Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

KineBench evaluates embodied world models by extracting 6D end-effector poses from generated videos and executing them in a physics simulator, avoiding the inverse-dynamics confound.

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 · deepseek-v4-flash

2026-08-01 11:25 UTC pith:IJ6R3IY3

load-bearing objection A genuinely useful IDM-free closed-loop benchmark with honest limitations, but the scaling claim and pose-error validation on generated frames need tightening before the numbers are taken at face value. the 4 major comments →

arxiv 2607.19876 v1 pith:IJ6R3IY3 submitted 2026-07-22 cs.RO cs.CV

KineBench: Benchmarking Embodied World Models via IDM-Free Kinematic Grounding

Zeyu Liu , Zhangzhe Zhu , Yang Zhang , Chenyou Fan , Chenjia Bai , Xuelong Li This is my paper
classification cs.RO cs.CV
keywords embodied world modelsclosed-loop evaluationkinematic grounding6D pose estimationinverse dynamics modelsSPARCmanipulability indexscaling laws
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.

KineBench sets out to make closed-loop evaluation of embodied world models trustworthy by removing the inverse dynamics model from the loop. Instead of learning a 2D-to-3D action extractor, it recovers 6D end-effector poses from each generated frame through a modular cascade of segmentation, depth estimation, and CAD-based pose tracking, then executes those poses in a physics simulator. The paper argues that this explicit kinematic grounding reduces the attribution ambiguity between world-model errors and action-extraction errors, so simulator success rates can be read as evidence about the world model. KineBench also adds two classical 3D kinematic metrics, SPARC for trajectory smoothness and the Maruyama Manipulability Index for kinematic feasibility, and organizes 20 manipulation tasks into four suites that probe execution, transfer, visual out-of-distribution robustness, and scaling. The empirical finding is a task-complexity-bounded nonlinear scaling trend, suggesting that more data and compute help less as task difficulty rises.

Core claim

The paper's central claim is that IDM-based closed-loop evaluation is confounded because learned action extractors are brittle on novel trajectories. KineBench replaces IDMs with an explicit geometric pipeline of segmentation, depth, and CAD-based 6D pose tracking, then executes the recovered poses in a physics simulator. On 20 tasks it reduces attribution ambiguity and reveals that frontier video models struggle with contact-rich dynamics, appearance-specific generalization, and task transfer, and that scaling gains are task-complexity-dependent.

What carries the argument

The explicit kinematic grounding pipeline: a cascade that segments the gripper in each generated frame, recovers metric depth, and aligns the gripper CAD model to produce 6D end-effector poses for simulator execution. Its rigid-body constraint acts as a low-pass filter on frame-level jitter while remaining sensitive to physical hallucinations. Two robot-centric diagnostics, SPARC and the Maruyama manipulability index, quantify trajectory smoothness and kinematic feasibility, giving a direct standard for physical understanding rather than visual plausibility.

Load-bearing premise

The load-bearing premise is that the segmentation-depth-pose cascade is accurate enough on generated frames that simulator failures reflect world-model errors rather than extractor errors; the roughly ten-degree rotational error is the point where this premise is most exposed.

What would settle it

Run KineBench on the same generated videos with simulator ground-truth depth versus the learned depth estimator, and check whether the relative ranking of models on orientation-sensitive tasks changes beyond the reported 1.5–3 cm translational and ten-degree rotational error margins. Alternatively, execute a model that scores high under KineBench using simulator ground-truth poses; if success collapses, the extractor was masking world-model failures.

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

If this is right

  • Closed-loop success rates under KineBench can be attributed primarily to world-model quality, not to action-extractor generalization, for tasks within the extractor's operational envelope.
  • SPARC flags motion-fluency failures in zero-shot or under-optimized models, while the manipulability index exposes generation that pushes the robot toward singular or unreachable configurations, providing complementary failure diagnoses.
  • The four-suite structure separates execution ability, task transfer, visual OOD robustness, and scaling, so a model's failure mode can be localized to a specific capability gap.
  • The observed task-complexity-bounded nonlinear scaling implies that uniform increases in data and compute will not reliably improve embodied world models on harder tasks, so scaling strategies must be task-aware.
  • Saturation of SPARC in fully fine-tuned models suggests that further gains for those models require semantic alignment rather than smoother motion generation.

