pith. sign in

arxiv: 2606.26855 · v1 · pith:B3AIYGQ5new · submitted 2026-06-25 · 💻 cs.RO

Humanoid-DART: Humanoid Loco-Manipulation using Diffusion-guided Augmentation through Relabeling and Tracking

Pranav Debbad , Kanish Thiagarajan , Victor Dh\'edin , Shafeef Omar , Majid Khadiv This is my paper

Pith reviewed 2026-06-26 04:44 UTC · model grok-4.3

classification 💻 cs.RO
keywords humanoid robotsloco-manipulationdiffusion modelsreinforcement learningimitation learningself-supervised learninggoal-conditioned policies
0
0 comments X

The pith

A framework using diffusion models to generate trajectories and reinforcement learning to track them allows humanoids to acquire loco-manipulation skills from sparse demonstrations with minimal supervision.

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

The paper aims to show that combining diffusion-based trajectory generation with reinforcement learning creates a self-supervised system capable of expanding humanoid behaviors starting from limited demonstrations. This is important because traditional imitation learning for complex robot tasks demands expensive data collection and ongoing human corrections for failures. By generating goal-conditioned trajectories with diffusion and having the policy track them, the method enables automatic exploration of the goal space. A sympathetic reader would care if this reduces the barrier to training versatile humanoid robots for manipulation while moving.

Core claim

Our approach combines diffusion-based trajectory generation with reinforcement learning, where the latter is used to track goal-conditioned trajectories produced by the diffusion model for a range of loco-manipulation skills. This self-supervised framework bootstraps from sparse demonstrations and progressively expands its behavioral repertoire, enabling the learning of a goal-conditioned policy that automatically explores the goal space with minimal expert supervision.

What carries the argument

Diffusion-guided augmentation through relabeling and tracking that produces and follows goal-conditioned trajectories to bootstrap policies.

If this is right

  • The framework scales to multiple humanoid loco-manipulation skills.
  • It enables a goal-conditioned policy to explore the goal space with minimal expert supervision.
  • Behavioral repertoire expands progressively from sparse initial demonstrations.
  • Ablation studies confirm effectiveness over state-of-the-art methods.

Where Pith is reading between the lines

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

  • This could be adapted for other robot types by modifying the trajectory tracking components.
  • Real robot experiments would test whether the simulated tracking success transfers to physical hardware.
  • The method may connect to other self-supervised learning techniques in robotics for data efficiency.

Load-bearing premise

The diffusion model produces trajectories that remain trackable by the RL policy across an expanding goal space without requiring additional human corrections or post-hoc filtering.

What would settle it

Demonstrating that the RL policy frequently fails to track newly generated diffusion trajectories for unseen goals, necessitating human intervention.

Figures

Figures reproduced from arXiv: 2606.26855 by Kanish Thiagarajan, Majid Khadiv, Pranav Debbad, Shafeef Omar, Victor Dh\'edin.

Figure 1
Figure 1. Figure 1: Real-world deployment: Humanoid-DART trajectories deployed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed Humanoid-DART pipeline. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of structured partial unmasking on local-to-global cross [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Task-space coverage over the pipeline. Bins coloured by discovery order (purple=earliest, yellow=latest); green stars are seed demos; dashed [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Family of learnt loco-manipulation skills for [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
read the original abstract

Imitating human demonstrations has emerged as a dominant paradigm for learning humanoid loco-manipulation policies. However, scaling these approaches remains challenging due to the high cost of collecting diverse demonstrations and the need for continual human intervention to correct policy failures. In this paper, we present a self-supervised framework that bootstraps from sparse demonstrations and progressively expands its behavioral repertoire, enabling the learning of a goal-conditioned policy that automatically explores the goal space with minimal expert supervision. Our approach combines diffusion-based trajectory generation with reinforcement learning, where the latter is used to track goal-conditioned trajectories produced by the diffusion model for a range of loco-manipulation skills. Through extensive ablation studies and comparisons with state-of-the-art methods, we demonstrate the effectiveness of our framework on multiple humanoid loco-manipulation skills.

