homer@UniKoblenz: Winning Team of the RoboCup Virtual @Home Open Platform League 2021

In this paper we present our approaches for solving this year’s virtual RoboCup @Home tasks with special focus on object detection and manipulation in the simulation environment Gazebo, which lead to our successful participation in the Open Platform League. There, we deployed the model of robot TIAGo Steel Edition, which is publicly provided by PAL Robotics. For object detection and pose estimation, we use a custom pixel-based clustering approach for segmenting potentially object supporting planes first. Instances of arbitrary and unknown objects can thus be detected and suitable grasping poses can be deduced. We perform object recognition by training our YOLO v3 network on a synthetic dataset of rendered YCB object images with Blender. A generalized grasping pipeline, which integrates extent and orientation information from our general object detection, performs planning and execution of combined arm and torso trajectory using MoveIt.