Pith. sign in

REVIEW 6 major objections 5 minor 43 references

GAUGE: A Measurement-Grounded Benchmark for Physical Fidelity in Simulation Engines and Video World Models

T0 review · 6 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims that physical fidelity in simulation and video generation must be measured per physical regime, because no tested physics engine is uniformly faithful and video world models can match the form of physical laws while…

desk verdict A genuinely useful benchmark package with a solid engine track, but the world-model headline overreaches an unvalidated SAM3 centroid pipeline. read the letter →

arxiv 2608.05948 v1 pith:5RT7TCY4 submitted 2026-08-06 cs.AI cs.CVcs.RO

classification cs.AIcs.CVcs.RO
keywords physicalfidelityphysicsenginesvideoworldmodelssimulationbenchmarkmotioncapturerigid-bodydynamicsdeformableobjectsphysical-lawconsistency
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 argues that physical fidelity cannot be judged from visual plausibility, a single trajectory error, or textbook-looking equations. It introduces GAUGE, a measurement-grounded benchmark of 22 real-world task families, and uses it to test three physics engines and six video world models. The results show that no engine is uniformly faithful: errors concentrate in impulsive contact, rapid textile motion, and volumetric deformation. For world models, the paper shows that generated videos can be well fitted by the expected physical equation form while recovering wrong accelerations, momentum transfer, and oscillation timing. If true, both simulator selection and world-model evaluation must be task-specific and must separate equation structure from parameter accuracy.

What carries the argument

The load-bearing object is the benchmark's generalized trajectory representation: rigid-body position $P(t)$, marker-wise Gaussian curvature $K(t)$ for textiles, and face areas $A(t)$ for volumetric bodies, computed frame-by-frame and compared with the real mean trajectory via RMSE and dynamic time warping. On the world-model side, the distinguishing mechanism is the separation of law-form consistency, measured by $R^2$ and quadratic form improvement, from parameter recovery, measured by fitted acceleration, period, and momentum-transfer efficiency. The SAM3-based centroid tracking with metric scaling from known object dimensions supplies the trajectory estimate. Together, these components let the benchmark localize a failure to a physical mechanism rather than reporting a single aggregate score.

What would settle it

Render a physics-engine rollout with known ground-truth accelerations and periods, run the same SAM3-based tracking and metric pipeline on the rendered video, and compare the recovered values with the known simulation values; if the recovery error is comparable to the world-model errors, the claim that models output wrong parameters is weakened, while accurate recovery would support it.

Watch

Extended reading notes

Core claim

GAUGE comprises approximately 1,560 motion-capture trials across 22 task families covering rigid bodies, cables, textiles, and volumetric deformable objects, recorded with sub-millimeter localization and paired with calibrated friction, restitution, cloth stiffness, Young's modulus, and Poisson's ratio. On the engine track, Isaac Sim, Genesis, and Newton were run on 14 task families with GAUGE's calibrated parameters and default solver settings, then compared against real mean trajectories using RMSE and dynamic time warping, plus task-specific observables such as longest stationary duration, momentum transfer efficiency, period, and energy loss. No engine dominated: the best bouncing-ball result was still 15.63 times the real-world RMSE baseline, Isaac Sim reached 128.26 times baseline on textile flinging, and deformable-body errors remained roughly an order of magnitude above real-world baselines. On the world-model track, six generative video models were prompted with the same initial frames and evaluated by recovering centroid trajectories with SAM3 segmentation, then testing dynamic error, coefficient of determination, and quadratic form improvement alongside fitted acceleration, period, and momentum transfer efficiency. The central result is that models can achieve high $R^2$ or low QFI while recovering parameters far from reality, such as a pendulum fit of $R^2 = 0.99$ with a 1.93-second period against a real 1.06 seconds, or a free-fall fit with $0.088\,\text{m}/\text{s}^2$ against $9.81\,\text{m}/\text{s}^2$.

Load-bearing premise

The world-model conclusions rest on the assumption that the 2D centroid positions extracted from generated videos by automatic segmentation and scaled by known object dimensions are accurate enough to estimate accelerations, periods, and momentum transfer; if segmentation bias or scale errors creep in, the reported parameter errors could be perception errors rather than physics errors.

