EndoWAM: A Grounded World-Action Model for Generalizable Endoscopic Navigation

Chi Kit Ng; Hongliang Ren; Huxin Gao; Jeremy Yuen-Chun Teoh; Jiaxi Wang; Jinsong Lin; Liangjing Shao; Wanhao Liu; Yin Wang; Zhiyong Xiong

REVIEW 3 major objections 4 minor 13 references

EndoWAM: A Grounded World-Action Model for Generalizable Endoscopic Navigation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read EndoWAM reports that grounding an endoscopic action policy in predicted future target regions—rather than current-frame targets—raises closed-loop navigation success to 80.2% on physical phantom trials, with 90.1% zero-shot generalization.

desk verdict Solid, well-ablated WAM-for-endoscopy paper that deserves refereeing, but the success metric appears to reuse the same detector that generates action and grounding labels, so the headline margins need an independent check before I'd trust the generalization claims. read the letter →

arxiv 2608.01221 v1 pith:NDFGMTAG submitted 2026-08-02 cs.RO

Jinsong Lin , Zikang Pan , Wanhao Liu , Chi Kit Ng , Liangjing Shao , Zihang Yu , Ziyu Wang , Yin Wang

show 5 more authors

Jiaxi Wang Jeremy Yuen-Chun Teoh Zhiyong Xiong Huxin Gao Hongliang Ren

This is my paper

classification cs.RO

keywords WorldActionModelEndoscopicnavigationFuturegroundingDiffusiontransformerVision-language-actionRoboticsurgeryZero-shotgeneralizationDiscreteexpert

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

EndoWAM argues that robotic endoscopic navigation improves when the policy is grounded in predicted future targets rather than current frames alone. It couples a pretrained video world model with a discrete action expert through a shared predictive representation: the action expert and a Grounding DiT both read intermediate denoising features, so the world model learns target-aware dynamics in one denoising pass. On physical phantom trials across ureteroscopy, esophagoscopy, and ERCP, the paper reports 80.2% closed-loop success, 53.1 points above the strongest baseline, and 90.1% zero-shot success under unseen environments and target shifts. If true, predictive future grounding—not current-frame grounding—is the key to robust control under occlusion and viewpoint change, at a control frequency fast enough for deployment.

What carries the argument

The load-bearing object is the shared predictive representation $H=\mathcal{H}(z_{t+1}^{\tau_v})$, extracted from an intermediate block of a pretrained video DiT after a single denoising pass. It encodes future latent dynamics without decoding pixel frames; the discrete action expert decodes action chunks from it, and the Grounding DiT reconstructs VAE latents of future target-region crops from it. The future-grounding objective injects target-aware supervision into intermediate world-model features, and because only one denoising pass runs at inference, the dynamics prior costs little latency.

What would settle it

Re-label the instructed target's future positions in the EndoMotion phantom videos by independent manual annotation instead of the detector-plus-Kalman OBB tracks; retrain and redeploy EndoWAM on the same 32-trial protocol. If manual and automatic tracks diverge on dark or occluded frames and success drops substantially, the reported margins rest on tracker-consistent targets rather than independently verifiable navigation.

Watch

Extended reading notes

Core claim

EndoWAM claims to be the first World Action Model for generalizable robotic endoscopic navigation. Its central proposal is future grounding: instead of localizing the instructed anatomical target only in the current observation, the model predicts task-relevant target regions in future observations from the intermediate denoising features of a video diffusion transformer. That shared predictive representation simultaneously conditions a discrete action expert that outputs multi-step three-degree-of-freedom commands and a Grounding DiT that reconstructs latent target crops from future frames. At deployment the grounding branch is dropped, so the future-grounded representation adds no extra in

Load-bearing premise

The entire training signal—action labels and future-target supervision—comes from Grounding DINO detections smoothed by a Kalman filter, so if those automatic target tracks are wrong on occluded or dark frames, the policy learns from corrupted targets and the success numbers, measured against the same tracker, could be overstated.

