When the demo was released, people had a lot of questions about how the cameras would deal with occlusions and products being covered up by bodies and odd angles. Recognizing and localizing and tracking a single person picking up individual items slowly is a much easier task than tracking many people in a crowd crisscrossing and blocking the camera constantly; that it is apparently failing in exactly the scaled up realistic scenarios where you would predict increased failures points to it being a genuine difficulty and not simply too few crowded scenes in the training set.

(I bet their deep learning team is having a lot of fun trying to figure out how to make RNNs rather than CNNs scale and train well...)