Pith. sign in

REVIEW 4 major objections 4 minor 39 references

This paper claims humanoid behavior models scale only when three coordinated choices—global-frame motion tracking, quantity-plus-diversity data, and a transformer with a spherical latent space—are adopted together, cutting global tracking e

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 23:56 UTC pith:QY7INM2C

load-bearing objection Promising BFM scaling recipe, but the data-quantity experiment conflates more data with more optimizer updates; still worth serious refereeing. the 4 major comments →

arxiv 2607.15163 v1 pith:QY7INM2C submitted 2026-07-16 cs.RO cs.AI

Scaling Behavior Foundation Model for Humanoid Robots

Weishuai Zeng , Kangning Yin , Xiaojie Niu , Shunlin Lu , Weixiang Zhong , Jiahe Chen , Feiyu Jia , Xiao Chen
show 10 more authors
Zirui Wang Furui Xu Ming Zhou Kailin Li Weinan Zhang He Wang Li Yi Dahua Lin Jiangmiao Pang Jingbo Wang
This is my paper
classification cs.RO cs.AI
keywords behavior foundation modelhumanoid controlwhole-body motion trackingreinforcement learningdata scalingtransformerlatent spacegeneralization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper is trying to establish a scaling recipe for behavior foundation models—pretrained controllers that turn motion specifications into humanoid actions. Its central claim is that scaling does not come from simply adding more reference motions or more parameters; it requires the learning task, the data distribution, and the architecture to change together. The task should be to reproduce integrated whole-body trajectories in the global frame, so root translation and pose are guided as one behavior. The data should be scaled on two separate axes: on-policy rollout volume for quantity, and diverse motion sources for coverage. With this recipe, the model reduces mean per-keypoint position error by over 10% in local mode and 82% in global mode compared with existing humanoid controllers.

Core claim

The discovery defended in the paper is that a behavior foundation model for humanoid robots scales when (1) motion tracking is posed as the reproduction of whole-body trajectories in the global frame rather than decoupled root and pose imitation; (2) training data is understood as the policy's own on-policy rollouts, so quantity comes from more parallel environments and longer horizons while diversity comes from expanding the reference-motion corpus; and (3) the policy network is a transformer whose goal embeddings live on a hypersphere, allowing structured behavior representations to emerge from the tracking objective alone. The empirical centerpiece is that a 3M-parameter version of this m

What carries the argument

The argument is carried by three mechanisms that work as one design. The first is the global-frame motion-tracking reward, which penalizes root position, body position and rotation, and velocities together, so that a root error is not masked by local re-anchoring. The second is the data recipe: effective training data are the on-policy rollouts, with quantity set by environment parallelism and rollout horizon, and diversity set by a curated corpus of 102 million retargeted motion frames; an occupancy-rate proxy (k-means over 64-dimensional behavioral features) is used to separate the two effects. The third is the Humanoid Transformer, which interleaves proprioception and action tokens, injec

Load-bearing premise

The scaling conclusions stand only if the occupancy-rate proxy (k-means clusters over 64-dimensional behavioral features, fit on the largest partition) is a faithful measure of behavioral coverage, and if increasing GPUs and rollout horizon increases pure data quantity rather than also changing the number of policy-update steps; the paper itself flags both caveats.

What would settle it

Train two models with identical settings except that the number of PPO updates per rollout batch is held constant while the total rollout quantity varies; if the gains disappear, the 'quantity' effect is really an optimization-steps effect. Alternatively, compare two corpora with the same occupancy rate but very different motion counts: if performance differs, the occupancy-rate measure is not capturing the diversity that matters.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Holding the architecture fixed, adding more reference motions of already-covered behaviors yields diminishing returns; the measurable gains come from adding sources that expand behavioral coverage, especially when the test set comes from new data sources.
  • A medium-size transformer can outperform a substantially larger MLP under the same training budget, so architecture choice, not parameter count alone, determines how well behavior-model capacity scales in this setting.
  • Because the same policy serves global control (with root localization) and local control (by re-anchoring the control signal to the robot's current root), one pretrained model can support multiple deployment modes without retraining.
  • Scaling on-policy data by width (more parallel environments) and depth (longer horizon) jointly is what helps; increasing either on its own does not produce consistent improvements.
  • The latent space is robust to moderate perturbations, meaning a high-level planner could operate on the behavior manifold rather than on raw joint targets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the paper's separation of quantity and diversity is correct, a practical consequence the authors leave implicit is that data collection effort for new humanoid skills should focus on sourcing novel behavior types, not on amassing more examples of known ones.
  • The occupancy-rate measure is fit to the largest training partition, so a natural extension would be to validate diversity against an independent, held-out behavior taxonomy; if coverage estimates shift under that change, the scaling conclusion is not yet a law.
  • The observed convergence of latent representations across control modes at larger scale hints that the shared manifold could be reused for few-shot skill acquisition, but it also warns that optimizing one control mode may trade off another in the same latent space.
  • A direct extension to test the global-frame claim: ablate only the root-position tracking reward on motions that differ mainly by root translation (walking forward versus marching in place) and see whether the global formulation is what resolves the ambiguity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes and evaluates a three-part scaling recipe for Behavior Foundation Models (BFMs) for humanoid control: (1) a motion-tracking learning paradigm that reproduces integrated whole-body reference trajectories in the global frame, (2) a data strategy that jointly scales on-policy rollout quantity and reference-motion diversity, and (3) the Humanoid Transformer architecture with a hyperspherical latent space. The authors train on a 102M-frame corpus, evaluate on simulation and real hardware, and report large reductions in Mean Per-Keypoint Position Error compared to existing humanoid controllers. The central claim is that coordinated scaling of these three ingredients yields a scalable general-purpose humanoid control foundation model.

