Pith. sign in

REVIEW 4 major objections 4 minor 33 references

The paper claims that a context-aware pose policy can make humanoid upper-body motion match human reference motion more closely than inverse kinematics and other baselines, while running in about 20 milliseconds per pose.

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 ↗

HuBE is a closed-loop pose-generation framework that produces context-appropriate, human-like upper-body motions for multiple humanoid robots, trained on an LLM-annotated dataset with bone-scaling augmentation.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A plausible, well-engineered pipeline for goal-conditioned human-like upper-body pose generation on humanoids, but the head-to-head superiority claim is not established because the baseline protocol is never specified. the 4 major comments →

arxiv 2508.19002 v1 pith:XTNQ4457 submitted 2025-08-26 cs.RO

HuBE: Cross-Embodiment Human-like Behavior Execution for Humanoid Robots

Shipeng Lyu , Fangyuan Wang , Weiwei Lin , Luhao Zhu , David Navarro-Alarcon , Guodong Guo This is my paper
classification cs.RO
keywords humanoid robotshuman-like motion generationbehavioral appropriatenesspose generationcross-embodimentinverse kinematicsdata augmentationbehavioral cloning
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 reading

The paper proposes HuBE, a method for making humanoid robots move in ways that are both human-like and appropriate to the situation. Its central claim is that a goal-conditioned pose policy, trained on human motion data enriched with detailed text descriptions of what is happening, can produce upper-body poses closer to human reference motion than existing baselines, while running about seventeen times faster than direct inverse kinematics. To transfer across different robot bodies, the method rescales human bone lengths to the robot skeleton during training and then executes the generated poses with a two-step inverse-kinematics solver. The paper also contributes HPose, a dataset of 6D joint poses with situational annotations. If correct, this would make context-aware, transferable human-like behavior practical for humanoid robots with different body proportions.

Core claim

HuBE's central claim is that human-likeness for humanoid robots has two components, similarity to human kinematics and appropriateness to the surrounding context, and that both can be learned by a single closed-loop policy that takes the robot's current upper-body state, the target poses of both hands and the head, and a textual description of the situation, and outputs the next upper-body pose. The paper argues that this removes the structural mismatch between motion generation and execution that occurs in open-loop methods, where a pose is first synthesized without considering the robot's actual configuration and then retargeted. By training on bone-scaled augmented data, the same policy t

What carries the argument

The load-bearing object is the imitation policy πθ(s, g, l) → a, trained by behavioral cloning, together with a training-and-execution loop that keeps the robot in the generation process. The policy receives the robot's current 11-joint upper-body state, a goal pose for the hands and head, and a language description of the situation; the language is encoded with a pretrained tokenizer and fused with a projected version of the state and goal. Hindsight training samples goals from later frames of the same trajectory, letting the policy reach arbitrary target poses rather than only the next recorded frame. Bone scaling rescales directed bone vectors in human mocap data to match a robot's fixed

Load-bearing premise

The comparison assumes the baselines were given the same task inputs as HuBE, namely the current robot state, the target end-effector pose, and the situational text; the paper does not describe how those baselines receive the goal, so if they were run without it, the reported head-to-head metrics measure different tasks and the central outperformance claim collapses.

What would settle it

Give the inverse-kinematics, motion-prediction, and imitation baselines exactly the same inputs HuBE receives, including the current robot state, the target hand-and-head pose, and the situational text, then measure human-similarity error on the same test actions. If the motion-prediction or imitation baseline matches or beats HuBE's reported human-similarity MPJPE under equal conditioning, the outperformance is an artifact of unequal task setup rather than a property of the framework.

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

If this is right

  • A single policy trained once can execute upper-body behaviors on more than one humanoid body without per-robot retraining or shape adaptation, as long as the robot's bone lengths fall within the augmented training distribution.
  • Motion generation and execution can share a closed loop, so the generated pose always respects the robot's current state, avoiding the human-likeness degradation typical of open-loop pipelines.
  • The same end-effector goal can produce different full-body poses depending on the extra textual context, enabling socially appropriate behavior in different situations with nearly identical kinematic targets.
  • At roughly 0.02 seconds per pose, the method is fast enough for online use, unlike search-based inverse kinematics, making human-like motion practical in real-time interaction loops.
  • The HPose dataset with 6D joint poses and situational annotations provides a training and evaluation resource for context-aware pose generation beyond this specific framework.