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

1 major / 0 minor

Summary. The paper presents Humanoid-DART, a self-supervised framework for humanoid loco-manipulation that starts from sparse human demonstrations. It uses a diffusion model to generate goal-conditioned trajectories and reinforcement learning to track those trajectories, thereby expanding the behavioral repertoire across multiple skills with minimal ongoing expert intervention. The approach is evaluated via ablations and comparisons to prior methods.

Significance. If the trackability of diffusion-generated trajectories holds across an expanding goal space, the framework could meaningfully lower the data-collection burden for complex humanoid behaviors. The diffusion-plus-RL pipeline is a plausible route to more autonomous skill expansion, and the manuscript's mention of extensive ablations provides a natural place to test the core assumption.

major comments (1)
  1. [Abstract] Abstract: the central self-supervised expansion claim rests on the diffusion model producing trajectories that the RL policy can track without additional human corrections or post-hoc filtering. No quantitative evidence (success rates, failure modes, or ablation on tracking reliability) is visible in the provided text, leaving the weakest assumption untested in the summary.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback on the abstract. We address the single major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central self-supervised expansion claim rests on the diffusion model producing trajectories that the RL policy can track without additional human corrections or post-hoc filtering. No quantitative evidence (success rates, failure modes, or ablation on tracking reliability) is visible in the provided text, leaving the weakest assumption untested in the summary.

    Authors: We agree that the abstract should more explicitly convey the quantitative support for the trackability assumption that underpins the self-supervised expansion claim. The full manuscript reports success rates, failure-mode breakdowns, and ablations on RL tracking of diffusion-generated trajectories (Sections 4.2–4.4 and Appendix C), but these metrics are not summarized in the abstract. We will revise the abstract to include concise quantitative statements on tracking reliability (e.g., average success rates across skills and the fraction of trajectories requiring no post-hoc filtering) so that the central claim is better supported in the summary. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper presents a pipeline that bootstraps from sparse human demonstrations using a diffusion model to generate trajectories, followed by RL policies to track them for goal-conditioned loco-manipulation. No equations, fitted parameters, or derivations are shown that reduce to their own inputs by construction. The method depends on external simulation environments and initial demos, with ablations and comparisons to SOTA methods providing independent validation. No self-citation chains, ansatzes smuggled via prior work, or self-definitional steps are evident in the described framework.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities; the central claim rests on the unstated premise that diffusion-generated trajectories are sufficiently realistic for RL tracking.

pith-pipeline@v0.9.1-grok · 5681 in / 1062 out tokens · 31495 ms · 2026-06-26T04:44:47.680725+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

29 extracted references · 4 canonical work pages

  1. [1]

    Learning human-to-humanoid real-time whole-body teleoperation,

    T. He, Z. Luo, W. Xiao, C. Zhang, K. Kitani, C. Liu, and G. Shi, “Learning human-to-humanoid real-time whole-body teleoperation,” arXiv preprint arXiv:2403.04436, 2024

    work page arXiv 2024

  2. [2]

    Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion,

    Q. Liao, T. E. Truong, X. Huang, Y . Gao, G. Tevet, K. Sreenath, and C. K. Liu, “Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion,” 2025

    2025

  3. [3]

    Unitracker: Learning universal whole-body motion tracker for humanoid robots,

    K. Yin, W. Zeng, K. Fan, M. Dai, Z. Wang, Q. Zhang, Z. Tian, J. Wang, J. Pang, and W. Zhang, “Unitracker: Learning universal whole-body motion tracker for humanoid robots,” 2025

    2025

  4. [4]

    Sonic: Supersizing motion tracking for natural humanoid whole-body control,

    Z. Luo, Y . Yuan, T. Wang, C. Li, F. Casta ˜neda, S. Chen, Z.-A. Cao, J. Li, D. Minor, Q. Ben, J. Park, D. Sami, Z. Wang, X. Da, R. Ding, C. Hogg, L. Song, E. Lim, E. Jeong, T. He, H. Xue, W. Xiao, S. Yuen, J. Kautz, Y . Chang, U. Iqbal, L. J. Fan, and Y . Zhu, “Sonic: Supersizing motion tracking for natural humanoid whole-body control,” 2026

