Robotics · arXiv:2610.33807 v1 · submitted 24 Oct 2026 · 114 references
RoboWorM: Evaluating Movement Realism, Consistency and Acceleration Affects in Embodied Video World Models
The paper says
“Video world models (VWMs) are rapidly emerging as scalable neural simulators for robotic manipulation. However, most benchmarks only evaluate them as open-world video generators, using Vision-Language Models (VLMs) that lack strict 3D spatial and kinematic reasoning.”
more of the abstract
To bridge this gap, we introduce RoboWorM, a comprehensive, deployment-aware benchmark that leverages known hardware priors such as robot kinematic trees and fixed camera configurations, to systematically examine structural and spatiotemporal fidelity. RoboWorM curates 1,605 episodes across single-arm, bimanual, and humanoid morphologies, spanning both Markovian and long-horizon non-Markovian tasks under single- and multi-view setups. To explicitly measure physical consistency, we propose two deterministic, VLM-free metrics: KIVA for intra-view kinematic stability, and RIGIS for multi-view geometric consistency. By exploiting intrinsic robotic invariants, KIVA improves hallucination detection alignment with human annotations by 25.8% relative to the strongest frontier VLMs, while RIGIS achieves a 130% higher Spearman correlation with human preferences. Furthermore, since real-world deployment requires low inference latency, we utilize RoboWorM to conduct the first large-scale audit of the quality-efficiency trade-off. Evaluating 9 state-of-the-art acceleration techniques across 3 leading world models (over 29,000 generated rollouts), we demonstrate how these acceleration methods can retain superficial visual quality while silently compromising deployment-critical physical constraints. Using our metrics, we achieve 2.3× maximum inference speed-up with minimal quality loss.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 27 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. RoboWorM delivers a principled, deterministic alternative to VLM-based evaluation for robotic video world models; the central claims hold, but the noisy human ground truth and unvalidated depth estimator introduce moderate caveats. This line is wrong
4 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 114 references say.
The paper says, at §3.4.2 RIGIS, Appendix B.2
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
RIGIS depends critically on per-view depth maps and camera poses predicted by VGGT-Ω (Wang et al., 2026). The paper provides no validation that VGGT-Ω produces accurate depth for the robot manipulation scenes in the benchmark (cluttered tabletops, reflective surfaces, thin grippers).
more of objection 1
Scale alignment (Eq. B.3) corrects only median log-ratio drift, not structured errors. Without evidence of VGGT-Ω's domain-specific accuracy, the reported Spearman ρ = 0.749 (and the claimed 130% improvement over VLMs) could reflect depth-estimator biases coincidentally agreeing with human perceptual preferences rather than genuine geometric consistency. The limitation statement (Section 5) does not mention this issue. The authors should either (a) provide a sanity check using known camera extrinsics and SfM point clouds on real sequences from the same domains, or (b) clearly bound the metric's validity and discuss how VGGT-Ω bias might affect the reported correlations.
The paper says, at §3.4.1 KIVA, Eq. B.1, Appendix D.3
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
KIVA's violation thresholds (γ_e) are calibrated per embodiment using only 9 human-labeled videos (3 per quality level) drawn from a held-out subset. Appendix D.3 reports that calibration improves macro overlap from 60.8% to 67.5% with ±2.8 standard deviation across 1000 random draws of the 9 videos.
more of objection 2
This calibration set is extremely small relative to the variation in the data (inter-annotator agreement is only 53.6% and 56.0% on the full set). The threshold choice may be unstable across different calibration draws, and the reported 25.8% improvement over VLM judges (73.5% vs 58.4% micro overlap) is based on this calibrated KIVA. The paper should assess whether the advantage holds under a more robust calibration procedure (e.g., a larger or different set of calibration videos) and discuss the sensitivity of the comparison to calibration set composition.
The paper says, at §4.1, Appendix D.1.4
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
The inter-annotator agreement on the human ground truth is low: 53.6% for embodied physics and 56.0% for multi-view consistency (Fig. D2c). This means the reference against which both KIVA/RIGIS and VLM judges are compared is itself noisy. The paper reports alignment metrics (Spearman, overlap, pairwise rank accuracy) against this noisy ground truth.
more of objection 3
The comparisons may be affected by annotator noise; for instance, a metric that systematically agrees with one annotator's bias could show inflated alignment. The paper should discuss the implications of the low agreement for the reliability of the human-alignment comparisons and consider whether the observed improvements are robust to annot cator label uncertainty (e.g., via Monte Carlo label noise injection).
The paper says, at §3.4.2 RIGIS, Appendix B.2
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
RIGIS uses VGGT-Ω, a learned video- geometry model, to estimate depth and camera poses. While the paper claims RIGIS is 'VLM-free', VGGT-Ω is still a learned estimator that may introduce biases. The metric's score S is a composite of geometric consistency and pose stability, but both depend on these estimates.
more of objection 4
The paper should clarify that RIGIS is not a pure geometric metric but a learned-surogate metric, and discuss how possible biases in VGGT-Ω (e.g., over-smoothing of edges, systematic depth underestimation) could affect the scores. This does not invalidate the metric but it should be transparent for users.
6 minor objections
The paper says, at Title
Pith says: minor objection 1 of 6
The word 'Affects' in the title should be 'Effects'. This appears to be a typographical error.
