Scroll

Scaling Behavior Foundation Model for Humanoid Robots

Technical Report

Weishuai Zeng1,6,7,* Kangning Yin2,7,* Xiaojie Niu7,* Shunlin Lu7 Weixiang Zhong7
Jiahe Chen3,7 Feiyu Jia7 Xiao Chen7 Zirui Wang3,7 Furui Xu7 Ming Zhou7 Kailin Li7
Weinan Zhang2,7 He Wang4,6 Li Yi5,6 Dahua Lin1,7 Jiangmiao Pang7 Jingbo Wang7

* Equal contribution

1. The Chinese University of Hong Kong 2. Shanghai Jiao Tong University 3. Zhejiang University 4. Peking University
5. Tsinghua University 6. Galbot 7. Shanghai Artificial Intelligence Laboratory
Paper arXiv Video Twitter Code
ScaleBFM overview figure

Abstract

Humanoid control requires natural whole-body coordination, precise real-time responses to control signals, and robust generalization across diverse environmental contexts, making it a cornerstone for generalist embodied agents. Behavior Foundation Models (BFMs) have recently emerged as a promising solution to address these challenges by leveraging large-scale behavioral data to achieve superior expressiveness, versatility and generalization. However, despite growing interest in scaling BFMs to further improve their capabilities, it remains unclear how key factors, including the learning paradigm, behavioral data and model architecture should be coordinated to enable effective scaling. In this work, we revisit the scaling recipe for BFMs and demonstrate that substantial performance gains can be achieved through the coordination of three core components: 1) the learning paradigm of motion tracking that reformulates diverse humanoid control problems as the reproduction of integrated whole-body behaviors in the global frame; 2) the strategic synergy between on-policy rollout quantity and reference motion diversity; and 3) the expressive and scalable model architecture termed the Humanoid Transformer that facilitates the natural emergence of structured behavioral representations. Through extensive experiments in both simulation and real-world deployment, we demonstrate that our approach yields significant improvements in control fidelity and task generalization, reducing Mean Per-Keypoint Position Error (MPKPE) on the test set by over 10% in local mode and 82% in global mode compared with existing humanoid controllers. These results establish BFM as a principled and effective foundation for scalable and general-purpose humanoid control.

Scaling On-policy Data

BONES Test Set
Ours Test Set
Root-and-Hand
Mode
Scaling on-policy data collection on the BONES Test Set under Root-and-Hand Mode
Scaling on-policy data collection on the Ours Test Set under Root-and-Hand Mode
Root-and-End-Effector
Mode
Scaling on-policy data collection on the BONES Test Set under Root-and-End-Effector Mode
Scaling on-policy data collection on the Ours Test Set under Root-and-End-Effector Mode
Root-and-Upper-Body
Mode
Scaling on-policy data collection on the BONES Test Set under Root-and-Upper-Body Mode
Scaling on-policy data collection on the Ours Test Set under Root-and-Upper-Body Mode
Whole-Body
Mode
Scaling on-policy data collection on the BONES Test Set under Whole-Body Mode
Scaling on-policy data collection on the Ours Test Set under Whole-Body Mode

Jointly scaling the width (parallel environments) and depth (rollout horizon) of on-policy data collection consistently improves BFM performance across diverse control modes and test sets, with the largest configuration achieving the strongest overall results. These gains likely arise from larger amounts of on-policy experience, which yield more reliable PPO optimization. In contrast, scaling either dimension alone does not consistently improve performance, suggesting that effective on-policy scaling requires a balanced increase in both width and depth rather than expanding either dimension in isolation.

