Body Schema Policy: Body as the Vision-Action Interface for Efficient Robot Loco-Manipulation

Lechen Zhang, Chenyuan Chen, Muxin Zhou, Zewen He, Yoshihiko NakamuraarXiv 2026 · arXiv

TL;DR

Body Schema Policy (BSP) is a novel imitation learning approach that organizes a robot's perception and action around its physical structure, much like the human "body schema." By embedding the robot's kinematic tree directly into the neural architecture, BSP achieves the performance of massive foundation models with 75x fewer parameters and significantly faster inference speeds.

What Problem Does This Paper Address?

Current state-of-the-art robot policies for whole-body loco-manipulation (tasks requiring simultaneous movement and precise handling) typically rely on "brute-force" scaling. These models, often called Vision-Language-Action (VLA) models, boast billions of parameters and require extensive pre-training and task-specific fine-tuning.

The authors argue that these generic architectures ignore a fundamental truth: a robot has a body. Standard policies treat actions as flat vectors of numbers, forcing the network to "re-learn" the robot's own geometry and kinematics from scratch for every task. In contrast, humans use a "body schema"—an internal representation of our body's structure and posture—to efficiently coordinate movement and focus our attention. BSP aims to bring this biological efficiency to robotics by building the body's structure directly into the vision-action interface.

Methods

BSP introduces three core components that move away from "flat" neural processing:

  1. Latent Body Schema: The policy represents the robot as a graph (a kinematic tree). Each joint and link in the robot's body is assigned its own "token" (a vector of features).
  2. Neural Screw Operator: To move information between these tokens, the authors use "screw motions"—a mathematical framework for describing rigid body movement. This operator propagates features along the kinematic tree, effectively teaching the network how the motion of one joint (like the shoulder) affects the position of another (like the hand).
  3. Body-Queried Vision: Instead of looking at the whole image at once, each body part "queries" the visual scene independently. For example, the hand token might focus on the object it needs to grasp, while the base token looks at the floor to navigate.
The BSP architecture organizes visual queries and action generation around the robot's latent body schema.

The entire system acts as a denoiser (similar to the technology behind stable diffusion), starting with a noisy guess of the required action and iteratively refining it based on visual evidence retrieved through the body schema.

Datasets

The authors evaluated BSP across three diverse settings:

  • SIMPLE Benchmark: Nine simulated tasks using the Unitree G1 humanoid, ranging from tabletop grasping to bimanual handovers and pushing chairs.
  • ManiSkill-HAB: A benchmark for mobile manipulation using a Fetch robot (a wheeled base with a single arm).
  • Real-world G1 Humanoid: Testing the policy on physical hardware for complex tasks like pushing a trolley and throwing items into a trash bin.

Key Insights and Ideas

  • Inductive Bias Beats Scale: By telling the model how a robot's body is connected, it doesn't need billions of parameters to "figure it out." This allows a 35-million-parameter model to compete with a 2.6-billion-parameter giant.
  • Emergent "Focal" vs. "Ambient" Vision: Without any explicit spatial supervision, the model naturally learned to use its body parts differently. The hands learned to focus on small, task-relevant details (focal vision), while the base learned to attend to the broader scene for navigation (ambient vision).
  • Kinematic Generalization: Standard MLPs and Transformers struggle to predict robot poses outside of their training data. Because BSP uses screw operators (which mirror real physics), it generalizes far better to unusual or extreme body postures.
Figure 1. From the human body schema to a robot’s latent body schema. Left: the human body schema links perception with the body and its movement. Middle: Body Schema Policy (BSP) organizes visual queries and action generation around features on the robot’s latent body schema. Right: success rate versus action chunking latency on simulated G1 whole-body tasks. IL: imitation learning; VLA: vision–language–action model; WAM: world-action model.

Key Results

  • Efficiency: On simulated G1 tasks, the 35M-parameter BSP matched the performance of the $\Psi_0$ foundation model (2.6B parameters) while being 75x smaller.
  • Speed: BSP is incredibly fast, returning action chunks in just 13.9 ms on an RTX 6000 Ada GPU—faster than any other policy compared in the study.
  • Performance: It nearly doubled the success rate of the popular "Diffusion Policy" on complex whole-body tasks.
  • Real-World Success: On a real Unitree G1, BSP outperformed both Diffusion Policy and the $\Psi_0$ foundation model, successfully completing multi-stage tasks like opening a trash bin with a foot while holding an object.

Limitations and Future Work

The authors note that BSP currently relies on the kinematic meaning of joint commands. If a robot is controlled via a "latent" space (where commands don't correspond to specific physical joints), the body schema is less effective.

Future directions include:

  • Cross-embodiment policies: Training a single BSP that can control different types of robots by simply swapping the underlying body graph.
  • Integration with VLAs: Using the efficient BSP architecture as a "head" for even larger vision-language models.
  • Skill Libraries: Creating compact libraries of "atomic" skills (like "grasp" or "walk") that can be composed for complex agentic behavior.

Summary updated

Abstract

Currently, robots can learn whole-body loco-manipulation through scale: the strongest policies have billions of parameters and still need task-specific fine-tuning. Humans have limited processing capacity, yet learn and act efficiently through a body schema that organizes what they see and how they move. We introduce Body Schema Policy (BSP), an imitation learning policy that builds such a schema into the vision-action interface. Its latent body schema places joint and link tokens on the robot's kinematic tree, where learned screw motions carry features between links. Through body-queried vision, each body part retrieves its own visual evidence, which conditions the denoising of its actions. With this inductive bias, BSP returns an action chunk faster than any policy we compare. Compact, body-specific visual readouts also emerge without spatial supervision: manipulation commands depend strongly on task objects, while locomotion commands respond more broadly to the scene. On nine simulated Unitree G1 whole-body tasks, BSP matches the $Ψ_0$ humanoid foundation model with $75\times$ fewer parameters and nearly doubles Diffusion Policy's success. The same holds on a real G1 and a Fetch mobile manipulator.
BibTeX
@misc{zhang2026body,
  author = {Lechen Zhang and Chenyuan Chen and Muxin Zhou and Zewen He and Yoshihiko Nakamura},
  title = {{Body Schema Policy: Body as the Vision-Action Interface for Efficient Robot Loco-Manipulation}},
  year = {2026},
  eprint = {2611.22713},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22713},
}
Figure