Body Schema Policy: Learning a Body-Aware Vision-Action Interface for Efficient Humanoid Loco-Manipulation

Inspired by the human body schema, we study robot morphology as a shared interface between visual perception and action generation. We introduce Body Schema Policy (BSP), a diffusion policy that organizes vision and action around the robot's body. Given parent-child connectivity, BSP learns body representations that jointly guide visual processing and action generation. Joint and link features retrieve spatial information from a pretrained visual encoder, connecting visual evidence to the movements being generated. On diverse simulation and real-world humanoid whole body loco manipulation tasks, BSP uses only 35M parameters and outperforms Diffusion Policy with 60M parameters on every task and achieves comparable aggregate success to the ${\Psi }_0$ foundation model with $75\times$ fewer parameters. These results support body schema as a practical architectural principle for learning compact and effective whole-body visuomotor policies.