Scaling Reference Motions

Composition of the large-scale human motion dataset aggregated from multiple sources
We construct a large-scale human motion dataset by aggregating over 102 million frames at 50 FPS from multiple sources, which are subsequently retargeted to the target humanoid.
BONES Test Set
Ours Test Set
Root-and-Hand
Mode
Scaling reference motions on the BONES Test Set under Root-and-Hand Mode
Scaling reference motions on the Ours Test Set under Root-and-Hand Mode
Root-and-End-Effector
Mode
Scaling reference motions on the BONES Test Set under Root-and-End-Effector Mode
Scaling reference motions on the Ours Test Set under Root-and-End-Effector Mode
Root-and-Upper-Body
Mode
Scaling reference motions on the BONES Test Set under Root-and-Upper-Body Mode
Scaling reference motions on the Ours Test Set under Root-and-Upper-Body Mode
Whole-Body
Mode
Scaling reference motions on the BONES Test Set under Whole-Body Mode
Scaling reference motions on the Ours Test Set under Whole-Body Mode

Scaling reference motions exhibits two distinct regimes. Homogeneous scaling, which increases the amount of in-domain data without substantially expanding behavioral coverage, provides only marginal performance improvements. In contrast, heterogeneous scaling, which introduces diverse data sources and significantly expands behavioral coverage, yields substantial gains on out-of-domain benchmarks while offering little additional benefit on in-domain evaluation. These results suggest that the effectiveness of reference-motion scaling depends not only on data quantity but also on behavioral coverage and its relevance to the target tasks, highlighting diverse behavioral patterns as a strong foundation for BFM pretraining.

Scaling Model

Architecture of the Humanoid Transformer
We introduce an expressive and scalable architecture termed the Humanoid Transformer, which facilitates the natural emergence of structured behavioral representations.
BONES Test Set
Ours Test Set
Root-and-Hand
Mode
Scaling model architecture on the BONES Test Set under Root-and-Hand Mode
Scaling model architecture on the Ours Test Set under Root-and-Hand Mode
Root-and-End-Effector
Mode
Scaling model architecture on the BONES Test Set under Root-and-End-Effector Mode
Scaling model architecture on the Ours Test Set under Root-and-End-Effector Mode
Root-and-Upper-Body
Mode
Scaling model architecture on the BONES Test Set under Root-and-Upper-Body Mode
Scaling model architecture on the Ours Test Set under Root-and-Upper-Body Mode
Whole-Body
Mode
Scaling model architecture on the BONES Test Set under Whole-Body Mode
Scaling model architecture on the Ours Test Set under Whole-Body Mode

The proposed Humanoid Transformer consistently outperforms the conventional MLP across diverse control modes, with a medium-sized Transformer already matching or exceeding the performance of a substantially larger MLP. These results demonstrate that a more expressive architecture provides a stronger backbone for BFM pretraining by utilizing model capacity more effectively. Nevertheless, scaling the Transformer exhibits diminishing returns and mode-dependent saturation, suggesting that performance is increasingly limited by optimization trade-offs across control modes rather than model capacity alone.

Latent Analysis

Relatively Small Model
Relatively Large Model
Root Mode
Bimanual Mode
Root-and-Hand Mode
End-Effector Mode
Root-and-End-Effector Mode
Upper-Body Mode
Root-and-Upper-Body Mode
Whole-Body Mode

The proposed Humanoid Transformer naturally learns structured latent representations without auxiliary objectives, exhibiting locality, global organization, and robustness to moderate perturbations. Moreover, increasing model capacity leads to convergence of latent representations across different control modes, suggesting that richer shared behavioral representations emerge as BFMs scale. This convergence may underlie the observed mode-dependent scaling behavior, where improvements in some control modes are accompanied by slight degradation in others due to stronger coupling within the shared latent space.

Conclusion

In this paper, we present a systematic investigation of the scaling behavior of Behavior Foundation Models for humanoid robots by revisiting three fundamental aspects: the learning paradigm, behavioral data, and model architecture. Through extensive empirical studies, we derive an effective scaling recipe that yields substantial improvements in control fidelity, execution robustness, and generalization capability. Nevertheless, the proposed framework has its limitations. Although we curate a unified control interface with diverse control modes, it remains unclear whether this interface is the most appropriate abstraction and how these modes should be integrated with future high-level policies. Consequently, the design of BFMs should continue to evolve alongside advances in high-level policy learning. Moreover, the current infrastructure for BFM pretraining remains relatively preliminary and constrained, limiting its scalability to broader settings. We hope that this work provides a foundation for systematic development of BFMs, ultimately contributing to the realization of general-purpose humanoid intelligence.

