Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Multi-Embodiment Robotic Retargeting via Guided Diffusion Model

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A single graph-conditioned diffusion model can retarget a reference motion to many robot skeletons at once, without motion data from the target robot.

desk verdict A plausible graph-conditioned retargeting idea whose diffusion formulation is mis-specified and whose validation is too weak to support the unified claims. read the letter →

arxiv 2505.20857 v2 pith:Q7SILNKW submitted 2025-05-27 cs.RO

Zhefeng Cao , Ben Liu , Shunpeng Yang , Sen Li , Wei Zhang , Hua Chen This is my paper · ORCID
classification cs.RO
keywords motionretargetingmulti-embodimentdiffusionmodelgraphconditioningenergy-basedguidancerobotskeletontransformerdecoderzero-shotgeneralization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that one diffusion-based motion generator can retarget a reference motion to many different robot bodies at once, even when the bodies have different joint counts, link lengths, and topologies, and even when no motion data exists for the target body. The key move is to condition the generator on a graph description of each robot's skeleton and to guide sampling with an energy term that rewards matching the reference motion's key joint positions after a body-size scaling. If this works as claimed, it would let a single model serve a library of legged robots instead of requiring per-robot retargeting code or paired motion datasets.

What carries the argument

The load-bearing object is the graph-conditioned denoising transformer decoder the authors call G-DReaM, which tokenizes every motion at the joint level so that variable-length skeletons fit one architecture. A robot skeleton is represented as a directed acyclic graph whose node attributes are joint rotation axes, edge attributes are link vectors, and adjacency matrix carries four relation types (none, self, parent, child); this graph enters the network through spatial attention with relational embeddings and through multi-head cross-attention that gives each condition its own parameters. The energy-based guidance is the kinematic loss $f_{kin}$, a combination of scaled keypoint similarity, forward-kinematics consistency, velocity tracking, and regularization, weighted by $\lambda = 10^4$ in the denoising objective, and the joint map $\eta$ becomes a temporal attention mask that decides which reference and target joints may influence each other.

What would settle it

Retarget one reference motion to a target skeleton twice, once with the correct joint map and once with a randomly permuted map, and compare the positional retargeting loss; if the permuted map reaches the same centimeter-level loss as the correct map, the correspondence term is not carrying the claimed workload. A second check: repeat the comparison on a skeleton whose leg-length ratio falls outside the 0.5 to 2 augmentation range and see whether foot-ground contact and joint-limit violations appear in the generated motions.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that a transformer-based denoising network, conditioned on a reference motion and on graph-encoded source and target skeletons, produces retargeted motions for heterogeneous embodiments without paired or target motion data. The denoiser is trained with a reconstruction prior drawn from the reference-motion distribution plus an energy-based guidance term built from kinematic retargeting losses: scaled keypoint similarity, forward-kinematic consistency, velocity matching, and a regularization term. Skeleton graphs store each joint's rotation axis, each link's parent-to-child vector, and a four-valued connectivity matrix; joint correspondence is supplied as a sparse attention mask. The experiments retarget motions between humanoids and between humanoids and a point-foot biped, and the authors report centimeter-level joint position errors, adaptation to new skeletons by adding their graphs to training, and zero-shot handling of unseen motions similar to the training set.

Load-bearing premise

The retargeting objective assumes that the manually chosen semantic joint correspondences (hip, knee, ankle, toe, shoulder, elbow, hand) and a single leg-length scaling factor capture how a motion should transfer from one body to another, so if those correspondences do not match the functional roles of joints across skeletons, the guidance pushes generation toward the wrong poses.

