Automatiserad batterisortering med datorseende och djupinlärning : Python-baserat system för objektigenkänning och sortering med visionsystem och robot

The recycling industry faces the challenge of rapidly and accurately sorting large volumes of cylindrical batteries. This study presents an AI-based vision system that integrates a GPU-accelerated YOLOv5 model for real-time detection of buttoncells and cylindrical batteries with a UR3 robot for automated picking. Affine transformations map image coordinates to the robot’s workspace, and an asynchronous queue architecture ensures continuous processing without unnecessary stops. The system achieved 98 \% precision and 96 \% recall, with [email protected] of 97 \% and [email protected]:0.95 of 81 \%, while maintaining inference times around 20 ms per frame. The average pick cycle time was 5.48 s with continously running conveyer belt and detection stability remained above 95 \% at conveyor speeds up to 345 mm/s and when handling tightly clustered objects at speeds up to 200 mm/s. These results demonstrate that coupling AI vision with collaborative robotics can deliver both high accuracy and throughput. The system’s modular architecture allows straightforward scaling to additional battery types and higher processing rates, highlighting its potential to improve efficiency, precision, and sustainability in recycling.