Where Pith is reading between the lines

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

  • The explicit-grounding principle could extend beyond gripper manipulation to other embodiments, replacing CAD-based pose tracking with keypoint or learned-correspondence trackers that offer similarly inspectable error channels.
  • Because the pipeline retains roughly ten degrees of rotational error, KineBench scores on orientation-sensitive tasks may still carry residual extractor variance; reporting per-task extractor error bounds alongside success rates would sharpen the attribution claim.
  • The scaling result suggests a testable hypothesis: mixing trajectory diversity at fixed compute will improve hard-task scaling more than homogeneous data volume, a prediction that could be checked on the released benchmark.
  • KineBench's closed-loop success and manipulability costs could serve as verifiable rewards for training world models, connecting this evaluation protocol directly to learning algorithms.

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 presents KineBench, a closed-loop benchmark for embodied world models (EWMs) that avoids learned inverse dynamics models (IDMs) by extracting 6D end-effector poses from generated videos with a modular cascade of fine-tuned YOLO segmentation, MoGeV2 depth estimation, and FoundationPose CAD-based pose tracking. The extracted poses are executed in ManiSkill3, yielding closed-loop success rates, and are also scored with SPARC and the Maruyama Manipulability Index as robot-centric kinematic diagnostics. The benchmark comprises 20 manipulation tasks in four suites (basic execution, task transfer, visual OOD, complexity scaling) and is used to evaluate several video-generation models. The paper reports model rankings, generalization trends, and a task-complexity-dependent nonlinear scaling pattern.

Significance. If the pose-extraction pipeline is sufficiently accurate on the actual frames produced by the evaluated EWMs, KineBench would be a valuable community resource: it offers an inspectable, IDM-free alternative to existing closed-loop evaluators, introduces established 3D kinematic metrics to generative-video assessment, and ships code and datasets. The controlled comparison against an IDM baseline on unseen simulator-rendered trajectories is a genuine strength, as is the modular design that makes intermediate perception outputs measurable. However, the central attribution claim—that failures can be read as evidence about the world model rather than the action extractor—depends on pose accuracy on generated videos, which the paper does not directly quantify. The evidence currently supports a benchmark that reduces IDM-specific ambiguity, but the degree of reduction is unquantified on the evaluation distribution.

major comments (4)
  1. [§4.2, Figs. 4–5] The pipeline is validated only on simulator-rendered trajectories with synthetic perturbations, not on frames actually generated by the evaluated EWMs. The paper reports a remaining rotational error of roughly ten degrees for the MoGeV2-based pipeline, which is large for precision tasks such as StackCube and OpenBoxHard. Since many success-rate differences in Table 1 are small (e.g., 56.32 vs. 46.17), this extractor error could confound model rankings. The authors should either measure pose error on generated frames from the evaluated models or provide a sensitivity analysis (e.g., adding controlled pose noise of 5°/10° and reporting changes in success rates) to bound the impact on the attribution claim.
  2. [Table 1] All closed-loop success rates are reported as single percentages without error bars, trial counts, or statistical tests. The scaling configs in Suite 3 (e.g., 44.83 vs. 47.88 vs. 73.33) may be within sampling noise without this information. The paper should report the number of episodes per task/model cell and confidence intervals or standard errors, and ideally a significance test for the main model comparisons. Without this, the cross-model and scaling conclusions are difficult to assess.
  3. [§3.2, Figs. 6–7] The SPARC definition in the text yields non-positive values (it is the negative arc length of the normalized Fourier spectrum), yet Fig. 6 plots positive 'Mean SPARC score' values around 0.5. Similarly, the Maruyama Manipulability Index is defined as a volume that is larger-better, but Fig. 7 says 'lower values indicate better manipulation quality.' The normalization and sign conventions used for the reported scores need to be stated precisely; as written, the two figures contradict the metric definitions and make the robot-centric metrics hard to interpret.
  4. [§4.4, Conclusion] The 'task-complexity-bounded nonlinear scaling' conclusion is based on a single architecture (Wan 2.1 1.3B) with no uncertainty estimates and no replication across other model families. While the authors acknowledge the need for broader experiments, the abstract and conclusion state this as empirical guidance for data-scaling strategies. The claim should be softened or supported with additional architectures and confidence intervals.
minor comments (4)
  1. [References] Reference [1] contains a typo: 'Rransactions' should be 'Transactions'.
  2. [Figure 6] The spelling is inconsistent: 'Hailuo-V2' appears in the text while the figure uses 'Hailluo-V2'. Unify names across all figures and tables.
  3. [§3.2 and Fig. 7] The term 'Manip score' is used without a clear definition; explain how the manipulability cost is computed from the index and how the robust min–max normalization is applied, so readers can reproduce the values.
  4. [§4.1 / Table 1] The Suite 3 column headers 'Scale 10', 'Scale 25', 'Scale 50' are not explained in the text; specify that these are trajectory counts and how they relate to the stated 10–100 training trajectories.