The paper says, at Figure 3 caption
Pith says: minor objection 2 of 6
The caption states 'The hand examples and numerical values, including the reference thresholds, are illustrative rather than empirical results.' This should be moved to the main text or figure caption should explicitly note that these are not actual data. It currently appears only in the figure caption, but some readers may miss it.
show 4 more
The paper says, at §3.4.1 KIVA
Pith says: minor objection 3 of 6
The description of the diagnostic diffusion model is somewhat terse; it would benefit from a reference to the appendix earlier. Currently the main text says 'definitions of the diffusion forward process, workspace transformations, and other parameters are deferred to Appendix B.1'. This is acceptable but the text should be more self-contained for readers who skip appendices.
The paper says, at Table 2
Pith says: minor objection 4 of 6
The table has a note 'FVD follows panel-specific protocols: multi-view rows share a per-camera reference, while single-view rows use per-configuration references and are read vertically.' This is important for interpretation but could be overlooked. Consider making the note more prominent.
The paper says, at Appendix A.2
Pith says: minor objection 5 of 6
The paragraph on 'Filtering' notes that episodes with missing streams are dropped rather than padded, and that 'failed and degenerate trajectories' are retained. This is appropriate for an audit benchmark, but it should be mentioned more prominently in the main text (e.g., in Section 3.1) to set expectations about evaluation difficulty.
The paper says, at Appendix B.1, Eq. (B.1)
Pith says: minor objection 6 of 6
The threshold calibration equation maximizes exact agreement. However, with only 9 videos per morphology, the optimization may overfit. The paper reports bootstrapped means but should also report the variance of the selected thresholds (γ₁, γ₂) across bootstrap draws.
What would settle it
For at least 50 real multi-view episodes from the benchmark that have ground-truth camera calibration (e.g., DROID's provided extrinsics, ALOHA's checkerboard calibration), compute (i) the RIGIS S_geo score using the VGGT-Ω pipeline, and (ii) a direct geometric consistency metric: per-view reprojection error of SfM-derived 3D points or, where available, ground-truth depth from the robot's known mesh. Compute the Spearman correlation between (i) and (ii). If ρ < 0.7, the depth estimator introduces sufficient noise that RIGIS's reported ρ = 0.749 with human preferences may be driven by artifacts rather than genuine geometric consistency.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 27 Oct 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
RoboWorM does something genuinely useful: replace VLM judges with deterministic metrics built from known hardware priors—kinematic trees, fixed cameras, forward kinematics. KIVA and RIGIS are the first metrics of their kind for robotic world models, and the paper implements them carefully across 1,605 episodes, five embodiments, and up to four views. The acceleration audit (nine methods, 29k rollouts) is a first and produces actionable findings: caching works well, token pruning does not. The paper is transparent about its limitations and unusually thorough in its appendices. I trust these numbers more than most in the field.
The central claim—that KIVA and RIGIS align better with human judgment than frontier VLMs—is supported. KIVA's 73.5% overlap versus 58.4% for the best VLM, and RIGIS's Spearman ρ of 0.749 versus 0.326, are real and sizable. The acceleration results (caching best, pruning worst) are credible.
Soft spots. First, the human ground truth: inter-annotator agreement is only 53.6% (embodied physics) and 56.0% (multi-view consistency). The paper adjudicates disagreements, but the target is noisy. This actually makes the relative advantage over VLMs harder to achieve, so it doesn't undermine the main claim. Still, the authors should report alignment on high-consensus cases (where both raters agree) to show the metrics separate clear good from clear bad.
Second, RIGIS relies on VGGT-Ω for depth and pose without validation on robotic scenes. If VGGT-Ω systematically underestimates depth for blurry generated frames, the correlation with human preferences could partly reflect depth-estimator artifacts, not genuine geometry. The paper's main claim is about human alignment, not absolute accuracy, so this doesn't sink it, but the authors should validate or at least caveat this more prominently. The stress-test concern is well-placed.
Third, KIVA threshold calibration on nine videos per embodiment is fragile. The paper is transparent, uses bootstrap uncertainty, and excludes calibration videos from reported agreement, but nine isn't enough for confident generalization across embodiments. Minor, given how honestly it's reported.
The acceleration timing has some missing cells and per-cell tuning that prevents cross-setting comparisons—acknowledged limitations, not flaws.
This paper deserves a serious referee. I'd recommend conditional acceptance with three requests: (1) report metric performance on high-consensus human labels, (2) validate or discuss VGGT-Ω calibration on robotic scenes, and (3) increase per-embodiment calibration samples for KIVA if possible.
The strongest objection
The RIGIS geometric consistency score S_geo (Eq. B.10) depends on per-view depth maps and camera poses predicted by VGGT-Ω (Wang et al., 2026). The paper provides no evidence that VGGT-Ω yields accurate depth for the robot manipulation scenes in the benchmark (cluttered tabletops, reflective surfaces, thin grippers). Scale alignment (Eq. B.3) corrects only median log-ratio drift per frame, not structured errors (e.g., systematic underestimation of depth for distant objects or over-smoothing of edges). If VGGT-Ω systematically underestimates depth for blurry or low-contrast generated frames, its per-view RGB and depth residuals (Eq. B.6–B.7) could be lower for exactly those frames that humans also rate poorly, producing spurious alignment. The paper reports Spearman ρ = 0.749 with human preferences, but without a ground-truth baseline for geometric consistency (e.g., from known camera extrinsics and SfM point clouds), this correlation cannot be attributed to genuine multi-view physics rather than VGGT-Ω's coincidental agreement with human perceptual biases. The authors acknowledge no such validation, and the limitation statement (Section 5) does not mention depth estimator fidelity. This is load-bearing because the central claim that RIGIS outperforms VLMs depends on the metric correctly measuring cross-view consistency.
