Thorbecke is naming the right bottleneck. End-effector dexterity, not locomotion or whole-body control, is what gates humanoid usefulness in the work cells operators actually care about. Tactile sensing, fingertip force control, and grasp generalization are where the engineering deficit sits, and they are also where the supply chain is thinnest. Most humanoid OEMs are still sourcing hands as a separate program from the body, and the parts count per hand (16 to 22 actuators in the high-end designs) drives a disproportionate share of BOM cost and failure modes.
The operator read: discount any humanoid pilot demo that does not show the robot manipulating deformable objects, variable-geometry parts, or anything requiring in-hand reorientation. Pick-and-place from a fixture to a bin is a solved problem with a $30K cobot arm. The reason BMW Spartanburg and GXO Flowery Branch matter as verification anchors is that they force the hand question into the open under customer acceptance terms.