Pith. sign in

REVIEW 3 major objections 5 minor 28 references

A generative world model turns hand gestures into robot training videos, so data collection no longer needs a physical robot.

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 ↗

T0 review · grok-4.5

2026-07-14 15:59 UTC pith:PZQRAOXX

load-bearing objection Solid systems paper that makes digital teleoperation real enough to train policies, but the pure-hardware-free claim is overstated because of re-anchoring. the 3 major comments →

arxiv 2607.06558 v2 pith:PZQRAOXX submitted 2026-07-07 cs.RO

RynnWorld-Teleop: An Action-Conditioned World Model for Digital Teleoperation

Haoyu Zhao , Xingyue Zhao , Hangyu Li , Biao Gong , Kehan Li , Siteng Huang , Xin Li , Deli Zhao
show 1 more author
Zhongyu Li
This is my paper
classification cs.RO
keywords digital teleoperationaction-conditioned world modelegocentric video generationrobot data engineimitation learningSim2Realdepth-aware skeletal conditioningvideo diffusion transformer
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.

Robot learning is starved for diverse demonstration data because every real teleoperation session ties an operator to specific hardware and a fixed workspace. This paper claims that the bottleneck can be removed by digital teleoperation: an operator streams hand poses into a robot-centric generative world model that, from a single reference image, synthesizes the egocentric video a robot would have seen. The same pose stream is retargeted into robot joint actions, producing complete state-action trajectories without ever moving a real robot. The authors instantiate the idea in RynnWorld-Teleop, which reaches real-time interactive rates and yields data that trains policies capable of zero-shot transfer to real bimanual tasks, and that further raises success rates when mixed with real demonstrations. If the claim holds, operator imagination, not laboratory hardware, becomes the limit on scalable robot data.

Core claim

Policies trained exclusively on trajectories synthesized by RynnWorld-Teleop achieve effective zero-shot transfer to real robots across dexterous bimanual tasks, and adding the same synthetic data to real demonstration sets consistently lifts success rates, establishing digital teleoperation as a high-fidelity, hardware-independent data engine.

What carries the argument

RynnWorld-Teleop: a depth-aware skeletal conditioning signal (joints rendered with camera-distance color and size) fed into a video Diffusion Transformer that is first pretrained on large-scale human egocentric video, then adapted to robot embodiments, and finally distilled into a causal streaming student that generates at 40+ FPS.

Load-bearing premise

A few thousand paired human-robot episodes plus depth cues in 2D skeletons are enough to close the embodiment and contact-physics gap so that the generated videos and retargeted actions form valid imitation data.

What would settle it

Train a policy solely on RynnWorld-Teleop rollouts for a new bimanual task whose objects and lighting never appear in the generator's training set, then measure real-robot success over dozens of trials; if success collapses relative to real-data baselines, the claim fails.

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

If this is right

  • Robot data collection scales with operator hours rather than robot availability or workspace resets.
  • A single reference image can instantiate new manipulation scenes without 3D assets or system identification.
  • Real demonstration sets can be enlarged by synthetic trajectories that improve policy success rates.
  • Policies can be trained entirely offline from generated data and still transfer zero-shot to hardware.
  • The same operator gesture stream can be retargeted to different robot embodiments after generation.

Where Pith is reading between the lines

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

  • If chunked re-anchoring can be replaced by fully generative long-horizon rollouts, digital teleoperation would eliminate the last remaining dependence on real camera frames during data collection.
  • Cross-embodiment conditioning on kinematic descriptors could remove the need for per-robot fine-tuning and turn the model into a fleet-wide data engine.
  • The same pipeline could supply closed-loop interactive simulators for online policy fine-tuning rather than only offline imitation datasets.

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

3 major / 5 minor

Summary. The paper proposes digital teleoperation: an operator hand-pose stream drives a robot-centric action-conditioned video world model (RynnWorld-Teleop) that synthesizes egocentric robot videos from a single reference image; the same poses are retargeted to 54-DoF robot actions, yielding imitation trajectories without moving a physical robot during collection. The system combines depth-aware skeletal rendering, progressive human-to-robot DiT training (large egocentric human pretraining then fine-tuning on 1,800 paired episodes), and streaming autoregressive distillation to a causal student that runs at 40+ FPS. Empirically, policies (DP, π0, π0.5) trained only on the generated data achieve non-trivial zero-shot real-robot success on four bimanual tasks, and mixing 300 synthetic episodes with 300 real episodes consistently raises success rates (most notably +20 points on lid placement). Ablations, t-SNE feature overlap, OOD reference-image edits, and standard video metrics (Table 4) are provided in support.