Significance. If the claims are upheld, the paper would provide a valuable empirical demonstration of how to scale humanoid behavior foundation models, with useful design principles for the learning objective, data curation, and architecture. Strengths include the large-scale dataset, multiple control modes, evaluation against external baselines (GMT, TWIST, SONIC), real-world deployment, and candid statements of limitations. The main load-bearing concern is that the on-policy data scaling experiment confounds data quantity with the number of PPO update steps; this needs to be addressed before the quantity-diversity synergy claim is fully supported.

major comments (4)
  1. [§4.2.1, App. B.4] The on-policy data scaling experiment confounds data quantity with the number of PPO gradient updates. App. B.4 states that "the number of mini-batches is adjusted according to the rollout horizon so that the mini-batch size remains constant." With learning epochs fixed, the number of gradient steps per epoch is proportional to the number of mini-batches, and hence to the total number of collected frames. The text in §4.2.1 attributes gains to "more reliable estimates of quantities involved in PPO optimization," but each gradient update still sees the same number of transitions; what increases is the number of updates (and, with GPU count, the compute budget). The observed trend therefore cannot be assigned to data quantity per update. A control experiment that holds the number of gradient updates fixed (e.g., by scaling mini-batch size while keeping update count constant) is necessary t
  2. [Abstract / Table 3] The headline numbers in the abstract—"over 10% in local mode and 82% in global mode"—are not tied to a single baseline in Table 3. For L-MPKPE on the BONES test set, the reduction relative to SONIC is (0.0436 − 0.0400)/0.0436 ≈ 8.3%, which is below 10%; relative to GMT/TWIST the global-mode reduction is about 86–87%, not 82%. The claimed percentages should be aligned with explicit baseline entries in Table 3, or the claim should be revised. As written, the headline cannot be verified from the benchmark table.
  3. [§4.2.2, App. B.5] The division of reference-motion scaling into "homogeneous" and "heterogeneous" regimes rests on the occupancy-rate proxy. The K-Means clusters are fit on the L partition, and the paper itself acknowledges (B.5) that the measure "is inherently dependent on the choice of reference motions used to construct the shared clustering space." Since XXS, XS, and S are all subsets of BONES-SEED while M and L add new sources, the occupancy increase from S to L may partly reflect the clustering support rather than an independent measure of behavioral diversity. Because the conclusion that only diversity expansion improves cross-source generalization is load-bearing, the proxy should be validated with alternative features, different cluster counts, or a leave-one-source-out analysis.
  4. [§4.2–§4.3, Table 3] All scaling curves and benchmark numbers appear to come from a single training run per configuration, despite PPO being stochastic. Several decisive comparisons are small (e.g., L-MPKPE values of 0.0400 vs. 0.0396, and BFM-Global vs. BFM-Local on the Ours test set), and some scaling curves are non-monotonic. Without multiple seeds, error bars, or a statistical test, the "clear margin" claims in §4.5 and the scaling conclusions are not quantitatively supported. At least 3–5 seeds should be reported for the main configurations, with confidence intervals for Table 3.
minor comments (4)
  1. [§3.4, Eq. (A.10)] The text describes the latent space as the "unit hypersphere," but RMSNorm(x) = γ · x / sqrt((1/d) Σ x_i²) has norm γ√d unless the scaling factor is explicitly removed. Clarify the exact normalization used for the unit-sphere claim.
  2. [Figures 4–9] In the provided version, several of the scaling figures have very small axis labels and legends, making the values difficult to read. The figures should be enlarged or the key numbers reproduced in tables.
  3. [Table 3 caption] The caption states that BFM-Bym is evaluated under the global control mode, but it is not immediately clear whether BFM-Local and BFM-Global are two different models or the same model evaluated in two modes. Please clarify the evaluation protocol for each row.
  4. [Throughout] A few typos and grammar issues appear (e.g., "All algorithmic and environmental configurations and kept fixed" in App. B.6, and "Detailed descriptions of the local and global control modes" in §4.5 without an explicit pointer to App. B.1). A light proofread would help.