Circularity Check

0 steps flagged

No circularity found: KineBench's success metrics are simulator-executed outputs, and SPARC/manipulability are standard definitions applied post hoc.

full rationale

KineBench's derivation chain runs generated video -> explicit 6D pose extraction -> physics-simulator execution -> task success, plus kinematic metrics computed from the extracted poses. None of these quantities is fitted to the paper's headline results. The fine-tuned YOLO and MoGeV2 components are perception tools; their parameters are not optimized to produce the reported success rates or scaling trends, so the benchmark does not reduce to its inputs. SPARC and the Maruyama Manipulability Index are classical robotics definitions (cited to Balasubramanian and Yoshikawa) applied to recovered trajectories; the paper's claim that they 'exhibit task- and model-dependent associations' with success is an empirical correlation, not a definitional equivalence. The paper explicitly acknowledges the pose pipeline's residual error ('The remaining rotational error of the MoGeV2-based pipeline is approximately ten degrees') and states it 'reduces' rather than removes attribution ambiguity ('we view the pipeline as reducing IDM-specific attribution ambiguity rather than providing error-free pose recovery'); these are validity caveats, not circular steps. Self-citations (HumanoidGen asset credit, multi-agent world-model papers) are not load-bearing for KineBench's evaluation design. No quoted equation or fitted parameter is equivalent by construction to a predicted result.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The benchmark introduces no new physical entities or forces. It relies on standard perception models and a simulator; the main free choices are metric parameters and normalization. The load-bearing assumptions are about the reliability of the pose extraction and the validity of simulator success as a proxy.

free parameters (2)
  • SPARC cutoff frequency omega_c = adaptive per trajectory; exact rule not specified in text
    The SPARC integral depends on omega_c; the paper calls it adaptive but does not give the selection rule, so the metric's scale is not fully pinned down.
  • Manip score normalization percentiles = 10th-90th robust min-max
    The manipulability index is normalized using chosen percentiles, which affects relative scores across models and suites.
axioms (4)
  • domain assumption FoundationPose on generated video frames produces physically meaningful 6D poses.
    The entire benchmark assumes CAD-constrained pose tracking suppresses local artifacts while remaining sensitive to physical hallucinations; Sec. 3.1.
  • domain assumption ManiSkill3 physics simulation is a valid proxy for physical plausibility and task success.
    Closed-loop success is defined by simulator execution; Sec. 3.3 and 4.1.
  • domain assumption Closed-loop execution success reflects EWM quality rather than pipeline error.
    This is the design goal of the IDM-free pipeline, but it is only partially validated in Sec. 4.2 on rendered trajectories.
  • domain assumption SPARC and Maruyama manipulability index are appropriate diagnostic metrics for video-generated motion.
    Adopted from robotics/clinical literature; their association with success is empirical, not derived (Sec. 3.2, Fig. 6-7).

pith-pipeline@v1.3.0-alltime-deepseek · 13143 in / 12276 out tokens · 126375 ms · 2026-08-01T11:25:25.080863+00:00 · methodology