Editorial extensions

If this is right

  • Simulator selection should be guided by task regime: Isaac Sim led on several rigid-contact tasks, Genesis on dynamic textiles and most foam deformations, Newton on selected bending cases, and no single engine should be treated as a default for all physical interactions.
  • World-model evaluations should report equation-form fit and physical-parameter recovery separately, since a high $R^2$ or low QFI can accompany a badly wrong acceleration, period, or momentum transfer.
  • Generated videos that look physically plausible cannot safely be used as training or policy-evaluation environments until the recovered acceleration, period, and contact behavior are checked against real measurements.
  • Prompt sensitivity is a measured effect: the negative-prompt condition improved some model-task pairs and worsened others, so benchmark protocols should fix prompt templates and report paired conditions.
  • The calibrated material parameters let a developer trace an engine's failure to a specific physical mechanism such as contact, friction, or stiffness rather than to overall visual quality.

Reading between the lines

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

  • An implicit next step is carrying the form-versus-parameter distinction into other generative domains, such as fluids, acoustics, or soft tissues, where a model may reproduce a governing equation's shape while getting its coefficients wrong.
  • A testable extension is to feed a rendered physics-engine rollout through the same SAM3 trajectory pipeline used in the paper; if recovered accelerations and periods deviate as much as the world-model outputs do, part of the reported parameter error is a perception artifact rather than a modeling error.
  • The complementary engine strengths suggest that a hybrid or learned solver switching integrators by regime could outperform any single engine on this benchmark, and GAUGE's per-task errors provide the baseline to test that.
  • The paper's stated limitation that the world-model track is limited to 2D rigid-body tasks implies a natural expansion: once 3D tracking is available, the curvature and face-area metrics from the engine track could be applied to generated textiles and deformables.
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

6 major / 5 minor

Summary. The paper introduces GAUGE, a benchmark of 22 real-world physics task families with motion-capture ground truth, calibrated material parameters, and task-specific observables. It evaluates three physics engines (Isaac Sim, Genesis, Newton) on 14 tasks using generalized trajectory errors and task-specific metrics, and evaluates six video world models on five rigid-body tasks using metrics that separate equation-form fit from parameter accuracy. The central findings are that no engine is uniformly faithful, with the largest errors in impulsive contact, rapid textile motion, and volumetric deformation, and that world models can produce trajectories with the expected equation form while recovering incorrect accelerations, momentum-transfer efficiencies, and oscillation periods. The manuscript is unusually explicit about thresholds, fallbacks, and limitations, and the appendices provide substantial detail on calibration and metric computation.

Significance. If the results hold, GAUGE would be a valuable resource for the community: it combines cross-regime real-world grounding, independent parameter calibration, and a separation between physical-law form and parameter accuracy that prior benchmarks largely conflate. The engine-track comparison, the prompt-sensitivity ablation, and the explicit disclosure of effective-parameter mappings are genuine strengths, and the paper avoids the circularity pitfall of fitting simulator parameters from the evaluated trajectories. However, the world-model conclusions are currently only as strong as the SAM3 centroid-to-metric tracking pipeline, which is not validated against known trajectories and has no reported seed variance or uncertainty propagation. In addition, several metric definitions and reported numbers are internally inconsistent. These issues are local and fixable, but they must be resolved before the abstract's world-model claim is fully supported.