Where Pith is reading between the lines

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

  • Editorial inference: because the random-frequency hindsight training accounts for a large share of the accuracy gains independently of the text, a follow-up should test HuBE on situations that the LLM annotator never paraphrased, to see whether appropriateness transfers beyond language variety.
  • Editorial inference: the reported head-to-head numbers depend on giving every baseline the same inputs. A clean rerun that conditions the inverse-kinematics, motion-prediction, and imitation baselines on the same goal pose and situational text would settle whether the margins are a property of the method or of the evaluation setup.
  • Editorial inference: the bone-scaling augmentation covers only upper-body fixed and floating bones; extending it to legs, hips, and spine could make the same policy produce full-body expressive behavior, which the paper explicitly identifies as future work.
  • Editorial inference: a deployment test on a humanoid whose kinematic parameters were excluded from the nine-robot augmentation set would separate true cross-embodiment generalization from memorization of the training robot shapes.
Share X Bluesky LinkedIn Reddit HN

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 HuBE, a bi-level closed-loop framework for generating human-like upper-body motions on humanoid robots. The method conditions a pose-generation policy on the current robot state, a next-frame end-effector goal pose, and a situational language annotation, then maps the generated poses to robot joint commands via a two-step constrained inverse-kinematics solver. To support training, the authors construct the HPose dataset by reannotating KIT, AMASS, and Motion-X with GPT-4o generated situational text, and they introduce a bone-scaling data augmentation procedure to adapt human mocap to heterogeneous humanoid skeletons. The paper reports experiments on three humanoid platforms and claims significant improvements over IK, siMLPe, and HumanPlus in motion similarity, behavioral appropriateness, and computational efficiency, as well as millimeter-level cross-embodiment compatibility.

Significance. If the empirical claims were fully substantiated, HuBE would be a useful contribution to context-aware humanoid motion generation: it explicitly couples situational semantics with goal-conditioned pose prediction, and the bone-scaling augmentation is a sensible way to expose a policy to multiple robot morphologies without per-robot retraining. The paper also presents a large reannotated dataset, which could be valuable to the community. However, the current manuscript does not provide enough evidence to support the headline claims. The baseline protocol in the comparative study is unspecified, the appropriateness metrics rely on language annotations that are generated from the very motion frames used as ground truth, and several tables contain entry-level errors and lack statistical measures. These issues affect the central conclusion and require substantial revision.