0 comments
read the original abstract

Evaluating the physical consistency of embodied world models(EWMs) is a critical open challenge. While closed-loop evaluation via simulator rollouts offers a more faithful assessment of physical plausibility than open-loop alternatives, existing frameworks almost exclusively rely on Inverse Dynamics Models(IDMs) for action extraction. Due to the intricate mapping from 2D pixel space to 3D kinematic space, the learned IDMs can be brittle to data outside their training distribution, resulting in unreliable action extraction from the generated videos with novel objects and scenarios. This creates an unavoidable attribution ambiguity between world model inaccuracies and extractor errors. To reduce this ambiguity, we present KineBench, an IDM-free closed-loop benchmark for EWMs, built upon an explicit kinematic grounding pipeline. Given a generated video, KineBench employs cascaded visual foundation models to directly extract 6D end-effector poses from individual frames, which are then executed in a physics simulator for closed-loop validation. Beyond execution-based task success, KineBench incorporates two classical 3D kinematic metrics--Spectral Arc Length (SPARC) and the Maruyama Manipulability Index--to characterize trajectory smoothness and kinematic feasibility from a robot-centric perspective. Built on 20 diverse manipulation tasks in ManiSkill3, KineBench evaluates EWMs across four progressive suites: basic execution, task transfer, visual out-of-distribution generalization, and complexity-conditioned scaling. Evaluation across frontier models reveals task-complexity-bounded nonlinear scaling in embodied video generation, providing empirical guidance for future data-scaling strategies.

Figures

Figures reproduced from arXiv: 2607.19876 by Chenjia Bai, Chenyou Fan, Xuelong Li, Yang Zhang, Zeyu Liu, Zhangzhe Zhu.

Figure 1
Figure 1. Figure 1: Comparison of evaluation paradigms for Embodied World Models. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: KineBench evaluation pipeline 3.1 Explicit Kinematic Grounding Pipeline The inherent generalization limitations of IDMs introduce significant confound￾ing factors into closed-loop evaluation, as discussed in Sec. 1. To reduce this bot￾tleneck, KineBench constructs a pipeline integrating 2D instance segmentation, monocular depth estimation (MoGeV2 [32]), and 6D pose tracking (Foundation￾Pose [34]). The over… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the KineBench Task and Benchmark Design [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Generalization and robustness comparison between KineBench and [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of pose and depth estimation errors under different [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Relationship between SPARC score and standardized success rate [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Manipulation score across different suites. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. WorldExam: Benchmarking World Models from Apparent Appearance to Inherent Reactivity

    cs.CV 2026-08 conditional novelty 6.0

    Across 1,474 cases and 20 models, WorldExam shows that video world models split along paradigm lines — camera-, action-, and language-driven models each dominate one capability, and none combines strong reactivity wit...

Reference graph

Works this paper leans on

44 extracted references · 16 linked inside Pith · cited by 1 Pith paper

  1. [1]

    IEEE Rransactions On Biomedical Engineering59(8), 2126–2136 (2011)

    Balasubramanian, S., Melendez-Calderon, A., Burdet, E.: A robust and sensitive metric for quantifying movement smoothness. IEEE Rransactions On Biomedical Engineering59(8), 2126–2136 (2011)

    2011

  2. [2]

    Bansal, H., Lin, Z., Xie, T., Zong, Z., Yarom, M., Bitton, Y., Jiang, C., Sun, Y., Chang, K.W., Grover, A.: Videophy: Evaluating physical commonsense for videogeneration.In:InternationalConferenceonLearningRepresentations(ICLR) (2025)

    2025

  3. [3]

    Brooks, T., Peebles, B., Holmes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video gener- ation models as world simulators (2024),https://openai.com/research/video- generation-models-as-world-simulators

    2024

  4. [4]

    In: Forty-first International Conference on Machine Learning (2024)

    Bruce, J., Dennis, M.D., Edwards, A., Parker-Holder, J., Shi, Y., Hughes, E., Lai, M., Mavalankar, A., Steigerwald, R., Apps, C., et al.: Genie: Generative interactive environments. In: Forty-first International Conference on Machine Learning (2024)

