Robotic Sorting of Batteries Using Visual Few-shot Learning and Fusion with Depth Data

Recycling electronic waste becomes more and more important to recover precious materials. Batteries in particular contain significant amounts of rare-earth elements. These can be extracted in specialized recycling methods provided that the batteries are pre-sorted. Toxic substances released during the sorting process can expose human workers to health risks. Therefore, we propose an automated sorting line with an industrial robot for batteries based on visual and depth sensors, whereas other approaches predominantly only use monomodal sensor input. We use few-shot learning methods for visual detection to not rely on large datasets. Additionally, the visual predictions are verified based on the processed depth data using expert knowledge. The expert knowledge is integrated using rules regarding the spatial dimensions of each category of batteries. We evaluate our application in a realistic test scenario. We find that the combination of deep learning and depth data processing ensures high accuracy for sorting applications with a reduced number of training samples.