The reader's weakest_assumption concerned KIVA threshold calibration and human agreement. My concern targets RIGIS's depth estimator validation, a different and earlier step in the evidence chain. If VGGT-Ω is inaccurate, the entire RIGIS metric loses its grounding, and the 130% improvement over VLMs could be artifactual. However, the paper still provides reasonable evidence that KIVA works under acknowledged calibration noise, so the overall verdict (CONDITIONAL) remains appropriate. The condition should be extended to require VGGT-Ω validation on the benchmark scenes. My concern does not warrant rejection because the paper's contribution of systematic acceleration auditing and the benchmark itself are valuable independent of the metric's exact alignment.
The referee's report
If the metrics are validated, RoboWorM provides a principled, deployment-aware alternative to VLM-based evaluation of video world models, filling a gap in assessing kinematic and multi-view consistency. The acceleration audit is the first systematic study of how inference speedups affect physical fidelity in robotic VWMs, which is practically important for deployment. The paper ships detailed appendices, calibration protocols, and an extension framework, supporting reproducibility and adoption.
Pith's reading of the paper (a machine's summary, not the authors')
Video world models (VWMs) are increasingly used as neural simulators for robotic manipulation, but existing evaluations rely on Vision-Language Models (VLMs) that lack strict 3D spatial and kinematic reasoning. The paper argues that for deployment, one already has known hardware priors (robot kinematic trees, fixed camera configurations) that can be exploited for rigorous evaluation. It introduces RoboWorM, a benchmark of 1,605 real-world episodes across multiple robot morphologies, and two deterministic metrics: KIVA for intra-view kinematic stability (measuring rigidity and jerk violations against real-video baselines) and RIGIS for multi-view geometric consistency (measuring depth reprojection consistency and camera pose stability). The central claim is that these metrics align with human judgments of physical realism significantly better than frontier VLM judges: KIVA improves hallucination detection alignment by 25.8% and RIGIS achieves a 130% higher Speerman correlation with human preferences. The benchmark also produces the first large-scale audit of inference acceleration techniques (caching, sparse attention, token pruning) applied to robotic world models, revealing that acceleration methods can preserve superficial visual quality while silently compromising deployment-critical physical constraints.
What carries the argument. KIVA (Kinematic Invariant Violation Assessment): a diagnostic diffusion model trained on real robot videos to predict 3D keypoints; on generated videos, out-of-distribution features cause high prediction variance, which is quantified as rigidity (link length deformation) and jerk (temporal discontinuities), normalized to the 95th percentile of real-video reconstruction errors. RIGIS (Robotic Inter-view Geometry & Intrinsic-Extrinsic Stability): a multi-view evaluation pipeline that uses a modified VGGT-Ω bacbone to estimate depth maps and camera poses, then computes geometric consistency (reprojection RGB+depth error in covariible regions) and camera pose stability (temporal variance of intrinsics and relative extrinsics), combined into a unified score in [0,1].
Core claim. The paper establishes that physical realism in robot video generation can be evaluated deterministically by leveraging known hardware priors—robot kinematic trees and fixed camera extrinsics—rather than relying on VLM judges that lack strict spatial reasoning. It introduces KIVA, which trains a diffusion model on real robot videos to predict 3D keypoints from visual features; on generated videos, out-of-distribution features cause high variance predictions, and the paper quantifies this as rigidity (link length violations) and jerk (temporal discontinuities), normalized to real-video references. And RIGIS, which uses a modified VGGT-Ω bacbone to estimate depth maps and camera poses from multi-view inputs, then measurers geometric consistency (reprojection error) and pose stability (temporal variance of intrinsics and extrinsics). The paper shows that these metrics achieve 25.8% better hallucination detection alignment and 130% higher Speerman correlation with human preferences than frontier VLMs, and that they can systematically reveal how acceleration methods (especially token pruning and sparse attention) silently degrade physical fidelity.
The premise it rests on. The ground truth human annotations against which both KIVA/RIGIS and VLM judges are compared have only 53.6% (embodied physics) and 56.0% (multi-view consistency) inters-annotator exact agreement, and the KIVA violation thresholds are calibrated per embodiment using only 9 human-labeled videos (3 per quality level), so the absolute alignment numbers (25.8% improvement, 73.5% overlap) may be unstable and may not generalize fully to other settings or annotation pools.
If the paper is right (Pith's inferences, not the authors')
Robotic world models can now be audited for physical realism without VLM judges, enabling cheaper, more scalable evaluation for deployment.
Acceleration methods (caching, sparse attention, token pruning) can be rank-ordered by their physical fidelity cost; caching emerges as the recommended choice for deployment.
The benchmark provides a standardized grid (1,605 episodes, 3 morphologies, single-/multi-view, Markovian/non-Markovian) that can be extended with new world models, embodiments, or acceleration techniques, each with a clear protocol.