BibTeX

@article{zeng2026scaling,
  title={Scaling Behavior Foundation Model for Humanoid Robots},
  author={Zeng, Weishuai and Yin, Kangning and Niu, Xiaojie and Lu, Shunlin and Zhong, Weixiang and Chen, Jiahe and Jia, Feiyu and Chen, Xiao and Wang, Zirui and Xu, Furui and Zhou, Ming and Li, Kailin and Zhang, Weinan and Wang, He and Yi, Li and Lin, Dahua and Pang, Jiangmiao and Wang, Jingbo},
  journal={arXiv preprint arXiv:2607.15163},
  year={2026}
}

From motion tracking to reusable behavior

BFM pretraining is instantiated as a goal-conditioned reinforcement learning problem, with humanoid whole-body motion tracking used as the proxy task. Motion tracking is not the final control objective; it turns behavior learning into a concrete and measurable control problem while keeping the control interface separate from one fixed form of command.

The tracking recipe requires the robot to reproduce reference motions as integrated whole-body trajectories in the global frame. Removing root-position tracking can give nearly identical guidance to behaviors with distinct global semantics, such as walking forward and marching in place, while decoupling root following from local pose tracking can compromise coordination between root and pose evolution.

Data scaling is not just more clips

Under PPO-based motion tracking, training data scale has two sides: quantity and diversity. Effective training data are the on-policy rollouts collected through environment interaction, whose scale is governed by environment parallelism and rollout horizon. Reference motions shape the rollout distribution by introducing a wider range of behavior patterns for imitation.

ScaleBFM builds its reference corpus from a large-scale human motion collection with 102 million frames at 50 FPS, aggregated from multiple open-source motion datasets and then retargeted to the target humanoid. The dataset figure below shows the sources and broad behavior coverage used for this reference motion diversity.

Reference motion dataset composition chart showing 102 million frames at 50 FPS across multiple motion sources
Reference motion diversity. The dataset figure shows how the 102M-frame motion collection combines multiple sources and behavior styles before retargeting them to the humanoid.

Scaling results

The experiments report success rate and tracking errors under diverse control modes. For on-policy data scaling, the width of interaction increases with the number of parallel GPUs, while the depth increases with the rollout horizon of each environment. Representative modes include the root-and-hand mode, root-and-end-effector mode, root-and-upper-body mode, and whole-body control mode.

The scaling results cover the BONES Test Set and the Ours Test Set under whole-body control mode. The evaluation protocol reports global and local tracking errors over the activated links specified by each control mode.

Scaling on-policy data collection on the BONES test set
BONES test set. Scaling on-policy data collection under whole-body control mode.
Scaling on-policy data collection on the Ours test set
Ours test set. The same scaling study on high-quality stylistic motions.

Behavior videos

The videos below follow the main behavior and control axes: dexterous manipulation, natural and agile locomotion, whole-body coordinated loco-manipulation, real-world tracking, global control mode, local control mode, and the eight curated control modes used to present the same reference behavior through diverse behavior specifications.

Whole-body mode

Whole-body mode is the most fully specified interface in the curated control set. It activates 14 target links, covering the pelvis, shoulders, elbows, hands, torso, hips, knees, and feet, so the target describes nearly the entire humanoid pose rather than only a sparse subset.

Boxing whole body global

Circle walk whole body global

Slow run whole body global

Upper body whole body global

Walk whole body globald

Takeaways

  • ScaleBFM is organized around three ingredients: the learning paradigm, behavioral data, and model architecture.
  • The learning paradigm is motion tracking formulated as reproduction of integrated whole-body trajectories in the global frame.
  • The control interface uses eight curated control modes, allowing the same underlying reference behavior to be presented through diverse behavior specifications.