major comments (4)
  1. [Section IV-C, Table IV] The head-to-head comparison against IK, siMLPe, and HumanPlus is not interpretable because the paper never states what inputs these baselines receive. HuBE's policy is trained on tuples (s_i, g_i, l_i, a_i), where g_i is the next-frame end-effector pose and l_i is a situational text. For the comparison to be fair, each baseline must be given the same current state, the same goal pose, and the same situational context. The paper does not describe how siMLPe (a motion-prediction model), HumanPlus (an imitation-from-video policy), or the IK solver are conditioned on the goal g and the text l. If these baselines were evaluated without the same goal or context, then the reported numbers, e.g., HuBE H-S MPJPE 0.0337 versus siMLPe 0.138 for Dancing, reflect task difficulty rather than method quality. This is a load-bearing omission for the central claim of 'significant improvement over SOTA bas
  2. [Section III-A and Section IV] The behavioral-appropriateness evaluation is circular. The situational text l is generated by GPT-4o from two frames of the very motion sequence that later serves as the ground-truth action a_i, and the same l is used both as a conditioning input during policy training and as the query for MM-Dist and R-Precision at evaluation time. This setup measures retrieval of in-distribution text-motion pairs, not whether the generated behavior is appropriate for an independently specified situation. To support the appropriateness claim, the authors should use held-out language annotations from unseen motions, ideally human-written or produced by an LLM without access to the target frames, and should consider a perceptual study with human raters. FMD is also defined on the premise that the dataset motions are appropriate, which is an assumption that should be stated and defended.
  3. [Table III] Table III contains several errors that prevent verification of the ablation study. The row labels 'MLP(w)0', 'MLP(w/o)1', and 'Transformer(w)3' appear to contain stray digits; the H-S MPJPE value '0.45' in the MLP(w/o) Random row is almost certainly a typo for 0.045, yet it is used as the reported result. More substantively, the table reports only point estimates with no error bars, no repeated seeds, and no statistical significance tests. Since the random-frequency ablation is the main justification for the hindsight-training design, the authors should provide variance estimates and, where possible, significance tests across seeds.
  4. [Section IV-A, Table II, Fig. 5, Table IV] The claim of 'millimeter-level cross-embodiment compatibility' is supported only by the bone-size stability results in Table II (three humanoids) and by Fig. 5, which shows retargeting accuracy without a quantitative table or error bars. Meanwhile, Table IV contains a garbled entry ('294' in the Playing/HumanPlus H-S MPJOE cell, presumably 2.94) and omits FMD, MM-Dist, and R-Precision for IK and HumanPlus, so the appropriateness comparison is effectively only against siMLPe. The paper should provide a complete, corrected table with standard deviations, and should report the retargeting-accuracy numbers that underlie Fig. 5 so that the cross-embodiment claim can be assessed.
minor comments (4)
  1. [Throughout] There are numerous typos and grammatical errors, e.g., 'an pivotal role' in the Introduction, 'Sequently' in the Introduction, and 'i.g.' instead of 'e.g.' in Section IV-A. The manuscript should be carefully proofread.
  2. [Section III-C, Eq. (5)] The notation in Eq. (5) is unclear: the forward-kinematics function should be written consistently (e.g., FK_i(J) rather than f_i^FK), and the second equation uses (J_torso^*, J_arms) as arguments without defining the concatenation order.
  3. [Section IV-B] The ablation study is performed on the KIT ML subset because of its 240 Hz frequency, but the full HPose dataset contains mixed frequencies from KIT, AMASS, and Motion-X. The text should explain how the mixed-frequency data are handled during training and whether the frequency ablation transfers to the full dataset.
  4. [Table III and Table IV] The table captions and labels should be fixed: Table III is quantitative, not merely 'qualitative'; the model labels ('Ous') in Table IV should be corrected to 'Ours'; and all numerical values should be formatted consistently with the same number of leading zeros.

Circularity Check

2 steps flagged

Appropriateness and cross-embodiment claims reduce to dataset construction: LLM-generated text labels come from the ground-truth motions used as training targets, and robot bone parameters R are injected into the pseudo-GT training set and then reported as 'millimeter-level' stability on those same robots.

specific steps
  1. self definitional [Section III-A (Data Definition) and Section IV (evaluation metrics: MM-Dist / R-Precision / FMD)]
    "To enrich the contextual situation l of human motions, we use the LLM (GPT-4o) to generate this description of given motion sequence with the annotation ˆl in the original dataset. ... the LLM generates a rich and detailed l ... · Multimodal Distance (MM-Dist): calculate average distances between situation context and generated motions. · R-Precision: quantify the semantic discriminability of generated motions by establishing a text-to-motion cross-modal retrieval task. ... This metric works based on the premise that the human demonstrated motion in dataset is appropriate."

    The contextual label l is produced by GPT-4o from the very ground-truth motion frames that the policy is trained to output. Behavioral appropriateness is then measured as MM-Dist/R-Precision between generated motions and l (and FMD assumes the dataset motion is appropriate). Because l is a paraphrase of the ground-truth action, any policy that reconstructs the ground truth will automatically achieve low MM-Dist and high R-Precision. The 'appropriateness' result is therefore defined into the data construction rather than tested against an independent situational context.

  2. fitted input called prediction [Section III-A (Data Augmentation, Eq. 1) and Section IV-A (Bone-size stability, Table II)]
    "we get the augmented dataset H′ (pseudo ground truth, Pseudo GT) through a bone scaling operation by updating the bone size with collected humanoid robot body parameters R = {ri, rk} ... R are collected from 9 typical humanoid robots, such as Unitree H1 and PAL's TALOS. ... we add the augmented the pseudo GT to re-train the policy. Fortunately, the generation error is optimized to the millimeter level."