The results imply that visual-quality metrics (FVD, CLIP score) are insufficient for evaluating robotic world models, as they miss kinematic and multi-view failures.
Long-horizon degradation (error accumulation over rollout depth) is quantitatively measured, informing design of context windows or restart strategies.
If KIVA and RIGIS are adopted by the community, the cost of evaluating robotic world models could drop dramatically (from VLM API costs to a single GPU run per rollout), enabling more iterative model development.
The paper's result that token pruning degrades physical fidelity most severely (13.4% NR loss) suggests that pruning methods that preserve structural rather than perceptual features may be needed for robotics.
The finding that caching (WorldCache, DiCache) preserves physical fidelity best while offering 1.7x mean speedup suggests a concrete design win: reuse of intermediate features from prior steps preserves the spatial structure that token pruning disrupts.
The low inters-annotator agreement (53.6%, 56.0%) on the human evaluation set that serves as ground truth for both KIVA/RIGIS and VLM comparisons suggests that the absolute alignment percentages (25.8%, 73.5%) are noisy and may not replicate exactly on other annotations sets; the relative ranking (our metrics > VLMs) is more robust than the absolute numbers.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Video world models (VWMs) are rapidly emerging as scalable neural simulators for robotic manipulation. However, most benchmarks only evaluate them as open-world video generators, using Vision-Language Models (VLMs) that lack strict 3D spatial and kinematic reasoning. To bridge this gap, we introduce RoboWorM, a comprehensive, deployment-aware benchmark that leverages known hardware priors such as robot kinematic trees and fixed camera configurations, to systematically examine structural and spatiotemporal fidelity. RoboWorM curates 1,605 episodes across single-arm, bimanual, and humanoid morphologies, spanning both Markovian and long-horizon non-Markovian tasks under single- and multi-view setups. To explicitly measure physical consistency, we propose two deterministic, VLM-free metrics: KIVA for intra-view kinematic stability, and RIGIS for multi-view geometric consistency. By exploiting intrinsic robotic invariants, KIVA improves hallucination detection alignment with human annotations by 25.8% relative to the strongest frontier VLMs, while RIGIS achieves a 130% higher Spearman correlation with human preferences. Furthermore, since real-world deployment requires low inference latency, we utilize RoboWorM to conduct the first large-scale audit of the quality-efficiency trade-off. Evaluating 9 state-of-the-art acceleration techniques across 3 leading world models (over 29,000 generated rollouts), we demonstrate how these acceleration methods can retain superficial visual quality while silently compromising deployment-critical physical constraints. Using our metrics, we achieve 2.3× maximum inference speed-up with minimal quality loss.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-27
The authors' main claim, as the classifier picked it out 0.95
“Evaluating 9 state-of-the-art acceleration techniques across 3 leading world models (over 29,000 generated rollouts), we demonstrate how these acceleration methods can retain superficial visual quality while silently compromising deployment-critical physical constraints.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.82 · states limitation 0.17 · human participants 0.19 · controlled comparison 0.58 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.93 · mentions code or data availability 0.05
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Video world models (VWMs) are rapidly emerging as scalable neural simulators for robotic manipulation. | background 1.00 | 0.04 | 0.02 | asserted 1.00 |
| However, most benchmarks only evaluate them as open-world video generators, using Vision-Language Models (VLMs) that lack strict 3D spatial and kinematic reasoning. | gap 0.97 | 0.07 | 0.03 | asserted 0.99 |
| To bridge this gap, we introduce RoboWorM, a comprehensive, deployment-aware benchmark that leverages known hardware priors such as robot kinematic trees and fixed camera configurations, to systematically examine structural and spatiotemporal fidelity. | contribution 0.99 | 0.90 | 0.05 | asserted 0.99 |
| RoboWorM curates 1,605 episodes across single-arm, bimanual, and humanoid morphologies, spanning both Markovian and long-horizon non-Markovian tasks under single- and multi-view setups. | method 0.78 | 0.92 | 0.03 | asserted 1.00 |
| To explicitly measure physical consistency, we propose two deterministic, VLM-free metrics: KIVA for intra-view kinematic stability, and RIGIS for multi-view geometric consistency. | method 0.77 | 0.94 | 0.05 | asserted 1.00 |
| By exploiting intrinsic robotic invariants, KIVA improves hallucination detection alignment with human annotations by 25.8% relative to the strongest frontier VLMs, while RIGIS achieves a 130% higher Spearman correlation with human preferences. | result 1.00 | 0.95 | 0.21 | asserted 1.00 |
| Furthermore, since real-world deployment requires low inference latency, we utilize RoboWorM to conduct the first large-scale audit of the quality-efficiency trade-off. | contribution 0.95 | 0.90 | 0.06 | asserted 0.96 |
| Evaluating 9 state-of-the-art acceleration techniques across 3 leading world models (over 29,000 generated rollouts), we demonstrate how these acceleration methods can retain superficial visual quality while silently compromising deployment-critical physical constraints. | result 0.47 | 0.95 | 0.32 | asserted 1.00 |
| Using our metrics, we achieve 2.3$\times$ maximum inference speed-up with minimal quality loss. | result 1.00 | 0.95 | 0.24 | asserted 1.00 |
What kind of paper the classifier reports this is: dataset or benchmark 1.00 · evidence from observation or experiment 0.71
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-27: 1 request, 7,184 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (6)
Figures are the authors', reproduced from the paper.
