Pith. sign in

REVIEW 4 major objections 3 minor 1 cited by

BiomechGPT: Extending Motion-Language Models to Clinical Motion Understanding

T0 review · 4 major / 3 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 This paper claims that a single multimodal motion-language model, trained on biomechanically parameterized movements, can answer detailed clinical questions about gait and movement.

desk verdict A genuinely useful clinical motion-LLM study, but the 'measuring walking' claim needs calibration data and the abstract overstates the dataset. read the letter →

arxiv 2505.18465 v2 pith:ZI2EE5WV submitted 2025-05-24 cs.CV

Ruize Yang , Ann Kennedy , R. James Cotton This is my paper
classification cs.CV
keywords biomechanicsmotion-languagemodelclinicalmotionunderstandingVQ-VAEtokenizationmarkerlesscapturetransferlearningrehabilitationmultimodalLLM
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 asks whether a single multimodal language model, trained on tokenized biomechanical movement, can answer the kinds of clinical questions a clinician would pose about how a person moves. The authors assembled a large dataset of clinical movement trials from hundreds of participants, many with gait impairments, and built BiomechGPT by fine-tuning a 4-billion-parameter language model on question-answer pairs over VQ-VAE motion tokens. Their central claim is that this model can solve detailed clinical tasks, reporting high F1 scores for activity, impairment, diagnosis, assistive device use, and fall history, plus strong correlations for walking speed and cadence. If that holds, a single flexible model could replace bespoke pipelines for many movement-analysis questions and scale with additional data. The paper also reports positive transfer learning, meaning adding tasks improves performance on existing ones.

What carries the argument

The load-bearing component is a VQ-VAE motion tokenizer that converts 34-dimensional joint-angle trajectories from a biomechanical model into discrete codebook tokens, plus a cross-format tokenizer that maps heterogeneous motion formats into the same latent space without paired data. The tokens are then treated as a foreign language: the language model's vocabulary is extended with motion tokens, and QLoRA supervised fine-tuning teaches the model to answer clinical questions from the token sequence. The tokenizer fixes one representation for all downstream tasks, which is what makes a single generalist model possible.

What would settle it

Re-run the assistive-device and fall-history tasks with a tokenizer that preserves wrist, forearm, and foot degrees of freedom and reduces reconstruction error; if F1 does not improve and diagnosis accuracy on a larger held-out set stays near chance, the claim that the tokenized representation carries the needed clinical information is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a multimodal biomechanics-language model trained on a large dataset of movements parameterized with a biomechanical model can solve detailed, clinically meaningful movement tasks. Concretely, BiomechGPT, a 4-billion-parameter model fine-tuned with QLoRA on roughly 27k tokenized motion-text pairs, outperforms a token-histogram XGBoost baseline on all classification tasks and achieves Pearson correlations of 0.92 for cadence and 0.91 for walking speed against instrumented-walkway ground truth. The authors interpret positive transfer as evidence that a generalist model can learn shared features of human movement across tasks.

Load-bearing premise

The clinical signal survives the full pipeline: markerless reconstruction, token compression with roughly 5.78 degrees per-joint RMSE, and the deliberate zeroing of wrist, forearm, and foot metatarsal degrees of freedom.

Editorial extensions

If this is right

  • A single model can replace task-specific pipelines for activity recognition, impairment detection, diagnosis, device use, fall history, and walking metrics.
  • Adding more annotated tasks improves performance on existing tasks, so denser annotation of the same motion dataset should raise accuracy across all tasks.
  • Larger base models improve results, so scaling model and dataset together is a plausible route to a rehabilitation foundation model.
  • The same tokenized representation enables a chat interface for clinicians to ask questions about individual movement trials.
  • Regression abilities emerge from standard language-model training without a dedicated regression head.