    2024

  5. [5]

    arXiv preprint arXiv:2506.21539 (2025)

    Cen, J., Yu, C., Yuan, H., Jiang, Y., Huang, S., Guo, J., Li, X., Song, Y., Luo, H., Wang, F., et al.: Worldvla: Towards autoregressive action world model. arXiv preprint arXiv:2506.21539 (2025)

    Pith/arXiv arXiv 2025

  6. [6]

    arXiv preprint arXiv:2512.15840 (2025)

    Chen, B., Zhang, T., Geng, H., Song, K., Zhang, C., Li, P., Freeman, W.T., Malik, J., Abbeel, P., Tedrake, R., et al.: Large video planner enables generalizable robot control. arXiv preprint arXiv:2512.15840 (2025)

    Pith/arXiv arXiv 2025

  7. [7]

    arXiv preprint arXiv:2509.22642 (2025)

    Chi, X., Jia, P., Fan, C.K., Ju, X., Mi, W., Zhang, K., Qin, Z., Tian, W., Ge, K., Li, H., et al.: Wow: Towards a world omniscient world model through embodied interaction. arXiv preprint arXiv:2509.22642 (2025)

    arXiv 2025

  8. [8]

    arXiv preprint arXiv:2601.15282 (2026)

    Deng, Y., Pan, Z., Zhang, H., Li, X., Hu, R., Ding, Y., Zou, Y., Zeng, Y., Zhou, D.: Rethinking video generation model for the embodied world. arXiv preprint arXiv:2601.15282 (2026)

    arXiv 2026

  9. [9]

    Advances in neural information processing systems36, 9156–9172 (2023)

    Du, Y., Yang, S., Dai, B., Dai, H., Nachum, O., Tenenbaum, J., Schuurmans, D., Abbeel, P.: Learning universal policies via text-guided video generation. Advances in neural information processing systems36, 9156–9172 (2023)

    2023

  10. [10]

    arXiv preprint arXiv:2601.04137 (2026)

    Fan, C.K., Chi, X., Ju, X., et al.: Wow, wo, val! a comprehensive embodied world model evaluation turing test. arXiv preprint arXiv:2601.04137 (2026)

    arXiv 2026

  11. [11]

    In: Advances in Neural Information Processing Systems 31, pp

    Ha, D., Schmidhuber, J.: Recurrent world models facilitate policy evolution. In: Advances in Neural Information Processing Systems 31, pp. 2451–2463. Curran As- sociates, Inc. (2018),https://papers.nips.cc/paper/7512-recurrent-world- models-facilitate-policy-evolution,https://worldmodels.github.io

    2018

  12. [12]

    arXiv preprint arXiv:1912.01603 (2019)

    Hafner, D., Lillicrap, T., Ba, J., Norouzi, M.: Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603 (2019)

    Pith/arXiv arXiv 1912

  13. [13]

    arXiv preprint arXiv:2010.02193 (2020)

    Hafner, D., Lillicrap, T., Norouzi, M., Ba, J.: Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193 (2020)

    Pith/arXiv arXiv 2010

  14. [14]

    Nature pp

    Hafner, D., Pasukonis, J., Ba, J., Lillicrap, T.: Mastering diverse control tasks through world models. Nature pp. 1–7 (2025)

    2025

  15. [15]

    arXiv preprint arXiv:2412.14803 (2024)

    Hu, Y., Guo, Y., Wang, P., Chen, X., Wang, Y.J., Zhang, J., Sreenath, K., Lu, C., Chen, J.: Video prediction policy: A generalist robot policy with predictive visual representations. arXiv preprint arXiv:2412.14803 (2024)

    Pith/arXiv arXiv 2024

  16. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) KineBench 17

    Huang, Z., He, Y., Yu, J., Zhang, F., Si, C., Jiang, Y., Zhang, Y., Wu, T., Jin, Q., Chanpaisit, N., Wang, Y., Chen, X., Wang, L., Lin, D., Qiao, Y., Liu, Z.: VBench: Comprehensive benchmark suite for video generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) KineBench 17

    2024

  17. [17]

    Advances in Neural Information Processing Systems38, 156210–156256 (2026)