References (114)
- [1] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your ViT but faster. In International Conference on Learning Representations, 2023 on Pith
- [2] Jiazi Bu, Pengyang Ling, Yujie Zhou, Yibin Wang, Yuhang Zang, Dahua Lin, and Jiaqi Wang. Dicache: Let diffusion model determine its own cache. In International Conference on Learning Representations, volume 2026, pp.\ 73778--73803, 2026 on Pith
- [3] Nhat Chung, Taisei Hanyu, Toan Nguyen, Huy Le, Frederick Bumgarner, Duy Minh Ho Nguyen, Khoa Vo, Kashu Yamazaki, Chase Rainwater, Tung Kieu, et al. Rethinking progression of memory state in robotic manipulation: An object-centric perspective. In Proceedings of the AAAI Confere... on Pith
- [4] Yufan Deng, Zilin Pan, Hongyu Zhang, Xiaojie Li, Ruoqing Hu, Yufei Ding, Yiming Zou, Yan Zeng, and Daquan Zhou. Rethinking video generation model for the embodied world. In Forty-third International Conference on Machine Learning, 2026 on Pith
- [5] Haoyi Duan, Hong-Xing Yu, Sirui Chen, Li Fei-Fei, and Jiajun Wu. Worldscore: A unified evaluation benchmark for world generation. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 27713--27724, 2025 on Pith
- [6] Shenyuan Gao, William Liang, Kaiyuan Zheng, Ayaan Naveed Malik, Seonghyeon Ye, Sihyun Yu, Wei-Cheng Tseng, Yuzhu Dong, Kaichun Mo, Chen-Hsuan Lin, Jiannan Xiang, Yuqi Xie, Ruijie Zheng, Dantong Niu, Pooya Jannaty, Jinwei Gu, Jun Zhang, Jitendra Malik, Pieter Abbeel, Ming-Yu Li... on Pith
- [7] Yanjiang Guo, Lucy Shi, Jianyu Chen, and Chelsea Finn. Ctrl-world: A controllable generative world model for robot manipulation. In International Conference on Learning Representations, volume 2026, pp.\ 6121--6138, 2026 on Pith
- [9] Taisei Hanyu, Nhat Chung, Huy Le, Toan Nguyen, Yuki Ikebe, Anthony Gunderman, Duy Nguyen Ho Minh, Khoa Vo, Tung Kieu, Kashu Yamazaki, et al. Slotvla: Towards modeling of object-relation representations in robotic manipulation. In 2026 IEEE International Conference on Robotics ... on Pith
- [10] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 conference on empirical methods in natural language processing, pp.\ 7514--7528, 2021 on Pith
- [11] Wenlong Huang, Yu-Wei Chao, Arsalan Mousavian, Ming-Yu Liu, Dieter Fox, Kaichun Mo, and Li Fei-Fei. Pointworld: Scaling 3d world models for in-the-wild robotic manipulation. The IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR), 2026 on Pith
- [12] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recogniti... on Pith
- [13] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pp.\ 492--518. Springer, 1992 on Pith
- [14] Joel Jang, Seonghyeon Ye, Zongyu Lin, Jiannan Xiang, Johan Bjorck, Yu Fang, Fengyuan Hu, Spencer Huang, Kaushil Kundalia, Yen-Chen Lin, Lo\" i c Magne, Ajay Mandlekar, Avnish Narayan, You Liang Tan, Guanzhi Wang, Jing Wang, Qi Wang, Yinzhen Xu, Xiaohui Zeng, Kaiyuan Zheng, Rui... on Pith
- [16] Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, Peter David Fagan, Joey Hejna, Masha Itkina, Marion Lepert, Yecheng Jason Ma, Patrick Tree Mill... on PithPith's reviewarXiv
- [17] Dacheng Li, Yunhao Fang, Yukang Chen, Shuo Yang, Shiyi Cao, Justin Wong, Michael Luo, Xiaolong Wang, Hongxu Yin, Joseph Gonzalez, et al. Worldmodelbench: Judging video generation models as world models. Advances in Neural Information Processing Systems, 38, 2026 a on Pith
- [19] Huiqiong Li, Jiayu Wang, Zhiting Mei, Anirudha Majumdar, Jingjing Chen, and Bin Zhu. Robotrustbench: Benchmarking the trustworthiness of video world models for robotic manipulation. In Findings of the Association for Computational Linguistics: EMNLP 2026, 2026 c . URL https://... on PithPith's reviewarXiv
- [20] Xingyang Li, Muyang Li, Tianle Cai, Haocheng Xi, Shuo Yang, Yujun Lin, Lvmin Zhang, Songlin Yang, Jinbo Hu, Kelly Peng, et al. Radial attention: O (n n) sparse attention with energy decay for long video generation. Advances in Neural Information Processing Systems, 38: 0 16822... on Pith
- [21] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR... on Pith
- [23] Ziyan Liu, Yeqiu Chen, Hongyi Cai, Tao Lin, Shuo Yang, Zheng Liu, and Bo Zhao. Vla-pruner: Temporal-aware dual-level visual token pruning for efficient vision-language-action inference. arXiv e-prints, pp.\ arXiv--2511, 2025 on Pith
- [24] Tung M Luu, Tri Ton, Duc M Nguyen, Xuan Thanh Nguyen, Nguyen Gia Binh, Quang Tan Nguyen, Vu Duong, Daniel Sonntag, Duy MH Nguyen, Vien Anh Ngo, et al. Progress-guided preference generation for aligning generative control policies. Conference on Robot Learning (CoRL), 2026 on Pith
show 80 more
- [25] Zhengyao Lyu, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K Wong. Fastercache: Training-free video diffusion model acceleration with high quality. In Y. Yue, A. Garg, N. Peng, F. Sha, and R. Yu (eds.), International Conference on Learning Representa... on Pith