Significance. If the pure-digital data engine claim holds, the work would meaningfully loosen the physical-teleoperation bottleneck that currently limits scale and diversity of robot imitation datasets. The technical stack is carefully engineered: depth-modulated 2D skeletons, distribution-aligned additive conditioning (Eq. 3), progressive cross-domain training, and a two-stage causal distillation that delivers interactive rates on a single H100 are concrete, reusable contributions. The pure-synthetic π0 numbers (Table 3) and the consistent augmentation gains supply the first quantitative evidence that a robot-centric generative world model can substitute for, and amplify, physical demonstrations. Open model and code links further raise the work’s utility. The result is therefore of clear interest to the robot-learning community provided the hardware-independence claim is made precise and the statistical support is tightened.

major comments (3)
  1. Sec. 4.1 (Mitigating Drift via Chunked Re-anchoring) states that after the first 81-frame chunk every subsequent chunk is re-initialized with the actual egocentric frame captured by the real robot camera as the new I_ref. Consequently the pure-synthetic rows of Table 3 (π0 trained on 0 Real + 300 RynnWorld-Teleop) still depend on real camera frames for every chunk after the first. This procedure re-introduces physical hardware into the very trajectories presented as hardware-free, so the abstract and introduction claims of data collection “independent of physical hardware” and “without ever moving a real robot” are stronger than the experimental pipeline supports. Either pure long-horizon generation without real re-anchoring must be demonstrated for the policy experiments, or the paradigm-level language must be revised to match the hybrid procedure actually used.
  2. Table 3 reports success rates over 35 consecutive trials with no standard errors, confidence intervals, or statistical tests. Several of the claimed gains (especially the pure-synthetic versus real comparisons and the +20-point lid-placement lifts) are therefore difficult to assess for reliability; with n=35 a few additional failures can change the ranking. Error bars or a non-parametric test should be added before the “consistently improves” and “effective zero-shot” statements can be treated as load-bearing.
  3. The robotic domain-adaptation stage (Sec. 3.4, Table 2) uses only 1,800 episodes spanning four tightly related tabletop tasks collected on a single embodiment. The claim that progressive fine-tuning plus depth-aware 2D skeletons closes the embodiment and contact-physics gap sufficiently for the generated videos to serve as valid imitation data therefore rests on a narrow distribution. Additional held-out tasks, objects, or a second robot platform would strengthen (or bound) the generalization argument that underpins the data-engine narrative.
minor comments (5)
  1. Fig. 5 t-SNE is described as showing “significant overlap”; a quantitative distribution-distance metric (e.g., MMD or FID on the same I3D features) would make the claim less subjective.
  2. Eq. (3) introduces a learnable gate α that is “initialized to a small value (e.g., 0.1)”; the final learned value (or its range across layers) is never reported and would help readers judge how much the control branch is actually used.
  3. The latency breakdown (Sec. 5.2) sums to ~100 % but does not state whether VAE encoding of the skeletal video is performed once per chunk or per frame; a short clarification would aid reproducibility of the 40 FPS figure.
  4. Several concurrent action-conditioned egocentric models (Hand2World, GeneratedReality, etc.) are cited; a short paragraph contrasting real-time robot-centric versus human-centric design choices would sharpen the related-work positioning.
  5. Typos / polish: “We introducedigital teleoperation” (abstract), missing spaces after periods in a few places, and inconsistent capitalization of task names in Table 2 versus the main text.

Circularity Check

0 steps flagged

Empirical robotics paper with standard training objectives and external real-robot metrics; no derivation reduces a claimed prediction to its own fitted inputs by construction.

full rationale

RynnWorld-Teleop is an engineering system paper: it defines a depth-aware skeletal conditioning (Sec. 3.2), progressive human-to-robot DiT training (Sec. 3.4), and streaming distillation (Sec. 3.5), then uses the resulting generator plus retargeting (Sec. 4.1) to produce imitation data. The training losses (conditional flow matching Eq. 2, causal MSE Eq. 4, DMD) are standard external objectives; the evaluation metrics (PSNR/SSIM/LPIPS/FVD on held-out EgoDex-Test and Robotic-Test, plus real-robot success rates over 35 trials in Table 3) are independently measured on physical hardware. No equation equates a free parameter fitted on one subset to a 'prediction' on a related quantity. Self-citations (e.g., concurrent Zhao et al. works) appear only as related-work context and are not load-bearing uniqueness theorems that force the central claims. The chunked re-anchoring procedure (Sec. 4.1) that re-uses real camera frames for long-horizon synthesis is a methodological limitation that weakens the pure-hardware-free rhetoric, but it is an explicit engineering choice, not a circular reduction of a derivation. The pipeline is therefore self-contained against external benchmarks; any residual score of 1 reflects only ordinary author self-citation volume, not circular reasoning.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central claim rests on standard video-diffusion and imitation-learning machinery plus a small set of domain assumptions about retargeting and embodiment transfer. Free parameters are ordinary training hyperparameters. No new physical entities are postulated; 'digital teleoperation' is a named paradigm, not a new force or particle. The largest unproved loads are that 2D depth-modulated skeletons plus limited real paired data suffice for transferable contact dynamics, and that re-anchoring with real frames is an acceptable engineering compromise rather than a contradiction of the hardware-free claim.

