Scaling Behavior Foundation Model for Humanoid Robots

TL;DR

ScaleBFM is a systematic scaling study for Behavior Foundation Models (BFMs) that control humanoid robots, casting the pretraining problem as PPO-based motion tracking over a large corpus of retargeted human reference motions. The authors identify three coordinated levers — the learning paradigm (whole-body trajectory tracking in the global frame, not decoupled root + local pose), behavioral data (jointly scaling on-policy rollout width × depth and heterogeneous reference-motion diversity), and model architecture (a “Humanoid Transformer” that outperforms MLP policies at matched compute) — and show that only their joint scaling improves control fidelity and generalization. Headline result: >10% MPKPE reduction in local mode and 82% in global mode vs prior humanoid controllers, validated in sim and on real hardware across eight curated control-interface modes (root-only, bimanual, root+hand, end-effector, root+end-effector, whole-body, plus local/global framings).

Key claims

Reformulating humanoid control as motion tracking of integrated whole-body trajectories in the global frame — rather than decoupling root position from local pose — is load-bearing: removing root-position tracking makes semantically distinct behaviors (walking forward vs marching in place) look identical to the policy, and decoupling root from pose compromises coordination [§Method].
On-policy data scaling requires jointly scaling width (parallel environments) and depth (rollout horizon); scaling either dimension alone does not consistently improve performance across control modes [§Scaling On-policy Data].
Reference-motion scaling exhibits two distinct regimes: homogeneous scaling (more in-domain data, similar behavior coverage) yields only marginal gains, while heterogeneous scaling (diverse data sources expanding behavior coverage) yields substantial OOD gains with little ID benefit [§Scaling Reference Motions].
A Transformer policy (Humanoid Transformer) consistently outperforms a conventional MLP across control modes at matched or lower parameter counts; a medium Transformer matches or exceeds a substantially larger MLP [§Scaling Model].
Transformer scaling exhibits diminishing returns and mode-dependent saturation — performance ceilings become limited by optimization trade-offs across control modes rather than model capacity alone [§Scaling Model].
The Humanoid Transformer learns structured latent representations without auxiliary objectives, with locality, global organization, and robustness to moderate perturbations; latent representations across different control modes converge as model capacity grows [§Latent Analysis].
The unified control interface exposes eight curated control modes (root-only, bimanual, root+hand, end-effector, root+end-effector, whole-body, in local and global framings) so the same underlying reference behavior can be re-presented under diverse specifications at inference time [§Method, §Full Video Set].
Headline numbers: >10% MPKPE reduction on the test set in local mode and 82% in global mode vs prior humanoid controllers, with real-world deployment demonstrated across dexterous manipulation, agile locomotion, and whole-body coordinated loco-manipulation [§Abstract, §Conclusion].
Reference-motion corpus: 102M frames at 50 FPS aggregated from multiple open-source human motion datasets and retargeted to the target humanoid — the substrate that heterogeneous-diversity scaling operates over [§Data Scaling].

Method

BFM pretraining is instantiated as a goal-conditioned reinforcement learning problem, with humanoid whole-body motion tracking used as the proxy task. The policy is trained with PPO to reproduce reference motions as integrated whole-body trajectories in the global frame — the “global-frame + joint-tracking” formulation is deliberate: removing root-position tracking creates ambiguity between behaviors with distinct global semantics (walking forward vs marching in place), and decoupling root following from local pose tracking compromises root-pose coordination.

Reference motions are aggregated from multiple open-source human motion datasets (102M frames at 50 FPS) and retargeted onto the target humanoid. On-policy rollouts are scaled along two independent dimensions: environment parallelism (width, tied to number of parallel GPUs) and rollout horizon (depth). The model side compares MLPs against a “Humanoid Transformer” backbone; scaling the Transformer’s parameter count is studied alongside the data axes.

At inference time, a unified control interface exposes eight curated modes (root-only, bimanual, root+hand, end-effector, root+end-effector, whole-body, each in local and global framings) so the same trained BFM can be conditioned on very sparse control signals (pelvis only) or dense whole-body specifications (14 activated links).

Results

Prior-work comparison: >10% MPKPE reduction on the test set in local mode; 82% reduction in global mode [§Abstract, §Conclusion].
On-policy data scaling: consistent improvements across all four evaluated control modes only when width × depth scale jointly; single-axis scaling does not consistently improve [§Scaling On-policy Data].
Reference-motion scaling: heterogeneous scaling drives OOD gains; homogeneous scaling drives only marginal gains, with little OOD benefit [§Scaling Reference Motions].
Architecture scaling: medium-size Transformer matches or exceeds a substantially larger MLP across all four modes; further Transformer scaling saturates in a mode-dependent way [§Scaling Model].
Real-world validation: demonstrated across styled walking, highly dynamic dances, fighting/sports, frequent-turn walking, dexterous manipulation, and whole-body coordinated loco-manipulation on real humanoid hardware, evaluated on BONES and Ours test sets [§Full Video Set].
Latent structure: the Humanoid Transformer exhibits emergent locality, global organization, and cross-mode representation convergence at larger capacities, without any auxiliary representation-learning objective [§Latent Analysis].

Why it’s interesting

ScaleBFM is the first filed instance where a scaling study — in the Kaplan/Chinchilla mold — is run explicitly on a humanoid whole-body control model rather than on a VLA action head. Two things sharpen its position on the wiki. First, the paper makes the data-diversity vs. data-volume distinction quantitative for RL-pretrained BFMs — heterogeneous > homogeneous is a claim that echoes EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data‘s “scale + diversity is the lever” on the manipulation side, and Sunday Robotics ran the Chinchilla scaling law with their VLA model (Tony Zhao follow-up)‘s (still unreleased) Chinchilla claim on the Sunday Robotics VLA, but on the whole-body-control problem where the prior consensus was that a well-tuned single-motion tracker was the ceiling. Second, the paper directly addresses the retargeting-quality bottleneck that Human-to-Robot Retargeting lists as an open question — its 102M-frame corpus is a retargeted-motion substrate, and the finding that heterogeneous motion diversity drives OOD generalization is evidence that retargeting quality is a lever proportional to source-corpus diversity, not proportional to per-clip fidelity.

Sits adjacent to ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting (physics-in-outer-loop retargeting + inner-loop RL tracker) on the same “motion-tracking-as-BFM-substrate” axis, but focuses on scaling laws rather than retargeting design; and complements the recipe-diverse VLA Models page by staking out the whole-body-control counterpart to that concept’s manipulation-recipe debates.