Editorial extensions

If this is right

If the reported numbers hold, predictive future grounding rather than current-frame grounding accounts for most of the gain: removing the Grounding DiT drops average success from 80.2% to 37.5%.
A single denoising pass can supply both a world-model prior and real-time control (7.5 Hz), making WAM-style policies feasible in time-sensitive endoscopy where prior WAMs run near 1 Hz.
The same target-aware predictive representation transfers zero-shot to unseen viewpoints, lumen reflectance, geometry, texture, and target shape, size, and color at 90.1% average success.
Roll-viewpoint augmentation at 45-degree intervals is a first-order training ingredient: without it, average success falls to 28.1%.
For the same vision-language backbone, the discretized action expert outperforms a diffusion action head (17.7% versus 8.3%), indicating action parameterization matters alongside predictive dynamics.

Reading between the lines

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

Because success is measured against targets detected by the same Grounding DINO pipeline that produced the training labels, the 90.1% zero-shot figure may partly reflect tracker consistency rather than anatomical understanding; an independent expert-labeled evaluation would separate the two.
The future-grounding recipe should transfer to other deformable, occluded visuomotor settings—bronchoscopy, colonoscopy, catheter or guidewire navigation—wherever a trackable target signal exists, but annotation quality would resurface as the bottleneck in each new anatomy.
The action expert currently emits three discrete degrees of freedom; extending the same shared predictive representation to higher-DoF instruments or continuous action distributions is a natural next test of generality.
If target annotations were available at multiple semantic levels (landmarks, lumen centerline, tool pose), the same grounding objective could encode richer future structure than a single oriented-bounding-box crop.

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.

Desk Editor's Note

read the letter

Read this one if you work on WAMs or surgical navigation. The genuine new thing is future grounding: conditioning a Grounding DiT on intermediate denoising features to reconstruct target latents in future frames, rather than grounding only current observations. That's a clean idea, clearly described, and the ablations back it: dropping the Grounding DiT costs 42.7 points, dropping viewpoint augmentation costs 52.1. The controlled variants (EG, CG, FG-Gen) are a sensible way to isolate the grounding paradigm. EndoMotion, 515 trajectories across three procedures with rotation augmentation, is a useful dataset contribution even if not released.

The weak spot is real and it's about the evaluation loop. Grounding DINO plus Kalman OBB tracks generate the action labels and the future-target crops used for supervision. Section 4.1 defines success as reaching the instructed target and maintaining it in view, but doesn't say who checks that. If the same detector is the judge, then policy, grounding branch, and success metric are all aligned to the same visual signatures. A detector that tracks a specular highlight or lumen edge rather than the anatomical target would teach that bias and then reward the policy for chasing it. That wouldn't fully explain the margins—baselines share the same action labels and would inherit the same bias—but it could inflate EndoWAM's advantage specifically because its future-grounding machinery is optimized to reproduce detector tracks. The paper needs either human verification of success, a held-out independent detector, or a per-trial target-localization error metric.

Minor but related: there are no confidence intervals or error bars on any success rate. 32 trials per condition gives a standard error around 8–9 percentage points per cell, so the per-procedure differences among EndoWAM variants are not all crisp, though the headline 80.2 vs 27.1 gap is large enough to survive. Also, the 'unseen viewpoint' test only varies in-plane roll; that's a narrower generalization claim than the abstract implies. No code or dataset release is mentioned, which matters for a paper whose contribution includes a dataset.

Overall: the central mechanism is plausible and well supported internally. The circularity concern is answerable and not necessarily a killer, but it's load-bearing enough that the paper needs a revision or at least an explicit limitation. I'd send it out. A good referee could ask for the success-criterion details and error bars in a normal round.

Referee Report

3 major / 4 minor

Summary. The paper proposes EndoWAM, described as the first World Action Model for generalizable robotic endoscopic navigation. The method couples a pretrained video diffusion transformer (Cosmos-Predict2.5-2B) with a discrete action expert and a 'Grounding DiT' that reconstructs future target-region latents from intermediate denoising features. The authors introduce EndoMotion, a dataset of 515 trajectories across ureteroscopy, esophagoscopy, and ERCP, with Grounding-DINO-derived target tracks and 8-fold roll augmentation. Experiments on physical phantoms report 80.2% average closed-loop success over 96 trials, outperforming the strongest baseline by 53.1 points, and 90.1% average zero-shot success under unseen environmental and target shifts. Ablations show that removing the Grounding DiT or viewpoint augmentation substantially lowers success. The central claim is that future grounding, not current-frame grounding, is the key to robust endoscopic navigation.