    The robot bone parameters R are the inputs that synthesize the pseudo-GT training examples. The bone-size stability experiment then measures generated bone lengths for three of those same robots (GR1, TALOS, G1 in Table II) and reports millimeter-level error. This verifies that the policy reproduced the bone lengths encoded during augmentation; it is not a prediction for unseen morphologies. The 'millimeter-level cross-embodiment compatibility' is forced by construction because the test robots' bone parameters were used to build the training distribution.

full rationale

The paper contains no self-citation chain or imported uniqueness theorem; the MLP/Transformer pose policy, hindsight training, and two-step CLIK execution are genuine engineering contributions with independent supervised-learning content, and the E-A/H-S similarity numbers are a legitimate check of held-out-motion reconstruction. However, two headline claims are partially circular. (1) Behavioral appropriateness: the situational text l is generated from the ground-truth motion frames used as targets, and the appropriateness metrics (MM-Dist, R-Precision, FMD) then measure closeness to that same ground truth, so the 'appropriateness' gain is a property of the annotation pipeline rather than an external contextual test. (2) Cross-embodiment compatibility: robot bone parameters R are used to synthesize pseudo-GT data, and the 'millimeter-level' bone-size stability is measured on robots from that same R; this is a training-distribution fit, not an out-of-morphology prediction. The comparative evaluation in Section IV-C also omits how IK, siMLPe, and HumanPlus are conditioned on goal g and text l; that threatens the head-to-head superiority claim but is a benchmarking-validity issue, not a circularity step. On balance, the central derivation is not entirely reduced to inputs; score 6 reflects partial circularity of the appropriateness and cross-embodiment evaluations.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

No new physical or theoretical entities are introduced. HPose is a dataset, not a postulated entity. The only fitted quantity is the learned policy, which is standard for machine learning systems.

free parameters (1)
  • Policy network parameters theta (MLP/Transformer) = learned from HPose dataset via Eq. 2
    All reported similarity metrics come from this fitted policy; specific weight values are not disclosed.
axioms (4)
  • domain assumption Human demonstrations in KIT, AMASS, Motion-X are behaviorally appropriate.
    The FMD metric 'works based on the premise that the human demonstrated motion in dataset is appropriate' (Section IV).
  • domain assumption The 11-joint upper-body reduction with quaternion rotations is sufficient to represent human-like behavior.
    Section III-A defines the simplified chain; the conclusion admits lower-body movements are excluded, so the full-body claim is not covered.
  • ad hoc to paper Scaling bone lengths from human to robot skeletons preserves the naturalness and appropriateness of the motion.
    Eq. 1 and Algorithm 1 produce Pseudo GT by repositioning joints along bone directions only; no joint-limit or dynamics constraints are included.
  • domain assumption GPT-4o descriptions are valid ground-truth context for motion.
    The LLM annotations are used both as training labels (Eq. 3) and as evaluation queries (MM-Dist, R-Precision), so their quality bounds all appropriateness numbers.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of HuBE: Cross-Embodiment Human-like Behavior Execution for Humanoid Robots." pith.science (2026). https://pith.science/paper/XTNQ4457

@misc{pith2026250819002,
  author       = {Pith},
  title        = {Pith review of: HuBE: Cross-Embodiment Human-like Behavior Execution for Humanoid Robots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XTNQ4457}},
  note         = {Machine review of arXiv:2508.19002}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Achieving both behavioral similarity and appropriateness in human-like motion generation for humanoid robot remains an open challenge, further compounded by the lack of cross-embodiment adaptability. To address this problem, we propose HuBE, a bi-level closed-loop framework that integrates robot state, goal poses, and contextual situations to generate human-like behaviors, ensuring both behavioral similarity and appropriateness, and eliminating structural mismatches between motion generation and execution. To support this framework, we construct HPose, a context-enriched dataset featuring fine-grained situational annotations. Furthermore, we introduce a bone scaling-based data augmentation strategy that ensures millimeter-level compatibility across heterogeneous humanoid robots. Comprehensive evaluations on multiple commercial platforms demonstrate that HuBE significantly improves motion similarity, behavioral appropriateness, and computational efficiency over state-of-the-art baselines, establishing a solid foundation for transferable and human-like behavior execution across diverse humanoid robots.