    Jing, Z., Yang, S., Ao, J., Xiao, T., Jiang, Y.G., Bai, C.: Humanoidgen: Data generation for bimanual dexterous manipulation via llm reasoning. Advances in Neural Information Processing Systems38, 156210–156256 (2026)

    2026

  18. [18]

    arXiv preprint arXiv:2001.08361 (2020)

    Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)

    Pith/arXiv arXiv 2001

  19. [19]

    arXiv preprint arXiv:2412.03603 (2024)

    Kong, W., Tian, Q., Zhang, Z., Min, R., Dai, Z., Zhou, J., Xiong, J., Li, X., Wu, B., Zhang, J., et al.: Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603 (2024)

    Pith/arXiv arXiv 2024

  20. [20]

    In: 2017 IEEE in- ternational conference on robotics and automation (ICRA)

    Landi, C.T., Ferraguti, F., Sabattini, L., Secchi, C., Fantuzzi, C.: Admittance con- trol parameter adaptation for physical human-robot interaction. In: 2017 IEEE in- ternational conference on robotics and automation (ICRA). pp. 2911–2916. IEEE (2017)

    2017

  21. [21]

    arXiv preprint arXiv:2512.00425 (2025)

    Le, M.Q., Zhu, Y., Kalogeiton, V., Samaras, D.: What about gravity in video generation? post-training newton’s laws with verifiable rewards. arXiv preprint arXiv:2512.00425 (2025)

    arXiv 2025

  22. [22]

    2, 2022-06-27

    LeCun, Y., et al.: A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review62(1), 1–62 (2022)

    2022

  23. [23]

    arXiv preprint arXiv:2402.17177 (2024)

    Liu, Y., Zhang, K., Li, Y., Yan, Z., Gao, C., Chen, R., Yuan, Z., Huang, Y., Sun, H., Gao, J., et al.: Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177 (2024)

    Pith/arXiv arXiv 2024

  24. [24]

    In: International Conference on Machine Learning (ICML) (2025)

    Meng, F., et al.: Towards world simulator: Crafting physical commonsense-based benchmark for video generation. In: International Conference on Machine Learning (ICML) (2025)

    2025

  25. [25]

    arXiv preprint arXiv:2510.25238 (2025)

    Qiao, Q., et al.: Vadb: A large-scale video aesthetic database with professional and multi-dimensional annotations. arXiv preprint arXiv:2510.25238 (2025)

    arXiv 2025

  26. [26]

    arXiv preprint arXiv:2410.18072 (2024)

    Qin, I., et al.: Worldsimbench: Towards video generation models as world simula- tors. arXiv preprint arXiv:2410.18072 (2024)

    Pith/arXiv arXiv 2024

  27. [27]

    arXiv preprint arXiv:2512.13507 (2025)

    Seedance, T., Chen, H., Chen, S., Chen, X., Chen, Y., Chen, Y., Chen, Z., Cheng, F., Cheng, T., Cheng, X., et al.: Seedance 1.5 pro: A native audio-visual joint generation foundation model. arXiv preprint arXiv:2512.13507 (2025)

    Pith/arXiv arXiv 2025

  28. [28]

    arXiv preprint arXiv:2510.09653 (2025)

    Team, U.: Ultralytics yolo evolution: An overview of yolo26, yolo11, yolov8, and yolov5. arXiv preprint arXiv:2510.09653 (2025)

    arXiv 2025

  29. [29]

    arXiv preprint arXiv:1812.01717 (2018)

    Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)

    Pith/arXiv arXiv 2018

  30. [30]

    Upadhyay, U., et al.: How close are world models to the physical world? arXiv preprint arXiv:2601.21282 (2026)

    arXiv 2026

  31. [31]

    arXiv preprint arXiv:2503.20314 (2025)

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.W., Chen, D., Yu, F., Zhao, H., Yang, J., et al.: Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314 (2025)

    Pith/arXiv arXiv 2025

  32. [32]

    Wang, R., Xu, S., Dong, Y., Deng, Y., Xiang, J., Lv, Z., Sun, G., Tong, X., Yang, J.: Moge-2: Accurate monocular geometry with metric scale and sharp details (2025), https://arxiv.org/abs/2507.02546

