Biomimetic Gaze–Force Shared Control for Collaborative Manipulation: An Upper-Limb Interface for Humanoid Robots

Biomimetic humanoid control can emulate not only human morphology and motion but also the sensorimotor organization that couples visual anticipation with physical correction during joint action. Inspired by this principle, we propose a gaze–force shared-control framework for collaborative manipulation under uncertain and changing intent. Gaze provides early evidence of the attended goal, while interaction force conveys local corrections through contact. The visual model includes an explicit non-target state, preventing looks toward the robot, motion path, or background from being misclassified as task goals. A probabilistic estimator fuses gaze and force evidence and weights goal-conditioned controllers derived from locally frozen linear-quadratic games. Their velocity outputs are blended with an admittance response: reliable and consistent cues increase assistance, whereas gaze–force conflict reduces it and returns authority to the human. Under bounded inputs, dissipative gains, and damped-Jacobian mapping, the velocity-generation layer remains bounded. Collaborative placement experiments demonstrate reliable rejection of non-target gaze, reduced human effort when cues agree, smooth adaptation to target changes, and rapid suppression of assistance during conflict. Although validated on a collaborative manipulator, the task-space framework provides a biomimetic upper-limb sensorimotor interface for future humanoid whole-body control.