pith. sign in

arxiv: 2606.22998 · v2 · pith:B5GE4J6Unew · submitted 2026-06-22 · 💻 cs.RO

TEXEDO : Test Time Scaling for Controller-aware Language-conditioned Humanoid Motion Generation

Jianuo Cao , Yuxin Chen , Yuzhen Song , Masayoshi Tomizuka , Chenran Li , Thomas Tian This is my paper

Pith reviewed 2026-06-26 08:49 UTC · model grok-4.3

classification 💻 cs.RO
keywords test-time scalinghumanoid motion generationlanguage-conditioned controldynamic feasibility verificationsemantic alignmentwhole-body trackingrobot deployment
0
0 comments X

The pith

Test-time selection using a feasibility verifier and semantic scorer produces more executable text-to-humanoid motions.

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

The paper establishes that a pretrained text-conditioned motion generator can be improved at inference time without retraining by sampling many candidate motions and keeping only the one that passes a dynamic feasibility check while scoring highest on text alignment. The feasibility check comes from a verifier trained on whole-body tracking data to predict whether a motion will succeed under the robot's controller, balance, and contact constraints. A sympathetic reader would care because language is an intuitive way to program humanoids yet most generated motions remain physically unrealizable; a lightweight selection step could turn existing generators into deployable tools. If the approach holds, robot programmers could obtain reliable motions simply by generating and filtering more candidates rather than collecting new robot-specific datasets or redesigning the generator.

Core claim

TEXEDO samples multiple motions from a fixed pretrained text-conditioned generator, applies a dynamic feasibility verifier distilled from whole-body tracking rollouts as a hard filter, and then selects the motion with highest semantic alignment inside the feasible set. Large-scale simulation and real-world tests on a Unitree G1 show consistent gains in both tracking fidelity and prompt adherence.

What carries the argument

TEXEDO test-time scaling pipeline that treats dynamic feasibility as a hard constraint and semantic alignment as the selection objective.

If this is right

  • Motions that pass the verifier are more likely to succeed under the actual whole-body controller without violating balance or actuation limits.
  • Semantic alignment can be optimized only among motions already known to be feasible, avoiding the trade-off between meaning and executability.
  • The same selection procedure works for any pretrained generator, so gains accrue without collecting new robot-specific training data.
  • Real-robot experiments confirm that the simulated verifier transfers to hardware, closing the sim-to-real gap for language-conditioned tasks.

Where Pith is reading between the lines

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

  • The framework could be applied to other robot morphologies by retraining only the feasibility verifier on that platform's tracking data.
  • If the verifier is cheap to evaluate, further scaling the number of candidates at test time may yield additional gains similar to test-time compute in language models.
  • Combining the verifier with online adaptation of the generator weights could produce a hybrid system that improves both at inference and through continued learning.

Load-bearing premise

The dynamic feasibility verifier distilled from whole-body tracking rollouts accurately predicts whether a generated motion will be physically executable by the robot's controller.

What would settle it

Deploy the motions selected by TEXEDO on the physical Unitree G1 and measure whether tracking error and success rate remain statistically indistinguishable from motions chosen by random sampling or by the original generator alone.

Figures

Figures reproduced from arXiv: 2606.22998 by Chenran Li, Jianuo Cao, Masayoshi Tomizuka, Thomas Tian, Yuxin Chen, Yuzhen Song.