Figures

Figures reproduced from arXiv: 2508.19002 by David Navarro-Alarcon, Fangyuan Wang, Guodong Guo, Luhao Zhu, Shipeng Lyu, Weiwei Lin.

Figure 1
Figure 1. Figure 1: Contextual semantics dictate human-like action execution by influencing behavioral appropriateness. Two cases to show the influ￾ence of the contextual situation for performing human-like actions. Case I (“hugging friend”) requires an open-arm posture with elevated elbows to express affection, while Case II (“boxing with friend”) demands guarded elbows for defensive intent, despite nearly identical end-effe… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the whole algorithm. This algorithm includes three parts, i.e., building dataset, model training and algorithm implementation. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The human skeletal system introduction. a) The composition of the [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The generation results for bone-size stability. The first row is the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The retargeting accuracy between generated actions and the robot’s [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: The results of ablation for observation input in our framework. The [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: An example of action-wise results (“Playing Kung Fu”) performed by the humanoid robot G1. Motion sequences are shown from left to right. From [PITH_FULL_IMAGE:figures/full_fig_p007_8.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

33 extracted references · 24 canonical work pages · 1 internal anchor

  1. [1]

    A human-like upper-limb motion planner: Generating naturalistic movements for humanoid robots,

    G. Gulletta, E. Silva, W. Erlhagen, R. Meulenbroek, M. Costa, and E. Bi- cho, “A human-like upper-limb motion planner: Generating naturalistic movements for humanoid robots,” International Journal of Advanced Robotic Systems, vol. 18, no. 2, p. 1729881421998585, 2021

    work page 2021

  2. [2]

    An unwillingness to act: behavioral appropriateness, situational constraint, and self-efficacy in shyness,

    G. Hill, “An unwillingness to act: behavioral appropriateness, situational constraint, and self-efficacy in shyness,” Journal of Personality, vol. 57, no. 4, pp. 871–890, 1989

    work page 1989

  3. [3]

    Dexmimicgen: Automated data generation for bimanual dexterous manipulation via imitation learning,

    Z. Jiang, Y . Xie, K. Lin, Z. Xu, W. Wan, A. Mandlekar, L. Fan, and Y . Zhu, “Dexmimicgen: Automated data generation for bimanual dexterous manipulation via imitation learning,” 2024

    work page 2024

  4. [4]

    Tantangan linguistik dalam pengimplementasian big data berbahasa indonesia pada robot humanoid: Tinjauan dan rekomendasi,

    S. Aisyah, T. Eugene, E. Nurhayati et al., “Tantangan linguistik dalam pengimplementasian big data berbahasa indonesia pada robot humanoid: Tinjauan dan rekomendasi,” Jurnal Pendidikan Bahasa dan Sastra Indonesia, vol. 1, no. 1, pp. 9–9, 2025

    work page 2025

  5. [5]

    Kinematic motion retargeting via neural latent optimization for learning sign language,

    H. Zhang, W. Li, J. Liu, Z. Chen, Y . Cui, Y . Wang, and R. Xiong, “Kinematic motion retargeting via neural latent optimization for learning sign language,” IEEE Robotics and Automation Letters , vol. 7, no. 2, pp. 4582–4589, 2022

    work page 2022

  6. [6]

    Bc-z: Zero-shot task generalization with robotic imitation learning,

    E. Jang, A. Irpan, M. Khansari, D. Kappler, F. Ebert, C. Lynch, S. Levine, and C. Finn, “Bc-z: Zero-shot task generalization with robotic imitation learning,” in Conference on Robot Learning . PMLR, 2022, pp. 991–1002. TABLE IV THE ACTION -WISE RESULTS FOR DIFFERENT METHODS MPJPE MPJOEAction Model E-A H-S↓ E-A H-S↓ FMD↓ MM-Dist↓ R-Precision↑ IK 0.00832 0 ...