Significance. If the reported results hold, the paper makes a substantive contribution: it demonstrates a concrete instantiation of a world-action model for endoscopy, introduces a reusable dataset, and provides a controlled comparison of grounding strategies. The ablations in Table 2 support the causal importance of the future-grounding branch and viewpoint augmentation, and the physical-phantom evaluation is a strength. The controlled variants (EG, CG, FG-Gen) are a thoughtful way to isolate the mechanism. However, the empirical claims are currently undermined by two load-bearing issues: the success metric may be entangled with the same detector used to generate training labels, and no uncertainty quantification is provided for any success rate. Addressing these issues is necessary before the 'generalizable navigation' claim can be accepted.

major comments (3)

[Sec. 3.3, Sec. 4.1] The evaluation protocol may be circular. In Sec. 3.3, Grounding DINO plus Kalman-filtered OBB tracks provide both the action labels (target-center offsets) and the future-target crops used for grounding supervision. In Sec. 4.1, a trial is successful if 'the endoscope reaches the instructed target and maintains it within the field of view,' but the manuscript never specifies how 'reaches' is measured in the physical phantom trials. If the same Grounding-DINO/Kalman tracker determines success, then the policy, all baselines, and the success metric share the same visual signature of the target; a systematic detector bias would be learned and then rewarded. This directly affects the central claim of generalizable navigation and the 80.2%/90.1% margins. The authors must state the success-determination mechanism and provide independent verification (e.g., human review, independent segmentatio
[Sec. 4.2, Tables 1-3] No uncertainty quantification is reported. With 32 trials per condition, a success rate of 56.3% vs 43.8% (esophagoscopy, EndoWAM vs EndoWAM-EG) has a binomial standard error on the difference of roughly 12.4 percentage points; the 12.5-point gap is not statistically meaningful. Even the headline 80.2% average over 96 trials should be accompanied by confidence intervals, and per-procedure claims should be flagged where not significant. The absence of CIs weakens the 'consistently outperforms' claim and the zero-shot results in Table 3.
[Sec. 4.5, Table 3] The zero-shot generalization result uses a two-stage protocol (lumen traversal + target localization) and the table footnote says it is 'not directly comparable' to the three-stage evaluation in Table 1, yet the abstract and Sec. 4.5 headline 90.1% without this caveat. Please report the protocol difference prominently in the abstract and clarify how many trials constitute each condition, and state whether the same success criteria and target-detection method apply.

minor comments (4)

[Appendices] Appendices A-G are referenced but not included in the manuscript; without them, details of the phantom setup, success criteria, and annotation pipeline are unverifiable. Please include them or state their availability.
[Eq. (1), Sec. 4.1] The shared representation H(z^tau_v_{t+1}) is defined abstractly in Eq. (1), but the connection to 'block-17 activation after a single denoising pass' is only given in Sec. 4.1. Define H's extraction point and dimension in Sec. 3.2.
[Sec. 3.2, Sec. 4.6] The phrase 'single denoising pass' should be made precise: how many flow steps are used at inference, with what deterministic schedule, and how does this relate to the training flow time tau_v?
[General] The provided PDF text contains many missing spaces between words, apparently a formatting artifact. The final manuscript must be re-typeset so that equations and prose are readable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical training/evaluation study; the use of the same detector for action labels and grounding supervision is a consistency of supervision, not a circular derivation.

full rationale

EndoWAM is an empirical systems paper: it constructs a dataset with Grounding DINO target tracks, derives discrete action labels and future-grounding supervision from those tracks, and trains a policy. This is a standard supervised-learning setup in which the target definition is shared between supervision signals; it does not constitute circularity unless the evaluation success criterion is shown to be the same detector output. The paper defines trial success as 'the endoscope reaches the instructed target and maintains it within the field of view' (Sec. 4.1) and does not state that success is scored by the same Grounding DINO/Kalman tracker used to generate training labels. Without such a statement, the shared-detector concern is a potential evaluation-bias issue, not an exhibited circular step. The only author-overlapping citation (BiliVLA, Lin et al. 2026) appears in related work and is not load-bearing; no uniqueness theorem, ansatz, or fitted parameter is imported from that work. All central comparisons are against external baselines on physical phantoms, and the claimed contributions are validated by closed-loop success rates rather than by derivation from the annotations themselves. Therefore no specific reduction of the paper's claims to its inputs can be quoted, and the circularity score is 0.

