Abstract
This paper describes the initial release of annotated datasets from the advanced robotics and artificial intelligence for critical raw materials recycling in the circular economy (XBAT+) project. Relatively small battery waste electrical and electronic equipment (WEEE) devices were sorted from cages of general WEEE during visits to recycling facilities and collection events. From 50 initially defined battery-WEEE categories, applying a frequency criterion reduced them to 15 categories comprising 421 battery-WEEE devices. Optical and x-ray imaging of battery-WEEE devices produced multiple images per device in some cases, resulting in 3045 static images. These were organised into high-quality, varying-quality, and RGB datasets, semi-automatically annotated, and split into training and test sets. Specifically, within each of the 15 categories, 20% of the images were allocated to the test sets, the rest to the training sets. Evaluation using YOLO26 and RF-DETR demonstrated promising mAP@50 and recall for supervised learning-based battery-WEEE identification and battery/nobattery detection, highlighting the potential of XBAT+ datasets to enable WEEE filtering and mitigate battery-based fire hazards in recycling facilities.