Scaling Behavior Foundation Model for Humanoid Robots

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 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.

• The paper presents a coordinated scaling recipe combining global-frame motion tracking, behaviorally diverse reference data, and Transformer architecture to improve humanoid whole-body control.
• The proposed 3M-parameter Transformer achieves 0.9677 success on the BONES benchmark and 0.9776 on cross-source tests, with global-frame G-MPKPE values of 0.0798 m and 0.0915 m, respectively.
• The results show that data diversity matters more than raw motion count, while moderate latent-space robustness, 50 Hz onboard inference, and eight masked-pose control modes support practical deployment on the Unitree G1.

This paper presents a systematic study of the scaling behavior of Behavior Foundation Models (BFMs) for humanoid whole-body control, identifying a coordinated recipe across three axes: the learning paradigm, the composition of training data, and the model architecture. The work is grounded in the observation that prior BFM efforts—such as distillation-based approaches like BFM4Humanoid and BeyondMimic, unsupervised RL methods like BFM-Zero, and motion-tracking-based pretraining as in SONIC—have explored scaling only in fragmented ways, without clarifying how these factors interact. The authors' central contribution is to make this coupling explicit and to demonstrate that coordinating it yields large gains in tracking fidelity and cross-domain generalization on the Unitree G1 platform.

Unified learning paradigm: global-frame motion tracking

The paper formulates BFM pretraining as goal-conditioned RL in which behavior is defined as a trajectory of proprioceptive states and actions, excluding goal states, which are treated as external specifications. Motion tracking serves purely as a proxy task for behavior learning rather than as the deployment-time control objective, which is what distinguishes a BFM from a conventional motion tracker. Training uses PPO with asymmetric actor-critic observations, where the critic receives privileged simulator state.

The most consequential design choice is the reward formulation: unlike BeyondMimic or SONIC, which either drop root-position tracking or decouple root following from pose tracking, the proposed model must reproduce reference motions as integrated whole-body trajectories in the global frame. The authors argue that removing root-position tracking makes behaviors with distinct global semantics (e.g., walking forward versus marching in place) nearly indistinguishable in the learning signal, while decoupling root and pose objectives compromises their coordination. This claim is supported by an ablation baseline (BFM-Bym) trained identically but with the BeyondMimic reward: the global-frame reward consistently reduces G-MPKPE relative to the ablation, indicating more coherent behavioral guidance. Because control signals are re-anchorable to the robot's current root state, the same pretrained model supports both global control (with root localization) and local control at deployment.

The control interface consists of masked whole-body target poses in root-relative Cartesian space, sampled from eight curated control modes spanning granularities from root-only to full 14-link whole-body specification. Unspecified links are naturally inpainted, allowing sparse commands such as end-effector targets to induce plausible whole-body behavior.

Data scaling: quantity versus diversity

A key conceptual clarification is that under PPO, the effective training data are the on-policy rollouts, whose quantity is governed by environment parallelism and rollout horizon—not by the number of reference motions, which instead shapes the behavioral distribution. Scaling experiments vary both dimensions across three levels (32/48/64 GPUs; rollout horizons 32/48/64), with 8192 environments per GPU. Jointly scaling width and depth produces consistent improvements, with the largest configuration best in nearly all settings, but scaling either dimension alone does not reliably help—suggesting an unresolved interplay between how much experience is collected and how it is accumulated per update.

For reference-motion scaling, the 102M-frame corpus (aggregated from LAFAN, AMASS, OMOMO, GRAB, SnapMoGen, FineDance, BONES-SEED, and Embody3D) is partitioned into five nested subsets. A K-Means occupancy analysis over a 64-dimensional behavioral feature space reveals two regimes: homogeneous scaling (XXS→S, all within BONES-SEED) leaves cluster occupancy essentially flat (~0.935–0.937), whereas heterogeneous scaling (S→L, adding OMOMO, GRAB, dance data, Embody3D) raises occupancy from 0.9365 to 0.9995. The empirical results align sharply with this analysis: homogeneous scaling yields only marginal gains even on the source-aligned BONES test set, while heterogeneous scaling produces little benefit on BONES but substantial gains on the cross-source test set (Xsens captures plus 100Style). The implication is direct: gains from more reference motions materialize only when scaling measurably expands behavioral coverage relevant to the target distribution—an important corrective to the common practice of equating dataset size with capability. The authors acknowledge that this coverage metric depends on the clustering support being defined by the full corpus, an assumption they flag as inherently untestable against truly long-tail behaviors.