major comments (6)
  1. [Sections 3.3.1, C.2.1, and 6] The world-model track rests entirely on the SAM3 centroid extraction and pixel-to-meter scaling pipeline described in Section 3.3.1 and C.2.1, but this pipeline is never validated against a known trajectory. Every quantitative world-model result (Table 4 and the fits in Eqs. 25, 36, and 45) inherits any segmentation bias, scale error, or depth drift from this pipeline. No seed variance, no uncertainty propagation, and no cross-check against manual or synthetic ground truth are reported. Section 6 explicitly states that 2D trajectories are insufficient and that correspondence and reconstruction uncertainty must be modeled before errors are attributed to the world model. The manuscript should therefore either provide such validation (for example, by running the identical pipeline on synthetic videos with known trajectories and reporting bias/variance) or substantially soften the abstract and conclusion claims that world models 'recover incorrect accelerations, momentum transfer, and oscillation timing.'
  2. [Section 3.3.2 Eq. (3) vs. Section C.3.3 Eq. (32)] The Dynamic Error is defined differently in the main text and the appendix. Eq. (3) defines DE as the mean absolute difference between m*x_ddot and the applied force, while Eq. (32) defines it as the mean positive part of (m*a - F_fric,max), i.e., a one-sided friction-exceedance penalty. These are not the same quantity, and the zero DE values in Table 4 do not indicate force-motion consistency under Eq. (3); the appendix itself notes that overly slow or stationary motion can also produce DE=0. The main text should either reconcile the two definitions or explicitly state that DE is a one-sided exceedance metric, not the general force-motion error promised in Section 3.3.2.
  3. [Section B.1, Eq. (17) and Table 3] The Energy Loss definition is internally inconsistent. Eq. (17) defines EL = (E_start - E_end)/E_start, for which zero is the ideal value for a conservative system, positive values indicate energy loss, and negative values indicate numerical energy gain. The text immediately after Eq. (17) says 'A value closer to 1 indicates better energy conservation,' which contradicts the equation and also contradicts the Table 3 note stating that zero is ideal for the pendulum EL row. The sentence in B.1 should be corrected to reflect the actual definition, and the pendulum EL discussion in Section 5.1 should be checked against the corrected interpretation.
  4. [Table 3 vs. Table 4 and Fig. 4] There are unexplained numeric inconsistencies in the reported baselines and results. The pendulum baseline period is 1.14 s in Table 3 but 1.06 s in Table 4 and Section C.3.6, even though both tables refer to the same real-world pendulum task; this directly changes the reported period error percentages in Section 5.2. In addition, the QFI values for negative-prompt rows in Fig. 4(a) and (b) do not match Table 4: for example, the wood slope-slider QFI for Cosmos3-Super-I2V with negative prompt is 569.36 in Table 4 but 174.0 in Fig. 4(a), and the bouncing-ball QFI for the same configuration is 12.50 in Table 4 but 3.1 in Fig. 4(b). These discrepancies must be resolved, as they are used to support the claim that prompt sensitivity is large and model-dependent.
  5. [Sections 3.3 and C.2.1] The world-model evaluation covers only five of the eight rigid-body tasks, yet the abstract and conclusion state the finding in general terms. The Appendix C.2.1 caption says that slope contact and non-smooth contact are excluded because of substantial 3D geometric deformations and discontinuous object states, and wall breaking is also absent from Table 4. This restriction is defensible, but it should be stated in the main text, and the conclusion 'current video world models exhibit ... incorrect accelerations, momentum-transfer efficiencies, oscillation periods' should be explicitly scoped to the five evaluated tasks. Otherwise readers may over-generalize the finding to all rigid-body scenarios.
  6. [Section A.2.2 and Table 3] The textile engine-track results depend on a chain of effective-parameter mappings with several ad hoc fallbacks (Eqs. 6, 10, and 11), including fixed engineering priors such as the 2.7 ratio and the 0.30/0.45 Poisson fallbacks. The authors disclose these choices, and the large error magnitudes in textile flinging and bending are unlikely to be erased by small parameter changes. Still, because the conclusion that 'the largest discrepancies arise in rapid textile motion' is based on these mappings, it would strengthen the paper to report a small sensitivity analysis over the fallback values, or at minimum to state in Section 5.1 how much of the observed textile error could plausibly be attributed to the parameter-mapping assumptions rather than to solver dynamics.
minor comments (5)
  1. [General] The paper uses both 'QF' and 'QFI' in Figure 4 and Table 4; the notation should be unified.
  2. [Section C.3.4] The bouncing-ball figure legend for the negative-prompt condition appears to show a different QFI value from the table; if the figure is generated from a different run, the discrepancy should be explained or corrected.
  3. [Section 3.2.1] The Hungarian-marker matching is stated to have an error below 1 cm, but the method for computing this error is not described; a one-sentence clarification would help.
  4. [Section 5.1] The phrase 'normalized periods of 1.10 and 1.09' in the pendulum discussion should explicitly state that these are ratios to the baseline period, consistent with the table header, to avoid confusion with seconds.
  5. [Section C.3.3] The claim that zero DE values are compatible with 'slightly slow, nearly stationary, or incorrectly synchronized motion' is important and should be reflected in the main-text discussion of the turntable results, where several DE entries are zero.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: engine parameters and world-model metrics are measurement outputs, not fitted inputs.