    work page 2022

  7. [7]

    Unsupervised motion retargeting for human-robot imitation,

    L. Annabi, Z. Ma, and S. Nguyen, “Unsupervised motion retargeting for human-robot imitation,” in Companion of the 2024 ACM/IEEE International Conference on Human-Robot Interaction , 2024, pp. 204– 208

    work page 2024

  8. [8]

    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

    Pith/arXiv arXiv 2024

  9. [9]

    The kit motion-language dataset,

    M. Plappert, C. Mandery, and T. Asfour, “The kit motion-language dataset,” Big data, vol. 4, no. 4, pp. 236–252, 2016

    work page 2016

  10. [10]

    Ex- pressive whole-body control for humanoid robots,

    X. Cheng, Y . Ji, J. Chen, R. Yang, G. Yang, and X. Wang, “Ex- pressive whole-body control for humanoid robots,” arXiv preprint arXiv:2402.16796, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Inverse kinematics for a 7-dof humanoid robotic arm with joint limit and end pose coupling,

    R. Dou, S. Yu, W. Li, P. Chen, P. Xia, F. Zhai, H. Yokoi, and Y . Jiang, “Inverse kinematics for a 7-dof humanoid robotic arm with joint limit and end pose coupling,” Mechanism and Machine Theory , vol. 169, p. 104637, 2022

    work page 2022

  12. [12]

    Humor: 3d human motion model for robust pose estimation,

    D. Rempe, T. Birdal, A. Hertzmann, J. Yang, S. Sridhar, and L. Guibas, “Humor: 3d human motion model for robust pose estimation,” in International Conference on Computer Vision (ICCV) , 2021

    work page 2021

  13. [13]

    Omnicontrol: Control any joint at any time for human motion generation,

    Y . Xie, V . Jampani, L. Zhong, D. Sun, and H. Jiang, “Omnicontrol: Control any joint at any time for human motion generation,” arXiv e- prints, pp. arXiv–2310, 2023

    work page 2023

  14. [14]

    Action-conditioned 3d human motion synthesis with transformer vae,

    M. Petrovich, M. Black, and G. Varol, “Action-conditioned 3d human motion synthesis with transformer vae,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 10 985–10 995

    work page 2021

  15. [15]

    Executing your commands via motion diffusion in latent space,

    X. Chen, B. Jiang, W. Liu, Z. Huang, B. Fu, T. Chen, and G. Yu, “Executing your commands via motion diffusion in latent space,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 18 000–18 010

    2023

  16. [16]

    Human motion diffusion as a generative prior,

    Y . Shafir, G. Tevet, R. Kapon, and A. Bermano, “Human motion diffusion as a generative prior,” in The Twelfth International Conference on Learning Representations , 2024

    work page 2024

  17. [17]

    Guided motion diffusion for controllable human motion synthesis,

    K. Karunratanakul, K. Preechakul, S. Suwajanakorn, and S. Tang, “Guided motion diffusion for controllable human motion synthesis,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 2151–2162

    2023

  18. [18]

    Whole-body humanoid robot locomotion with human reference,

    Q. Zhang, P. Cui, D. Yan, J. Sun, Y . Duan, G. Han, W. Zhao, W. Zhang, Y . Guo, A. Zhanget al., “Whole-body humanoid robot locomotion with human reference,” arXiv preprint arXiv:2402.18294 , 2024

    Pith/arXiv arXiv 2024

  19. [19]

    I-CTRL: Imitation to Control Humanoid Robots Through Constrained Reinforcement Learning

    Y . Yan, E. Mascaro, T. Egle, and D. Lee, “I-ctrl: Imitation to control humanoid robots through constrained reinforcement learning,” arXiv preprint arXiv:2405.08726, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    Okami: Teaching humanoid robots manipulation skills through single video imitation,

    J. Li, Y . Zhu, Y . Xie, Z. Jiang, M. Seo, G. Pavlakos, and Y . Zhu, “Okami: Teaching humanoid robots manipulation skills through single video imitation,” arXiv preprint arXiv:2410.11792 , 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Open-television: Teleoperation with immersive active visual feedback,

