Clutter, occlusion, and the bin-picking problem
A lab demo usually shows a robot picking a single, isolated, well-lit object sitting on a clean table, and that's a meaningfully easier problem than what a real warehouse presents. In practice, objects are often stacked, partially hidden behind each other, or piled together in a bin, a scenario specific enough that it has its own name: bin picking. Occlusion means the segmentation and pose estimation stages from earlier in this course have to work with incomplete information, guessing at an object's full shape and orientation from only the visible portion. Clutter compounds this further, since multiple overlapping items make it harder to tell where one object ends and the next begins, and a grasp aimed at one item can easily collide with or disturb its neighbors.
This is precisely why bin picking has remained a benchmark problem in robotics research for years rather than being solved once and forgotten. Companies moving product through warehouses need reliable performance across enormous variety, not just accuracy on a curated set of test objects, and small failure rates that look fine in a research paper translate into real lost throughput at scale.