Reading between the lines

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

  • Restoring the wrist, forearm, and foot degrees of freedom that were zeroed out would likely sharpen assistive-device classification, since hand-held devices leave cues in those joints; this is a testable variation of the paper's setup.
  • The cross-format tokenizer's ability to pool disparate datasets without paired data makes the approach portable to home or smartphone capture, provided reconstruction quality stays high enough to preserve clinical signal.
  • The very small test sets for TUG, FSST, and fall history mean the reported numbers are unstable, so a larger held-out cohort is the natural next test of the central claim.
  • If positive transfer holds as tasks are added, the model may eventually draw on its language pretraining for zero-shot or open-ended clinical reasoning, though the paper does not demonstrate that.
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 / 3 minor

Summary. The paper presents BiomechGPT, a multimodal motion-language model built by fine-tuning Gemma (1B and 4B) with QLoRA to consume VQ-VAE tokenized 34-joint biomechanical trajectories and answer clinical questions. The authors curated roughly 34 hours of markerless motion-capture data from about 467 participants, many with gait impairments, and generated prompt-answer pairs for activity recognition, impairment detection, diagnosis, assistive-device classification, fall history, and regression tasks (cadence, walking speed, TUG, FSST). They report strong classification F1 scores (activity 0.95, impairment 0.88), high Pearson correlations for cadence and speed (0.92 and 0.91 against instrumented-walkway ground truth), positive transfer when training on more tasks, and improvements over a token-histogram XGBoost baseline. The paper argues that a single biomechanics-language model can serve as a generalist tool for clinical motion analysis.

Significance. If the results hold, the paper provides a valuable demonstration that a fine-tuned LLM can jointly handle multiple clinical motion-understanding tasks on biomechanically parameterized inputs, with positive transfer across tasks and scaling with model size. Strengths include participant-level train/test splitting, external walkway ground truth for speed and cadence, multiple training seeds, a non-LLM baseline, and explicit reporting of tokenizer reconstruction error and limitations. These design choices make the central classification results credible. However, the headline claim of 'measuring walking' is currently supported only by Pearson correlations, which do not establish absolute accuracy, and the small test sets for several regression tasks undermine the strength of the claims. The abstract/body inconsistencies also need correction before the paper can be accepted.

major comments (4)
  1. [§3.2, Table 3, Fig. 4] The claim that BiomechGPT can 'measure walking speed' is not supported by the input representation: Section 3.2 states that the encoder discards pelvis position and velocity, and Section 3.1 states that body-shape parameters β are collected but not used. Absolute speed in m/s is the product of cadence and stride length, and stride length in meters requires body-segment scale; with no scale in the input, the model can at best estimate cadence and a normalized stride length scaled by an assumed leg length. A constant multiplicative error would leave Pearson's r unchanged, so the reported r=0.91 (Table 3) does not establish absolute accuracy. Please report slope, intercept, RMSE, and limits of agreement (or Bland-Altman), or reframe the task as cadence/normalized speed.
  2. [§4.2, Table 1, Table 3] The sentence 'In all cases, the regression slopes were highly significant (p<1e-10)' is arithmetically inconsistent with the test sizes in Table 1: FSST has 7 test samples and TUG has 18 test samples, for which r=0.90 and r=0.68 cannot yield p<1e-10. Report the actual test statistics, degrees of freedom, and p-values for each regression, and interpret TUG/FSST correlations cautiously given the 8- and 5-participant test sets.
  3. [Abstract vs. §1, §3.1, §3.2] The abstract states the dataset contains 71 hours from 750 participants and that a 'cross-format tokenizer' pools a second dataset without paired data, but the main text reports 34 hours from 467 participants (Introduction and Discussion) and Section 3.2 describes only a standard VQ-VAE tokenizer; no cross-format method, second dataset, or paired/unpaired pooling is defined anywhere in the body. These discrepancies need to be resolved or the abstract corrected, because the cross-format tokenizer is advertised as a contribution and the data scale is a central claim.
  4. [§4.1, Table 2] The fall-prediction evaluation rests on only 2 test participants (33 samples), and the authors note that test performance 'frequently exceeded the training performance'; with two participants, the reported F1=0.83 cannot be separated from participant-level idiosyncrasy. Either report per-participant results and confidence intervals or remove the strong fall-prediction claim from the abstract and conclusions.
minor comments (3)
  1. [§3.1] The first paragraph contains an incomplete sentence: 'such as which also involve sitting in a chair' should be completed or removed.
  2. [§4.2] The text refers to 'For Squared Step Test'; the standard name is 'Four Square Step Test' (as used in Table 1). Please harmonize the terminology.
  3. [§4.2, Fig. 4] The regression figures would be more informative with the line of identity, the fitted slope and intercept, and the RMSE; currently the reader cannot distinguish calibration from rank-order agreement.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: all reported results are empirical evaluations on held-out participants, and the walking speed/cadence ground truth is an external instrumented-walkway measurement; the paper's self-citations are methodological rather than load-bearing.