    X. Cheng, J. Li, S. Yang, G. Yang, and X. Wang, “Open-television: Teleoperation with immersive active visual feedback,” arXiv preprint arXiv:2407.01512, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Humanplus: Humanoid shadowing and imitation from humans,

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

    Pith/arXiv arXiv 2024

  23. [23]

    Judgments of social appropriateness by children and adolescents with autism,

    K. Loveland, D. Pearson, B. Tunali-Kotoski, J. Ortegon, and M. Gibbs, “Judgments of social appropriateness by children and adolescents with autism,” Journal of Autism and Developmental Disorders , vol. 31, pp. 367–376, 2001

    work page 2001

  24. [24]

    Learning socially appropriate robot approaching behavior toward groups using deep reinforcement learning,

    Y . Gao, F. Yang, M. Frisk, D. Hemandez, C. Peters, and G. Castel- lano, “Learning socially appropriate robot approaching behavior toward groups using deep reinforcement learning,” in 2019 28th IEEE interna- tional conference on robot and human interactive communication (RO- MAN). IEEE, 2019, pp. 1–8

    work page 2019

  25. [25]

    Perceived appropriateness and intention to use service robots in tourism,

    S. Ivanov and C. Webster, “Perceived appropriateness and intention to use service robots in tourism,” in Information and Communication Tech- nologies in Tourism 2019: Proceedings of the International Conference in Nicosia, Cyprus, January 30–February 1, 2019 . Springer, 2019, pp. 237–248

    work page 2019

  26. [26]

    Perceived appropriateness: A novel view for remediating per- ceived inappropriate robot navigation behaviors,

    Y . Zhou, “Perceived appropriateness: A novel view for remediating per- ceived inappropriate robot navigation behaviors,” in Companion of the 2023 ACM/IEEE International Conference on Human-Robot Interaction, 2023, pp. 781–783

    work page 2023

  27. [27]

    Amass: Archive of motion capture as surface shapes,

    N. Mahmood, N. Ghorbani, N. Troje, G. Pons-Moll, and M. Black, “Amass: Archive of motion capture as surface shapes,” in Proceedings of the IEEE/CVF international conference on computer vision , 2019, pp. 5442–5451

    work page 2019

  28. [28]

    Motion-x: A large-scale 3d expressive whole-body human motion dataset,

    J. Lin, A. Zeng, S. Lu, Y . Cai, R. Zhang, H. Wang, and L. Zhang, “Motion-x: A large-scale 3d expressive whole-body human motion dataset,” Advances in Neural Information Processing Systems , vol. 36, pp. 25 268–25 280, 2023

    work page 2023

  29. [29]

    Reinforcement learning: An introduction,

    R. Sutton, “Reinforcement learning: An introduction,” A Bradford Book, 2018

    work page 2018

  30. [30]

    Behavioral cloning from observa- tion,

    F. Torabi, G. Warnell, and P. Stone, “Behavioral cloning from observa- tion,” arXiv preprint arXiv:1805.01954 , 2018

    Pith/arXiv arXiv 2018

  31. [31]

    Hindsight experience replay,

    M. Andrychowicz, F. Wolski, A. Ray, J. Schneider, R. Fong, P. Welinder, B. McGrew, J. Tobin, P. Abbeel, and W. Zaremba, “Hindsight experience replay,” Advances in neural information processing systems , vol. 30, 2017

    work page 2017

  32. [32]

    Back to mlp: A simple baseline for human motion prediction,

    W. Guo, Y . Du, X. Shen, V . Lepetit, X. Alameda-Pineda, and F. Moreno- Noguer, “Back to mlp: A simple baseline for human motion prediction,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2023, pp. 4809–4819

    work page 2023

  33. [33]

    Motiondiffuse: Text-driven human motion generation with diffusion model,

    M. Zhang, Z. Cai, L. Pan, F. Hong, X. Guo, L. Yang, and Z. Liu, “Motiondiffuse: Text-driven human motion generation with diffusion model,” IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 6, pp. 4115–4128, 2024

    work page 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.