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

The paper introduces a new training objective (future grounding) and a dataset, but no new physical entities. Its claims rest on the transferability of a pretrained video prior, the accuracy of an automatic detector for all training labels, and the representativeness of roll-only phantom trials.

free parameters (5)

lambda_v, lambda_r (loss weights) = not disclosed
Balance weights in Eq. 6 between action, video, and grounding losses; set by hand and not reported in main text.
w_a (per-axis class weights) = not disclosed
Weights in the weighted cross-entropy for discrete action prediction (Eq. 4); chosen by hand.
tau_v (denoising time for shared representation) = block-17 activation after a single denoising pass (tau_v not numeric)
The intermediate latent state used as H in Eq. 1; a design choice that conditions both heads.
Rotation augmentation set = 8 angles at 45-degree intervals
Data augmentation that expands the dataset; the evaluation uses unaligned roll offsets, so the model never sees the exact angles at test.
Action bin discretization = not disclosed
Discrete categorical actions per axis; the resolution of the action space is part of the task definition.

assumptions (4)

domain assumption Cosmos-Predict2.5-2B provides a video dynamics prior that transfers to endoscopic scenes after LoRA fine-tuning
The world model is initialized from this pretrained DiT and adapted with LoRA; if the prior does not transfer, the predictive features would not capture endoscopic dynamics (Sec. 3.2).
domain assumption Grounding DINO OBB detections, after Kalman filtering, are accurate enough to derive both action labels and future-target supervision
The action labels and grounding targets both come from this pipeline; annotation errors propagate into both supervision signals (Sec. 3.3).
standard math Rectified flow training objective (Eq. 3) is a valid way to train the video world model
Standard training objective borrowed from diffusion/flow literature; not original to this paper.
domain assumption Physical phantom trials with roll-offset viewpoint sampling are representative of endoscopic navigation challenges
The evaluation protocol (Sec. 4.1) uses 32 trials per procedure with in-plane roll variations; if real-world tissue contact and deformation differ, closed-loop results may not transfer.

how reviews work

0 comments

Cite this review

Pith. "Pith review of EndoWAM: A Grounded World-Action Model for Generalizable Endoscopic Navigation." pith.science (2026). https://pith.science/paper/NDFGMTAG

@misc{pith2026260801221,
  author       = {Pith},
  title        = {Pith review of: EndoWAM: A Grounded World-Action Model for Generalizable Endoscopic Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NDFGMTAG}},
  note         = {Machine review of arXiv:2608.01221}
}

read the original abstract

Autonomous endoscopic navigation can reduce clinicians' operational burden, yet robust control remains challenging due to tissue deformation, transient occlusions, and rapidly changing viewpoints. Existing learning-based policies typically predict actions from current observations without explicitly modeling future dynamics, limiting their robustness and reliability in safety-critical settings. World Action Models (WAMs) offer a promising alternative by coupling predictive visual dynamics with action generation, but extending them to robotic endoscopy remains challenging due to limited training data, restricted viewpoint diversity, deformable anatomy, and high inference latency. We present EndoWAM, which is, to our knowledge, the first WAM for generalizable robotic endoscopic navigation. EndoWAM introduces future grounding, which predicts task-relevant target regions in future observations from intermediate denoising features of a video world model. Specifically, EndoWAM couples a lightweight diffusion transformer for future target-region prediction with a discrete action expert through a shared predictive representation. This design injects target-aware supervision into predictive dynamics modeling, improving robustness to visual degradation and viewpoint changes while enabling real-time control in a single denoising pass. We further introduce EndoMotion, a robotic endoscopic motion dataset spanning three anatomically distinct procedures: ureteroscopy, esophagoscopy, and endoscopic retrograde cholangiopancreatography (ERCP). EndoWAM consistently outperforms all baselines and alternative grounding strategies, while demonstrating strong zero-shot generalization to unseen viewpoints, environments, and targets. These results establish EndoWAM as a predictive, target-grounded framework for accurate, generalizable, and long-horizon navigation in visually constrained endoscopic environments.

