Vision models expanded what robots could reliably see
For most of industrial robotics history, a robot could only reliably handle objects it was explicitly programmed to expect: precise positions, known shapes, consistent lighting. Modern computer vision, particularly deep learning based object detection and pose estimation, changed that meaningfully. A bin-picking robot can now identify and grasp a specific part from a jumbled pile of mixed, randomly oriented items, something that would have required rigid fixturing or manual sorting a decade earlier. This alone expanded automation into tasks like mixed-order picking and quality inspection that used to require a human's eyes.
This is a real, already-deployed shift, not a future promise. Warehouse pick-and-pack operations, defect detection on production lines, and produce sorting all lean on trained vision models today, and they've moved the frontier of 'what a robot can reliably see and handle' considerably further than fixed, pre-programmed vision systems ever managed.