    2026

  5. [5]

    AMASS: Archive of motion capture as surface shapes,

    N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. J. Black, “AMASS: Archive of motion capture as surface shapes,” in International Conference on Computer Vision, pp. 5442–5451, Oct. 2019

    2019

  6. [6]

    Omniretarget: Interaction-preserving data generation for humanoid whole-body loco-manipulation and scene interaction,

    L. Yang, X. Huang, Z. Wu, A. Kanazawa, P. Abbeel, C. Sferrazza, C. K. Liu, R. Duan, and G. Shi, “Omniretarget: Interaction-preserving data generation for humanoid whole-body loco-manipulation and scene interaction,” 2025

    2025

  7. [7]

    Dynaretarget: Dynamically-feasible retargeting using sampling-based trajectory optimization,

    V . Dhedin, I. Taouil, S. Omar, D. Yu, K. Tao, A. Dai, and M. Khadiv, “Dynaretarget: Dynamically-feasible retargeting using sampling-based trajectory optimization,” 2026

    2026

  8. [8]

    Deep imitation learning for humanoid loco-manipulation through human teleoperation,

    M. Seo, S. Han, K. Sim, S. H. Bang, C. Gonzalez, L. Sentis, and Y . Zhu, “Deep imitation learning for humanoid loco-manipulation through human teleoperation,” inIEEE-RAS International Conference on Humanoid Robots, 2023

    2023

  9. [9]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” 2020

    2020

  10. [10]

    Human motion diffusion model,

    G. Tevet, S. Raab, B. Gordon, Y . Shafir, D. Cohen-Or, and A. H. Bermano, “Human motion diffusion model,” 2022

    2022

  11. [11]

    Diffuse-cloc: Guided diffusion for physics-based character look-ahead control,

    X. Huang, T. Truong, Y . Zhang, F. Yu, J. P. Sleiman, J. Hodgins, K. Sreenath, and F. Farshidian, “Diffuse-cloc: Guided diffusion for physics-based character look-ahead control,”ACM Transactions on Graphics, vol. 44, no. 4, p. 1–12, 2025

    2025

  12. [12]

    Dream- control: Human-inspired whole-body humanoid control for scene interaction via guided diffusion,

    D. Kalaria, S. S. Harithas, P. Katara, S. Kwak, S. Bhagat, S. Sastry, S. Sridhar, S. Vemprala, A. Kapoor, and J. C.-K. Huang, “Dream- control: Human-inspired whole-body humanoid control for scene interaction via guided diffusion,” 2025

    2025

  13. [13]

    Kimodo: Scaling controllable human motion generation,

    D. Rempe, M. Petrovich, Y . Yuan, H. Zhang, X. B. Peng, Y . Jiang, T. Wang, U. Iqbal, D. Minor, M. de Ruyter, J. Li, C. Tessler, E. Lim, E. Jeong, S. Wu, E. Hassani, M. Huang, J.-B. Yu, C. Chung, L. Song, O. Dionne, J. Kautz, S. Yuen, and S. Fidler, “Kimodo: Scaling controllable human motion generation,”arXiv:2603.15546, 2026

    work page arXiv 2026

  14. [14]

    Curriculum learning for reinforcement learning domains: A framework and survey,

    S. Narvekar, B. Peng, M. Leonetti, J. Sinapov, M. E. Taylor, and P. Stone, “Curriculum learning for reinforcement learning domains: A framework and survey,”Journal of Machine Learning Research, vol. 21, no. 181, pp. 1–50, 2020

    2020

  15. [15]

    Learning humanoid standing-up control across diverse postures,

    T. Huang, J. Ren, H. Wang, Z. Wang, Q. Ben, M. Wen, X. Chen, J. Li, and J. Pang, “Learning humanoid standing-up control across diverse postures,” 2025

    2025

  16. [16]