    Pith/arXiv arXiv 2025

  33. [33]

    IEEE transactions on image processing 13(4), 600–612 (2004)

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600–612 (2004)

    2004

  34. [34]

    In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition

    Wen, B., Yang, W., Kautz, J., Birchfield, S.: Foundationpose: Unified 6d pose esti- mation and tracking of novel objects. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 17868–17879 (2024) 18 Zeyu Liu*, Zhangzhe Zhu*, Yang Zhang* (equal contribution)

    2024

  35. [35]

    arXiv preprint arXiv:2511.18870 (2025)

    Wu, B., Zou, C., Li, C., Huang, D., Yang, F., Tan, H., Peng, J., Wu, J., Xiong, J., Jiang, J., et al.: Hunyuanvideo 1.5 technical report. arXiv preprint arXiv:2511.18870 (2025)

    Pith/arXiv arXiv 2025

  36. [36]

    arXiv preprint arXiv:2310.06114 (2023)

    Yang, M., Du, Y., Ghasemipour, K., Tompson, J., Schuurmans, D., Abbeel, P.: Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114 (2023)

    Pith/arXiv arXiv 2023

  37. [37]

    arXiv preprint arXiv:2602.15922 (2026)

    Ye, S., et al.: World action models are zero-shot policies. arXiv preprint arXiv:2602.15922 (2026)

    Pith/arXiv arXiv 2026

  38. [38]

    The international journal of Robotics Research4(2), 3–9 (1985)

    Yoshikawa, T.: Manipulability of robotic mechanisms. The international journal of Robotics Research4(2), 3–9 (1985)

    1985

  39. [39]

    arXiv preprint arXiv:2510.11512 (2025)

    Yuan, J., et al.: Likephys: Evaluating intuitive physics understanding in video diffusion models via likelihood preference. arXiv preprint arXiv:2510.11512 (2025)

    arXiv 2025

  40. [40]

    In: The Fourteenth International Conference on Learning Representations (2026), https://openreview.net/forum?id=yDmb7xAfeb

    Zhang, J., Jiang, M., Dai, N., Lu, T., Uzunoglu, A., Zhang, S., Wei, Y., Wang, J., Patel, V.M., Liang, P.P., Khashabi, D., Peng, C., Chellappa, R., Shu, T., Yuille, A., Du, Y., Chen, J.: Wow!: World models in a closed-loop world. In: The Fourteenth International Conference on Learning Representations (2026), https://openreview.net/forum?id=yDmb7xAfeb

    2026

  41. [41]

    Trans- actions on Machine Learning Research (2025),https://openreview.net/forum? id=xT8BEgXmVc

    Zhang, Y., Bai, C., Zhao, B., Yan, J., Li, X., Li, X.: Decentralized transformers with centralized aggregation are sample-efficient multi-agent world models. Trans- actions on Machine Learning Research (2025),https://openreview.net/forum? id=xT8BEgXmVc

    2025

  42. [42]

    In: The Thirty- ninthAnnualConferenceonNeuralInformationProcessingSystems(2025),https: //openreview.net/forum?id=rRxFIOoEeF

    Zhang, Y., Li, X., Ye, J., Qiu, S., Qu, D., Li, X., Zhang, C., Bai, C.: Revisiting multi-agent world modeling from a diffusion-inspired perspective. In: The Thirty- ninthAnnualConferenceonNeuralInformationProcessingSystems(2025),https: //openreview.net/forum?id=rRxFIOoEeF

    2025

  43. [43]

    arXiv preprint arXiv:2512.06963 (2025)

    Zheng, Y., et al.: Videovla: Video generation as generalizable visual planning for robotic manipulation. arXiv preprint arXiv:2512.06963 (2025)

    arXiv 2025

  44. [44]

    arXiv preprint arXiv:2506.18241 (2025)

    Zhu, Y., et al.: Ewmbench: A comprehensive evaluation benchmark for embodied world models. arXiv preprint arXiv:2506.18241 (2025)

    arXiv 2025