Figures

Figures reproduced from arXiv: 2608.01221 by the authors.

**Figure 1.** Overview of EndoWAM for autonomous navigation across diverse endoscopic tasks. The policy maps the current observation and instruction to action chunks that control the three degrees of freedom of a flexible endoscope. intervention induces complex and rapidly evolving visual dynamics, while transient occlusions further degrade visual observations, posing significant challenges to vision-based navigation methods. Re… view at source ↗

**Figure 2.** EndoWAM architecture. The Video DiT models future latent dynamics from the current frame and instruction. Its hidden states condition the Grounding DiT to reconstruct future target regions and the action expert to generate an action chunk. RoboGround (Huang et al. 2025) uses LISA (Lai et al. 2024) to segment target and background regions as auxiliary inputs, while VIP (Li et al. 2025c) crops and enlarges target reg… view at source ↗

**Figure 3.** Conceptual comparison of different paradigms in EndoWAM. (a) EndoWAM-EG performs explicit input grounding by injecting a grounding-expert crop of the current target. (b) EndoWAM-CG performs CoT grounding by regressing future OBBs from the shared predictive representation. (c) EndoWAM-FG-Gen jointly generates future frames and target regions through the world model’s generative pathway. (d) EndoWAM (Ours) performs… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Closed-loop deployment trajectories. Representative EndoWAM rollouts across ureteroscopy, esophagoscopy, and ERCP, covering the three navigation phases: entrance cannulation, lumen traversal, and target localization. 4 Experiments We design our experiments around the f…

**Figure 5.** Figure 5: Attention visualization for Grounding DiT ablation. Without the Grounding DiT, the policy attends diffusely to specular highlights and mucosal textures, whereas the full model concentrates its attention around the task-relevant target even under darkness and partial …

**Figure 6.** Figure 6: Zero-shot deployment examples. Representative observations from EndoWAM rollouts under unseen crossenvironment shifts in lumen reflectance, geometry, and texture, and cross-target shifts in target shape, size, and color. and partial occlusion. These results show that…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 1 canonical work pages

[2]

arXiv:2506.09985

V-JEPA 2: Self-supervised video models enable understanding, prediction and planning. arXiv:2506.09985. Bai, S.; Cai, Y.; Chen, R.; Chen, K.; Chen, X.; Cheng, Z.; Deng,L.;Ding,W.;Gao,C.;Ge,C.;etal.2025. Qwen3-VL technical report. arXiv:2511.21631. Bi,H.;Tan,H.;Xie, S.;Wang,Z.;Huang,S.;Liu,H.;Zhao, R.; Feng, Y.; Xiang, C.; Rong, Y.; et al

arXiv 2025
[3]

arXiv:2512.13030

Motus: A unified latent action world model. arXiv:2512.13030. Bjorck, J.; Castañeda, F.; Cherniadev, N.; Da, X.; Ding, R.; Fan, L.; Fang, Y.; Fox, D.; Hu, F.; Huang, S.; et al

arXiv
[4]

arXiv:2503.14734

GR00T N1: An open foundation model for generalist hu- manoid robots. arXiv:2503.14734. Black,K.;Brown,N.;Driess,D.;Esmail,A.;Equi,M.;Finn, C.; Fusai, N.; Groom, L.; Hausman, K.; Ichter, B.; et al. 2024.π 0: A Vision-Language-Action Flow Model for Gen- eral Robot Control. arXiv:2410.24164. Buxbaum, J. L.; Freeman, M.; Amateau, S. K.; Chalhoub, J.M.;Coelho-...

arXiv 2024
[5]

arXiv:2505.03233

GraspVLA: A grasping foundation model pre-trained on billion-scale synthetic action data. arXiv:2505.03233. Duan,X.;Xie,D.;Zhang,R.;Li,X.;Sun,J.;Qian,C.;Song, X.; and Li, C

arXiv
[7]