full rationale

GAUGE's derivation chain is self-contained. In the physics-engine track, simulator inputs are calibrated physical parameters obtained from independent characterization tests (Section A.2.1: inclined-plane friction, air-bearing restitution, Style3D cloth tests, DIC for Young's modulus and Poisson's ratio), and evaluation compares simulated generalized trajectories against real motion capture. No simulator parameter is fitted from the evaluated trajectories, so the reported engine errors are measurement outputs. In the world-model track, the model receives only an initial frame and prompt; SAM3 segmentation and centroid tracking (Eq. 18) yield pixel trajectories, converted to metric coordinates using known object dimensions (Eqs. 22, 33), and accelerations, MTE, and periods are fitted from those recovered trajectories (Eqs. 25, 36, 38-39, 45). These fitted quantities are generated-video measurements, not inputs to the world model, so the observed dissociation between equation form (low QFI / high R2) and parameter accuracy is an empirical finding rather than a construction. The paper's self-references are limited to the project page and its own benchmark table; no load-bearing claim rests on a self-citation. Section 6's limitation that 2D trajectories are insufficient and that reconstruction uncertainty must be modeled is a measurement-validity caveat for the world-model track, not a circular step: it weakens attribution of errors to the world model, but does not make the evaluation reduce to its own inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical entities, forces, or conservation laws are introduced. The free parameters are evaluation thresholds and disclosed engineering priors in parameter conversion; the axioms are measurement and projection assumptions on which the empirical claims rest.

free parameters (4)
  • LSD stationarity thresholds = epsilon_z = 0.5 mm, v_th = 5 mm/s
    Hand-set thresholds in Eq. (15) that define stationary intervals for the Newton's cradle LSD metric; they directly affect engine scores.
  • Newton VBD isotropic ratio = 2.7
    Eq. (11) uses 2.7 as a fixed engineering prior when K_s/K_sh is outside [2,4]; the paper calls it not a theoretical constant.
  • Textile Poisson fallback = nu = 0.30 or 0.45
    Eq. (6) substitutes fixed values when measured Poisson ratio falls outside the supported [0, 0.5) range, altering solver stiffness.
  • World-model pixel-to-meter scale = per-task known object dimensions (0.05 m cube, 0.06 m ball)
    Section C.2.1; assumed known object size converts 2D centroids to metric positions, and annotation errors propagate into fitted acceleration and period.
assumptions (5)
  • domain assumption Marker masses and retroreflective markers do not measurably alter object dynamics.
    Section 4.1 states marker masses are negligible; the real baseline trajectories assume markers do not perturb the physics.
  • domain assumption The motion capture system provides sub-millimeter 3D localization at 180 Hz, sufficient to treat recorded trajectories as ground truth.
    Section 4.1; all real baselines and parameter calibrations depend on this accuracy.
  • domain assumption For world-model evaluation, the object stays at roughly constant depth relative to the camera, so a single metric scale converts pixel centroids to world trajectories.
    Sections 3.3.1 and C.2.1; this underpins the fitted acceleration, period, and MTE values.
  • ad hoc to paper Measured anisotropic textile properties can be represented by effective isotropic solver parameters matched at small strain.
    Appendix A.2.2; the mapping uses free-transverse uniaxial assumptions and fixed fallbacks, and Genesis has no input channel for shear stiffness.
  • ad hoc to paper The angular deficit at mesh vertices is a valid generalized trajectory for textile deformation, although it is not normalized by area and hence is not a true Gaussian curvature.
    Appendix B.1 explicitly acknowledges the missing length^-2 dimension, yet uses this quantity for RMSE/DTW comparisons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GAUGE: A Measurement-Grounded Benchmark for Physical Fidelity in Simulation Engines and Video World Models." pith.science (2026). https://pith.science/paper/5RT7TCY4

@misc{pith2026260805948,
  author       = {Pith},
  title        = {Pith review of: GAUGE: A Measurement-Grounded Benchmark for Physical Fidelity in Simulation Engines and Video World Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5RT7TCY4}},
  note         = {Machine review of arXiv:2608.05948}
}
read the original abstract

Physics engines facilitate large-scale training and evaluation for embodied intelligence, while generative video world models are emerging as implicit simulators of future states and interactions. However, existing evaluations of physical fidelity are often conducted in isolation and rely heavily on perceptual similarity or human judgments, providing limited insight into which physical principles or parameters are violated. We introduce GAUGE, a real-world-grounded diagnostic benchmark for jointly evaluating how numerical simulators and generative video world models reproduce or deviate from real-world physics. It comprises 22 controlled task families covering rigid bodies, flexible cables, textiles, and volumetric deformable objects. Grounded in real-world trajectories and paired with calibrated physical metadata, uncertainty annotations, and task-specific observables, these tasks cover fundamental physical processes including collision, friction, momentum transfer, oscillation, self-contact, and deformation across diverse materials and conditions. We benchmark Isaac Sim, Genesis, and Newton on 14 task families using generalized trajectory errors, and evaluate 6 image-to-video models on 5 rigid-body tasks by testing physical-law consistency and the temporal stability of inferred parameters. Our results reveal no uniformly faithful physics engine, with the largest discrepancies arising in impulsive contact, rapid textile motion, and volumetric deformation. We further find that video world models can produce trajectories with the expected equation form while recovering incorrect accelerations, momentum transfer, and oscillation timing. GAUGE lays the groundwork for developing more physically faithful simulators and world models for embodied intelligence.

Figures

Figures reproduced from arXiv: 2608.05948 by the authors.

Figure 1
Figure 1. Overview of GAUGE. Controlled real-world experiments provide calibrated physical metadata and motion-capture ground truth for rigid bodies, textiles, and volumetric deformable bodies. The physics-engine track (left) compares trajectories from matched simulated scenes with real measurements. The video world-model track (right) generates future motion from an initial frame and prompt, then evaluates the tracked motion… view at source ↗
Figure 2
Figure 2. Generalized trajectories used for physics-engine evaluation. (a) Rigid-body motion is [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Real-world motion-capture system and representative tracking configurations. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Representative physical-law fits for generated videos. Displacement is plotted against [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Overview of GAUGE’s 22 standardized real-world task families [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Geometry variants of the nonsmooth contact task. [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Experimental characterization and validation of task-specific physical parameters. [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Representative generalized trajectories used for physics-engine evaluation. Red curves and [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]
Figure 9
Figure 9. Figure 9: Example of qualitative trajectory visualization for physical realism evaluation. The last three [PITH_FULL_IMAGE:figures/full_fig_p033_9.png]
Figure 10
Figure 10. Figure 10: Slope-slider trajectories for the three materials. Points are the retained, unfiltered projected [PITH_FULL_IMAGE:figures/full_fig_p036_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 15 canonical work pages

  1. [1]

    Validating robotics simulators on real-world impacts, 2021

    Brian Acosta, William Yang, and Michael Posa. Validating robotics simulators on real-world impacts, 2021. doi: 10.48550/ARXIV.2110.00541

  2. [2]

    Videophy-2: A challenging action-centric physical commonsense evaluation in video generation, 2025

    Hritik Bansal, Clark Peng, Yonatan Bitton, Roman Goldenberg, Aditya Grover, and Kai-Wei Chang. Videophy-2: A challenging action-centric physical commonsense evaluation in video generation, 2025. doi: 10.48550/ARXIV.2503.06800

  3. [3]

    Benchmarking the sim-to-real gap in cloth manipulation.IEEE Robotics and Automation Letters,9(3), Mar 2024

    David Blanco-Mulero, Oriol Barbany, Gokhan Alcan, Adrià Colomé, Carme Torras, and Ville Kyrki. Benchmarking the sim-to-real gap in cloth manipulation.IEEE Robotics and Automation Letters,9(3), Mar 2024. doi: 10.1109/lra.2024.3360814

  4. [4]

    Genie: Generative interactive environments, 2024

    Jake Bruce, Michael Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, Yusuf Aytar, Sarah Bechtle, Feryal Behbahani, Stephanie Chan, Nicolas Heess, Lucy Gonzalez, Simon Osindero, Sherjil Ozair, Scott Reed, Jingwei Zhang, Konrad Zolna, Jeff Clune, Nando de Freitas, Satinder Si...

  5. [5]

    Sam 3: Segment anything with concepts, 2025

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman Rädle, Triantafyllos Afouras, Effrosyni Mavroudi, Katherine Xu, Tsung-Han Wu, Yu Zhou, Liliane ...

  6. [6]

    14 Benchmarking simulated robotic manipulation through a real world dataset.IEEE Robotics and Automation Letters,5(1), Jan 2020

    Jack Collins, Jessie McVicar, David Wedlock, Ross Brown, David Howard, and Jurgen Leitner. 14 Benchmarking simulated robotic manipulation through a real world dataset.IEEE Robotics and Automation Letters,5(1), Jan 2020. doi: 10.1109/lra.2019.2953663

  7. [7]

    Genesis world: Pbd cloth solver, 2026, https://github.com/ Genesis-Embodied-AI/genesis-world

    Genesis AI Team. Genesis world: Pbd cloth solver, 2026, https://github.com/ Genesis-Embodied-AI/genesis-world

  8. [8]

    Hirani, Mathieu Desbrun, and Peter Schröder

    Eitan Grinspun, Anil N. Hirani, Mathieu Desbrun, and Peter Schröder. Discrete shells, 2003. doi: 10.2312/SCA03/062-067

Show all 43 references
  1. [9]

    phyworldbench

    Jing Gu, Xian Liu, Yu Zeng, Ashwin Nagarajan, Fangrui Zhu, Daniel Hong, Yue Fan, Qianqi Yan, Kaiwen Zhou, Ming-Yu Liu, and Xin Eric Wang. "phyworldbench": A comprehensive evaluation of physical realism in text-to-video models, 2025. doi: 10.48550/ARXIV.2507.13428

  2. [10]

    Omnifysics: Towards physical intelligence evolution via omni-modal signal processing and network optimization,

    Minghao Han, Dingkang Yang, Yue Jiang, Yizhou Liu, and Lihua Zhang. Omnifysics: Towards physical intelligence evolution via omni-modal signal processing and network optimization,

  3. [11]

    Real garment benchmark (rgbench): A comprehensive benchmark for robotic garment manipulation featuring a high-fidelity scalable simulator, 2025

    Wenkang Hu, Xincheng Tang, Yanzhi E, Yitong Li, Zhengjie Shu, Wei Li, Huamin Wang, and Ruigang Yang. Real garment benchmark (rgbench): A comprehensive benchmark for robotic garment manipulation featuring a high-fidelity scalable simulator, 2025. doi: 10.48550/ARXIV. 2511.06434

  4. [12]

    InProceedings of the 2nd Workshop on World Models: Understanding, Modelling and Scaling at ICLR 2026, 2026

    SwarnimJainandShangzheWu.RigidBench: EvaluatingRigid-BodyPhysicsinVideoGeneration Models. InProceedings of the 2nd Workshop on World Models: Understanding, Modelling and Scaling at ICLR 2026, 2026

  5. [13]

    How far is video generation from world model: A physical law perspective, 2024

    Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective, 2024. doi: 10.48550/ARXIV.2411.02385

  6. [14]

    Iris: A real-world benchmark for inverse recovery and identification of physical dynamic systems from monocular video, 2026

    Rasul Khanbayov, Mohamed Rayan Barhdadi, Erchin Serpedin, and Hasan Kurban. Iris: A real-world benchmark for inverse recovery and identification of physical dynamic systems from monocular video, 2026. doi: 10.48550/ARXIV.2603.16432

  7. [15]

    H. W. Kuhn. The hungarian method for the assignment problem.Naval Research Logistics Quarterly,2(1-2), Mar 1955. doi: 10.1002/nav.3800020109

  8. [16]

    Pisa experi- ments: Exploring physics post-training for video diffusion models by watching stuff drop, 2025

    Chenyu Li, Oscar Michel, Xichen Pan, Sainan Liu, Mike Roberts, and Saining Xie. Pisa experi- ments: Exploring physics post-training for video diffusion models by watching stuff drop, 2025. doi: 10.48550/ARXIV.2503.09595

  9. [17]

    Gonzalez, Ion Stoica, Song Han, and Yao Lu

    Dacheng Li, Yunhao Fang, Yukang Chen, Shuo Yang, Shiyi Cao, Justin Wong, Michael Luo, Xiaolong Wang, Hongxu Yin, Joseph E. Gonzalez, Ion Stoica, Song Han, and Yao Lu. Worldmod- elbench: Judging video generation models as world models, 2025. doi: 10.48550/ARXIV.2502. 20694

  10. [18]

    Evaluating real-world robot manipulation policies in simulation,

    Xuanlin Li, Kyle Hsu, Jiayuan Gu, Karl Pertsch, Oier Mees, Homer Rich Walke, Chuyuan Fu, Ishikaa Lunawat, Isabel Sieh, Sean Kirmani, Sergey Levine, Jiajun Wu, Chelsea Finn, Hao Su, Quan Vuong, and Ted Xiao. Evaluating real-world robot manipulation policies in simulation,

  11. [19]

    Xpbd: position-based simulation of compliant constrained dynamics

    Miles Macklin, Matthias Müller, and Nuttapong Chentanez. Xpbd: position-based simulation of compliant constrained dynamics. InProceedings of the 9th International Conference on Motion in Games, Oct 2016. 15

  12. [20]

    A user’s guide to calibrating robotics simulators, 2020

    Bhairav Mehta, Ankur Handa, Dieter Fox, and Fabio Ramos. A user’s guide to calibrating robotics simulators, 2020. doi: 10.48550/ARXIV.2011.08985

  13. [21]

    Michelis, and Robert K

    Manuel Mekkattu, Mike Y. Michelis, and Robert K. Katzschmann. Sors: A modular, high-fidelity simulator for soft robots, 2025. doi: 10.48550/ARXIV.2512.15994

  14. [22]

    Towards world simulator: Crafting physical commonsense- based benchmark for video generation, 2024

    Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense- based benchmark for video generation, 2024. doi: 10.48550/ARXIV.2410.05363

  15. [23]

    Do generative video models understand physical principles?, 2025

    Saman Motamed, Laura Culp, Kevin Swersky, Priyank Jaini, and Robert Geirhos. Do generative video models understand physical principles?, 2025. doi: 10.48550/ARXIV.2501.09038

  16. [24]

    Journal of Visual Communication and Image Representation,18(2), Apr 2007

    MatthiasMüller, BrunoHeidelberger, MarcusHennix, andJohnRatcliff.Positionbaseddynamics. Journal of Visual Communication and Image Representation,18(2), Apr 2007. doi: 10.1016/j. jvcir.2007.01.005

  17. [25]

    Newton physics: Vbd cloth implementation, 2026,https://github

    Newton Developers. Newton physics: Vbd cloth implementation, 2026,https://github. com/newton-physics/newton

  18. [26]

    NVIDIA, :, Aditi, Niket Agarwal, Arslan Ali, Jon Allen, Martin Antolini, Adeline Aubame, Alisson Azzolini, Junjie Bai, Maciej Bala, Yogesh Balaji, Josh Bapst, Aarti Basant, Mukesh Beladiya, Mohammad Qazim Bhat, Zaid Pervaiz Bhat, Dan Blick, Vanni Brighella, Han Cai, Tiffany Ca...

  19. [27]

    Omni physics: Surface deformable bodies, 2026, https://docs.omniverse

    NVIDIA. Omni physics: Surface deformable bodies, 2026, https://docs.omniverse. nvidia.com/kit/docs/omni_physics/110.1/dev_guide/deformables/ deformable_bodies.html

  20. [28]

    Katzschmann, and Stelian Coros

    Jan Obrist, Miguel Zamora, Hehui Zheng, Ronan Hinchet, Firat Ozdemir, Juan Zarate, Robert K. Katzschmann, and Stelian Coros. Pokeflex: A real-world dataset of volumetric deformable objects for robotics, 2024. doi: 10.48550/ARXIV.2410.07688

  21. [29]

    doi: 10.48550/ARXIV.2606.02800

  22. [30]

    Worldgym: World model as an environment for policy evaluation, 2025

    Julian Quevedo, Ansh Kumar Sharma, Yixiang Sun, Varad Suryavanshi, Percy Liang, and Sherry Yang. Worldgym: World model as an environment for policy evaluation, 2025. doi: 10.48550/ARXIV.2506.00613

  23. [31]

    Physical validation of simulators in com- puter graphics: a new framework dedicated to slender elastic structures and frictional contact

    Victor Romero, Mickaël Ly, Abdullah Haroon Rasheed, Raphaël Charrondière, Arnaud Lazarus, Sébastien Neukirch, and Florence Bertails-Descoubes. Physical validation of simulators in com- puter graphics: a new framework dedicated to slender elastic structures and frictional conta...

  24. [32]

    Scalable real2sim: Physics-aware asset generation via robotic pick-and-place setups, 2025

    Nicholas Pfaff, Evelyn Fu, Jeremy Binagia, Phillip Isola, and Russ Tedrake. Scalable real2sim: Physics-aware asset generation via robotic pick-and-place setups, 2025. doi: 10.48550/ARXIV. 2503.00370

  25. [33]

    Team Seedance, De Chen, Liyang Chen, Xin Chen, Ying Chen, Zhuo Chen, Zhuowei Chen, Feng Cheng, Tianheng Cheng, Yufeng Cheng, Mojie Chi, Xuyan Chi, Jian Cong, Qinpeng Cui, Fei Ding, Qide Dong, Yujiao Du, Haojie Duanmu, Junliang Fan, Jiarui Fang, Jing Fang, Zetao Fang, Chengjian...

  26. [34]

    Stable neo-hookean flesh simulation

    Breannan Smith, Fernando De Goes, and Theodore Kim. Stable neo-hookean flesh simulation. ACM Transactions on Graphics,37(2), Mar 2018. doi: 10.1145/3180491

  27. [35]

    Tim Rädsch, Yuki M Asano, Hilde Kuehne, Stefan Bauer, Priyank Jaini, Robert Geirhos, and Carsten T. Lüth. Physics-iq verified, 2026. doi: 10.48550/ARXIV.2606.18943

  28. [36]

    Reconciling reality through simulation: A real-to-sim-to-real approach for robust manipulation, 2024

    Marcel Torne, Anthony Simeonov, Zechu Li, April Chan, Tao Chen, Abhishek Gupta, and Pulkit Agrawal. Reconciling reality through simulation: A real-to-sim-to-real approach for robust manipulation, 2024. doi: 10.48550/ARXIV.2403.03949

  29. [37]

    Worldbench: Disam- biguating physics for diagnostic evaluation of world models, 2026

    Rishi Upadhyay, Howard Zhang, Jim Solomon, Ayush Agrawal, Pranay Boreddy, Shruti Satya Narayana, Yunhao Ba, Alex Wong, Celso M de Melo, and Achuta Kadambi. Worldbench: Disam- biguating physics for diagnostic evaluation of world models, 2026. doi: 10.48550/ARXIV.2601. 21282

  30. [38]

    Diffcloud: Real-to-sim from point clouds with differentiable simulation and rendering of deformable objects, 2022

    Priya Sundaresan, Rika Antonova, and Jeannette Bohg. Diffcloud: Real-to-sim from point clouds with differentiable simulation and rendering of deformable objects, 2022. doi: 10.48550/ARXIV. 2204.03139

  31. [39]

    Daydreamer: World models for physical robot learning, 2022

    Philipp Wu, Alejandro Escontrela, Danijar Hafner, Ken Goldberg, and Pieter Abbeel. Daydreamer: World models for physical robot learning, 2022. doi: 10.48550/ARXIV.2206.14176

  32. [40]

    Nonsmooth Contact

    Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng, Shenyuan Gao, Sihyun Yu, George Kurian, Suneel Indupuru, You Liang Tan, Chuning Zhu, Jiannan Xiang, Ayaan Malik, Kyungmin Lee, William Liang, Nadun Ranawaka, Jiasheng Gu, Yinzhen Xu, Guanzhi Wang, Fengyuan Hu, Avnish Narayan, Johan Bjor...

  33. [41]

    Wan: Open and advanced large-scale video generative models, 2025

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pande...

  34. [2024]

    doi: 10.48550/ARXIV.2405.05941

  35. [2026]

    doi: 10.48550/ARXIV.2602.07064

Pith tools

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