Humanoid Self-Representation
Learning which body is the self—and how that body occupies 3D space—from sensorimotor experience.
Humanoid robots that share physical spaces with people first need to determine which visible body is their own. We learn this distinction from proprioceptive–visual correspondence alone, without identity labels, CAD geometry, URDF, or a predefined kinematic model.
The learned self–other distinction then supplies supervision for a predictive 3D self-model that maps joint configurations to body occupancy. The model supports target reaching, collision-aware motion planning, and human-to-robot motion retargeting on a 29-DoF humanoid robot.
This work treats bodily self-representation as a capability that can emerge from experience, rather than a geometric model that must be engineered in advance.
Project page · arXiv · PDF