Circularity Check

0 steps flagged

No significant circularity: empirical scaling study with held-out benchmarks; noted confounds are validity/limitation issues, not circular derivations.

full rationale

The paper is an empirical systems study, not a derivation. Its central claims—that coordinated motion tracking, scaled on-policy rollouts plus diverse references, and the Humanoid Transformer reduce MPKPE—are evaluated on held-out BONES and cross-source (Xsens/100Style) test sets and against external baselines (GMT, TWIST, SONIC), so they do not reduce to fitted parameters or to a self-citation chain. The main potential self-reference is the B.5 occupancy proxy: clusters are fit on the L partition and then all partitions are scored against that same clustering. The paper explicitly acknowledges this dependency ('this quantitative measure is inherently dependent on the choice of reference motions used to construct the shared clustering space'), and the proxy is used only descriptively to name scaling regimes; the actual performance gains are measured independently on held-out benchmarks, so the conclusion is not forced by the proxy. Self-citations ([6], [3], [21], etc.) are contextual related-work and are not load-bearing for the scaling results. The B.4 design—where 'the number of mini-batches is adjusted according to the rollout horizon so that only the total amount of collected on-policy data varies across experiments, while the mini-batch size remains constant'—means that scaling width/depth also scales the number of optimizer updates; this is a real internal-validity confound and should be addressed with a fixed-update-count control, but it is a causal-attribution/correctness risk rather than a circular reduction of the paper's claims to its inputs. No uniqueness theorem, no fitted-input-called-prediction step, and no ansatz-smuggled-via-citation chain appears.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The paper's central claims rest on several unverified domain assumptions: global-frame motion tracking is a sufficient proxy for behavior learning; retargeted human motion data covers humanoid behavior; the hand-crafted occupancy measure captures diversity; and PPO rollouts are the correct notion of training data. Reward and sampling hyperparameters are hand-tuned without sensitivity analysis. No new physical entities are introduced.

free parameters (7)
  • Reward weights = 0.5, 1.0, 1.0, 1.0, 1.0, -0.1, -10.0, 1.0
    Hand-tuned in Table 5; no sensitivity analysis; changes reward balance and thus all scaling conclusions.
  • Reward kernel widths (sigma) = 0.3, 0.3, 0.4, 1.0, 3.14
    Hand-set in Table 5; control the sharpness of each tracking reward term.
  • Adaptive sampling meta-parameters = beta=0.999, T_eval=200, w_min=0.03, w_max=1.0
    Hand-set in Section 3.3; determine the resampling distribution and therefore the effective data distribution.
  • Occupancy clustering setup = k=2000, 200k sampled features, 64-dim feature vector
    Choice of K-Means and features in B.5; the homogeneous vs heterogeneous scaling interpretations are measured through this proxy.
  • Domain randomization ranges = mu_s~U[0.3,1.6], COM offsets, hand mass, velocity perturbations
    Hand-chosen in Table 6; robustness and sim-to-real transfer depend on them.
  • Success/failure threshold = 0.5 m
    Used for episode termination, adaptive sampling success, and evaluation success rate; changing it changes all reported rates and the diversity analysis.
  • PPO entropy coefficient and initial noise std = 5e-3, 0.8
    Hand-set in Table 7; affect exploration and policy behavior.
axioms (6)
  • domain assumption Goal states are excluded from the definition of behavior (Eq. 3.1)
    Underlies the BFM premise that behavior is a trajectory over proprioception and actions; goal states are treated as external specifications.
  • domain assumption Motion tracking is a valid proxy task for learning reusable behaviors
    Section 3.1/3.2; the entire pretraining objective assumes imitating reference motions yields generalizable skills.
  • domain assumption Retargeted human motion data provides sufficient behavior coverage for a humanoid
    Section 3.3; no guarantee that human data maps to feasible or diverse humanoid motions after retargeting.
  • domain assumption On-policy rollouts are the effective training data in PPO-based BFM pretraining
    Section 3.3; framing data scaling as quantity (GPUs/horizon) vs diversity (reference motions) depends on this premise.
  • ad hoc to paper The occupancy-rate clustering measures behavioral diversity
    B.5; the paper admits the measure depends on the chosen reference partition, yet the main scaling interpretations use it.
  • domain assumption Global-frame integrated tracking gives less ambiguous supervision than root-decoupled tracking
    Section 3.2 reward design; asserted, but not independently quantified beyond the BFM-Bym ablation.

pith-pipeline@v1.3.0-alltime-deepseek · 50786 in / 13564 out tokens · 105567 ms · 2026-08-01T23:56:28.196415+00:00 · methodology

0 comments
read the original 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.

