A research team spanning The Chinese University of Hong Kong, Shanghai Jiao Tong University, Zhejiang University, Peking University, Tsinghua University, Galbot, and the Shanghai AI Lab has published a landmark study on scaling Behavior Foundation Models (BFMs) for humanoid robots. The paper, "Scaling Behavior Foundation Model for Humanoid Robots" (arXiv:2607.15163), identifies a three-component recipe that yields substantial improvements in control fidelity and task generalization.
The Three-Component Scaling Recipe
The team demonstrates that BFM scaling requires coordinated tuning of three pillars, rather than simply throwing more parameters at the problem:
- Motion Tracking Paradigm — Reformulating diverse humanoid control problems as the reproduction of integrated whole-body behaviors in the global coordinate frame, rather than decoupled root and pose tracking.
- Data-Diversity Co-Scaling — The strategic synergy between on-policy rollout quantity (governed by environment parallelism and rollout horizon) and reference motion diversity (the breadth of behavior patterns used for imitation).
- Humanoid Transformer Architecture — An expressive, scalable model architecture that naturally emerges structured behavioral representations.
Results: 82% Reduction in Global Tracking Error
The scaled BFM reduces Mean Per-Keypoint Position Error (MPKPE) on the test set by over 10% in local mode and a remarkable 82% in global mode compared with existing humanoid controllers. Global mode is the harder problem — it requires the robot to track both root displacement and full-body posture as an integrated whole.
Key findings from the scaling experiments:
- On-policy data scaling requires balanced increases in both width (parallel environments) and depth (rollout horizon). Scaling either dimension alone does not consistently improve performance.
- Reference motion diversity follows two regimes: homogeneous scaling (more in-domain data) yields only marginal gains, while heterogeneous scaling (introducing diverse data sources) substantially improves out-of-domain benchmarks.
- Humanoid Transformer consistently outperforms conventional MLP baselines, with a medium-sized Transformer already matching or exceeding a substantially larger MLP.
102 Million Frames of Motion Data
The team constructed a large-scale human motion dataset by aggregating over 102 million frames at 50 FPS from multiple sources, subsequently retargeted to the target humanoid platform. This is among the largest motion datasets used for humanoid BFM pretraining to date.
The paper covers an impressive range of demonstrated behaviors across four categories: dexterous manipulation (flower arrangement, fruit collection, plugging chargers), natural and agile locomotion (boxing, cartwheels, tai chi, running), whole-body coordinated loco-manipulation (laundry, cooking, badminton, making beds), and versatile control under sparse constraints.
Latent Space Analysis
Perhaps most interesting from a scientific perspective is the latent analysis. The Humanoid Transformer naturally learns structured latent representations without auxiliary objectives — exhibiting locality, global organization, and robustness to perturbations. 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 explain the mode-dependent scaling behavior: improvements in some control modes are accompanied by slight degradation in others due to stronger coupling within the shared latent space.
Implications for the Field
The study establishes BFM as a principled and effective foundation for scalable, general-purpose humanoid control. By demonstrating clear scaling laws with empirical measurements, the research provides a roadmap for future work — much in the way scaling laws guided progress in large language models.
However, the authors note limitations: the interface design for integrating BFMs with high-level policies remains an open question, and current pretraining infrastructure remains preliminary. The team hopes this work provides a foundation for the systematic development of BFMs toward general-purpose humanoid intelligence.
Paper: arXiv:2607.15163
Project Page: scalebfm.github.io
Code: GitHub
