Skip to content

Scaling Behavior Foundation Model for Humanoid Robots

🕒 Published (v1): 2026-07-16 16:08 UTC · Source: Arxiv · link

Why this paper was selected

Scaling behavior foundation model for humanoid embodied agents; foundational for generalist robots

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper identifies a principled scaling recipe for Behavior Foundation Models (BFMs) for humanoid robots, coordinating three components: a global-frame motion-tracking learning paradigm, a synergy between on-policy rollout quantity and reference motion diversity, and a new Humanoid Transformer architecture. The resulting BFM achieves significant gains in whole-body control fidelity, reducing MPKPE by >10% (local) and 82% (global) versus existing humanoid controllers.

Problem

Existing humanoid controllers are largely task-specific, requiring per-task reward engineering that limits generalization. BFMs have emerged as a unifying approach, but it is unclear how the three key ingredients—learning paradigm, training data, and model architecture—should be coordinated to enable effective scaling. Prior work (e.g., SONIC) scales individual factors in isolation without explaining their coupling.

Method

Learning Paradigm. Humanoid control is formulated as a goal-conditioned MDP where the proxy task is whole-body motion tracking in the global frame (not root-relative), optimized with PPO. Reproducing reference motions as integrated global trajectories eliminates behavioral ambiguity (e.g., walking vs. marching-in-place) and provides coherent full-body reward signals. A masked whole-body target-pose control interface with eight modes (\(M = \{m_0, \ldots, m_7\}\)) allows a single policy to respond to diverse behavioral specifications at inference.

Training Data. Under PPO, effective training data are on-policy rollouts, not reference motion count alone. Scalability requires quantity (number of parallel environments Ă— rollout horizon) and diversity (breadth of the reference corpus). The reference dataset aggregates 102M frames at 50 FPS from eight open-source datasets (LAFAN, AMASS, OMOMO, GRAB, SnapMoGen, FineDance, BONES-SEED, Embody3D), retargeted to the target humanoid via a two-stage skeleton-alignment + frame-wise IK pipeline. An adaptive sampling scheme re-weights difficult trajectories exponentially:

\[w_t = \begin{cases} w_{t-1}/\beta^{T_{\text{eval}}} & \text{if failed} \\ w_{t-1} \cdot \beta^{T_{\text{eval}}} & \text{if succeeded} \end{cases}\]

with \(\beta=0.999\), clipped to \([0.03, 1.0]\).

Model Architecture (Humanoid Transformer). Proprioceptive states, goal states, and actions are extended into temporal windows, encoded by modality-specific tokenizers, and interleaved as a context sequence. A learnable query token aggregates context via self-attention. Goal states are injected via cross-attention. RMSNorm normalizes goal embeddings onto a bounded hypersphere, inducing a structured spherical latent space for behavioral intentions purely from motion-tracking supervision—no auxiliary regularization objectives.

Key Contributions

  • Identification of global-frame motion tracking as the principled and scalable learning paradigm for BFMs, distinguishing it from root-relative or decoupled designs.
  • Formal clarification that BFM data scaling is driven by the synergy of on-policy rollout quantity and reference motion diversity, not raw reference motion count.
  • Humanoid Transformer: a scalable architecture with cross-attention goal conditioning and RMSNorm-induced spherical latent space that yields structured behavioral representations without auxiliary losses.
  • A 102M-frame open-source retargeted motion corpus covering locomotion, dexterous manipulation, and loco-manipulation.
  • Sim-and-real validation: >10% MPKPE reduction in local mode and 82% in global mode over existing humanoid controllers.

Results

  • MPKPE (local mode): >10% reduction vs. existing humanoid controllers on the test set.
  • MPKPE (global mode): 82% reduction vs. existing humanoid controllers on the test set.
  • Validated in both simulation and real-world deployment across dexterous manipulation, locomotion, and whole-body loco-manipulation tasks.
  • (No per-baseline ablation numbers are available in the provided text excerpt.)

Limitations

  • Quantitative per-baseline breakdown (e.g., vs. BFM4Humanoid, SONIC individually) is not reported in the provided text.
  • Relies on PPO; the paper acknowledges alternative RL algorithms exist but does not systematically compare them.
  • Motion retargeting introduces an embodiment gap; quality degrades for motions with large human–humanoid morphological mismatch.
  • Data diversity is bounded by available open-source human motion datasets; highly specialized or contact-rich behaviors may be underrepresented.
  • Real-world experiments are described qualitatively; quantitative real-robot MPKPE or task-success metrics are not detailed in the provided excerpt.

Relevance to Agentic AI / LLM Agents

BFMs serve as the low-level behavioral backbone for generalist embodied agents—analogous to how language models serve as the backbone for LLM-based agents—and the scaling laws studied here directly inform how to build reliable robot "actuator layers" for agentic systems. The paper's masked control interface and spherical latent space design mirror goal-conditioning and representation-learning patterns prominent in language-agent architectures, suggesting shared design principles across modalities. For researchers tracking agentic AI, this work establishes that scaling embodied controllers requires coordinated attention to paradigm, data, and architecture—the same three axes debated in LLM scaling—and provides a concrete recipe. The structured behavioral representations also open a path toward high-level agent planners composing low-level BFM skills, a key open challenge in hierarchical embodied agents.