Figures

Figures reproduced from arXiv: 2607.15163 by Dahua Lin, Feiyu Jia, Furui Xu, He Wang, Jiahe Chen, Jiangmiao Pang, Jingbo Wang, Kailin Li, Kangning Yin, Li Yi, Ming Zhou, Shunlin Lu, Weinan Zhang, Weishuai Zeng, Weixiang Zhong, Xiao Chen, Xiaojie Niu, Zirui Wang.

Figure 1
Figure 1. Figure 1: Our Behavior Foundation Model enables versatile humanoid behaviors across a broad spectrum of tasks, including dexterous manipulation, natural and agile locomotion and whole-body coordinated loco￾manipulation, which highlights its potential as a foundation for general-purpose humanoid control. Website: scalebfm.github.io Abstract Humanoid control requires natural whole-body coordination, precise real-time … view at source ↗
Figure 2
Figure 2. Figure 2: Reference Motion Dataset Composition. We construct a large-scale human motion dataset by aggre￾gating over 102 million frames at 50 FPS from multiple sources, which are subsequently retargeted to the target humanoid. Random samples are visualized to demonstrate its broad coverage of diverse behavior patterns. more consistent behavior acquisition. Details on reward components are provided in Appendix A.2. D… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of Model Architecture. We introduce an expressive and scalable architecture termed the Humanoid Transformer, which facilitates the natural emergence of structured behavioral representations. 3.4 Model Architecture We finally introduce the Humanoid Transformer, an expressive and scalable architecture tailored for BFM pretraining. To provide richer temporal information for behavior learning, we firs… view at source ↗
Figure 4
Figure 4. Figure 4: Results of scaling on-policy data collection on the BONES Test Set under whole-body control mode. Best outcomes with the highest success rate and lowest tracking errors are highlighted with red circles. previously unseen motions, including 839 sequences captured using Xsens [57] and 810 sequences from 100Style [58]. Together, these cross-source sequences comprise around 4.5 million frames at 50 FPS. 4.2 Sc… view at source ↗
Figure 5
Figure 5. Figure 5: Results of scaling on-policy data collection on the Ours Test Set under whole-body control mode. Best outcomes with the highest success rate and lowest tracking errors are highlighted with red circles. experimental designs adopted to ensure a fair comparison, are further provided in Appendix B.4. As shown by the results, jointly scaling the width and depth dimensions substantially improves perfor￾mance acr… view at source ↗
Figure 6
Figure 6. Figure 6: Results of scaling reference motions on the BONES Test Set under whole-body mode. Best outcomes corresponding to the highest success rate and lowest tracking errors are highlighted with red circles. All other training configurations are kept fixed across experiments, and the optimal on-policy data con￾figuration is adopted throughout. We still report results of two test sets under four representative contr… view at source ↗
Figure 7
Figure 7. Figure 7: Results of scaling reference motions on the Ours Test Set under whole-body mode. Best outcomes corresponding to the highest success rate and lowest tracking errors are highlighted with red circles. The behavioral coverage analysis naturally divides reference-motion scaling into two distinct regimes: homogeneous scaling increases the amount of reference motions while leaving the measured behav￾ioral coverag… view at source ↗
Figure 8
Figure 8. Figure 8: Results of scaling model architectures on the BONES Test Set under whole-body control mode. Best outcomes with the highest success rate and lowest tracking errors are highlighted with red circles. ing set with fine-grained coverage of domain-specific behaviors. 2) Since no finite dataset can exhaust the space of possible behaviors, rare and long-tail behavioral patterns will inevitably remain absent. A mor… view at source ↗
Figure 9
Figure 9. Figure 9: Results of scaling model architectures on the Ours Test Set under whole-body control mode. Best outcomes with the highest success rate and lowest tracking errors are highlighted with red circles. As shown in the results, the Humanoid Transformer generally achieves higher success rates and lower tracking errors than the MLP across diverse experiments. Notably, the medium-sized Humanoid Trans￾former already … view at source ↗
Figure 10
Figure 10. Figure 10: Visualization of the Latent Space. As introduced in Section 3.4, the latent space of the Humanoid Transformer is designed as a con￾tinuous manifold on the hypersphere. Through￾out this paper, we focus on the unit hypersphere, which is obtained by removing the multiplicative scaling factor in the RMSNorm module and there￾fore remains consistent across different layers. To examine the locality and global or… view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of Latent-space Visualizations across Multiple Control Modes for Models of Differ [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Results of scaling on-policy data collection on the BONES Test Set under root-and-hand control mode. Best results with the highest success rate and lowest tracking errors are highlighted with red circles.  [PITH_FULL_IMAGE:figures/full_fig_p030_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Results of scaling on-policy data collection on the Ours Test Set under root-and-hand control mode. Best results with the highest success rate and lowest tracking errors are highlighted with red circles. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Results of scaling on-policy data collection on the BONES Test Set under root-and-end-effector mode. Best results with the highest success rate and lowest tracking errors are highlighted with red circles.  [PITH_FULL_IMAGE:figures/full_fig_p031_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Results of scaling on-policy data collection on the Ours Test Set under root-and-end-effector mode. Best results with the highest success rate and lowest tracking errors are highlighted with red circles. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Results of scaling on-policy data collection on the BONES Test Set under root-and-upper-body mode. Best results with the highest success rate and lowest tracking errors are highlighted with red circles.  [PITH_FULL_IMAGE:figures/full_fig_p032_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Results of scaling on-policy data collection on the Ours Test Set under root-and-upper-body mode. Best results with the highest success rate and lowest tracking errors are highlighted with red circles. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Results of scaling reference motions on BONES Test Set under root-and-hand mode. Best results corresponding to the highest success rate and lowest tracking errors are highlighted with red circles [PITH_FULL_IMAGE:figures/full_fig_p034_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Results of scaling reference motions on Ours Test Set under root-and-hand mode. Best results corresponding to the highest success rate and lowest tracking errors are highlighted with red circles. 34 [PITH_FULL_IMAGE:figures/full_fig_p034_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Results of scaling reference motions on BONES Test Set under root-and-end-effector mode. Best results corresponding to the highest success rate and lowest tracking errors are highlighted with red circles [PITH_FULL_IMAGE:figures/full_fig_p035_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Results of scaling reference motions on Ours Test Set under root-and-end-effector mode. Best results corresponding to the highest success rate and lowest tracking errors are highlighted with red circles. 35 [PITH_FULL_IMAGE:figures/full_fig_p035_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Results of scaling reference motions on BONES Test Set under root-and-upper-body mode. Best results corresponding to the highest success rate and lowest tracking errors are highlighted with red circles [PITH_FULL_IMAGE:figures/full_fig_p036_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Results of scaling reference motions on Ours Test Set under root-and-upper-body mode. Best results corresponding to the highest success rate and lowest tracking errors are highlighted with red circles. 36 [PITH_FULL_IMAGE:figures/full_fig_p036_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Results of scaling model architectures on BONES Test Set under root-and-hand mode. Best results corresponding to the highest success rate and lowest tracking errors are highlighted with red circles [PITH_FULL_IMAGE:figures/full_fig_p038_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Results of scaling model architectures on Ours Test Set under root-and-hand mode. Best results corresponding to the highest success rate and lowest tracking errors are highlighted with red circles. 38 [PITH_FULL_IMAGE:figures/full_fig_p038_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Results of scaling model architectures on BONES Test Set under root-and-end-effector mode. Best results corresponding to the highest success rate and lowest tracking errors are highlighted with red circles [PITH_FULL_IMAGE:figures/full_fig_p039_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Results of scaling model architectures on Ours Test Set under root-and-end-effector mode. Best results corresponding to the highest success rate and lowest tracking errors are highlighted with red circles. 39 [PITH_FULL_IMAGE:figures/full_fig_p039_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Results of scaling model architectures on BONES Test Set under root-and-upper-body mode. Best results corresponding to the highest success rate and lowest tracking errors are highlighted with red circles [PITH_FULL_IMAGE:figures/full_fig_p040_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Results of scaling model architectures on Ours Test Set under root-and-upper-body mode. Best results corresponding to the highest success rate and lowest tracking errors are highlighted with red circles. 40 [PITH_FULL_IMAGE:figures/full_fig_p040_29.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 19 linked inside Pith