Editorial extensions

If this is right

  • Retargeting no longer needs paired motion data or per-robot optimization code; a single model can produce motions for multiple skeletons from one reference.
  • Non-homeomorphic pairs, such as a biped following only the lower-body part of a humanoid reference, can be handled by dropping missing corresponding joints from the joint map.
  • A pretrained model can be adapted to a new robot in about 5 hours by adding its skeleton graph to the training conditions, without supplying new motion data.
  • The same model handles unseen motions similar to its training distribution, such as new walking or running clips, with only a modest increase in retargeting error.
  • Controlling which joints are marked as corresponding changes the generated pose in a predictable way, so the correspondence map can be used as an interactive control input.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension the paper leaves implicit is to learn the joint correspondence map itself during training rather than taking it from a manually chosen semantic list; the attention-mask formulation is differentiable and would support such learning.
  • The same energy-based guidance could absorb hard physical constraints, such as joint limits, foot-ground contact, and self-collision, as additional energy terms, which would directly address the kinematic violations the authors acknowledge in their limitation section.
  • If the method scales, it points toward a single embodiment-conditioned motion prior that could be shared across simulation and real robots, making retargeting a side input rather than a per-robot engineering task.
  • A testable extension is to measure whether retargeting quality degrades smoothly with increasing topological distance of the target skeleton; the current evaluation covers humanoids plus one biped, so quadrupeds or manipulators with different end-effector roles would stress the correspondence assumptions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes G-DReaM, a graph-conditioned diffusion model for retargeting reference motions across multiple robot embodiments. The method encodes skeleton topology and geometry as graph conditions, trains a transformer-based denoiser, and uses an energy-based term (kinematic losses) to avoid requiring target-embodiment motion data. Experiments are reported on humanoid and biped robots, including homeomorphic and non-homeomorphic pairs, with additional tests for skeleton generalization and new-embodiment adaptation.

Significance. If the approach worked as claimed, it would be a meaningful step toward a unified motion-retargeting framework that does not require paired motion data for every embodiment. The proposed graph-conditioned architecture and the idea of injecting kinematic losses as guidance are potentially interesting. However, the current manuscript does not convincingly establish the central claim: the evaluation metric is the training objective, no baselines are reported, the theoretical justification of the energy-guided training is not a valid derivation, and the 'zero-shot' claims are inconsistent with the experimental protocol. These issues are load-bearing rather than presentation-level, so the contribution as presented is not yet substantiated.

major comments (4)
  1. [§3.2, Eq. (1)/(7)] The training objective in Eq. (7) is not a valid energy-guided diffusion formulation. The first term trains the denoiser to reconstruct source-embodiment samples x, while the second term evaluates f_kin(D_θ) via forward kinematics on the target skeleton g_X. For a non-homeomorphic pair (e.g., a humanoid to the Tron1 biped, which lacks arm joints), no single output tensor can simultaneously reconstruct the source motion and satisfy the kinematic objective on a different topology. The manuscript asserts in §3.2 that this 'aligns the learned score function with the desired score function −∇_x f_kin(x)', but this is not a derivation: minimizing a weighted sum of a denoising loss and a kinematic penalty during training does not imply that the resulting model samples from p(x) ∝ p_ref(x) exp(−E(x)). Consequently, the mechanism that is supposed to enable retargeting without target motion data is not established, and this is a load-bearing flaw for the central claim.
  2. [§4.2, Table 1, Eq. (9)] The reported evaluation metric, 'Positional MSE Loss', is exactly the L_similar term in the training loss (Eq. (9)), which is minimized with weight w0 = 100 and an overall guidance weight λ = 10^4. Reporting low values of the training objective does not independently validate retargeting quality; it is partially circular. Moreover, the paper provides no baselines, so the Conclusion's phrase 'achieving comparable retargeting metric' is unsupported—there is nothing to compare against. The tables also present single numbers without variance or confidence intervals, which is problematic for a stochastic diffusion-based generator.
  3. [§4.2, Appendix D.1, Limitations] The 'zero-shot' claims are overstated and inconsistent with the experimental protocol. The link-length and joint-correspondence perturbations in §4.2 are within the augmentation ranges described in Appendix B.2 (link lengths scaled 0.5–2.0, random removal of joint correspondences), so these are within-distribution tests rather than zero-shot generalization. The new-embodiment experiments in Appendix D.1 require 5 hours of additional training, and the Limitations section explicitly states that 'our model currently works only on embodiments it has encountered during training.' The paper should align its claims with these facts and avoid calling the adaptation procedure 'zero-shot'.
  4. [§4.2, Table 1 and Table 3] The quantitative evaluation is insufficiently specified: Table 1 does not state which reference motions were used, how many motion clips were evaluated per embodiment, or which embodiment pairs were tested. In particular, the non-homeomorphic case (humanoid to Tron1) is only demonstrated qualitatively in Figure 3, with no quantitative error for that pair. Without clear evaluation protocols and quantitative results for non-homeomorphic retargeting, the central claim of handling heterogeneous skeletons is not substantiated.
minor comments (6)
  1. [Abstract and §1] There are grammatical errors, e.g., 'with a customized attention mechanisms' in the Abstract and 'In recently years' at the start of §1; these should be corrected.
  2. [Figure 1] The caption contains a typo: 'retargetd' should be 'retargeted'.
  3. [§3.3] The sentence 'So the the useful information of the reference motion within each condition is related to the corresponding joint map η' contains a duplicated 'the' and is awkwardly phrased.
  4. [§4.3] The phrase 'on a the same GPU' should read 'on the same GPU'; also, 'This lightweight adaptation' is a sentence fragment.
  5. [Tables 1, 3, 5] The units are inconsistent: Table 1 reports cm^2, while Tables 3 and 5 report m^2. The text in §4.2 says errors are 'at the centimeter level', but values like 169.6 cm^2 correspond to about 13 cm RMSE; please clarify the units and the interpretation.
  6. [References] Reference [34] is a Hugging Face dataset URL; please provide a more formal citation or verify the accessibility and version of the dataset.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline retargeting metric is the same L_similar term minimized in the training guidance loss, so the reported positional-MSE numbers are partially baked in; independent generalization checks keep the circularity partial.

