Towards Embodiment Scaling Laws in Robot Locomotion

Bo Ai^1*† Liu Dai^1* Nico Bohlinger^4* Dichen Li^1*

Tongzhou Mu¹ Zhanxin Wu³ K. Fay¹ Henrik I. Christensen¹ Jan Peters^4,5 Hao Su^1,2

(* equal contribution, † corresponding author)

¹University of California San Diego, USA ²Hillbot Inc, USA

³Cornell University, USA

⁴Technical University of Darmstadt, Germany

⁵German Research Center for AI (DFKI); Robotics Institute Germany; hessian.AI, Germany

One Model, Two Worlds, Many Embodiments

TLDR: We train a single locomotion policy on ~1,000 procedurally generated robots and uncover embodiment scaling laws that enable generalization to diverse embodiments in simulation and the real world.

Overview

This work investigates embodiment scaling laws in robotics, hypothesizing that training a single control policy on a larger number of diverse robot embodiments improves its ability to generalize to unseen ones.

Generating ~1000 Robots

To study the effects of embodiment scaling, we procedurally generate GENBOT-1K dataset consisting if approximately 1,000 varied robot embodiments, including humanoids, quadrupeds, and hexapods, with different geometry, topology, and kinematics.

Humanoid

Quadruped

Hexapod

Cross-Embodiment Learning

We train policies using a single model architecture capable of handling diverse observation and action spaces on different random subsets of embodiments to uncover embodiment scaling laws.

Embodiment Scaling Laws

Training generalist locomotion policies on subsets of GENBOT-1K shows that generalization to unseen robots improves steadily as the number of training embodiments increases.

Key Observations
More training embodiments → better generalization to unseen embodiments (C1–C4)
Harder embodiments require more embodiments to saturate generalization (C1 vs C2–C3)
Cross-morphology training improves generalization (C4 vs C5–C7)
Embodiment scaling >> pure data scaling for embodiment generalization (C4 vs C8)

Real-world Deployment

We confirm our results, by zero-shot transferring the learned unified policy to the Unitree Go2 quadruped and the Unitree H1 humanoid robots in the real world. All results shown are using one single policy.

Forward

Backward

Sideward

Pushes

100% joint limits on every knee

60% joint limits on right rear knee

20% joint limits on right rear knee

60% joint limits on front left knee

40% joint limits on front left knee

20% joint limits on front left knee

Gravel

Grass

Pavement

Cobblestone

Acknowledgments

This research is funded by the NSF AI-Center TILOS, the Hillbot Embodied AI Fund, the National Science Centre Poland (Weave programme UMO-2021/43/I/ST6/02711), and by the German Science Foundation (DFG) (grant number PE 2315/17-1).

Co-author Hao Su is the CTO for Hillbot and receives income. The terms of this arrangement have been reviewed and approved by the University of California, San Diego, in accordance with its conflict of interest policies.

We thank the German Research Center for AI (DFKI), Research Department: Systems AI for Robot Learning, for lending the Unitree Go2 and Unitree H1 robots.

Finally, we thank Oleg Kaidanov (DFKI, TU Darmstadt) for his continuous help with the real-world robot deployment.

Citation

@misc{ai2025embodimentscalinglaws,
    title={Towards Embodiment Scaling Laws in Robot Locomotion},
    author={Bo Ai and Liu Dai and Nico Bohlinger and Dichen Li and Tongzhou Mu and Zhanxin Wu and K. Fay and Henrik I. Christensen and Jan Peters and Hao Su},
    year={2025},
    eprint={2505.05753},
    archivePrefix={arXiv},
    primaryClass={cs.RO},
    url={https://arxiv.org/abs/2505.05753}, 
}

This website was inspired by Kevin Zakka's and Brent Yi's and builds on Nico Bohlinger's and Bo Ai's.