- [26] Tomohiro Motoda, Masaki Murooka, Ryoichi Nakajo, Muhammad A. Muttaqien, Koshi Makihara, Hanbit Oh, Keisuke Shirai, Floris Erich, Ryo Hanai, and Yukiyasu Domae. Aist-bimanual manipulation, 2025 on Pith
- [27] Umair Nawaz, Ahmed Heakl, Ufaq Khan, Abdelrahman Shaker, Salman Khan, and Fahad Shahbaz Khan. Worldcache: Content-aware caching for accelerated video world models. In European Conference on Computer Vision (ECCV), 2026. URL https://arxiv.org/abs/2603.22286 on PitharXiv
- [28] Duc Minh Nguyen, Nghiem Tuong Diep, Binh Gia Nguyen, Trong-Bao Ho, Doanh Le, Tan Q Nguyen, Thien-Loc Ha, Nhiem Tran, Bao Thach, Nhat X Tran, et al. Foca: Future-oriented conditioning for data-efficient vision-language-action adaptation. International Conference on Machine Lear... on Pith
- [33] Abby O’Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE Internat... on Pith
- [34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp... on Pith
- [37] Kele Shao, Keda Tao, Can Qin, Haoxuan You, Yang Sui, and Huan Wang. Holitom: Holistic token merging for fast video large language models. Advances in Neural Information Processing Systems, 38: 0 135547--135570, 2026 on Pith
- [38] Oriane Sim \'e oni, Huy V. Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seung Eun Yi, Michael Ramamonjisoa, Francisco Massa, Daniel HAZIZA, Luca Wehrstedt, Jianyuan Wang, Timoth \'e e Darcet, Th \'e o Moutakanni, Leone... on Pith
- [39] Hoai-Chau Tran, Duy Nguyen, Duy M Nguyen, TrungTin Nguyen, Ngan Le, Pengtao Xie, Daniel Sonntag, James Zou, Binh T Nguyen, and Mathias Niepert. Accelerating transformers with spectrum-preserving token merging. Advances in Neural Information Processing Systems, 37: 0 30772--30810, 2024 on Pith
- [40] Tuan Tran Anh, Duy MH Nguyen, Hoai-Chau Tran, Michael Barz, Khoa D Doan, Roger Wattenhofer, Vien Ngo, Mathias Niepert, Daniel Sonntag, and Paul Swoboda. How many tokens do 3d point cloud transformer architectures really need? Advances in Neural Information Processing Systems, ... on Pith
- [42] Khoa Vo, Taisei Hanyu, Yuki Ikebe, Trong Thang Pham, Nhat Chung, Minh Nhat Vu, Duy Nguyen Ho Minh, Anh Nguyen, Anthony Gunderman, Chase Rainwater, et al. Clutter-resistant vision-language-action models through object-centric and geometry grounding. arXiv e-prints, pp.\ arXiv--... on Pith
- [44] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pp.\ 2555--2563, 2023 on Pith
- [45] Jianyuan Wang, Minghao Chen, Shangzhan Zhang, Nikita Karaev, Johannes Sch\"onberger, Patrick Labatut, Piotr Bojanowski, David Novotny, Andrea Vedaldi, and Christian Rupprecht. Vggt- . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), ... on Pith
- [46] Haoyu Wu, Jingyi Xu, Hieu Le, and Dimitris Samaras. Importance-based token merging for efficient image and video generation. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 4983--4995. IEEE, 2025 a on Pith
- [47] Haoyu Wu, Diankun Wu, Tianyu He, Junliang Guo, Yang Ye, Yueqi Duan, and Jiang Bian. Geometry forcing: Marrying video diffusion and 3d representation for consistent world modeling. In International Conference on Learning Representations, volume 2026, pp.\ 119186--119209, 2026 a on Pith
- [50] Haocheng Xi, Shuo Yang, Yilong Zhao, Chenfeng Xu, Muyang Li, Xiuyu Li, Yujun Lin, Han Cai, Jintao Zhang, Dacheng Li, Jianfei Chen, Ion Stoica, Kurt Keutzer, and Song Han. Sparse video-gen: Accelerating video diffusion transformers with spatial-temporal sparsity. In Forty-secon... on Pith
- [51] Adina Yakefu, Bin Xie, Chongyang Xu, Enwen Zhang, Erjin Zhou, Fan Jia, Haitao Yang, Haoqiang Fan, Haowei Zhang, Hongyang Peng, Jing Tan, Junwen Huang, Kai Liu, Kaixin Liu, Kefan Gu, Qinglun Zhang, Ruitao Zhang, Saike Huang, Shen Cheng, Shuaicheng Liu, Tiancai Wang, Tiezhen Wan... on PitharXiv
- [52] Shuo Yang, Haocheng Xi, Yilong Zhao, Muyang Li, Jintao Zhang, Han Cai, Yujun Lin, Xiuyu Li, Chenfeng Xu, Kelly Peng, et al. Sparse videogen2: Accelerate video generation with sparse attention via semantic-aware permutation. Advances in Neural Information Processing Systems, 38... on Pith
- [54] Evelyn Zhang, Jiayi Tang, Xuefei Ning, and Linfeng Zhang. Training-free and hardware-friendly acceleration for diffusion models via similarity-based token pruning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 9878--9886, 2025 on Pith
- [57] Dian Zheng, Ziqi Huang, Hongbo Liu, Kai Zou, Yinan He, Fan Zhang, Yuanhan Zhang, Jingwen He, Wei-Shi Zheng, Yu Qiao, and Ziwei Liu. VBench-2.0 : Advancing video generation benchmark suite for intrinsic faithfulness. arXiv preprint arXiv:2503.21755, 2025 on PithPith's reviewarXiv
- [59] Bolya, Daniel and Fu, Cheng-Yang and Dai, Xiaoliang and Zhang, Peizhao and Feichtenhofer, Christoph and Hoffman, Judy , booktitle=. Token Merging: Your on Pith
- [60] Dicache: Let diffusion model determine its own cache , author=. International Conference on Learning Representations , volume= on Pith