Figure 1
Figure 1. Figure 1: TEXEDO overview. Given a language prompt, the TEXT stage samples multiple whole-body motions from a text￾conditioned motion generator. The SEE stage evaluates each candidate with two grounded verifiers: a dynamic score that predicts controller executability and a semantic score that measures text-motion alignment. The DO stage selects a deployable motion by first filtering out dynamically infeasible candid… view at source ↗
Figure 2
Figure 2. Figure 2: Dual verifier design. Rdyn estimates dynamic feasi￾bility from the motion alone and Rtext measures text-motion alignment in a learned embedding space. B. Dynamic Feasibility And Semantic Alignment Evaluation (SEE) What turns a pool of candidate motions into a single deployable trajectory is, ultimately, a definition of motion quality. The SEE stage answers this question with two scalars assigned to every c… view at source ↗
Figure 3
Figure 3. Figure 3: Best-of-N curves for dynamic feasibility on FSQ-GPT. Subplots report Succ, Empjpe-l, Eacc, Evel, and Q∗ as N increases from 1 to 32. Across metrics, Rdyn-only consistently outperforms Random and closely approaches Oracle. IV. EXPERIMENTS We evaluate TEXEDO as a deployable test-time scaling system for language-conditioned humanoid motion generation along three axes: test-time selection performance, zero-sho… view at source ↗
Figure 4
Figure 4. Figure 4: Semantic alignment verifier turns larger candidate pools into better prompt alignment. For each prompt, Rtext￾only selects the candidate with the highest Rtext score; the selected motions improve consistently with increasing N when evaluated by an independent VLM Judge. B. Effectiveness Of TEXEDO For Test-Time Scaling Effectiveness of individual verifiers. We first isolate the effect of each verifier under… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison. Representative case studies in MuJoCo simulation. Top-left: a success case of TEXEDO, where the generated motion is dynamically feasible and semantically aligned. Top-right: a failure case of Rdyn-only produces a stable but semantically generic motion. Bottom-left: a failure case of Rtext-only selects a semantically aligned but untrackable motion. Bottom-right: a failure case of Bas… view at source ↗
Figure 6
Figure 6. Figure 6: Real-robot execution results on the Unitree G1. Across diverse prompt types spanning locomotion, upper-body gestures, and their combinations, TEXEDO consistently selects motions that execute stably and faithfully realize the intended semantics on physical hardware. better preserve the language instruction. The proposed filter￾then-rerank rule combines these complementary signals to select motions that impr… view at source ↗
read the original abstract

Text-conditioned motion generation is a promising interface for programming humanoid robots, yet current generators are often trained on human motion datasets retargeted to robot morphologies. Although such data provides rich semantic and kinematic priors, it fails to capture the nuances of whole-body tracking controllers, including balance, contact dynamics, actuation limits, and controller-specific failure modes. As a result, generated motions can be semantically plausible but difficult or impossible for the robot to execute. We introduce TEXEDO, a test-time scaling framework for humanoid motion generation that improves motion quality without requiring a stronger underlying generator. Given a text prompt, TEXEDO samples multiple candidate motions from a pretrained text-conditioned generator and selects the best motion that is both executable and task-aligned. The reward model combines a dynamic feasibility verifier, distilled from whole-body tracking rollouts to predict physical executability, with a semantic alignment verifier that measures text-motion alignment in a learned co-embedding space. Our pipeline treats dynamic feasibility as a hard constraint and semantic alignment as the selection objective within the feasible set. Through large-scale simulation studies and real-world deployment on a Unitree G1 humanoid robot, we show that TEXEDO consistently improves both tracking fidelity and text alignment. These results demonstrate that grounded verification is an effective path toward deployable language-guided humanoid motion generation. Project website: https://jianuocao.github.io/TEXEDO/

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces TEXEDO, a test-time scaling framework for text-conditioned humanoid motion generation. It samples multiple candidate motions from a pretrained generator and selects the best one by treating dynamic feasibility (via a verifier distilled from whole-body tracking rollouts) as a hard constraint and semantic alignment (in a learned co-embedding space) as the selection objective within the feasible set. The method is claimed to improve tracking fidelity and text alignment, with supporting evidence from large-scale simulation studies and real-world deployment on a Unitree G1 robot.

Significance. If the central assumption holds, the approach provides a practical, training-free way to adapt existing motion generators to robot-specific controller constraints such as balance and contact dynamics. The real-robot evaluation on the Unitree G1 is a positive aspect for deployability claims. However, the absence of any quantitative metrics, baselines, or experimental details in the abstract prevents assessment of effect sizes or whether the gains are meaningful relative to prior work.

major comments (2)
  1. [Abstract] Abstract: the central claim of consistent improvements in tracking fidelity and text alignment rests on the dynamic feasibility verifier (distilled only from whole-body tracking rollouts) correctly classifying executability for motions sampled from the text-conditioned generator. No evidence is provided that generator outputs lie inside the rollout distribution; systematic misclassification would directly invalidate the selection step and the reported gains.
  2. [Abstract] Abstract: the manuscript states that 'large-scale simulation studies and real-world deployment' demonstrate the improvements, yet supplies no metrics, baselines, success rates, or statistical details. This omission makes the soundness of the empirical claims impossible to evaluate from the provided text.
minor comments (1)
  1. [Abstract] The abstract refers to a 'reward model' combining the two verifiers, but the selection procedure is described only at a high level; a precise statement of how the hard constraint is enforced (e.g., filtering before ranking or joint optimization) would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We agree that the abstract requires strengthening to better support the claims with quantitative details and will revise it accordingly. Below we respond point-by-point to the major comments.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of consistent improvements in tracking fidelity and text alignment rests on the dynamic feasibility verifier (distilled only from whole-body tracking rollouts) correctly classifying executability for motions sampled from the text-conditioned generator. No evidence is provided that generator outputs lie inside the rollout distribution; systematic misclassification would directly invalidate the selection step and the reported gains.

