Page - 51 - in Joint Austrian Computer Vision and Robotics Workshop 2020

the box is done by asynchronously opening the box such that the model logs spread randomly on the floor. The floor area designed as log picking area can be shielded during a box emptying process to prevent the logs from spreading too wide in the area. With the project goal of the crane being able to autonomously store the logs in the box, this automated emptying process enables an endless cycle for automated training refining the AI based procedureswithout supervision. The crane is controlled at a high level by an external PC which is connected via Ethernet to a HAWE-ESX control unit. The ESX controls the hydraulic pistons and sends the signals of the wire-rope and custom angular sensors via Ethernet back to the host PC. The PC also receives data from two cameras mounted on the fix and movable part of the crane as well as from five IMUs mounted on each of the crane joints. These sensors will serve for automated model creation as we assume to not have CAD drawings of every crane in a retrofitting process. Theoverallconnectivityschematic isshown inFig.3. Figure 2. Crane model system with automated elements for continuous learningwithout human intervention. Figure 3. Overview on the connectivity of the model crane, the sensors, and the externalPC. 2.1.RetrofittableSensors Automating machinery in the wood sector is challenging since not only the sensors that enable autonomy need to be equipped ideally without disassembling the machine, they also need to be autarkic in terms of energy, and withstand very harsh environments. Thus, robust magnetic angular positionsensorsfollowing[1]suitablefor retrofitting and wireless operation have been integrated on the cranemodel. Theycaneasilybeadaptedfordifferent joint geometries. The basic architecture is shown in Fig. 4 together with the lab setup (currently with wired CAN). In addition, capacitive sensors Figure4.TheexperimentalsensorandCADcoilgeometry on the rotary joint of the end-effector: The coil PCB and signal processing circuitry is mounted to the non-rotating head whereas the conductive plate is mounted on the rotating shaft. The conductive counterpart consists of a 3Dprintedholder andwrappedcopper foil. following [2] are integrated in the end-effector to augment themachinerywithasense for loggrasping quality (Fig.5). Thecraneandsensorsare simulated Figure 5. Left: V-REP model. Bottom right: gripper design. Top right: photograph of the gripper prototype including the sensorelementswireless electronics. in V-REP. There, the communication and control are tested using V-REP/ROS and V-REP/Python bridge. The simulation also serves as an environment for AI training of the crane controls and for optimizations on sensor placement following [3]. A video of the simulation frameworkcanbe found in [4] 51