Automation of collaborative robots using vision and tactile sensing for quality inspection

The industry is undergoing a transformation. This shift involves not only the automation of repetitive, simple tasks like pick-and-place operations but also tasks that were once thought to require human supervision or decision-making. Increasing flexibility in programming collaborative robots, advancements in sensor technology and their integration, growing computing power, developments in AI, and paradigm shifts in human perspectives have all accelerated this transformation. In this context, this dissertation investigates the incorporation of robots into a field still dominated by human labor (despite its repetitive and tedious nature) namely, quality inspection. The field of quality inspection presents a wide range of research opportunities, including robotics, sensor fusion, data processing, AI applications, human-robot collaboration, and more. In this dissertation, two distinct quality inspection concepts are explored. The challenges associated with each concept are different from each other, as are the corresponding methodologies used to address them. The first concept is aimed at solutions for automatic robot trajectory generation for surface defect inspection. While deep learning based methods for image processing have significantly advanced defect and anomaly detection in images, the automation of robot trajectory generation for scanning free-form surfaces in a flexible and efficient manner is still in its early stages. In this dissertation, after exploring the use of reinforcement learning in a simulation environment, a real-world robot trajectory generation method is introduced. This method employs Dynamic Movement Primitives (DMPs) to generate an initial trajectory and then iteratively refines and optimizes it based on the quality of captured images in an episodic manner. The second concept focuses on the inspection of tactile surface properties in the industry. Unlike visual quality inspection, this method is less widespread. However, due to the inherently subjective nature of human decision-making, it poses challenges in consistency and reliability. This subjectivity can, in turn, lead to undesirable variations in the production of high-quality goods. This dissertation aims to introduce a robot-based tactile inspection method capable of classifying surfaces with even subtle differences in their tactile properties, while also providing objective and quantifiable tactile metrics. For this purpose, data collected using a soft, human-inspired artificial finger is processed by a deep learning-based time series classification method called InceptionTime. This dissertation contributes to the advancement of automated quality inspection in industrial settings by addressing both visual and tactile quality inspection. The proposed approaches offer to enhance the efficiency of using robot by using learning-based trajectory optimization (for visual inspection) and deep learning-based classification (for tactile inspection). These approaches aim to offer more reliable, objective and scalable solutions to the traditional inspection methods.