Kim,J.W.;Zhao,T.Z.;Schmidgall,S.;Deguet,A.;Krieger, M.; Finn, C.; and Krieger, A

SRT-H: A hierarchical framework for autonomous surgery via language-conditioned imitation learning.Science Robotics, 10(104): eadt5254. Kim,J.W.;Zhao,T.Z.;Schmidgall,S.;Deguet,A.;Krieger, M.; Finn, C.; and Krieger, A. 2024a. Surgical robot transformer (SRT): Imitation learning for surgical tasks. arXiv:2407.12998. Kim, M. J.; Gao, Y.; Lin, T.-Y.; Lin, Y.-...

arXiv 2026
[9]

BiliVLA: Scene-Aware Vision-Language-Action Model with Reinforcement Learning for Autonomous Biliary Endoscopic Navigation

BiliVLA: Scene-aware vision-language-action model with reinforcement learning for autonomous biliary endoscopic navigation. arXiv:2606.23531. Liu, S.; Zeng, Z.; Ren, T.; Li, F.; Zhang, H.; Yang, J.; Jiang, Q.; Li, C.; Yang, J.; Su, H.; et al

work page Pith review arXiv
[10]

Ma,T.;Zheng,J.;Wang,Z.;Jiang,C.;Cui,A.;Liang,J.;and Yang,S.2026.Dit4dit:Jointlymodelingvideodynamicsand actions for generalizable robot control

Sur- gical embodied intelligence for generalized task autonomy in laparoscopic robot-assisted surgery.Science Robotics, 10(104): eadt3093. Ma,T.;Zheng,J.;Wang,Z.;Jiang,C.;Cui,A.;Liang,J.;and Yang,S.2026.Dit4dit:Jointlymodelingvideodynamicsand actions for generalizable robot control. arXiv:2603.10448. Moghani,M.;Nelson,N.;Ghanem,M.;Diaz-Pinto,A.;Hari, K.;A...

arXiv 2026
[12]

arXiv:2508.10333

Recon- VLA: Reconstructive vision-language-action model as ef- fective robot perceiver. arXiv:2508.10333. Tian, Q.; Liao, H.; Huang, X.; Chen, J.; Zhang, Z.; Yang, B.;Ourselin,S.;andLiu,H.2024. DD-VNB:Adepth-based dual-loop framework for real-time visually navigated bron- choscopy. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems...

arXiv 2024

Show all 13 references

[13]

arXiv:2603.17240

Gigaworld-policy: An efficient action-centered world–model. arXiv:2603.17240. Zawalski,M.;Chen,W.;Pertsch,K.;Mees,O.;Finn,C.;and Levine, S
[2022]

InInternational Conference on Learning Representations (ICLR)

LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Representations (ICLR). Huang, H.; Chen, X.; Chen, Y.; Li, H.; Han, X.; Wang, Z.; Wang,T.;Pang,J.;andZhao,Z.2025. RoboGround:Robotic manipulationwithgroundedvision–languagepriors. InPro-...

2025 arXiv
[2024]

Singh,A.;Shi,W.;andWang,M.D.2023

General-purpose foundation models for increased autonomy in robot-assisted surgery.Nature Ma- chine Intelligence, 6(11): 1275–1283. Singh,A.;Shi,W.;andWang,M.D.2023. Autonomoussoft tissue retraction using demonstration-guided reinforcement learning. arXiv:2309.00837. Song, W.;...

2023 arXiv
[2025]

arXiv:2501.03575

Cosmos world foundation model platform for physical AI. arXiv:2501.03575. Alian,A.;Zari,E.;Wang,Z.;Franco,E.;Avery,J.P.;Runci- man,M.;Lo,B.;yBaena,F.R.;andMylonas,G.2023. Cur- rentengineeringdevelopmentsforroboticsystemsinflexible endoscopy.Techniques and Innovations in Gastro...

2023 arXiv
[2026]

arXiv:2601.21998

Causal world modeling for robot control. arXiv:2601.21998. Li, S.; Gao, Y.; Sadigh, D.; and Song, S. 2025a. Unified video action model. arXiv:2503.00200. Li, S.; Wang, J.; Dai, R.; Ma, W.; Ng, W. Y.; Hu, Y.; and Li, Z. 2025b. RoboNurse-VLA: Robotic scrub nurse system basedonVi...

