Spin Robotics contributed to this development by providing a robotic screwdriver that was fully integrated into the Universal Robot arm from CSEM, the partner with whom we have been working most closely and who also provided the footage essential to this project. AVL supported the use case with a battery unit adapted to ensure safe experimentation. Using the low-level code examples delivered by SPIN, CSEM implemented the functions required for complete script-based control of the robotic screwdriver.

A deep learning pipeline was developed to automate screw detection, beginning with the generation of thousands of high-quality synthetic images in Blender. Each image was paired with precise positional labels, providing a strong foundation for model training. The model was then refined using real images captured from the use-case battery to better reflect real-world conditions.

The detected screws were tracked across video frames, enabling multiple viewpoints. This multiframe approach is essential for accurately estimating screw positions in physical space through triangulation. Once the screw locations are identified, the robot autonomously moves toward each one, allowing the robotic screwdriver to remove them sequentially. Detection, triangulation, and removal are fully integrated into an automated workflow that operates without human intervention.