollisionenvironment isa3Doccupancygridexcluding robot parts. Known and recognized objects are represented as geometric primitives or as mesh mod- els of the objects. In [1] the authors present a pick and place approach where they have to deal with knownandunknown objects, clutteredworkspaceandnoisy sensordata. 3. TargetEnvironmentandSystem For the a proof-of-concept implementation of the proposed order picking system we use the robot Baxter from Rethink Robotics (see Fig. 1a). It is a two-arm robot with internal sensors such as cameras and proximity sensors in the wrists. In order to get a global overview of the environment we added a RGBD camera on top. The environment Baxter operates in is shown in Figure 5b. It is a mock-upofa typicalmanual storage. 2 138