The first step in many industrial manufacturing processes is the provision of individual parts that need to be retrieved from bins, sorted and placed in a predefined matter. With the increasing needfor flexible automation, robot-based bin picking systems are a promising solution for this.

This thesis provides a comprehensive overview of the aspects and challenges involved in robot bin picking. With the aim of automating the handling of small electronic components, a new binpicking system is designed, developed and realised based on the current state of the art.

The use of AI enables fast and robust detection and classification of objects, but requires large amounts of training data. Therefore, a pipeline for generating synthetic training is developed using the 3D software Blender. Photorealistic renderings based on CAD data and textures allow training of AI models that can perform detection, classification and segmentation tasks in real industrial environments. The successful reduction of the synthetic-to-real domain gap represents a major success not only for the bin picking system. Other AI-supported industrial applications can also benefit from the data generated via this pipeline in the future.

Image processing for the bin picking system is implemented as a two-stage process. A neural network trained on synthetic data performs 2D object detection and classification. A combination of RANSAC and ICP feature matching algorithms perform 3D pose estimation based on 2D regions of interest.

Implementation and tests of the system are carried out on a demonstrator in an industrial robotcell. A 6-axis robot with parallel gripper and a state-of-the-art structured light 3D camera are used for this purpose. The latest 3D camera hardware with particularly high resolution enables the detection, localisation and precise oriented gripping of small objects. This has so far been given little consideration in the context of industrial bin picking and represents a technical innovation.The integration of commercial gripping and path planning software enables collision-free handling of the objects by the robot. The bin picking system is validated using real scenarios with electronic components and welding nuts. The effectiveness of the system is demonstrated for a number of simplified bin picking scenarios.This work lays the foundation for the optimisation, further development and industrial testing of the bin picking system. Individual modules from this work can also be used for other industrial machine vision applications.