Supporting infrastructure includes a two-stage retargeting pipeline (skeleton alignment via SMPL shape or BVH offset optimization, then sequential frame-wise IK), termination at 0.5 m global deviation, reference state initialization at the termination point, and clamped adaptive sampling ($\beta=0.999$, weights clipped to $[0.03, 1.0]$).

Architecture: the Humanoid Transformer

The proposed Humanoid Transformer tokenizes temporal windows of proprioception and actions into an interleaved context sequence processed by self-attention with RoPE, injects future goal tokens through cross-attention, and predicts actions or values from a learnable query token that attends over—but is not attended to by—the context. RMSNorm projects goal embeddings onto a hypersphere, inducing a structured latent space shaped solely by the tracking objective, without auxiliary regularization losses. The actor uses five consecutive future frames plus one stochastically sampled offset in $[5,32]$ (dynamically adjustable at deployment to absorb latency); the critic uses fixed exponentially spaced offsets $\{0,1,2,4,8,16,32\}$.

Scaling experiments compare MLP backbones (3.05M and 11.86M parameters) against Transformer variants from 0.41M to 9.91M parameters. The medium Transformer (3M parameters) matches or exceeds the substantially larger MLP, and further capacity growth yields diminishing returns—evidence that architectural expressiveness, not raw parameter count, drives effective capacity utilization here. However, scaling does not uniformly improve all control modes; some saturate early, which the authors attribute to inter-mode learning-difficulty imbalance and optimization trade-offs within the shared latent space.

Latent-space structure

Qualitative visualization shows that latent trajectories for individual motions are locally smooth on the unit hypersphere, and globally organized: opposing intentions (forward/backward walking, left/right crouching) occupy separated regions preserving directional relationships. Robustness is quantified by rotating latent vectors along random directions: success rate degrades only modestly up to 20° perturbations (0.9717→0.9646 on BONES; 0.9836→0.9400 on Ours), supporting the claim that behavioral interpretation tolerates moderate latent noise. Notably, increasing model capacity drives convergence of latent representations across different control modes, offering a mechanistic explanation for the observed trade-offs among modes during architecture scaling—that improved alignment in one mode can come at slight cost to others sharing the latent space.

Benchmark results

Against off-the-shelf controllers GMT, TWIST, and SONIC, plus the BFM-Bym ablation, the 3M-parameter model achieves strong margins. On the source-aligned BONES test set it reaches a 0.9677 success rate with G-MPKPE of 0.0798 m, versus 0.9239 for SONIC (whose training set may overlap this benchmark, a caveat the authors note) and below 0.45 for GMT and TWIST. Generalization gaps are larger on the cross-source test set: GMT and TWIST collapse to success rates near 0.08, SONIC reaches 0.5937, while the proposed model attains 0.9776 with G-MPKPE of 0.0915 m. Relative to existing humanoid controllers overall, the paper reports MPKPE reductions exceeding 10% in local mode and 82% in global mode. Real-world deployment runs onboard inference at 50 Hz via TensorRT atop a 200 Hz PD loop, with timestamp-adjusted future-frame indexing compensating communication latency, and supports online switching among all eight control modes.

Limitations and open questions

The authors are explicit about several constraints. First, whether the eight-mode masked-pose interface is the right abstraction—and how it should integrate with future high-level policies—remains open. Second, the scaling study is limited relative to LLM-scale investigations: distributed training infrastructure for humanoid pretraining is fragile, and onboard compute caps model size if 50 Hz inference must be preserved alongside high-level policy headroom. Third, the behavioral-coverage metric presupposes that the full corpus provides adequate clustering support, an assumption that cannot be verified against unobserved long-tail behaviors. Fourth, the mechanism behind width–depth interactions in on-policy data collection, and the mode-coupling dynamics in the shared latent space, are described but not fully characterized.

Conclusion

This paper reframes BFM scaling as a coordination problem among learning paradigm, data quantity–diversity synergy, and architecture, rather than a matter of enlarging any single factor. Its principal empirical findings—that global-frame integrated tracking yields more coherent guidance than decoupled rewards, that reference-motion scaling pays off only when it expands measured behavioral coverage on relevant distributions, and that a moderately sized Transformer outperforms larger MLPs while inducing structured, robust latents without auxiliary objectives—together constitute a concrete, reproducible recipe for general-purpose humanoid foundation controllers.