free parameters (5)
  • gating scalar α initialization = 0.1
    Initialized to 0.1 and learned; controls how strongly the pose branch is injected into the pretrained DiT (Eq. 3).
  • LoRA rank = 64
    Rank-64 adapters used in one training track; chosen by authors, not derived.
  • chunk length for re-anchoring = 81 frames
    81-frame segments used both for training clips and for drift mitigation; matches EgoDex slicing but is a free design choice for long-horizon generation.
  • DMD generative / critic learning rates = 2e-6 / 5e-7
    2e-6 and 5e-7 chosen for the adversarial distillation stage; affect final 4-step student quality.
  • translational workspace scale factor s = 1.5 (example)
    Maps operator workspace to robot workspace in the retargeting chain (Appendix B); example value 1.5 given.
axioms (5)
  • standard math Standard conditional flow-matching and distribution-matching distillation objectives produce a usable causal video generator when trained on the described data.
    Invoked throughout Sec. 3.1 and 3.5; taken from Lipman et al. and Yin et al. without re-derivation.
  • domain assumption A 21-joint depth-modulated 2D skeleton rendered into VAE latent space is a sufficient action interface for fine-grained bimanual manipulation.
    Core design of Sec. 3.2; no proof that 2D depth cues recover all necessary 3D contact information.
  • domain assumption Human egocentric pretraining followed by fine-tuning on 1,800 real robot episodes transfers manipulation priors across the embodiment gap.
    Sec. 3.4 progressive training; ablation shows collapse without pretraining, but the transfer itself is assumed sufficient for policy-grade data.
  • domain assumption Damped-least-squares IK with null-space shoulder prior yields joint actions that remain valid when paired with generated pixels for imitation learning.
    Appendix B retargeting pipeline; success of downstream policies is taken as evidence rather than independently verified kinematics fidelity.
  • ad hoc to paper Periodic re-anchoring with real camera frames does not invalidate the claim of hardware-independent digital teleoperation for the reported experiments.
    Sec. 4.1 engineering choice that re-introduces real robot images into long-horizon synthetic trajectories used for policy training.
invented entities (3)
  • digital teleoperation paradigm no independent evidence
    purpose: Names the decoupling of operator time from physical robot and workspace via a generative world model.
    Introduced in abstract and Sec. 1 as the framing contribution; evaluated only inside this paper's system.
  • depth-aware skeletal representation (color- and size-modulated joints) no independent evidence
    purpose: Supplies explicit 3D cues inside a 2D conditioning video for the DiT.
    Sec. 3.2 and Fig. 2; novel rendering choice relative to mask-based baselines, but still a design, not a new physical object.
  • RynnWorld-Teleop (robot-centric causal DiT student) no independent evidence
    purpose: Concrete 40+ FPS action-conditioned generator that emits robot videos from hand poses.
    The system instantiation; independent evidence would require third-party re-use or external benchmarks beyond the authors' four tasks.

pith-pipeline@v1.1.0-grok45 · 23416 in / 3966 out tokens · 48535 ms · 2026-07-14T15:59:13.174544+00:00 · methodology

0 comments
read the original abstract

Scaling robot learning requires massive, diverse trajectory data, yet collection is currently bottlenecked by physical teleoperation, where every demonstration binds operator time to specific hardware and workspaces. We introduce digital teleoperation, a paradigm that decouples data collection from physical constraints by replacing the real robot with a generative world model. In this framework, an operator's hand-pose stream drives a robot-centric generative world model to synthesize high-fidelity egocentric videos from a single reference image. The recorded pose stream serves as an embodiment-agnostic action label transferable to any target robot via standard retargeting, yielding complete state-action trajectories for imitation learning independent of physical hardware. We instantiate this paradigm in RynnWorld-Teleop, a system that integrates depth-aware skeletal conditioning, progressive human-to-robot training on a video Diffusion Transformer, and streaming autoregressive distillation. This pipeline compresses the generative process into a single-pass inference, enabling 40+ FPS, real-time interactive generation on a single H100 GPU. Policies trained exclusively on RynnWorld-Teleop-generated data achieve effective zero-shot Sim2Real transfer across dexterous and diverse bimanual tasks. Moreover, augmenting real-world datasets with our digitally teleoperated data consistently improves success rates, demonstrating that RynnWorld-Teleop serves as a high-fidelity, scalable data engine for the next generation of robotic agents.

