How does a humanoid robot keep its balance while carrying a box?

When a humanoid lifts a box, the table stops supporting its weight. Holding the box close reduces the turning effect of its weight. To take a step, the robot must manage the change in support as one foot leaves the floor. How does it learn to make that correction? Experiments in Beijing and Oregon trace the path from a camera image and a changing foot load to a new command at a joint. A separate study follows a robot as it lifts a basket from a side platform, carries it to a table and sets it down. The researchers count a success only when the basket rests securely on the tabletop. A robot can stay upright and still fail its delivery. To judge what the AI contributes, we need to know which movements it controls and whether it recovers when the next box carries its weight differently.