  1. [1]

    Learning humanoid standing-up control across diverse postures.arXiv preprint arXiv:2502.08378,

    [1]Tao Huang, Junli Ren, Huayi Wang, Zirui Wang, Qingwei Ben, Muning Wen, Xiao Chen, Jianan Li, and Jiangmiao Pang. Learning humanoid standing-up control across diverse postures.arXiv preprint arXiv:2502.08378,

    Pith/arXiv arXiv

  2. [3]

    A unified and general humanoid whole-body controller for versatile locomotion.arXiv preprint arXiv:2502.03206,

    [3]Yufei Xue, Wentao Dong, Minghuan Liu, Weinan Zhang, and Jiangmiao Pang. A unified and general humanoid whole-body controller for versatile locomotion.arXiv preprint arXiv:2502.03206,

    Pith/arXiv arXiv

  3. [4]

    Homie: Hu- manoid loco-manipulation with isomorphic exoskeleton cockpit.arXiv preprint arXiv:2502.13013,

    [4]Qingwei Ben, Feiyu Jia, Jia Zeng, Junting Dong, Dahua Lin, and Jiangmiao Pang. Homie: Hu- manoid loco-manipulation with isomorphic exoskeleton cockpit.arXiv preprint arXiv:2502.13013,

    Pith/arXiv arXiv

  4. [5]

    Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills.arXiv preprint arXiv:2502.01143,

    [5]Tairan He, Jiawei Gao, Wenli Xiao, Yuanhang Zhang, Zi Wang, Jiashun Wang, Zhengyi Luo, Guanqi He, Nikhil Sobanbab, Chaoyi Pan, et al. Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills.arXiv preprint arXiv:2502.01143,

