Motion Retargeting from Humans to Humanoids

Human motion capture is a rich source of training data for humanoids, but a robot's body is not a human's. Its arms may be shorter, its joints may bend less, and its proportions differ. Retargeting adapts the motion, and how you do it decides whether the robot's hand ends up where the human's did.

Method Robot arm length vs human Robot elbow limit

Two ways to retarget

Joint copying sends the human's joint angles straight to the robot, $\mathbf{q}_r = \mathbf{q}_h$, clipped to the robot's limits. It keeps the style of the motion but puts the hand in the wrong place when the limbs differ in length. Position matching solves inverse kinematics so the robot's hand follows the human's hand, scaled to the robot's size:

$$\mathbf{q}_r = \arg\min_{\mathbf{q}}\ \big\lVert f_r(\mathbf{q}) - s\,f_h(\mathbf{q}_h)\big\rVert^2 \quad \text{within the robot's joint limits}$$

What to notice

This robot's upper arm is relatively longer and its forearm shorter than the human's. Copying angles onto those proportions distorts the hand's path, and where the elbow limit is reached, the path flattens. Matching positions keeps the path's shape, but where the human hand goes out of the robot's reach, the robot falls short and may twist into odd poses. Real retargeting, like the methods used to train humanoid whole-body policies, mixes both and adds contact and balance constraints so feet and hands touch the world in the right places.