    WoCoCo: Learning whole- body humanoid control with sequential contacts,

    C. Zhang, W. Xiao, T. He, and G. Shi, “WoCoCo: Learning whole- body humanoid control with sequential contacts,”arXiv preprint arXiv:2406.06005, 2024

    work page arXiv 2024

  17. [17]

    Learning a family of motor skills from a single motion clip,

    S. Lee, S. Lee, Y . Lee, and J. Lee, “Learning a family of motor skills from a single motion clip,”ACM Trans. Graph., vol. 40, July 2021

    2021

  18. [18]

    Parc: Physics-based augmentation with reinforcement learning for character controllers,

    M. Xu, Y . Shi, K. Yin, and X. B. Peng, “Parc: Physics-based augmentation with reinforcement learning for character controllers,” in Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers, SIGGRAPH Conference Papers ’25, (New York, NY , USA), Association for Com- puting Machinery, 2025

    2025

  19. [19]

    Universal humanoid motion representations for physics-based control,

    Z. Luo, J. Cao, J. Merel, A. Winkler, J. Huang, K. Kitani, and W. Xu, “Universal humanoid motion representations for physics-based control,” 2024

    2024

  20. [20]

    On the continuity of rotation representations in neural networks,

    Y . Zhou, C. Barnes, J. Lu, J. Yang, and H. Li, “On the continuity of rotation representations in neural networks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5745–5753, 2019

    2019

  21. [21]

    Scalable diffusion models with transform- ers,

    W. Peebles and S. Xie, “Scalable diffusion models with transform- ers,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 4195–4205, 2023

    2023

  22. [22]

    Diffusion forcing: Next-token prediction meets full- sequence diffusion,

    B. Chen, D. M. Monso, Y . Du, M. Simchowitz, R. Tedrake, and V . Sitzmann, “Diffusion forcing: Next-token prediction meets full- sequence diffusion,” 2024

    2024

  23. [23]

    Masked- mimic: Unified physics-based character control through masked mo- tion inpainting,

    C. Tessler, Y . Guo, O. Nabati, G. Chechik, and X. B. Peng, “Masked- mimic: Unified physics-based character control through masked mo- tion inpainting,” 2024

    2024

  24. [24]

    Simgenhoi: Physically realistic whole-body humanoid-object interaction via generative modeling and reinforcement learning,

    Y . Lin, Y . Xie, J. Xie, Y . Huang, R. Wang, J. Lv, Y . Ma, and X. Zuo, “Simgenhoi: Physically realistic whole-body humanoid-object interaction via generative modeling and reinforcement learning,”arXiv preprint arXiv:2508.14120, 2025

    work page arXiv 2025

  25. [25]

    RePaint: Inpainting using denoising diffusion prob- abilistic models,

    A. Lugmayr, M. Danelljan, A. Romero, F. Yu, R. Timofte, and L. Van Gool, “RePaint: Inpainting using denoising diffusion prob- abilistic models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 11461–11471, 2022

    2022

  26. [26]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” 2022

    2022

  27. [27]

    DeepMimic: Example-guided deep reinforcement learning of physics-based charac- ter skills,

    X. B. Peng, P. Abbeel, S. Levine, and M. van de Panne, “DeepMimic: Example-guided deep reinforcement learning of physics-based charac- ter skills,” inACM Transactions on Graphics (SIGGRAPH), vol. 37, 2018

    2018

  28. [28]

    Hindsight experience replay,

    M. Andrychowicz, F. Wolski, A. Ray, J. Schneider, R. Fong, P. Welin- der, B. McGrew, J. Tobin, P. Abbeel, and W. Zaremba, “Hindsight experience replay,” inAdvances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  29. [29]

    mjlab: A Lightweight Framework for GPU-Accelerated Robot Learn- ing,

    K. Zakka, Q. Liao, B. Yi, L. Le Lay, K. Sreenath, and P. Abbeel, “mjlab: A Lightweight Framework for GPU-Accelerated Robot Learn- ing,” 2026. APPENDIX To keep the exact settings reproducible and machine- readable, we provide all remaining hyperparameters as Python configuration files as supplementary material: •the motion-generator backbone and training s...

    2026