    Pith/arXiv arXiv

  5. [6]

    Behavior foundation model for humanoid robots.arXiv preprint arXiv:2509.13780,

    [6]Weishuai Zeng, Shunlin Lu, Kangning Yin, Xiaojie Niu, Minyue Dai, Jingbo Wang, and Jiangmiao Pang. Behavior foundation model for humanoid robots.arXiv preprint arXiv:2509.13780,

    arXiv

  6. [7]

    Beyondmimic: From motion tracking to versatile humanoid control via guided dif- fusion.arXiv preprint arXiv:2508.08241,

    [9]Qiayuan Liao, Takara E Truong, Xiaoyu Huang, Yuman Gao, Guy Tevet, Koushil Sreenath, and C Karen Liu. Beyondmimic: From motion tracking to versatile humanoid control via guided dif- fusion.arXiv preprint arXiv:2508.08241,

    Pith/arXiv arXiv

  7. [8]

    Bfm-zero: A promptable behavioral foundation model for humanoid control using unsupervised reinforcement learning.arXiv preprint arXiv:2511.04131,

    [10]Yitang Li, Zhengyi Luo, Tonghe Zhang, Cunxi Dai, Anssi Kanervisto, Andrea Tirinzoni, Haoyang Weng, Kris Kitani, Mateusz Guzek, Ahmed Touati, et al. Bfm-zero: A promptable behavioral foundation model for humanoid control using unsupervised reinforcement learning.arXiv preprint arXiv:2511.04131,

    arXiv

  8. [9]

    Sonic: Supersizing motion tracking for natural hu- manoid whole-body control.arXiv preprint arXiv:2511.07820,

    [11]Zhengyi Luo, Ye Yuan, Tingwu Wang, Chenran Li, Sirui Chen, Fernando Castaneda, Zi-Ang Cao, Jiefeng Li, David Minor, Qingwei Ben, et al. Sonic: Supersizing motion tracking for natural hu- manoid whole-body control.arXiv preprint arXiv:2511.07820,

    Pith/arXiv arXiv

  9. [10]

    Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,

    [12]Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray , Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,

    Pith/arXiv arXiv 2001

  10. [16]

    Flow policy gradients for robot control

    [19]Brent Yi, Hongsuk Choi, Himanshu Gaurav Singh, Xiaoyu Huang, Takara E Truong, Carmelo Sfer- razza, Yi Ma, Rocky Duan, Pieter Abbeel, Guanya Shi, et al. Flow policy gradients for robot control. arXiv preprint arXiv:2602.02481,

    arXiv

  11. [17]

    Gmt: General motion tracking for humanoid whole-body control.arXiv preprint arXiv:2506.14770,

    [20]Zixuan Chen, Mazeyu Ji, Xuxin Cheng, Xuanbin Peng, Xue Bin Peng, and Xiaolong Wang. Gmt: General motion tracking for humanoid whole-body control.arXiv preprint arXiv:2506.14770,

    Pith/arXiv arXiv

  12. [18]

    Omnixtreme: Breaking the generality barrier in high- dynamic humanoid control.arXiv preprint arXiv:2602.23843,

    [22]Yunshen Wang, Shaohang Zhu, Peiyuan Zhi, Yuhan Li, Jiaxin Li, Yong-Lu Li, Yuchen Xiao, Xingxing Wang, Baoxiong Jia, and Siyuan Huang. Omnixtreme: Breaking the generality barrier in high- dynamic humanoid control.arXiv preprint arXiv:2602.23843,

    arXiv

  13. [19]

    Twist: Teleoperated whole-body imitation system.arXiv preprint arXiv:2505.02833,

    [24]Yanjie Ze, Zixuan Chen, Joao Pedro Araújo, Zi-ang Cao, Xue Bin Peng, Jiajun Wu, and C Karen Liu. Twist: Teleoperated whole-body imitation system.arXiv preprint arXiv:2505.02833,

    Pith/arXiv arXiv

  14. [20]

    Twist2: Scalable, portable, and holistic humanoid data collection system.arXiv preprint arXiv:2511.02832,

    [25]Yanjie Ze, Siheng Zhao, Weizhuo Wang, Angjoo Kanazawa, Rocky Duan, Pieter Abbeel, Guanya Shi, Jiajun Wu, and C Karen Liu. Twist2: Scalable, portable, and holistic humanoid data collection system.arXiv preprint arXiv:2511.02832,

    arXiv

  15. [21]

    Physhsi: Towards a real-world generalizable and natural humanoid- scene interaction system.arXiv preprint arXiv:2510.11072,

    21 [26]Huayi Wang, Wentao Zhang, Runyi Yu, Tao Huang, Junli Ren, Feiyu Jia, Zirui Wang, Xiaojie Niu, Xiao Chen, Jiahe Chen, et al. Physhsi: Towards a real-world generalizable and natural humanoid- scene interaction system.arXiv preprint arXiv:2510.11072,

    arXiv

  16. [22]