  1. fitted input called prediction [Appendix A, Eq (8)-(9); Sec 4.2, Table 1; Sec 3.4]
    "Lsimilar = Σ_{{(i,j)|ηij=1}} w0∥FK( ˆX0, gX , i)−P(M, α, j)∥^2 ... In our settings, we strongly rely on the guidance, hence set it to λ= 10^4 in all of our experiments. ... Table 1: ... Positional MSE Loss /cm^2"

    The evaluation metric reported in Table 1 is exactly the L_similar term of the energy-based guidance defined in Eq (9), using the same manually specified joint correspondences η and the same scaling factor α. The training loss in Eq (7) minimizes λ f_kin, and with w0=100 and λ=10^4 this term dominates training. Therefore the headline 'comparable retargeting metric' is the model's own training objective, not an independent measure of retargeting quality. The reported centimeter-level positional errors are, by construction, the quantity the network was directly trained to minimize; they cannot by themselves validate the retargeting claim.

full rationale

The paper's central mechanism is not justified by self-citation: prior-work references are external and no load-bearing uniqueness theorem is imported from the authors' own publications. However, one important circular step exists in the evaluation. The model is trained to minimize L_similar, the positional MSE between FK on the target skeleton and scaled reference keypoints, and the same L_similar is then reported as the retargeting metric in Table 1. Since the guidance weight is set to λ=10^4 and w0=100, the reported numbers largely reflect the optimizer minimizing its own loss rather than an independent test. That said, the paper does evaluate on unseen motions, link-length modifications, and newly added skeleton graphs, which are not identical to the training conditions and provide some independent evidence of generalization. These independent checks prevent the conclusion from reducing entirely to the training loss, so the appropriate score is partial circularity rather than full equivalence.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced; the graph representation, joint map, and guidance losses are constructs for conditioning the denoiser. The free parameters are the hand-set guidance and loss weights, plus the heuristic leg-length scaling factor that directly shapes the retargeting objective.

free parameters (4)
  • guidance weight λ = 10^4
    Balances denoising prior and kinematic guidance in Eq (1); set to 10^4 in all experiments, no sensitivity analysis.
  • kinematic loss weights w0, w1, w2, w3 = 100, 1, 900, 1
    Weights in Eqs (9)-(12) for position, consistency, velocity, and norm losses; hand-tuned, no ablation.
  • leg-length scaling factor α = l_des/l_ref
    Global scaling factor in Eq (9) mapping reference body scale to target; chosen as ratio of leg lengths, a simple heuristic that may be wrong for non-homeomorphic pairs.
  • temporal window size T_w = 31
    Attention mask horizon in temporal attention; set to 31 frames, affects temporal consistency but no study of its effect.
assumptions (4)
  • domain assumption Manually specified semantic keypoint correspondences and the leg-length scale α are sufficient to define correct retargeting across heterogeneous, including non-homeomorphic, skeletons.
    Entered in L_similar (Eq 9); if correspondence is wrong, guidance is wrong. Paper acknowledges manual correspondence as a limitation.
  • ad hoc to paper A diffusion denoiser trained on source-embodiment motions with a large guidance weight (λ=1e4) learns the target conditional distribution without ever seeing target motion data.
    Core mechanism in Section 3.2; no theoretical guarantee or baseline comparison supports it.
  • domain assumption The graph attributes (joint axis, link vector, adjacency type) at zero pose capture the embodiment's kinematics for retargeting.
    Section 3.1 representation; assumes these features are sufficient and that padding joints to a max T,J does not harm learning.
  • standard math Forward kinematics computed from predicted joint angles is exact and differentiable, so the losses can be minimized via backpropagation.
    Used in Eqs (9)-(12); this is standard robot kinematics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Embodiment Robotic Retargeting via Guided Diffusion Model." pith.science (2026). https://pith.science/paper/Q7SILNKW

