Siemens turns to CEA-List for new vision technologies for robotics and automated control

The purpose of CEA-List’s partnership with Siemens on computer vision for robotics was to improve robots’ understanding of their environment and the resulting actions. The potential industrial use cases for the research are wide, from manufacturing and agricultural robotics to sorting waste and dismantling equipment—any job that requires adaptability, precision, and controlled cycle times. In this AI-enabled system, robotic manipulation and visual inspection of the part are combined in a way that is particularly effective for grasping the part optimally and holding it in front of a camera for a reliable visible defect inspection.

The partnership produced three new technologies:

• HC6D zero-shot object location in 3D spaces (patented);
• Calculation of dynamic object input configurations, regardless of initial position;
• One-class anomaly detection using a learning-based “normality modeling” technique to detect defects of an unknown nature. This approach is much more frugal than conventional approaches (<200 images instead of >1000+ images).

CEA-List’s Phosphor software was used to integrate the technologies, which performed well in tests on several object inspection scenarios. The focus of the partnership will now shift to applying the technologies developed to the fine manipulation of deformable objects with multi-finger grippers, and generalizing the technologies to other use cases.

Contributed to the article

• Fabrice Mayran de Chamisso, vision & AI research engineer
• Boris Meden, vision & AI engineer-researcher