    Omniretarget: Interaction-preserving data generation for hu- manoid whole-body loco-manipulation and scene interaction.arXiv preprint arXiv:2509.26633,

    [27]Lujie Yang, Xiaoyu Huang, Zhen Wu, Angjoo Kanazawa, Pieter Abbeel, Carmelo Sferrazza, C Karen Liu, Rocky Duan, and Guanya Shi. Omniretarget: Interaction-preserving data generation for hu- manoid whole-body loco-manipulation and scene interaction.arXiv preprint arXiv:2509.26633,

    Pith/arXiv arXiv

  17. [23]

    Deep whole-body parkour.arXiv preprint arXiv:2601.07701,

    [28]Ziwen Zhuang, Shaoting Zhu, Mengjie Zhao, and Hang Zhao. Deep whole-body parkour.arXiv preprint arXiv:2601.07701,

    arXiv

  18. [24]

    Hdmi: Learning interactive humanoid whole-body control from human videos.arXiv preprint arXiv:2509.16757,

    [29]Haoyang Weng, Yitang Li, Nikhil Sobanbabu, Zihan Wang, Zhengyi Luo, Tairan He, Deva Ra- manan, and Guanya Shi. Hdmi: Learning interactive humanoid whole-body control from human videos.arXiv preprint arXiv:2509.16757,

    arXiv

  19. [25]

    Humanx: Toward agile and generalizable humanoid interaction skills from human videos.arXiv preprint arXiv:2602.02473,

    [30]Yinhuai Wang, Qihan Zhao, Yuen Fui Lau, Runyi Yu, Hok Wai Tsui, Qifeng Chen, Jingbo Wang, Jiangmiao Pang, and Ping Tan. Humanx: Toward agile and generalizable humanoid interaction skills from human videos.arXiv preprint arXiv:2602.02473,

    arXiv

  20. [26]

    Ul- tra: Unified multimodal control for autonomous humanoid whole-body loco-manipulation.arXiv preprint arXiv:2603.03279,

    [31]Xialin He, Sirui Xu, Xinyao Li, Runpei Dong, Liuyu Bian, Yu-Xiong Wang, and Liang-Yan Gui. Ul- tra: Unified multimodal control for autonomous humanoid whole-body loco-manipulation.arXiv preprint arXiv:2603.03279,

    arXiv

  21. [27]

    Uniact: Unified motion generation and action streaming for humanoid robots.arXiv preprint arXiv:2512.24321,

    [32]Nan Jiang, Zimo He, Wanhe Yu, Lexi Pang, Yunhao Li, Hongjie Li, Jieming Cui, Yuhan Li, Yizhou Wang, Yixin Zhu, et al. Uniact: Unified motion generation and action streaming for humanoid robots.arXiv preprint arXiv:2512.24321,

    arXiv

  22. [28]

    Textop: Real-time interactive text-driven humanoid robot motion generation and control.arXiv preprint arXiv:2602.07439,

    [33]Weiji Xie, Jiakun Zheng, Jinrui Han, Jiyuan Shi, Weinan Zhang, Chenjia Bai, and Xuelong Li. Textop: Real-time interactive text-driven humanoid robot motion generation and control.arXiv preprint arXiv:2602.07439,

    arXiv

  23. [29]

    Heracles: Bridging precise tracking and generative synthesis for general humanoid control.arXiv preprint arXiv:2603.27756,

    [34]Zelin Tao, Zeran Su, Peiran Liu, Jingkai Sun, Wenqiang Que, Jiahao Ma, Jialin Yu, Jiahang Cao, Pihai Sun, Hao Liang, et al. Heracles: Bridging precise tracking and generative synthesis for general humanoid control.arXiv preprint arXiv:2603.27756,

    arXiv

  24. [30]

    Robostriker: Hierarchical decision-making for autonomous humanoid boxing.arXiv preprint arXiv:2601.22517,

    [35]Kangning Yin, Zhe Cao, Wentao Dong, Weishuai Zeng, Tianyi Zhang, Qiang Zhang, Jingbo Wang, Jiangmiao Pang, Ming Zhou, and Weinan Zhang. Robostriker: Hierarchical decision-making for autonomous humanoid boxing.arXiv preprint arXiv:2601.22517,

    arXiv

  25. [31]

    Learning athletic humanoid tennis skills from imperfect human motion data.arXiv preprint arXiv:2603.12686,

    [36]Zhikai Zhang, Haofei Lu, Yunrui Lian, Ziqing Chen, Yun Liu, Chenghuai Lin, Han Xue, Zicheng Zeng, Zekun Qi, Shaolin Zheng, et al. Learning athletic humanoid tennis skills from imperfect human motion data.arXiv preprint arXiv:2603.12686,

    arXiv

  26. [33]

