The steadily increasing demand for electric vehicles (EVs) has led to a heightened need for efficient, automated battery disassembly systems for sustainable recycling. However, the variability of battery designs and the insufficient availability of training data pose significant challenges for machine learning in the context of disassembly automation. In this study, we propose an image-based approach that integrates deep learning and structured reasoning to enable the autonomous generation of disassembly sequences for EV battery packs. A YOLOv8-based pipeline for object detection and instance segmentation is trained using two different approaches to data augmentation: conventional image transformations are compared with synthetic image generation using Segment Anything Models (SAM). Object-specific augmentation using SAM leads to higher precision in object recognition than general, conventional augmentation techniques. Structural relationships between components are then derived using both bounding box heuristics and pixel-level segmentation masks. This enables reliable extraction of spatial and connection data. The information extracted from the image data can then be further processed to derive well-founded and adaptable disassembly sequences. This process is possible without the use of CAD models or extensive prior knowledge. The experimental results showed improved recognition accuracy by using SAM-based data augmentation compared to data augmentation with conventional image transformation (mean Average Precision at an Intersection-over-Union threshold of 0.5 (mAP@0.5) increase: 2.0 %) and compared to not using augmentation techniques (mAP@0.5 increase: 3.3 %) on a small-sized dataset containing 190 images (original size, + 50 images for SAM-based augmentation; + 160 images for conventional augmentation).
Baucks et al. (Thu,) studied this question.