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [2]

arXiv:2506.09985

V-JEPA 2: Self-supervised video models enable understanding, prediction and planning. arXiv:2506.09985. Bai, S.; Cai, Y.; Chen, R.; Chen, K.; Chen, X.; Cheng, Z.; Deng,L.;Ding,W.;Gao,C.;Ge,C.;etal.2025. Qwen3-VL technical report. arXiv:2511.21631. Bi,H.;Tan,H.;Xie, S.;Wang,Z.;Huang,S.;Liu,H.;Zhao, R.; Feng, Y.; Xiang, C.; Rong, Y.; et al

arXiv 2025

[2] [3]

arXiv:2512.13030

Motus: A unified latent action world model. arXiv:2512.13030. Bjorck, J.; Castañeda, F.; Cherniadev, N.; Da, X.; Ding, R.; Fan, L.; Fang, Y.; Fox, D.; Hu, F.; Huang, S.; et al

arXiv

[3] [4]

arXiv:2503.14734

GR00T N1: An open foundation model for generalist hu- manoid robots. arXiv:2503.14734. Black,K.;Brown,N.;Driess,D.;Esmail,A.;Equi,M.;Finn, C.; Fusai, N.; Groom, L.; Hausman, K.; Ichter, B.; et al. 2024.π 0: A Vision-Language-Action Flow Model for Gen- eral Robot Control. arXiv:2410.24164. Buxbaum, J. L.; Freeman, M.; Amateau, S. K.; Chalhoub, J.M.;Coelho-...

arXiv 2024

[4] [5]

arXiv:2505.03233

GraspVLA: A grasping foundation model pre-trained on billion-scale synthetic action data. arXiv:2505.03233. Duan,X.;Xie,D.;Zhang,R.;Li,X.;Sun,J.;Qian,C.;Song, X.; and Li, C

arXiv

[5] [7]

Kim,J.W.;Zhao,T.Z.;Schmidgall,S.;Deguet,A.;Krieger, M.; Finn, C.; and Krieger, A

SRT-H: A hierarchical framework for autonomous surgery via language-conditioned imitation learning.Science Robotics, 10(104): eadt5254. Kim,J.W.;Zhao,T.Z.;Schmidgall,S.;Deguet,A.;Krieger, M.; Finn, C.; and Krieger, A. 2024a. Surgical robot transformer (SRT): Imitation learning for surgical tasks. arXiv:2407.12998. Kim, M. J.; Gao, Y.; Lin, T.-Y.; Lin, Y.-...

arXiv 2026

[6] [9]

BiliVLA: Scene-Aware Vision-Language-Action Model with Reinforcement Learning for Autonomous Biliary Endoscopic Navigation

BiliVLA: Scene-aware vision-language-action model with reinforcement learning for autonomous biliary endoscopic navigation. arXiv:2606.23531. Liu, S.; Zeng, Z.; Ren, T.; Li, F.; Zhang, H.; Yang, J.; Jiang, Q.; Li, C.; Yang, J.; Su, H.; et al

work page Pith review arXiv

[7] [10]

Ma,T.;Zheng,J.;Wang,Z.;Jiang,C.;Cui,A.;Liang,J.;and Yang,S.2026.Dit4dit:Jointlymodelingvideodynamicsand actions for generalizable robot control

Sur- gical embodied intelligence for generalized task autonomy in laparoscopic robot-assisted surgery.Science Robotics, 10(104): eadt3093. Ma,T.;Zheng,J.;Wang,Z.;Jiang,C.;Cui,A.;Liang,J.;and Yang,S.2026.Dit4dit:Jointlymodelingvideodynamicsand actions for generalizable robot control. arXiv:2603.10448. Moghani,M.;Nelson,N.;Ghanem,M.;Diaz-Pinto,A.;Hari, K.;A...

arXiv 2026

[8] [12]

arXiv:2508.10333

Recon- VLA: Reconstructive vision-language-action model as ef- fective robot perceiver. arXiv:2508.10333. Tian, Q.; Liao, H.; Huang, X.; Chen, J.; Zhang, Z.; Yang, B.;Ourselin,S.;andLiu,H.2024. DD-VNB:Adepth-based dual-loop framework for real-time visually navigated bron- choscopy. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems...

arXiv 2024