    Snapmogen: Human motion generation from expressive texts.arXiv preprint arXiv:2507.09122,

    [46]Chuan Guo, Inwoo Hwang, Jian Wang, and Bing Zhou. Snapmogen: Human motion generation from expressive texts.arXiv preprint arXiv:2507.09122,

    arXiv

  27. [34]

    Embody 3d: A large- scale multimodal motion and behavior dataset.arXiv preprint arXiv:2510.16258,

    [49]Claire McLean, Makenzie Meendering, Tristan Swartz, Orri Gabbay , Alexandra Olsen, Rachel Ja- cobs, Nicholas Rosen, Philippe de Bree, Tony Garcia, Gadsden Merrill, et al. Embody 3d: A large- scale multimodal motion and behavior dataset.arXiv preprint arXiv:2510.16258,

    arXiv

  28. [35]

    Retargeting matters: General motion retargeting for humanoid motion tracking.arXiv preprint arXiv:2510.02252,

    [50]Joao Pedro Araujo, Yanjie Ze, Pei Xu, Jiajun Wu, and C Karen Liu. Retargeting matters: General motion retargeting for humanoid motion tracking.arXiv preprint arXiv:2510.02252,

    arXiv

  29. [38]

    Glu variants improve transformer.arXiv preprint arXiv:2002.05202,

    [59]Noam Shazeer. Glu variants improve transformer.arXiv preprint arXiv:2002.05202,

    Pith/arXiv arXiv 2002

  30. [64]

    & #%! % # $ ') %

    is adopted throughout. We still report results of two test sets under four representative control modes: root-and-hand mode (figure 18 and 19), root-and-end-effector mode (figure 20 and 21), root-and-upper-body mode (figure 22 and 23), and whole-body mode (figure 6 and 7). We evaluate the checkpoint obtained after 22,200 training epochs. To provide contex...

    2048

  31. [2017]

    Fasttd3: Simple, fast, and capable reinforcement learning for humanoid control.arXiv preprint arXiv:2505.22642,

    [18]Younggyo Seo, Carmelo Sferrazza, Haoran Geng, Michal Nauman, Zhao-Heng Yin, and Pieter Abbeel. Fasttd3: Simple, fast, and capable reinforcement learning for humanoid control.arXiv preprint arXiv:2505.22642,

    Pith/arXiv arXiv

  32. [2018]

    Holomotion-1 technical report.arXiv preprint arXiv:2605.15336,

    [16]Maiyue Chen, Kaihui Wang, Bo Zhang, Xihan Ma, Zhiyuan Yang, Yi Ren, Qijun Huang, Zihao Zhu, Yucheng Wang, and Zhizhong Su. Holomotion-1 technical report.arXiv preprint arXiv:2605.15336,

    Pith/arXiv arXiv

  33. [2019]

    Image generation with a sphere encoder.arXiv preprint arXiv:2602.15030,

    [53]Kaiyu Yue, Menglin Jia, Ji Hou, and Tom Goldstein. Image generation with a sphere encoder.arXiv preprint arXiv:2602.15030,

    arXiv

  34. [2020]

    Emergent abilities of large language models.arXiv preprint arXiv:2206.07682,

    20 [13]Jason Wei, Yi Tay , Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yo- gatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models.arXiv preprint arXiv:2206.07682,

    Pith/arXiv arXiv

  35. [2021]

    Zero-shot whole-body humanoid control via behav- ioral foundation models.arXiv preprint arXiv:2504.11054,

    [41]Andrea Tirinzoni, Ahmed Touati, Jesse Farebrother, Mateusz Guzek, Anssi Kanervisto, Yingchen Xu, Alessandro Lazaric, and Matteo Pirotta. Zero-shot whole-body humanoid control via behav- ioral foundation models.arXiv preprint arXiv:2504.11054,

    Pith/arXiv arXiv

  36. [2022]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127,

    [14]Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127,

    Pith/arXiv arXiv

  37. [2024]

    Isaac lab: A gpu-accelerated simulation framework for multi-modal robot learning.arXiv preprint arXiv:2511.04831,

    [55]Mayank Mittal, Pascal Roth, James Tigue, Antoine Richard, Octi Zhang, Peter Du, Antonio Serrano- Munoz, Xinjie Yao, René Zurbrügg, Nikita Rudin, et al. Isaac lab: A gpu-accelerated simulation framework for multi-modal robot learning.arXiv preprint arXiv:2511.04831,

    Pith/arXiv arXiv

  38. [2025]

    Learning getting-up policies for real- world humanoid robots.arXiv preprint arXiv:2502.12152,

    [2]Xialin He, Runpei Dong, Zixuan Chen, and Saurabh Gupta. Learning getting-up policies for real- world humanoid robots.arXiv preprint arXiv:2502.12152,

    Pith/arXiv arXiv

  39. [2026]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    [17]John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    Pith/arXiv arXiv