- [61] DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos , author=. Forty-third International Conference on Machine Learning , year= on Pith
- [62] Ctrl-world: A controllable generative world model for robot manipulation , author=. International Conference on Learning Representations , volume= on Pith
- [63] Clipscore: A reference-free evaluation metric for image captioning , author=. Proceedings of the 2021 conference on empirical methods in natural language processing , pages= on Pith
- [64] Vbench: Comprehensive benchmark suite for video generative models , author=. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2024 , organization= on Pith
- [65] Robust estimation of a location parameter , author=. Breakthroughs in statistics: Methodology and distribution , pages=. 1992 , publisher= on Pith
- [66] DreamGen: Unlocking Generalization in Robot Learning through Video World Models , author =. Proceedings of The 9th Conference on Robot Learning , pages =. 2025 , editor = on Pith
- [67] Robowm-bench: A benchmark for evaluating world models in robotic manipulation , author=. arXiv preprint arXiv:2604.19092 , year= on PithPith's reviewarXiv
- [68] DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset , author=. 2025 , eprint= on Pith
- [69] Worldmodelbench: Judging video generation models as world models , author=. Advances in Neural Information Processing Systems , volume= on Pith
- [70] Pisa: Piecewise sparse attention is wiser for efficient diffusion transformers , author=. arXiv preprint arXiv:2602.01077 , year= on PithPith's reviewarXiv
- [71] RoboTrustBench: Benchmarking the Trustworthiness of Video World Models for Robotic Manipulation , author=. Findings of the Association for Computational Linguistics: EMNLP 2026 , year= on Pith
- [72] Li, Xingyang and Li, Muyang and Cai, Tianle and Xi, Haocheng and Yang, Shuo and Lin, Yujun and Zhang, Lvmin and Yang, Songlin and Hu, Jinbo and Peng, Kelly and others , journal=. Radial Attention: on Pith
- [73] Evalcrafter: Benchmarking and evaluating large video generation models , author=. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2024 , organization= on Pith
- [74] Lyu, Zhengyao and Si, Chenyang and Song, Junhao and Yang, Zhenyu and Qiao, Yu and Liu, Ziwei and Wong, Kwan-Yee K , booktitle =. FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality , url = on Pith
- [75] Tomohiro Motoda and Masaki Murooka and Ryoichi Nakajo and Muhammad A. Muttaqien and Koshi Makihara and Hanbit Oh and Keisuke Shirai and Floris Erich and Ryo Hanai and Yukiyasu Domae , title = on Pith
- [76] Fast approximate nearest neighbors with automatic algorithm configuration , author=. International conference on computer vision theory and applications , volume=. 2009 , organization= on Pith
- [77] WorldCache: Content-Aware Caching for Accelerated Video World Models , author =. European Conference on Computer Vision (ECCV) , eprint = on Pith
- [78] Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0 , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization= on Pith
- [79] Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization= on Pith
- [80] H2R-Bench: Benchmarking human-to-robot manipulation video generation in world models , author=. arXiv preprint arXiv:2608.13049 , year= on PithPith's reviewarXiv
- [81] Holitom: Holistic token merging for fast video large language models , author=. Advances in Neural Information Processing Systems , volume= on Pith
- [82] Oriane Sim. Transactions on Machine Learning Research , issn=. 2026 , url= on Pith
- [83] Towards accurate generative models of video: A new metric & challenges , author=. arXiv preprint arXiv:1812.01717 , year= on PithPith's reviewarXiv
- [84] Exploring clip for assessing the look and feel of images , author=. Proceedings of the AAAI conference on artificial intelligence , volume= on Pith
- [85] Wang, Jianyuan and Chen, Minghao and Zhang, Shangzhan and Karaev, Nikita and Sch\"onberger, Johannes and Labatut, Patrick and Bojanowski, Piotr and Novotny, David and Vedaldi, Andrea and Rupprecht, Christian , title =. Proceedings of the IEEE/CVF Conference on Computer Vision ... on Pith
- [86] Importance-based token merging for efficient image and video generation , author=. 2025 IEEE/CVF International Conference on Computer Vision (ICCV) , pages=. 2025 , organization= on Pith
- [87] Geometry forcing: Marrying video diffusion and 3d representation for consistent world modeling , author=. International Conference on Learning Representations , volume= on Pith
- [88] Multiworld: Scalable multi-agent multi-view video world models , author=. arXiv preprint arXiv:2604.18564 , year= on PithPith's reviewarXiv
- [89] Robocoin: An open-sourced bimanual robotic data collection for integrated manipulation , author=. arXiv preprint arXiv:2511.17441 , year= on PithPith's reviewarXiv