Review history (3 revisions) →

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

28 extracted references · 19 linked inside Pith

  1. [1]

    Cosmos world foundation model platform for physical ai.arXiv preprint arXiv:2501.03575,

    Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai.arXiv preprint arXiv:2501.03575,

    Pith/arXiv arXiv

  2. [2]

    World simulation with video foundation models for physical ai.arXiv preprint arXiv:2511.00062,

    Arslan Ali, Junjie Bai, Maciej Bala, Yogesh Balaji, Aaron Blakeman, Tiffany Cai, Jiaxin Cao, Tianshi Cao, Eliza- beth Cha, Yu-Wei Chao, et al. World simulation with video foundation models for physical ai.arXiv preprint arXiv:2511.00062,

    Pith/arXiv arXiv

  3. [3]

    V-jepa 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985,

    Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, et al. V-jepa 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985,

    Pith/arXiv arXiv

  4. [4]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164,

    Pith/arXiv arXiv

  5. [5]

    Rt-1: Robotics transformer for real-world control at scale

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakr- ishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817,

    Pith/arXiv arXiv

  6. [6]

    Lome: Learning human-object manipulation with action-conditioned egocentric world model.arXiv preprint arXiv:2603.27449,

    Quankai Gao, Jiawei Yang, Qiangeng Xu, Le Chen, and Yue Wang. Lome: Learning human-object manipulation with action-conditioned egocentric world model.arXiv preprint arXiv:2603.27449,

    arXiv

  7. [7]

    World models can leverage human videos for dexterous manipulation

    Raktim Gautam Goswami, Amir Bar, David Fan, Tsung-Yen Yang, Gaoyue Zhou, Prashanth Krishnamurthy, Michael Rabbat, Farshad Khorrami, and Yann LeCun. World models can leverage human videos for dexterous manipulation. arXiv preprint arXiv:2512.13644,

    arXiv

  8. [8]

    Articulat3d: Reconstructing articulated digital twins from monocular videos with geometric and motion constraints

    Lijun Guo, Haoyu Zhao, Xingyue Zhao, Rong Fu, Linghao Zhuang, Siteng Huang, Zhongyu Li, and Hua Zou. Articulat3d: Reconstructing articulated digital twins from monocular videos with geometric and motion constraints. arXiv preprint arXiv:2603.11606,

    Pith/arXiv arXiv

  9. [9]

    World models.arXiv preprint arXiv:1803.10122, 2(3):440,

    David Ha and Jürgen Schmidhuber. World models.arXiv preprint arXiv:1803.10122, 2(3):440,

    Pith/arXiv arXiv

  10. [10]

    Egosim: Egocentric world simulator for embodied interaction generation.arXiv preprint arXiv:2604.01001,

    Jinkun Hao, Mingda Jia, Ruiyan Wang, Xihui Liu, Ran Yi, Lizhuang Ma, Jiangmiao Pang, and Xudong Xu. Egosim: Egocentric world simulator for embodied interaction generation.arXiv preprint arXiv:2604.01001,

    Pith/arXiv arXiv

  11. [11]

    Egodex: Learning dexterous manipulation from large-scale egocentric video.arXiv preprint arXiv:2505.11709,

    Ryan Hoque, Peide Huang, David J Yoon, Mouli Sivapurapu, and Jian Zhang. Egodex: Learning dexterous manipulation from large-scale egocentric video.arXiv preprint arXiv:2505.11709,

    Pith/arXiv arXiv

  12. [12]

    Self forcing: Bridging the train-test gap in autoregressive video diffusion.arXiv preprint arXiv:2506.08009,

    Xun Huang, Zhengqi Li, Guande He, Mingyuan Zhou, and Eli Shechtman. Self forcing: Bridging the train-test gap in autoregressive video diffusion.arXiv preprint arXiv:2506.08009,

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2504.16054,

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al.π0.5: A vision-language-action model with open-world generalization. arXiv preprint arXiv:2504.16054,

    Pith/arXiv arXiv

  14. [14]

    Pyramidal flow matching for efficient video generative modeling.arXiv preprint arXiv:2410.05954,

    Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling.arXiv preprint arXiv:2410.05954,

    arXiv

  15. [15]

    Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246,

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246,

    Pith/arXiv arXiv

  16. [16]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603,

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603,

    Pith/arXiv arXiv

  17. [17]

    Masquerade: Learning from in-the-wild human videos using data-editing.arXiv preprint arXiv:2508.09976, 2025a

    Marion Lepert, Jiaying Fang, and Jeannette Bohg. Masquerade: Learning from in-the-wild human videos using data-editing.arXiv preprint arXiv:2508.09976, 2025a. Marion Lepert, Jiaying Fang, and Jeannette Bohg. Phantom: Training robots without robots using only human videos, 2025.arXiv preprint arXiv:2503.00779, 2025b. Gen Li, Bo Zhao, Jianfei Yang, and Laur...

    Pith/arXiv arXiv 2025

  18. [18]

    Egocontrol: Controllable egocentric video generation via 3d full-body poses.arXiv preprint arXiv:2511.18173,

    Enrico Pallotta, Sina Mokhtarzadeh Azar, Lars Doorenbos, Serdar Ozsoy, Umar Iqbal, and Juergen Gall. Egocontrol: Controllable egocentric video generation via 3d full-body poses.arXiv preprint arXiv:2511.18173,

    arXiv

  19. [19]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. InProc. of International Conference on Learning Representations, volume 2025, pages 28085–28128,

    2025

  20. [20]

    Mitty: Diffusion-based human-to-robot video generation

    Yiren Song, Cheng Liu, Weijia Mao, and Mike Zheng Shou. Mitty: Diffusion-based human-to-robot video generation. arXiv preprint arXiv:2512.17253,

    arXiv

  21. [21]

    Playerone: Egocentric world simulator

    Yuanpeng Tu, Hao Luo, Xi Chen, Xiang Bai, Fan Wang, and Hengshuang Zhao. Playerone: Egocentric world simulator. arXiv preprint arXiv:2506.09995,

    arXiv

  22. [22]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025a

    Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025a. Yuang Wang, Chao Wen, Haoyu Guo, Sida Peng, Minghan Qin, Hujun Bao, Xiaowei Zhou, and Ruizhen Hu. Precise action-to-video generation...

    Pith/arXiv arXiv

  23. [23]

    Generated reality: Human-centric world simulation using interactive video generation with hand and camera control.arXiv preprint arXiv:2602.18422,

    Linxi Xie, Lisong C Sun, Ashley Neall, Tong Wu, Shengqu Cai, and Gordon Wetzstein. Generated reality: Human-centric world simulation using interactive video generation with hand and camera control.arXiv preprint arXiv:2602.18422,

    arXiv

  24. [24]

    X-humanoid: Robotize human videos to generate humanoid videos at scale.arXiv preprint arXiv:2512.04537,

    Pei Yang, Hai Ci, Yiren Song, and Mike Zheng Shou. X-humanoid: Robotize human videos to generate humanoid videos at scale.arXiv preprint arXiv:2512.04537,

    arXiv

  25. [25]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072,

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072,

    Pith/arXiv arXiv

  26. [26]

    Twist: Teleoperated whole-body imitation system.arXiv preprint arXiv:2505.02833,

    Yanjie Ze, Zixuan Chen, Joao Pedro Araújo, Zi-ang Cao, Xue Bin Peng, Jiajun Wu, and C Karen Liu. Twist: Teleoperated whole-body imitation system.arXiv preprint arXiv:2505.02833,

    Pith/arXiv arXiv

  27. [27]

    Hfgs: 4d gaussian splatting with emphasis on spatial and temporal high-frequency components for endoscopic scene reconstruction.arXiv preprint arXiv:2405.17872,

    Haoyu Zhao, Xingyue Zhao, Lingting Zhu, Weixi Zheng, and Yongchao Xu. Hfgs: 4d gaussian splatting with emphasis on spatial and temporal high-frequency components for endoscopic scene reconstruction.arXiv preprint arXiv:2405.17872,

    Pith/arXiv arXiv

  28. [28]

    Smap: Self- supervised motion adaptation for physically plausible humanoid whole-body control.arXiv preprint arXiv:2505.19463, 2025a

    Haoyu Zhao, Sixu Lin, Qingwei Ben, Minyue Dai, Hao Fei, Jingbo Wang, Hua Zou, and Junting Dong. Smap: Self- supervised motion adaptation for physically plausible humanoid whole-body control.arXiv preprint arXiv:2505.19463, 2025a. Haoyu Zhao, Hao Wang, Xingyue Zhao, Hao Fei, Hongqiu Wang, Chengjiang Long, and Hua Zou. Physsplat: Efficient physics simulatio...

    Pith/arXiv arXiv