full rationale

The paper's derivation chain is: multiview video -> markerless biomechanical reconstruction (modified LocoMujoco/OpenSim, fit 'similarly to [24]') -> VQ-VAE tokenized joint angles -> QLoRA fine-tuned Gemma -> task outputs. No step defines its output in terms of its input. Classification labels (activity, impairment, diagnosis, assistive device, falls history) are external clinical/administrative annotations on the collected dataset, not functions of the model or tokenizer, so the benchmarks are internal but not circular. Regression targets (cadence, walking speed, TUG, FSST) come from an instrumented walkway or manual timing, and the paper explicitly notes that this ground truth is 'independent of our motion-capture system'; thus the reported Pearson correlations are not fitted-inputs-renamed-as-predictions. The only self-citations ([3], [24], [31]) involve the first author's markerless motion-capture methodology and are used as a practical reconstruction recipe, not as an unverified uniqueness theorem or forced ansatz, so they do not make the central claim circular. The paper's own flagged limitations—zeroing wrist/metatarsal DOFs 'adversely impacted' assistive-device classification (Sec. 3.2), tokenizer RMSE of 5.78 degrees and possible loss of 'subtle motions differences' (Sec. 4.3), and very small test sets for falls (n=2), TUG (n=8), and FSST (n=5)—are validity, calibration, and reproducibility concerns, not circularity. Likewise, the skeptic concern that absolute walking speed is unidentifiable from scale-free joint-angle tokens is an input-information or calibration issue, not a case where the prediction reduces to its training target by construction. The abstract's mention of a cross-format tokenizer and pooled second dataset is not described in the main text, which is a completeness inconsistency, but it does not create a derivation-is-input loop.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The central results rest on a five-stage pipeline: markerless reconstruction, 34-DOF joint-angle parameterization, VQ-VAE tokenization, Gemma-3 fine-tuning, and template-based answer parsing. None of the stages is independently validated in this paper except the tokenizer's aggregate RMSE, and the clinical labels are of unstated reliability. The ledger records the hand-chosen hyperparameters that set the operating point plus the abstract's unimplemented cross-format tokenizer claim. No new physical entities are introduced.