- [90] Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity , author=. Forty-second International Conference on Machine Learning , year= on Pith
- [91] RoboChallenge: Large-scale Real-robot Evaluation of Embodied Policies , author=. 2025 , eprint= on Pith
- [92] Sparse videogen2: Accelerate video generation with sparse attention via semantic-aware permutation , author=. Advances in Neural Information Processing Systems , volume= on Pith
- [93] MiraBench: Evaluating Action-Conditioned Reliability in Robotic World Models , author=. arXiv preprint arXiv:2605.29360 , year= on PithPith's reviewarXiv
- [94] Training-free and hardware-friendly acceleration for diffusion models via similarity-based token pruning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume= on Pith
- [95] MultiView-Bench: A Diagnostic Benchmark for World-Centric Multi-View Integration in VLMs , author=. arXiv preprint arXiv:2607.08970 , year= on PithPith's reviewarXiv
- [96] Zheng, Dian and Huang, Ziqi and Liu, Hongbo and Zou, Kai and He, Yinan and Zhang, Fan and Zhang, Yuanhan and He, Jingwen and Zheng, Wei-Shi and Qiao, Yu and Liu, Ziwei , journal= on Pith
- [97] Worldscore: A unified evaluation benchmark for world generation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages= on Pith
- [98] RoboPhys-3D: A Comprehensive Embodied World Model Evaluation via 3D Reconstruction , author=. 2026 , eprint= on Pith
- [99] Rethinking video generation model for the embodied world , author=. Forty-third International Conference on Machine Learning , year= on Pith
- [100] Sea-raft: Simple, efficient, accurate raft for optical flow , author=. European Conference on Computer Vision , pages=. 2024 , organization= on Pith
- [101] Zhang, Guozhen and Liu, Chunxu and Cui, Yutao and Zhao, Xiaotong and Ma, Kai and Wang, Limin , title =. Proceedings of the 38th International Conference on Neural Information Processing Systems , articleno =. 2024 , isbn = on Pith
- [102] Grounding dino: Marrying dino with grounded pre-training for open-set object detection , author=. European conference on computer vision , pages=. 2024 , organization= on Pith
- [103] Zhou Wang and Bovik, A.C. and Sheikh, H.R. and Simoncelli, E.P. , journal=. Image quality assessment: from error visibility to structural similarity , year= on Pith
- [104] The unreasonable effectiveness of deep features as a perceptual metric , author=. 2018 IEEE/CVF conference on computer vision and pattern recognition , pages=. 2018 , organization= on Pith
- [105] Ewmbench: Evaluating scene, motion, and semantic quality in embodied world models , author=. arXiv preprint arXiv:2505.09694 , year= on PithPith's reviewarXiv
- [106] Pai-bench: A comprehensive benchmark for physical ai , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages= on Pith
- [107] Rethinking progression of memory state in robotic manipulation: An object-centric perspective , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume= on Pith
- [108] How Many Tokens Do 3D Point Cloud Transformer Architectures Really Need? , author=. Advances in Neural Information Processing Systems , volume= on Pith
- [109] HSC-VLA: Hierarchical Scene-Clearing for Robust Bimanual Manipulation in Dense Clutter , author=. arXiv preprint arXiv:2603.07484 , year= on PitharXiv
- [110] Faster-WAM: Efficient inference-time future conditioning for robust world action models , author=. arXiv preprint arXiv:2608.04404 , year= on PithPith's reviewarXiv
- [111] Simfoundry: Modular and automated scene generation for policy learning and evaluation , author=. arXiv preprint arXiv:2606.28276 , year= on PithPith's reviewarXiv
- [112] Clutter-resistant vision-language-action models through object-centric and geometry grounding , author=. arXiv e-prints , pages= on Pith
- [113] GloVLA: Let Geometry Move and Local VLA Interact for Robust Object-Centric Manipulation in Unstructured Environments , author=. arXiv preprint arXiv:2609.06256 , year= on PithPith's reviewarXiv
- [114] vla. cpp: A Unified Inference Runtime for Vision-Language-Action Models , author=. arXiv preprint arXiv:2606.08094 , year= on PithPith's reviewarXiv
- [115] Accelerating transformers with spectrum-preserving token merging , author=. Advances in Neural Information Processing Systems , volume= on Pith
- [116] Vla-pruner: Temporal-aware dual-level visual token pruning for efficient vision-language-action inference , author=. arXiv e-prints , pages= on Pith
- [117] Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think , author=. arXiv preprint arXiv:2606.20246 , year= on PithPith's reviewarXiv
- [118] Pointworld: Scaling 3d world models for in-the-wild robotic manipulation , author=. The IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR) , year= on Pith
The first 100 of 114 references are listed here. All of them are in this paper's citation record.
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment