Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Failure Detection for Surgical Robot Imitation Policies via Flow-Matching World Modeling

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Trained only on successful executions, a flow-matching world model detects surgical-robot failures by scoring whether observed visual changes match commanded actions, reaching a 96.6% detection rate at a 1.3% false-alarm rate on wrist-camer

desk verdict FoMo-FD is a well-ablated extension of world-model failure detection with a genuinely new scoring idea; the headline wrist-view result is strong but rests on evaluation-set selection and one false alarm, and the method is inherently blind to visually silent failures. read the letter →

arxiv 2607.27511 v1 pith:P4TFIKU5 submitted 2026-07-29 cs.RO

Zhefeng Huang , Yilin Cai , Ankit Patel , Mohammad Hajiha , Brendan Browne , Yue Chen This is my paper
classification cs.RO
keywords surgicalroboticsfailuredetectionflowmatchingworldmodelimitationlearningconformalpredictionanomalylatentdynamics
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 is trying to establish that low-level execution failures of a surgical imitation-learning policy can be detected without any failure demonstrations, using only a model of how the visual scene normally evolves under commanded actions. The proposed detector, FoMo-FD, learns a flow-matching world model that predicts the endpoint of an eight-step visual-action window from the preceding visual history and action chunk, then scores the observed endpoint by how far it lands from the model's Gaussian base after inverse transport. Across four surgical manipulation tasks—two in simulation, two on a physical surgical robot—with twenty failure modes, the wrist-camera version detects 96.6% of failed rollouts while alarming on only 1.3% of successful ones, outperforming observation-level anomaly scores and a prediction-error variant of the same model. If correct, window-level visual-action consistency is a sufficient failure signal for these settings, and failure monitoring can be built from the abundant successful-case data rather than from scarce failure examples.

What carries the argument

Inverse-transport nonconformity scoring: the observed endpoint latent is pushed back through the learned flow-matching ODE to the Gaussian base, and the squared normalized norm of the transported point serves as the nonconformity score. The flow-matching world model itself—an action-conditioned transformer over compressed latent visual tensors from a frozen vision foundation model—predicts the endpoint latent of an L=4-history, K=4-horizon window, so each score summarizes eight steps of visual-action consistency. Thresholds come from conformal calibration on successful policy rollouts, avoiding any need for failure data.

What would settle it

Run a failure mode that changes no pixel in any monitored view—for example, a gripper force drop with identical camera frames, or a disturbance outside the camera's field of view. If such failures go undetected while scores stay below the calibrated threshold, the claim that visual-action consistency alone is sufficient for these settings is refuted.

Watch

Extended reading notes

Core claim

The central claim is that a failure in surgical manipulation manifests as an inconsistency between the commanded action and the observed visual transition over a short window, and that this inconsistency is measurable by inverse-transport nonconformity in a learned latent dynamics model. Concretely, the paper trains a flow-matching vector field to transport Gaussian noise to the endpoint latent of a four-step future window, conditioned on four steps of visual history and the intervening action chunk, using only successful demonstrations. At test time it integrates the learned ODE backward from the observed endpoint and scores the squared normalized norm of the resulting base-space point; hig

Load-bearing premise

Every failure must show up as a visual-action inconsistency in the monitored camera within the eight-step window; if a failure leaves the pixels indistinguishable from normal variation, the score cannot rise.

Editorial extensions

If this is right

  • Failure detectors can be trained from successful demonstrations alone, plus a small calibration set of successful policy rollouts, removing the need for collecting failure examples.
  • Window-level visual-action consistency is more discriminative than per-observation anomaly scores, with the wrist-view score reaching 96.6% detection compared with roughly 43–45% for observation-level baselines.
  • The same world model scored by endpoint prediction error is markedly weaker (52.2%), so the inverse-transport scoring rule, not just the learned dynamics, carries the detection signal.
  • Camera geometry bounds performance: the wrist view (96.6%) strongly outperforms the fixed view (45.9%), while the fixed view still improves over its best baseline by 21.4% relative.
  • Longer endpoint horizons accumulate deviations: increasing the prediction horizon from one to four steps raises wrist-view detection from 42.2% to 96.6% at low false-alarm rates.

Reading between the lines

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

  • The method's success hinges on pixel visibility; failures with no visual footprint in any monitored camera (e.g., gripper force decay or an out-of-view disturbance) would likely be invisible to this score, suggesting multi-modal sensing as a natural extension.
  • Because only a latent encoder and action-conditioned dynamics are needed, the same inverse-transport score could transfer to other generative manipulation policies and non-surgical settings, not just the transformer-based policy evaluated here.
  • The endpoint-horizon ablation implies a latency-sensitivity trade-off for online use: longer horizons give higher detection but delay alarms; the paper's timing experiment reports 13.98 Hz scoring, but online intervention is left untested.
  • If the deployment policy's distribution shifts after calibration, conformal thresholds calibrated on successful rollouts may lose their coverage guarantee; a drift-adaptive or progress-conditioned threshold would be needed.
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

3 major / 5 minor

Summary. The paper proposes FoMo-FD, a runtime failure-detection method for surgical robot imitation policies. A flow-matching world model is trained only on successful expert demonstrations to predict the endpoint latent of a short visual-action window from the preceding latent history and commanded action chunk. At test time, the observed endpoint latent is transported backward under the learned flow and scored by its squared base-space norm; thresholds are set via conformal calibration on successful policy rollouts. The method is evaluated on four tasks (two simulated, two physical dVRK) with twenty failure modes, comparing against two observation-level baselines (logpZO, RND) and a prediction-error variant of the same world model. The wrist-view WM-NC detector reports 96.6% overall FDR at 1.3% FAR, outperforming the baselines and the fixed-view detector.

Significance. If the reported results are reliable, FoMo-FD is a significant step toward failure detection without failure demonstrations: it learns nominal visual-action dynamics from successes alone, uses a principled conformal calibration, and demonstrates a large effect over strong baselines on the wrist view. The paper includes several good practices: disjoint calibration and evaluation rollouts, per-view reporting, adaptation of baselines to the same VAE features, and qualitative score-trace analysis. The central idea—window-level visual-action consistency as a sufficient failure signal—is falsifiable and practically valuable. However, the headline performance is weakened by methodological issues in configuration selection and in the fixed-view comparison, as detailed below.

major comments (3)
  1. [§V-A, §V-B (Figs. 5 and 6)] The hyperparameters K=4 and the Tok+Mod action-conditioning mode are selected after observing their wrist-view FDR on the exact evaluation episodes used to report the headline 96.6% result. §V states that 'the remaining architecture, calibration protocol, and evaluation episodes' are fixed during the ablation, meaning the same episodes serve for model selection and for the final reported number. This makes the headline a selected maximum over configurations, not an unbiased estimate; the comparison with baselines (which use only a fixed, pre-specified configuration) is therefore conservative in the baselines' favor but inflated for WM-NC. A disjoint validation set (or nested calibration) is needed before the reported FDR can be taken at face value.
  2. [§IV-B, Table I] The claimed fixed-view improvement of WM-NC over the strongest fixed-view baseline is confounded by FAR. Table I reports fixed-view WM-NC at 45.9% FDR / 10.0% FAR, while the strongest fixed-view baseline (RND) achieves 37.8% FDR / 0.0% FAR. Comparing FDR at a ten-fold higher FAR is not a like-for-like comparison; a detector that fires on one in ten successful rollouts will trivially detect more failures. The paper's statement that WM-NC improves over the fixed-view baseline by 21.4% relative is misleading. The authors should report ROC curves, or at minimum compare FDR at matched FAR values (e.g., using threshold interpolation or a fixed-FAR evaluation).
  3. [§VI (limitations) and abstract] The paper does not disclose that the method is, by construction, blind to failures that do not manifest as a visible change in the monitored camera within the H=L+K=8-step window. All evidence flows through a single view's DINOv2/β-VAE latent (Eq. 2); there is no proprioceptive, force, or multi-view fused channel. The evaluation's 20 failure modes are all reported to be visible in the wrist camera (as the qualitative frames in Fig. 4 suggest), so the sufficiency claim is not tested against visually silent failures (e.g., grip-force decay with no visible change, or a disturbance outside the view). This is a scope restriction, not an internal inconsistency, but it should be stated prominently in the limitations and the abstract's general 'failure detection' should be qualified as 'visually observable failures'.
minor comments (5)
  1. [§IV-A, Table I] The 'No Failure' row for Shunt Insertion shows 6/20 false alarms for fixed-view WM-NC, i.e., 30% FAR on that task. This is not discussed; the overall FAR aggregates over tasks and hides this task-specific weakness. Please comment on why the fixed view is so unreliable on shunt insertion.
  2. [§III-B, Eq. (10)] With N=19 and α=0.05, the conformal threshold is the maximum calibration score (the 19th order statistic). This is a valid but extremely conservative choice; it may explain the low FAR but also makes the detector sensitive to a single outlier calibration rollout. Consider reporting the actual quantile used and discussing robustness of the threshold choice.
  3. [Abstract] Typo: 'a96.6%failure detection rate' missing spaces. Also in §III-A, 'V AE' should be 'VAE'.
  4. [§V-B, Fig. 6] The fixed-view FDR curve is non-monotonic (42.2% → ~50% → ~60% → ~46%?); the text says 'weaker and less monotonic' but does not explain the decrease at K=4. A sentence of interpretation would help.
  5. [§IV-A] The statement that the detector 'processes scoring windows at 13.98 Hz' is based on a desktop GPU (RTX 5090). This is not representative of clinical hardware; please label this as a peak desktop measurement and note the dependence on the encoder and ODE solver tolerance.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: training, calibration, and evaluation are disjoint, and the inverse-transport score is not defined in terms of failure labels.

full rationale

FoMo-FD trains a flow-matching world model exclusively on expert successful demonstrations (Eqs. 2, 5, 7), obtains thresholds from successful policy rollouts via conformal calibration (Eq. 10), and evaluates on held-out failure rollouts (§IV-A: 'policy rollouts are used for threshold calibration and evaluation, not for training the world model'). The nonconformity score (Eqs. 8–9) is the squared normalized base-space norm of the inverse-transported observed endpoint latent; it is a predictive-consistency measure, not a quantity fitted to or defined by failure occurrences. The endpoint-prediction formulation (Z_t, A_t → z_{t+K}) is a genuine forecasting task, and the detection target (failure vs. success) is never used in the world-model objective or threshold selection. The inverse-flow scoring criterion is explicitly credited to an external prior work [13] with no author overlap with the present paper, so it is independent support rather than a self-citation chain. The paper states its own limitations (§VI) regarding calibration rollouts and time-varying thresholds; the additional boundary that only failures visually apparent in the monitored camera within the H=L+K window can be detected is a sensing-scope limitation, not a circular derivation. No load-bearing step reduces to a self-definition, a fitted input renamed as a prediction, or an ansatz smuggled in via self-citation. Hence the derivation is self-contained and no circularity is present.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

No failure demonstrations enter training or calibration; the method's fitted quantities are hyperparameters and design choices, not failure-fitted parameters. The load-bearing external inputs are: standard flow-matching mathematics (Eqs. 4-5), the empirical validity of inverse-transport as a nonconformity score (borrowed from [13], not proven), conformal exchangeability (partially violated in the paper's own fixed-view data), and the premise that failures are visible in the monitored camera's latent stream.

free parameters (9)
  • Endpoint horizon K = 4 (wrist FDR 42.2% at K=1 to 96.6% at K=4)
    Selected by ablation (Fig. 6) run on the same evaluation episodes used to report the headline FDR.
  • Action-conditioning mode = cross-attention + global modulation (Tok+Mod)
    Selected by ablation (Fig. 5) on evaluation episodes; wrist FDR 87.2% (no action) to 96.6%.
  • Camera view for headline results = wrist (fixed view: 45.9% FDR, 10% FAR)
    Best of two evaluated views; both views reported but abstract highlights wrist numbers.
  • VAE weights beta, gamma = beta=0.01, gamma=0.5
    Hand-set in §IV-A; not ablated.
  • World-model loss weights lambda_FM, lambda_aux = 1.0, 0.5
    Hand-set (§III-A); not ablated.
  • Latent history length L = 4
    Hand-set; not ablated.
  • Base-distribution scale sigma = unstated
    N(0, sigma^2 I) in Eqs. (4), (9); sigma directly scales the score but its value is never given.
  • Conformal miscoverage alpha and calibration size N = alpha=0.05, N=19
    Standard; with N=19 the threshold is the max calibration score, limiting FAR control.
  • Latent geometry / DINOv2 variant = 16x16x8 channels; DINOv2 variant unstated
    Carried from VFMF; backbone size (ViT-S/B/L) unspecified, affects features.
assumptions (4)
  • domain assumption Inverse transport of the observed endpoint under the learned flow field lands near the Gaussian base if and only if the window is consistent with nominal dynamics (Eqs. 8-9).
    Asserted ('should lie near the Gaussian base distribution'), credited to FAILDetect [13]; flow matching learns a mean-field velocity, so the inversion is a heuristic whose discriminative validity is empirical.
  • domain assumption Exchangeability of calibration and deployment rollouts for the conformal threshold (Eq. 10).
    Authors flag it in §VI; fixed-view shunt insertion fires on 6/20 successful rollouts (Table I), evidence the assumption is violated in at least one view.
  • domain assumption DINOv2 + beta-VAE latents preserve failure-relevant visual evidence of the monitored view.
    Scores derive only from one camera stream; failures invisible in that stream cannot be detected (§IV-B).
  • domain assumption World model trained on expert demonstrations transfers to ACT policy rollouts with only threshold recalibration.
    §IV-A: detector trained on demonstrations, calibrated on policy rollouts; the distribution shift is absorbed by the conformal threshold rather than by retraining.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Failure Detection for Surgical Robot Imitation Policies via Flow-Matching World Modeling." pith.science (2026). https://pith.science/paper/P4TFIKU5

@misc{pith2026260727511,
  author       = {Pith},
  title        = {Pith review of: Failure Detection for Surgical Robot Imitation Policies via Flow-Matching World Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P4TFIKU5}},
  note         = {Machine review of arXiv:2607.27511}
}
read the original abstract

Imitation learning has shown increasing promise for autonomous robotic surgery, yet safe deployment remains challenging due to the safety-critical nature of surgical tasks and the complexity and variability of surgical environments. Failure detection is therefore an essential safeguard, but its development remains difficult due to the challenges of scarce failure data, highly variable manipulation dynamics, and the need to balance missed detections against disruptive false alarms. To address these challenges, we introduce FoMo-FD (Flow-Matching World Model for Failure Detection), a failure detection method that learns nominal short-horizon visual dynamics with an action-conditioned flow-matching world model. FoMo-FD scores the inverse-transport nonconformity of observed endpoint latents, enabling window-level detection of visual-action inconsistencies without requiring failure demonstrations. Detection thresholds are obtained by conformal calibration on successful executions, yielding task-specific alarms without assuming future failure types. We evaluate FoMo-FD on four surgically relevant manipulation tasks with twenty failure modes across simulation and real-world experiments using the da Vinci Research Kit (dVRK). Results show that FoMo-FD outperforms observation-level anomaly baselines and a prediction-error variant of the same world model, with the wrist-camera view achieving the strongest performance, including a 96.6% failure detection rate (FDR) at a 1.3% false alarm rate (FAR).

Figures

Figures reproduced from arXiv: 2607.27511 by the authors.

Figure 1
Figure 1. Overview of the proposed FoMo-FD framework. Multi-view observations are encoded by a frozen DINOv2 backbone and an offline-trained VAE [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Representative observations from the four tasks. Each panel shows [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative wrist-view score traces across tasks and detector variants. Rows indicate tasks and columns indicate detectors. Blue dashed curves [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Representative failure episodes using the proposed WM-NC wrist-view score. For each example, the score trace is paired with wrist-camera [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Action conditioning modes are denoted as NoAct, Mod, Tok, [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 3 linked inside Pith

  1. [1]

    Robotic constrained imitation learning for the peg transfer task in fundamentals of laparo- scopic surgery,

    K. Kawaharazuka, K. Okada, and M. Inaba, “Robotic constrained imitation learning for the peg transfer task in fundamentals of laparo- scopic surgery,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 606–612

    2024

  2. [2]

    Surgical robot transformer (srt): Imitation learning for surgical tasks,

    J. W. Kim, T. Z. Zhao, S. Schmidgall, A. Deguet, M. Kobilarov, C. Finn, and A. Krieger, “Surgical robot transformer (srt): Imitation learning for surgical tasks,” inConference on Robot Learning. PMLR, 2025, pp. 130–144

    2025

  3. [3]

    Srt-h: A hierarchical framework for autonomous surgery via language-conditioned imitation learning,

    J. W. Kim, J.-T. Chen, P. Hansen, L. X. Shi, A. Goldenberg, S. Schmidgall, P. M. Scheikl, A. Deguet, B. M. White, D. R. Tsai, et al., “Srt-h: A hierarchical framework for autonomous surgery via language-conditioned imitation learning,”Science robotics, vol. 10, no. 104, p. eadt5254, 2025

    2025

  4. [4]

    Suturebot: A precision framework & benchmark for autonomous end-to-end suturing,

    J. Haworth, J.-T. Chen, N. Nelson, J. W. Kim, M. Moghani, C. Finn, and A. Krieger, “Suturebot: A precision framework & benchmark for autonomous end-to-end suturing,”Advances in Neural Information Processing Systems, vol. 38, 2026

    2026

  5. [5]

    Real-time context-aware detection of unsafe events in robot-assisted surgery,

    M. S. Yasar and H. Alemzadeh, “Real-time context-aware detection of unsafe events in robot-assisted surgery,” in2020 50th Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN). IEEE, 2020, pp. 385–397

    2020

  6. [6]

    A confidence-based supervised-autonomous control strategy for robotic vaginal cuff closure,

    M. Kam, H. Saeidi, M. H. Hsieh, J. U. Kang, and A. Krieger, “A confidence-based supervised-autonomous control strategy for robotic vaginal cuff closure,” in2021 IEEE international conference on robotics and automation (ICRA). IEEE, 2021, pp. 12 261–12 267

    2021

  7. [7]

    Early failure detection in autonomous surgical soft-tissue manipula- tion via uncertainty quantification,

    J. Thompson, R. Koe, A. Le, G. Goodman, D. S. Brown, and A. Kuntz, “Early failure detection in autonomous surgical soft-tissue manipula- tion via uncertainty quantification,”arXiv preprint arXiv:2501.10561, 2025

    arXiv 2025

  8. [8]

    Safe uncertainty-aware learning framework for robotic suturing,

    W. P. Empleo, Y . Kim, H. Kim, T. R. Savarimuthu, and I. Iturrate, “Safe uncertainty-aware learning framework for robotic suturing,” IEEE Transactions on Medical Robotics and Bionics, vol. 8, no. 1, pp. 41–53, 2026

    2026

Show all 32 references
  1. [9]

    Unpacking failure modes of generative policies: Runtime monitoring of consistency and progress,

    C. Agia, R. Sinha, J. Yang, Z. Cao, R. Antonova, M. Pavone, and J. Bohg, “Unpacking failure modes of generative policies: Runtime monitoring of consistency and progress,” inConference on Robot Learning. PMLR, 2025, pp. 689–723

    2025

  2. [10]

    Code-as-monitor: Constraint-aware visual programming for reactive and proactive robotic failure detection,

    E. Zhou, Q. Su, C. Chi, Z. Zhang, Z. Wang, T. Huang, L. Sheng, and H. Wang, “Code-as-monitor: Constraint-aware visual programming for reactive and proactive robotic failure detection,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 6919– 6929

    2025

  3. [11]

    Safe: Multitask failure detection for vision-language- action models,

    Q. Gu, Y . Ju, S. Sun, I. Gilitschenski, H. Nishimura, M. Itkina, and F. Shkurti, “Safe: Multitask failure detection for vision-language- action models,”Advances in Neural Information Processing Systems, vol. 38, pp. 40 041–40 076, 2026

    2026

  4. [12]

    Failure prediction at runtime for generative robot policies,

    R. R ¨omer, A. Kobras, L. Worbis, and A. Schoellig, “Failure prediction at runtime for generative robot policies,”Advances in Neural Infor- mation Processing Systems, vol. 38, pp. 7631–7670, 2026

    2026

  5. [13]

    Can We Detect Failures Without Failure Data? Uncertainty-Aware Runtime Failure Detection for Imitation Learning Policies,

    C. Xu, T. K. Nguyen, E. Dixon, C. Rodriguez, P. Miller, R. Lee, P. Shah, R. A. Ambrus, H. Nishimura, and M. Itkina, “Can We Detect Failures Without Failure Data? Uncertainty-Aware Runtime Failure Detection for Imitation Learning Policies,” inProceedings of Robotics: Science an...

    2025

  6. [14]

    Rc-nf: Robot-conditioned normalizing flow for real-time anomaly detection in robotic manipulation,

    S. Zhou, B. Zhu, J. Yang, X. Zhao, J. Chen, and Y .-G. Jiang, “Rc-nf: Robot-conditioned normalizing flow for real-time anomaly detection in robotic manipulation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 43 050–43 060

    2026

  7. [15]

    Founda- tional world models accurately detect bimanual manipulator failures,

    I. R. Ward, M. Ho, H. Liu, A. Feldman, J. Vincent, L. Kruse, S. Cheong, D. Eddy, M. J. Kochenderfer, and M. Schwager, “Founda- tional world models accurately detect bimanual manipulator failures,” arXiv preprint arXiv:2603.06987, 2026

    2026

  8. [16]

    Conformal prediction: A gentle introduction,

    A. N. Angelopoulos and S. Bates, “Conformal prediction: A gentle introduction,”F oundations and Trends in Machine Learning, vol. 16, no. 4, pp. 494–591, 2023

    2023

  9. [17]

    An open-source research kit for the da vinci® surgical system,

    P. Kazanzides, Z. Chen, A. Deguet, G. S. Fischer, R. H. Taylor, and S. P. DiMaio, “An open-source research kit for the da vinci® surgical system,” in2014 IEEE international conference on robotics and automation (ICRA). IEEE, 2014, pp. 6434–6439

    2014

  10. [18]

    World models for general surgical grasping,

    H. Lin, B. Li, C. W. Wong, J. Rojas, X. Chu, and K. W. S. Au, “World models for general surgical grasping,” inProceedings of Robotics: Science and Systems, Delft, Netherlands, July 2024

    2024

  11. [19]

    Surgical vision world model,

    S. Koju, S. Bastola, P. Shrestha, S. Amgain, Y . R. Shrestha, R. P. Poudel, and B. Bhattarai, “Surgical vision world model,” inMICCAI Workshop on Data Engineering in Medical Imaging. Springer, 2025, pp. 1–10

    2025

  12. [20]

    Cosmos-h-surgical: Learning surgical robot policies from videos via world modeling,

    Y . He, P. Guo, M. Xu, Z. Li, A. Myronenko, D. Imans, B. Liu, D. Yang, M. Gu, Y . Ji, Y . Jin, R. Zhao, B. Shen, and D. Xu, “Cosmos-h-surgical: Learning surgical robot policies from videos via world modeling,” 2025

    2025

  13. [21]

    Towards suturing world models: Learning predictive models for robotic surgical tasks,

    M. K. Turkcan, M. Ballo, F. Filicori, and Z. Kostic, “Towards suturing world models: Learning predictive models for robotic surgical tasks,”

  14. [22]

    Cosmos-surg-dvrk: world foundation model-based automated online evaluation of surgical robot policy learning,

    L. Zbinden, N. Nelson, J.-T. Chen, X. Chen, J. W. Kim, M. Azizian, A. Krieger, and S. Huver, “Cosmos-surg-dvrk: world foundation model-based automated online evaluation of surgical robot policy learning,”IEEE Robotics and Automation Letters, 2026

    2026

  15. [23]

    How far are surgeons from surgical world models? a pilot study on zero-shot surgical video generation with expert assessment,

    Z. Chen, Q. Xu, J. Wu, B. Yang, Y . Zhai, G. Guo, J. Zhang, Y . Ding, N. Navab, and J. Luo, “How far are surgeons from surgical world models? a pilot study on zero-shot surgical video generation with expert assessment,” 2025. [Online]. Available: https://arxiv.org/abs/2511.01775

    2025

  16. [24]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khali- dov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby,et al., “Dinov2: Learning robust visual features without supervision,”Transactions on Machine Learning Research Journal, 2024

    2024

  17. [25]

    Dino-wm: World models on pre-trained visual features enable zero-shot planning,

    G. Zhou, H. Pan, Y . Lecun, and L. Pinto, “Dino-wm: World models on pre-trained visual features enable zero-shot planning,” inInternational Conference on Machine Learning. PMLR, 2025, pp. 79 115–79 135

    2025

  18. [26]

    Back to the features: Dino as a foundation for video world models,

    F. Baldassarre, M. Szafraniec, B. Terver, V . Khalidov, F. Massa, Y . LeCun, P. Labatut, M. Seitzer, and P. Bojanowski, “Back to the features: Dino as a foundation for video world models,” 2025. [Online]. Available: https://arxiv.org/abs/2507.19468

    2025 arXiv

  19. [27]

    Dino- foresight: Looking into the future with dino,

    E. Karypidis, I. Kakogeorgiou, S. Gidaris, and N. Komodakis, “Dino- foresight: Looking into the future with dino,”Advances in Neural Information Processing Systems, vol. 38, pp. 163 779–163 811, 2026

    2026

  20. [28]

    Vfmf: World modeling by forecasting vision foundation model features,

    G. Boduljak, Y . Lan, C. Rupprecht, and A. Vedaldi, “Vfmf: World modeling by forecasting vision foundation model features,” 2025. [Online]. Available: https://arxiv.org/abs/2512.11225

    2025

  21. [29]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  22. [30]

    Sufia-bc: Generating high quality demonstration data for visuomotor policy learning in surgical subtasks,

    M. Moghani, N. Nelson, M. Ghanem, A. Diaz-Pinto, K. Hari, M. Az- izian, K. Goldberg, S. Huver, and A. Garg, “Sufia-bc: Generating high quality demonstration data for visuomotor policy learning in surgical subtasks,” in2025 IEEE International Conference on Robotics and Automati...

    2025

  23. [31]

    Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware,” inProceedings of Robotics: Science and Systems, Daegu, Republic of Korea, July 2023

    2023

  24. [2025]

    Available: https://arxiv.org/abs/2503.12531

    [Online]. Available: https://arxiv.org/abs/2503.12531

    arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.