free parameters (6)
  • VQ-VAE codebook size K = 512
    Hand-chosen; Section 4.3 notes varying token count between 512 and 1024 changed results only within refit noise. All downstream F1 and correlation numbers depend on this codebook.
  • VQ-VAE downsampling factor l = 4
    Hand-chosen; the paper tested l=16 and saw no systematic effect within noise (Section 4.3). Determines the token sequence length fed to the LLM.
  • Commitment loss weight beta_commit = 0.02
    Standard VQ-VAE choice from prior work; affects token usage entropy and the reported 5.78-degree reconstruction RMSE.
  • Training epochs = 1.2
    Chosen because 'early experiments showed evidence of overfitting much beyond one epoch' (Section 3.3); a model-selection choice affecting every reported number.
  • QLoRA rank, alpha, dropout = 32, 32, 0.1 (ablation: 16, 16, 0.005)
    Follows MotionLLM; the rank-16 variant showed no clear impact on the 1B model (Section 4.3), indicating a hand-tuned hyperparameter rather than a fitted constant.
  • Zeroed joint degrees of freedom = Wrist flexion, deviation, supination; foot metatarsal
    Preprocessing choice in Section 3.2 that the paper links to degraded assistive-device classification; it removes information from the model input.
assumptions (5)
  • domain assumption Markerless multiview reconstruction fitted to the modified LocoMujoco/OpenSim biomechanical model yields accurate 34-DOF joint-angle trajectories for all participants.
    Invoked in Section 3.1 ('fit our markerless motion capture data similarly to [24]'); no within-dataset validation against marker-based ground truth is reported, so reconstruction error is unquantified for these participants.
  • domain assumption Clinical ground-truth labels (impairment status, diagnosis, assistive device, falls history) are correct.
    Labels are the supervision signal used to generate the Q-A dataset (Section 3.1). How falls history was collected (for example self-report) and inter-rater reliability are not described; noisy labels would make reported F1 an upper bound.
  • domain assumption The VQ-VAE codebook with 5.78 degrees per-joint RMSE preserves task-relevant movement differences.
    Explicitly flagged by the authors in Section 4.3: reconstruction error 'might fail to encode some subtle motions differences,' which would cap downstream task performance. This is the weakest link between raw data and all reported results.
  • domain assumption Gemma 3's pretrained language knowledge transfers to clinical biomechanics question answering.
    The model inherits Gemma's priors without control. The XGBoost comparison shows the motion signal contributes, but no ablation isolates how much of the diagnosis and device performance comes from language-prior class imbalances rather than from motion.
  • domain assumption Parsing the model's generated text yields faithful numerical answers for regression tasks.
    Section 4.2: 'we extracted the numerical output from BiomechGPT-generated responses.' Parsing errors or format drift would bias the reported Pearson correlations, and no parse-failure rate is reported.
invented entities (1)
  • Cross-format tokenizer pooling heterogeneous motion formats without paired data
    purpose: Claimed in the abstract to enable incorporating a second dataset and pooling annotations across datasets; absent from the method and results sections.
    Section 5.1 describes multiple shared-latent VQ-VAEs as future work ('We also plan to extend BiomechGPT's movement vocabulary by developing multiple VQ-V AEs'). The abstract reports the tokenizer as an accomplished design. No implementation, evaluation, or second dataset appears anywhere in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BiomechGPT: Extending Motion-Language Models to Clinical Motion Understanding." pith.science (2026). https://pith.science/paper/ZI2EE5WV

@misc{pith2026250518465,
  author       = {Pith},
  title        = {Pith review of: BiomechGPT: Extending Motion-Language Models to Clinical Motion Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZI2EE5WV}},
  note         = {Machine review of arXiv:2505.18465}
}
read the original abstract

Advances in markerless motion capture are making high-quality biomechanical data increasingly accessible, creating a growing need for scalable downstream analytics. Building a bespoke pipeline for each analysis task is time-consuming, motivating models that can flexibly handle diverse clinical questions within a single framework. Recent work has shown that fine-tuning language models to accept tokenized motion as an additional modality enables descriptive captioning of movement, raising the question of whether these models are also capable of clinically relevant motion understanding, where diverse tasks and annotations provide a natural testbed. We investigate whether such a multimodal motion--language model can answer detailed, clinically meaningful questions about movement. We collected 71 hours of biomechanical data from 750 participants, many with movement impairments, performing tasks commonly used in clinical assessment. To further expand the training dataset, we designed a cross-format tokenizer that directly encodes motion data from heterogeneous formats into a shared latent space without paired data, allowing a second dataset to be incorporated and enabling pooling annotations across datasets. From these tokenized representations, we constructed a multimodal dataset of motion-related question--answer pairs and used it to train BiomechGPT, a multimodal biomechanics--language model. BiomechGPT achieves competitive performance across a range of clinically relevant tasks, with performance scaling with both dataset and model size. It offers a new way for clinicians and researchers to interact with biomechanical data and represents a promising direction for rehabilitation-focused movement analysis. Project page: https://intelligentsensingandrehabilitation.github.io/BiomechGPT/

Figures

Figures reproduced from arXiv: 2505.18465 by the authors.

Figure 1
Figure 1. Screenshot from our chat interface, which allows selecting a trial from our motion dataset [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Confusion matrix for activity classification from one of our 4B models trained on all of the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Confusion matrix for impaired movement classification (top left) and diagnosis classification [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Regression for walking cadence (left) and and speed (right) from a 4B parameter model [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in 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. Markerless Motion Capture in Routine Clinical Upper Limb Assessments: Validity and Insights Beyond Ordinal Scoring

    cs.CV 2026-07 conditional novelty 5.0 of 10

    Routine three-camera markerless motion capture yields accurate upper-limb kinematics and construct-valid metrics that add specificity and post-ceiling sensitivity beyond ARAT ordinal scores.

Reference graph

Works this paper leans on

33 extracted references · 12 canonical work pages · cited by 1 Pith paper

  1. [1]

    Uhlrich, Antoine Falisse, Łukasz Kidzi ´nski, Julie Muccini, Michael Ko, Akshay S

    Scott D. Uhlrich, Antoine Falisse, Łukasz Kidzi ´nski, Julie Muccini, Michael Ko, Akshay S. Chaudhari, Jennifer L. Hicks, and Scott L. Delp. Opencap: Human Movement Dynamics from Smartphone Videos.PLOS Computational Biology, 19(10):e1011462, oct 19 2023. ISSN 1553-7358. doi: 10.1371/journal.pcbi.1011462. URL https://journals.plos.org/ ploscompbiol/article...

    doi:10.1371/journal.pcbi.1011462 2023

  2. [2]

    Kanko, Elise K

    Robert M. Kanko, Elise K. Laende, Elysia M. Davis, W. Scott Selbie, and Kevin J. Deluzio. Concurrent Assessment of Gait Kinematics Using Marker-Based and Markerless Motion Capture.Journal of Biomechanics, 127:110665, oct 11 2021. ISSN 1873-2380. doi: 10.1016/j.jbiomech.2021.110665

    arXiv 2021

  3. [3]

    Peiffer, Kunal Shah, Shawana Anarwala, Kayan Abdou, and R

    J.D. Peiffer, Kunal Shah, Shawana Anarwala, Kayan Abdou, and R. James Cotton. Fusing Uncalibrated IMUs and Handheld Smartphone Video to Reconstruct Knee Kinematics. In2024 10th IEEE RAS/EMBS International Conference for Biomedical Robotics and Biomechatronics (BioRob), pages 1275–1282, 9 2024. doi: 10.1109/BioRob60516.2024.10719724. URL https://ieeexplore...

    arXiv 2024

  4. [4]

    MotionLLM: Multimodal Motion-Language Learning with Large Language Models, 2024

    Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chi-Keung Tang. MotionLLM: Multimodal Motion-Language Learning with Large Language Models, 2024. URL http://arxiv.org/ abs/2405.17013

    arXiv 2024

  5. [5]

    Motiongpt: Human motion as a foreign language.Advances in Neural Information Processing Systems, 36, 2024

    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, 2024

    work page 2024

  6. [6]

    Neural Discrete Representation Learning

    Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural Discrete Representation Learning. InAdvances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. URL https://proceedings.neurips.cc/paper_files/paper/2017/hash/ 7a98af17e63a0ac09ce2e96d03992fbc-Abstract.html

    work page 2017

  7. [7]

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: A Skinned Multi-Person Linear Model.ACM Transactions on Graphics, 34(6):1–16, oct 26 2015. ISSN 07300301. doi: 10.1145/2816795.2818013. URL http://dl.acm.org/ citation.cfm?doid=2816795.2818013. [Online; accessed 2019-09-26]. 11

    arXiv 2015

  8. [8]

    Karen Liu, and Michael J

    Marilyn Keller, Keenon Werling, Soyong Shin, Scott Delp, Sergi Pujades, C. Karen Liu, and Michael J. Black. From Skin to Skeleton: Towards Biomechanically Accurate 3d Digital Humans.ACM Transactions on Graphics, 42(6):253:1–253:12, dec 5 2023. ISSN 0730-

    work page 2023

Show all 33 references
  1. [9]

    Myosuite – A Contact-Rich Simulation Suite for Musculoskeletal Motor Control

    Vittorio Caggiano, Huawei Wang, Guillaume Durandau, Massimo Sartori, and Vikash Ku- mar. Myosuite – A Contact-Rich Simulation Suite for Musculoskeletal Motor Control. http://arxiv.org/abs/2205.13600, may 26 2022. URL http://arxiv.org/abs/2205.13600. [Online; accessed 2023-12-28]

    2022 arXiv

  2. [10]

    Locomujoco: A Comprehensive Imitation Learning Benchmark for Locomotion

    Firas Al-Hafez, Guoping Zhao, Jan Peters, and Davide Tateo. Locomujoco: A Comprehensive Imitation Learning Benchmark for Locomotion. http://arxiv.org/abs/2311.02496, nov 30 2023. URLhttp://arxiv.org/abs/2311.02496. [Online; accessed 2023-12-26]

    2023 arXiv

  3. [11]

    Bleu: A Method for Automatic Evaluation of Machine Translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: A Method for Automatic Evaluation of Machine Translation. In Pierre Isabelle, Eugene Charniak, and Dekang Lin, editors, Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, p...

  4. [12]

    Krumholz, Jure Leskovec, Eric J

    Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M. Krumholz, Jure Leskovec, Eric J. Topol, and Pranav Rajpurkar. Foundation models for gener- alist medical artificial intelligence.Nature, 616(7956):259–265, 2023. ISSN 1476-

    2023

  5. [13]

    Sara Mahdavi, Bradley Green, Ewa Dominowska, Blaise Aguera Arcas, Joelle Barral, Dale Webster, Greg S

    Tao Tu, Shekoofeh Azizi, Danny Driess, Mike Schaekermann, Mohamed Amin, Pi-Chuan Chang, Andrew Carroll, Charles Lau, Ryutaro Tanno, Ira Ktena, Anil Palepu, Basil Mustafa, Aakanksha Chowdhery, Yun Liu, Simon Kornblith, David Fleet, Philip Mansfield, Sushant Prakash, Renee Wong,...

    2024

  6. [14]

    A Survey on Deep Learning for Human Activity Recognition.ACM Comput

    Fuqiang Gu, Mu-Huan Chung, Mark Chignell, Shahrokh Valaee, Baoding Zhou, and Xue Liu. A Survey on Deep Learning for Human Activity Recognition.ACM Comput. Surv., 54(8): 177:1–177:34, 2021. ISSN 0360-0300. doi: 10.1145/3472290. URL https://doi.org/10. 1145/3472290

    2021 doi

  7. [15]

    Intelligent recognition of human activities using deep learning techniques.PLOS One, 20(4): e0321754, 2025

    Shazab Bashir, Arfan Jaffar, Muhammad Rashid, Sheeraz Akram, and Sohail Masood Bhatti. Intelligent recognition of human activities using deep learning techniques.PLOS One, 20(4): e0321754, 2025. ISSN 1932-6203. doi: 10.1371/journal.pone.0321754

    2025 doi

  8. [16]

    Espay, Matteo Bologna, and Lazzaro di Biase

    Pasquale Maria Pecoraro, Luca Marsili, Alberto J. Espay, Matteo Bologna, and Lazzaro di Biase. Computer Vision Technologies in Movement Disorders: A Systematic Review.Movement Disorders Clinical Practice, 2025. ISSN 2330-1619. doi: 10.1002/mdc3.70123

    2025 doi

  9. [17]

    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. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5142–5151. doi: 10.1109/CVPR52688.2022. 00509

  10. [18]

    Bermano, and Daniel Cohen-Or

    Guy Tevet, Brian Gordon, Amir Hertz, Amit H. Bermano, and Daniel Cohen-Or. MotionCLIP: Exposing Human Motion Generation to CLIP Space. InComputer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXII, pages 358–374. Springe...

    2022 doi

  11. [19]

    Black, and Gül Varol

    Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating Diverse Human Motions from Textual Descriptions. InComputer Vision – ECCV 2022: 17th European 12 Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXII, pages 480–497. Springer-Verlag, 2022...

    2022 doi

  12. [20]

    TM2T: Stochastic and Tokenized Modeling for the Reciprocal Generation of 3D Human Motions and Texts

    Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. TM2T: Stochastic and Tokenized Modeling for the Reciprocal Generation of 3D Human Motions and Texts. InComputer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXV, pages 580...

    2022 doi

  13. [21]

    QLoRA: Efficient Finetuning of Quantized LLMs.Advances in Neural Information Processing Systems, 36: 10088–10115, 2023

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA: Efficient Finetuning of Quantized LLMs.Advances in Neural Information Processing Systems, 36: 10088–10115, 2023. URL https://proceedings.neurips.cc/paper_files/paper/ 2023/hash/1feb87871436031bdc0f2beaa...

    2023

  14. [22]

    MotionGPT-2: A General-Purpose Motion-Language Model for Motion Generation and Understanding, 2024

    Yuan Wang, Di Huang, Yaqi Zhang, Wanli Ouyang, Jile Jiao, Xuetao Feng, Yan Zhou, Pengfei Wan, Shixiang Tang, and Dan Xu. MotionGPT-2: A General-Purpose Motion-Language Model for Motion Generation and Understanding, 2024. URL http://arxiv.org/abs/ 2410.21747

    2024 arXiv

  15. [23]

    AGIR: Assessing 3D Gait Impairment with Reasoning based on LLMs, 2025

    Diwei Wang, Cédric Bobenrieth, and Hyewon Seo. AGIR: Assessing 3D Gait Impairment with Reasoning based on LLMs, 2025. URLhttp://arxiv.org/abs/2503.18141

    2025 arXiv

  16. [24]

    James Cotton

    R. James Cotton. Differentiable Biomechanics Unlocks Opportunities for Markerless Motion Capture. InIEEE International Consortium for Rehabilitation Robotics. arXiv, 2025. URL http://arxiv.org/abs/2402.17192. [Online; accessed 2024-02-28]

    2025 arXiv

  17. [25]

    Hamner, Ajay Seth, and Scott L

    Samuel R. Hamner, Ajay Seth, and Scott L. Delp. Muscle Contributions to Propulsion and Support during Running.Journal of Biomechanics, 43(14):2709–2716, oct 19 2010. ISSN 1873-2380. doi: 10.1016/j.jbiomech.2010.06.025

    2010 doi

  18. [26]

    Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni...

    2023

  19. [27]

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex...

    2024 arXiv

  20. [28]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, An- thony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Ma...

    2020 arXiv

  21. [29]

    Wiltschko, Tatsuya Tsukahara, Ayman Zeine, Rockwell Anyoha, Winthrop F

    Alexander B. Wiltschko, Tatsuya Tsukahara, Ayman Zeine, Rockwell Anyoha, Winthrop F. Gillis, Jeffrey E. Markowitz, Ralph E. Peterson, Jesse Katon, Matthew J. Johnson, and Sandeep Robert Datta. Revealing the structure of pharmacobehavioral space through mo- tion sequencing.Natu...

    2020 doi

  22. [30]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, page 785–794, New York, NY , USA, 2016. Association for Computing Machinery. ISBN 978145034...

    2016

  23. [31]

    {motion_placeholder} Identify the activity shown here

    Pouyan Firouzabadi, Wendy Murray, Anton R Sobinov, J.D. Peiffer, Kunal Shah, Lee E Miller, and R. James Cotton. Biomechanical Arm and Hand Tracking with Multiview Markerless Motion Capture. In2024 10th IEEE RAS/EMBS International Conference for Biomedical Robotics and Biomecha...

    2024

  24. [301]

    URL https://dl.acm.org/doi/10.1145/3618381

    doi: 10.1145/3618381. URL https://dl.acm.org/doi/10.1145/3618381. [Online; accessed 2023-12-12]

    2023 doi

  25. [4687]

    URL https://www.nature.com/articles/ s41586-023-05881-4

    doi: 10.1038/s41586-023-05881-4. URL https://www.nature.com/articles/ s41586-023-05881-4

    doi

Pith tools

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