@misc{pith2026250520857,
  author       = {Pith},
  title        = {Pith review of: Multi-Embodiment Robotic Retargeting via Guided Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q7SILNKW}},
  note         = {Machine review of arXiv:2505.20857}
}
read the original abstract

Motion retargeting for specific robot from existing motion datasets is one critical step in transferring motion patterns from human behaviors to and across various robots. However, inconsistencies in topological structure, geometrical parameters as well as joint correspondence make it difficult to handle diverse embodiments with a unified retargeting architecture. In this work, we propose a novel unified graph-conditioned diffusion-based motion generation framework for retargeting reference motions across diverse embodiments. The intrinsic characteristics of heterogeneous embodiments are represented with graph structure that effectively captures topological and geometrical features of different robots. Such a graph-based encoding further allows for knowledge exploitation at the joint level with a customized attention mechanisms developed in this work. For lacking ground truth motions of the desired embodiment, we utilize an energy-based guidance formulated as retargeting losses to train the diffusion model. As one of the first cross-embodiment motion retargeting methods in robotics, our experiments validate that the proposed model can retarget motions across heterogeneous embodiments in a unified manner. Moreover, it demonstrates a certain degree of generalization to both diverse skeletal structures and similar motion patterns.

Figures

Figures reproduced from arXiv: 2505.20857 by the authors.

Figure 1
Figure 1. Our G-DReaM (Graph-conditional Diffusion Retargeting across Multiple Embodiments) can be uniformly applied to heterogeneous embodiments without requiring their motion data, even when their skeletons are non-homeomorphic. All the motions here are the retargetd results from the same reference. Abstract: Motion retargeting for specific robot from existing motion datasets is one critical step in transferring motion patt… view at source ↗
Figure 2
Figure 2. Overview of G-DReaM. The denoising network is based on a transformer decoder with a noisy motion input Xt under the conditions of the reference motion M and graphs gM,X = {ϕM,X v , ϕM,X e , ψM,X}. The input motion is tokenized at the joint level, where the base joint and other joints are embedded by independent encoders. Then spatial and temporal attentions extract the relationships between all joints and the chrono… view at source ↗
Figure 3
Figure 3. The retargeting motion across multiple robotic embodiments. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Validation of the skeleton generalization from link length aspect. The red links are the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Validation of the skeleton generalization from joint correspondence aspect. The red joints [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Model performance during adaptation on new embodiments. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Retargeted motions (Fight) of both original and new embodiments after adaptation. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Retargeted motions (Kick) of both original and new embodiments after adaptation. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Retargeted motions (Stand up) of both original and new embodiments after adaptation. [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Retargeted motions (Throw) of both original and new embodiments after adaptation. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Retargeted motions (Walk) of both original and new embodiments after adaptation. [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Motion retargeting for the unseen motion (Jump). [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Motion retargeting for the unseen motion (Kick). [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]
Figure 14
Figure 14. Figure 14: Motion retargeting for the unseen motion (Run). [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]
Figure 15
Figure 15. Figure 15: Motion retargeting for the unseen motion (Walk). [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. EquiFusion: Kinematics-Agnostic Human Motion Prediction via Equivariant Latent Diffusion

    cs.CV 2026-07 accept novelty 7.5 of 10

    A permutation-equivariant latent diffusion model treats skeleton connectivity as input, enabling the first kinematics-agnostic stochastic human motion predictor that generalizes zero-shot to unseen and partial skeletons.

Reference graph

Works this paper leans on

42 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    B. v. Marum, A. Shrestha, H. Duan, P. Dugar, J. Dao, and A. Fern. Revisiting reward design and evaluation for robust humanoid standing and walking. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11256–11263, 2024. doi:10. 1109/IROS58592.2024.10802680

    arXiv 2024

  2. [2]

    Z. Li, X. B. Peng, P. Abbeel, S. Levine, G. Berseth, and K. Sreenath. Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control.The International Journal of Robotics Research, page 02783649241285161, 2024

    2024

  3. [3]

    Radosavovic, T

    I. Radosavovic, T. Xiao, B. Zhang, T. Darrell, J. Malik, and K. Sreenath. Real-world humanoid locomotion with reinforcement learning.Science Robotics, 9(89):eadi9579, 2024

    2024

  4. [4]

    Jiang, Z

    K. Jiang, Z. Fu, J. Guo, W. Zhang, and H. Chen. Learning whole-body loco-manipulation for omni-directional task space pose tracking with a wheeled-quadrupedal-manipulator.IEEE Robotics and Automation Letters, 2024

    work page 2024

  5. [5]

    T. He, Z. Luo, W. Xiao, C. Zhang, K. Kitani, C. Liu, and G. Shi. Learning human-to-humanoid real-time whole-body teleoperation. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8944–8951. IEEE, 2024

    2024

  6. [6]

    T. He, Z. Luo, X. He, W. Xiao, C. Zhang, W. Zhang, K. Kitani, C. Liu, and G. Shi. Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning.arXiv preprint arXiv:2406.08858, 2024

    arXiv 2024

  7. [7]

    M. Ji, X. Peng, F. Liu, J. Li, G. Yang, X. Cheng, and X. Wang. Exbody2: Advanced expressive humanoid whole-body control.arXiv preprint arXiv:2412.13196, 2024

    arXiv 2024

  8. [8]

    Ayusawa and E

    K. Ayusawa and E. Yoshida. Motion retargeting for humanoid robots based on simultaneous morphing parameter identification and motion optimization.IEEE Transactions on Robotics, 33(6):1343–1357, 2017

    work page 2017

Show all 42 references
  1. [9]

    T. Yoon, D. Kang, S. Kim, M. Ahn, S. Coros, and S. Choi. Spatio-temporal motion retargeting for quadruped robots.arXiv preprint arXiv:2404.11557, 2024

    2024 arXiv

  2. [10]

    Z. Fu, Q. Zhao, Q. Wu, G. Wetzstein, and C. Finn. Humanplus: Humanoid shadowing and imitation from humans.arXiv preprint arXiv:2406.10454, 2024

    2024 arXiv

  3. [11]

    S. Choi, M. K. Pan, and J. Kim. Nonparametric motion retargeting for humanoid robots on shared latent space. InRobotics: science and systems, 2020

    2020

  4. [12]

    S. Choi, M. J. Song, H. Ahn, and J. Kim. Self-supervised motion retargeting with safety guarantee. In2021 IEEE International Conference on Robotics and Automation (ICRA), pages 8097–8103. IEEE, 2021. 9

    2021

  5. [13]

    K. Hu, C. Ott, and D. Lee. Online human walking imitation in task and joint space based on quadratic programming. In2014 IEEE International Conference on Robotics and Automation (ICRA), pages 3458–3464. IEEE, 2014

    2014

  6. [14]

    Gomes, V

    W. Gomes, V . Radhakrishnan, L. Penco, V . Modugno, J.-B. Mouret, and S. Ivaldi. Humanoid whole-body movement optimization from retargeted human motions. In2019 IEEE-RAS 19th International Conference on Humanoid Robots (Humanoids), pages 178–185. IEEE, 2019

    2019

  7. [15]

    T. Li, J. Won, A. Clegg, J. Kim, A. Rai, and S. Ha. Ace: Adversarial correspondence em- bedding for cross morphology motion retargeting from human to nonhuman characters. In SIGGRAPH Asia 2023 Conference Papers, pages 1–11, 2023

    2023

  8. [16]

    Aberman, P

    K. Aberman, P. Li, D. Lischinski, O. Sorkine-Hornung, D. Cohen-Or, and B. Chen. Skeleton- aware networks for deep motion retargeting.ACM Transactions on Graphics (TOG), 39(4): 62–1, 2020

    2020

  9. [17]

    Y . Yan, E. V . Mascaro, and D. Lee. Imitationnet: Unsupervised human-to-robot motion retar- geting via shared latent space. In2023 IEEE-RAS 22nd International Conference on Humanoid Robots (Humanoids), pages 1–8. IEEE, 2023

    2023

  10. [18]

    T. Wang, R. Liao, J. Ba, and S. Fidler. Nervenet: Learning structured policy with graph neural networks. InInternational conference on learning representations, 2018

    2018

  11. [19]

    Huang, I

    W. Huang, I. Mordatch, and D. Pathak. One policy to control them all: Shared modular policies for agent-agnostic control. InInternational Conference on Machine Learning, pages 4455–4464. PMLR, 2020

    2020

  12. [20]

    Hamilton, Z

    W. Hamilton, Z. Ying, and J. Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017

    2017

  13. [21]

    I. Gat, S. Raab, G. Tevet, Y . Reshef, A. H. Bermano, and D. Cohen-Or. Anytop: Character animation diffusion with any topology, 2025. URLhttps://arxiv.org/abs/2502.17327

    2025 arXiv

  14. [22]

    W. Park, W. Chang, D. Lee, J. Kim, and S.-w. Hwang. Grpe: Relative positional encoding for graph transformer.arXiv preprint arXiv:2201.12787, 2022

    2022 arXiv

  15. [23]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  16. [24]

    Song and S

    Y . Song and S. Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019

    2019

  17. [25]

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based gen- erative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  18. [26]

    Karras, M

    T. Karras, M. Aittala, T. Aila, and S. Laine. Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35:26565–26577, 2022

    2022

  19. [27]

    Dhariwal and A

    P. Dhariwal and A. Nichol. Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780–8794, 2021

    2021

  20. [28]

    Ho and T

    J. Ho and T. Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  21. [29]

    Janner, Y

    M. Janner, Y . Du, J. B. Tenenbaum, and S. Levine. Planning with diffusion for flexible behavior synthesis.arXiv preprint arXiv:2205.09991, 2022. 10

    2022 arXiv

  22. [30]

    C. Lu, H. Chen, J. Chen, H. Su, C. Li, and J. Zhu. Contrastive energy prediction for exact energy-guided diffusion sampling in offline reinforcement learning. InInternational Confer- ence on Machine Learning, pages 22825–22855. PMLR, 2023

    2023

  23. [31]

    C. Guo, S. Zou, X. Zuo, S. Wang, W. Ji, X. Li, and L. 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

  24. [32]

    Tevet, S

    G. Tevet, S. Raab, B. Gordon, Y . Shafir, D. Cohen-Or, and A. H. Bermano. Human motion diffusion model.arXiv preprint arXiv:2209.14916, 2022

    2022 arXiv

  25. [33]

    X. Chen, B. Jiang, W. Liu, Z. Huang, B. Fu, T. Chen, and G. Yu. Executing your commands via motion diffusion in latent space. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18000–18010, 2023

    2023

  26. [34]

    H. Lv. Lafan1 retargeting dataset, n.d. URLhttps://huggingface.co/datasets/ lvhaidong/LAFAN1_Retargeting_Dataset

  27. [35]

    [Online]

    Unitree g1 robot. [Online]. Available:https://www.unitree.com/cn/g1

  28. [36]

    [Online]

    Unitree h1 robot. [Online]. Available:https://www.unitree.com/cn/h1

  29. [37]

    [Online]

    Limx biped robot tron1. [Online]. Available:https://www.limxdynamics.com/en/tron1

  30. [38]

    [Online]

    Limx humanoid cl series. [Online]. Available:https://www.limxdynamics.com/en/

  31. [39]

    [Online]

    Atlas. [Online]. Available:https://bostondynamics.com/

  32. [40]

    [Online]

    Talos. [Online]. Available:https://pal-robotics.com/robot/talos/

  33. [41]

    [Online]

    Cassie. [Online]. Available:https://www.agilityrobotics.com/

  34. [42]

    [Online]

    Valkyrie. [Online]. Available:https://www.nasa.gov/podcasts/ houston-we-have-a-podcast/valkyrie/. 11 Appendix A Training Loss We revisit the loss for the energy-based guided diffusion model here: min θ Ex∼pref(x)En∼N(0,σ 2I)) ∥Dθ(x+n, σ, C)−x∥2 2 +λ∥f kin(Dθ(x+n, σ, C))∥2 2 .(...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.