    Authors: We acknowledge this is a substantive concern: the abstract does not explicitly demonstrate that motions from the text-conditioned generator lie within the distribution of whole-body tracking rollouts used to distill the verifier. The full manuscript reports that TEXEDO yields measurable gains in tracking fidelity and text alignment across simulation and real-robot experiments, which provides indirect support for the verifier's utility on generator outputs. However, we agree that direct evidence of distribution overlap or out-of-distribution robustness would strengthen the argument. In revision we will update the abstract to reference the empirical validation and add a short discussion or ablation on verifier generalization. revision: yes

  2. Referee: [Abstract] Abstract: the manuscript states that 'large-scale simulation studies and real-world deployment' demonstrate the improvements, yet supplies no metrics, baselines, success rates, or statistical details. This omission makes the soundness of the empirical claims impossible to evaluate from the provided text.

    Authors: The referee is correct that the current abstract is too high-level and lacks the quantitative details needed to assess effect sizes. The full paper contains these results in the experiments section, including simulation metrics on tracking error and success rates as well as real-world deployment outcomes on the Unitree G1. We will revise the abstract to incorporate key numbers, baseline comparisons, and statistical details so that the claims can be evaluated directly from the abstract text. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper describes a test-time selection pipeline that samples from a pretrained generator and filters using a verifier distilled from separate whole-body tracking rollouts. No equations, self-definitional relations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The verifier training distribution and generator sampling are treated as distinct inputs with an explicit (if unverified) assumption of coverage; this does not reduce the claimed improvement to a tautology or self-referential fit by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review limited to abstract; no free parameters, axioms, or invented entities can be identified.

pith-pipeline@v0.9.1-grok · 5796 in / 1035 out tokens · 21796 ms · 2026-06-26T08:49:15.261492+00:00 · methodology

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

26 extracted references · 7 linked inside Pith

  1. [1]

    BONES-SEED: Skeletal everyday em- bodiment dataset, 2026

    Bones Studio. BONES-SEED: Skeletal everyday em- bodiment dataset, 2026. URL https://huggingface.co/ datasets/bones-studio/seed

    2026

  2. [2]

    Claw: Composable language-annotated whole-body mo- tion generation.arXiv preprint arXiv:2604.11251, 2026

    Jianuo Cao, Yuxin Chen, and Masayoshi Tomizuka. Claw: Composable language-annotated whole-body mo- tion generation.arXiv preprint arXiv:2604.11251, 2026

    Pith/arXiv arXiv 2026

  3. [3]

    Reimagination with test-time observation interven- tions: Distractor-robust world model predictions for visual model predictive control.arXiv preprint arXiv:2506.16565, 2025

    Yuxin Chen, Jianglan Wei, Chenfeng Xu, Boyi Li, Masayoshi Tomizuka, Andrea Bajcsy, and Ran Tian. Reimagination with test-time observation interven- tions: Distractor-robust world model predictions for visual model predictive control.arXiv preprint arXiv:2506.16565, 2025

    arXiv 2025

  4. [4]

    Scaling instruction-finetuned language models.Journal of Ma- chine Learning Research, 25(70):1–53, 2024

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models.Journal of Ma- chine Learning Research, 25(70):1–53, 2024

    2024

  5. [5]

    Generating diverse and natural 3d human motions from text

    Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5152–5161, 2022

    2022

  6. [6]

    Momask: Generative masked modeling of 3d human motions

    Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900–1910, 2024

    1900

  7. [7]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Motiongpt: Human motion as a foreign language.Advances in Neural Information Processing Systems, 36:20067–20079, 2023

    Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language.Advances in Neural Information Processing Systems, 36:20067–20079, 2023

    2023

  9. [9]

    Robomonkey: Scaling test-time sampling and verification for vision-language-action models.arXiv preprint arXiv:2506.17811, 2025

    Jacky Kwok, Christopher Agia, Rohan Sinha, Matt Fout- ter, Shulu Li, Ion Stoica, Azalia Mirhoseini, and Marco Pavone. Robomonkey: Scaling test-time sampling and verification for vision-language-action models.arXiv preprint arXiv:2506.17811, 2025

    arXiv 2025

  10. [10]

    Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion.arXiv preprint arXiv:2508.08241, 2025

    Qiayuan Liao, Takara E Truong, Xiaoyu Huang, Yu- man Gao, Guy Tevet, Koushil Sreenath, and C Karen Liu. Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion.arXiv preprint arXiv:2508.08241, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Let’s verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. InInternational Conference on Learn- ing Representations, volume 2024, pages 39578–39601, 2024

    2024

  12. [12]

    Perpetual humanoid control for real-time simulated avatars

    Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10895–10904, 2023

    2023

  13. [13]

    Sonic: Supersizing motion tracking for natural humanoid whole-body con- trol.arXiv preprint arXiv:2511.07820, 2025

    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 humanoid whole-body con- trol.arXiv preprint arXiv:2511.07820, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Amass: Archive of motion capture as surface shapes

    Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. InProceedings of the IEEE/CVF international conference on computer vision, pages 5442–5451, 2019

    2019

  15. [15]

    Finite scalar quantization: Vq-vae made simple

    Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. InInternational Conference on Learn- ing Representations, volume 2024, pages 51772–51783, 2024

    2024

  16. [16]

    Learning to reason with LLMs

    OpenAI. Learning to reason with LLMs. OpenAI Blog, 2024. https://openai.com/index/ learning-to-reason-with-llms/

    2024

  17. [17]

    Amp: Adversarial motion priors for stylized physics-based character control.ACM Transac- tions on Graphics (ToG), 40(4):1–20, 2021

    Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control.ACM Transac- tions on Graphics (ToG), 40(4):1–20, 2021

    2021

  18. [18]

    Kimodo: Scaling controllable human motion generation.arXiv preprint arXiv:2603.15546, 2026

    Davis Rempe, Mathis Petrovich, Ye Yuan, Haotian Zhang, Xue Bin Peng, Yifeng Jiang, Tingwu Wang, Umar Iqbal, David Minor, Michael de Ruyter, et al. Kimodo: Scaling controllable human motion generation.arXiv preprint arXiv:2603.15546, 2026

    arXiv 2026

  19. [19]

    Scaling llm test-time compute optimally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314, 2024

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Ku- mar. Scaling llm test-time compute optimally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Heracles: Bridging precise tracking and generative synthesis for general humanoid control.arXiv preprint arXiv:2603.27756, 2026

    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, 2026

    arXiv 2026

  21. [21]

    Human motion diffusion model.arXiv preprint arXiv:2209.14916, 2022

    Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model.arXiv preprint arXiv:2209.14916, 2022

    Pith/arXiv arXiv 2022

  22. [22]

    Realign: text-to- motion generation via step-aware reward-guided align- ment

    Wanjiang Weng, Xiaofeng Tan, Junbo Wang, Guo-Sen Xie, Pan Zhou, and Hongsong Wang. Realign: text-to- motion generation via step-aware reward-guided align- ment. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 10621–10629, 2026

    2026

  23. [23]

    From foresight to forethought: Vlm-in-the-loop policy steering via latent alignment.arXiv preprint arXiv:2502.01828, 2025

    Yilin Wu, Ran Tian, Gokul Swamy, and Andrea Ba- jcsy. From foresight to forethought: Vlm-in-the-loop policy steering via latent alignment.arXiv preprint arXiv:2502.01828, 2025

    arXiv 2025

  24. [24]

    Generating human motion from textual descrip- tions with discrete representations

    Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descrip- tions with discrete representations. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730–14740, 2023

    2023

  25. [25]

    Mo- tiondiffuse: Text-driven human motion generation with diffusion model.IEEE transactions on pattern analysis and machine intelligence, 46(6):4115–4128, 2024

    Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Mo- tiondiffuse: Text-driven human motion generation with diffusion model.IEEE transactions on pattern analysis and machine intelligence, 46(6):4115–4128, 2024

    2024

  26. [26]

    least-bad

    Zewei Zhang, Kehan Wen, Michael Xu, Junzhe He, Chenhao Li, Takahiro Miki, Clemens Schwarke, Chong Zhang, Xue Bin Peng, and Marco Hutter. Learning whole-body humanoid locomotion via motion generation and motion tracking.arXiv preprint arXiv:2604.17335, 2026. APPENDIXA RELATEDWORK Language-conditioned motion generation.A prominent line of work represents co...

    Pith/arXiv arXiv 2026