Body Schema Policy: Body as the Vision-Action Interface for Efficient Robot Loco-Manipulation
- arXiv
- Ask alphaXiv
- Project page
- Humanoids and mobile manipulation
- Imitation learning and data collection
TL;DR Body Schema Policy is a small diffusion policy that denoises actions on tokens placed along the robot's kinematic tree, each picking out its own image patches. On nine simulated Unitree G1 whole-body tasks it matches the 2.63B-parameter Ψ0 model (221 vs 220 of 270 episodes) with 75x fewer parameters, and succeeds in 16 of 20 trials on a real G1.
What they do
The strongest whole-body loco-manipulation policies have billions of parameters and still need per-task fine-tuning. Body Schema Policy (BSP) is an imitation-learning policy that instead builds the robot’s body structure into the vision-action interface, after the body schema of human motor control. Tokens on the robot’s kinematic tree pass features between links, each queries the camera image for its own evidence, and each joint’s action is denoised from its own token.
How it works
- Latent body schema. Joint tokens, link tokens and one global token are rebuilt at every denoising step from the noisy action chunk, the step and the proprioceptive state.
- Neural Screw Operator. Features move from parent to child links through learned screw motions, in the same compositional form forward kinematics uses for link poses.
- Body-queried vision. Joint and global tokens cross-attend to DINOv2 patch tokens. The queries depend on the current action hypothesis, so what each part looks at can change during denoising.
- Body-aware denoising. One shared temporal decoder predicts the noise of each action coordinate from that coordinate’s token; off-tree commands such as base velocity read the global token. Training uses Diffusion Policy’s DDPM objective.
Results
- SIMPLE, nine G1 whole-body tasks, 30 episodes each: 221 successes with 35M parameters, versus 220 for Ψ0 (2.63B), 177 for Patch Policy and 111 for Diffusion Policy. It leads every method under changed object and robot poses (71 vs 63 for Ψ0) and returns a chunk in 13.9 ms versus 148.2 ms for Ψ0.
- ManiSkill-HAB, Fetch: 1.6-1.8x the success of the better matched-size baseline on all three subtasks; level with π0.5 on opening the fridge and ahead on the other two.
- Real G1, 15 demos and 10 trials per task: 16/20 successes versus 13/20 for Ψ0 and 3/20 for Diffusion Policy.
- Ablation on three tasks: with a pooled image vector, replacing the body schema with Diffusion Policy’s U-Net drops successes from 60 to 24 of 90.
Limitations
Limited compute kept SIMPLE training and evaluation to few seeds. BSP relies on the kinematic meaning of joint commands; with whole-body controllers commanded in a latent space, such as SONIC and BFM-Zero, the body schema is less effective. The real-robot test covers two tasks with 10 trials each.
Videos
Figures




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.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
