Pith. sign in

REVIEW 5 major objections 7 minor 34 references

RCM-ACT: Imitation Learning with Dynamic RCM Calibration for Autonomous Intraocular Foreign Body Removal

T0 review · 5 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that RCM-ACT, an imitation learning framework with episode-start Remote Center of Motion calibration, achieves end-to-end autonomous intraocular ring grasp-and-place in an eye phantom with 0.686 mm mean grasp deviation and

desk verdict Plausible hardware demo, but the paper does not actually show that RCM calibration is what helps: the key control is missing, and the success rates are inconsistent across abstract, table, and text. read the letter →

arxiv 2508.19191 v3 pith:HNLMEH55 submitted 2025-08-26 cs.RO

Yue Wang , Wenjie Deng , Haotian Xue , Di Cui , Yiqi Chen , Mingchuan Zhou , Haochao Ying , Jian Wu This is my paper
classification cs.RO
keywords imitationlearningintraocularforeignbodyremovaleyesurgicalrobotRemoteCenterofMotioncalibrationactionchunkingtransformervisuomotorpolicyautonomousmanipulationstereomicroscopy
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 show that an imitation learning policy can drive an eye surgical robot through a complete foreign-body ring removal task—grasping a small ring and placing it onto a larger one inside an eye phantom—without teleoperation and without a depth sensor. The obstacle is that the instrument's pivot point (Remote Center of Motion) shifts between demonstrations, so kinematic data live in inconsistent coordinate frames, and motion scaling makes raw controller signals non-transferable. The proposed fix is RCM-ACT: a one-time, episode-level calibration that computes a rotation and translation from three fixed fiducial points and realigns all trajectory data to a global frame, then trains an action-chunking transformer on stereo images plus instrument kinematics. Reported results include a mean 3D grasp deviation of 0.686 mm, but the full-task success rate is stated inconsistently (11/20 in the abstract, 5/10 in Table I, 3 of 5 in the deployment text). If the approach holds up, it suggests that software-defined RCM calibration can replace hardware recalibration for submillimeter intraocular autonomy.

What carries the argument

The load-bearing mechanism is the RCM dynamic calibration module: three fixed, non-coplanar fiducial points reached under the same RCM constraint yield a rotation matrix R_t and translation d_t, and every timestep's pose is realigned by p' = R_t^T(p - d_t). This software-defined spatial anchoring converts RCM-corrupted kinematic data into a single global frame, which then lets a standard ACT-style encoder-decoder—latent variable z, action chunking, MSE reconstruction plus KL regularization—learn one policy across demonstrations. A second piece is the actuator-level action representation, which decouples the policy from the surgeon's motion scaling factor.

What would settle it

Deliberately shift the RCM entry point by a known amount (e.g., 0.5 mm) after the episode-start calibration and run the trained policy. If full-task success and grasp deviation remain at baseline levels, the one-time calibration claim is unsupported; if they degrade markedly, the claim is supported. Similarly, add controlled noise to the three fiducial measurements and measure whether grasp deviation degrades proportionally.

Watch

Extended reading notes

Core claim

The central claim is that adding RCM dynamic calibration to the ACT imitation learning architecture solves the coordinate-frame inconsistency caused by RCM point variation in eye surgery. At the start of each episode the instrument is moved to three fixed non-coplanar points; solving p_i(t) = R_t p_i(0) + d_t yields a rigid transform that is then applied in reverse to every observation and proprioceptive reading, anchoring the whole episode in a global coordinate system C0. Because actions are defined as 5D actuator-level target states (x, y, z, theta, gripper) rather than raw controller inputs, the policy is also invariant to motion scaling factors. Trained on 30 expert demonstrations with

Load-bearing premise

The whole method hinges on the assumption that measuring three fixed points once at the start of an episode—under the same RCM constraint as surgery—yields a rotation and translation that stays valid for every subsequent timestep, so any later drift or fiducial mislocalization corrupts the policy's coordinate frame.

Editorial extensions

If this is right

  • If the results replicate, eye surgical robots could perform a complete foreign-body removal subtask autonomously in phantom models, with the surgeon supervising rather than teleoperating every motion.
  • Software-defined RCM calibration would remove the need to physically recalibrate or re-register the robot when the instrument is reinserted or the entry point shifts.
  • Because the policy is trained on stereo images and instrument kinematics without depth sensors, the same recipe could transfer to other intraocular manipulation tasks that use microscope views.
  • The reported 0.686 mm mean grasp deviation is at the scale needed for touching millimeter-sized intraocular structures, making the framework a candidate basis for more complex vitreoretinal automation.

Reading between the lines

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

  • A natural extension the paper does not test is continuous RCM estimation: tracking the scleral entry point in the stereo images every few timesteps could catch mid-episode drift that the one-time calibration would miss.
  • The three-fiducial touch procedure is a practical bottleneck: automating fiducial detection from the microscope view would make the calibration safer and faster for clinical use.
  • The inconsistent success counts (11/20 vs 5/10 vs 3/5) suggest the headline effect size is sensitive to how trials are counted; a pre-registered, larger trial set would settle the true full-task success rate.
  • If offline calibration is replaced by online alignment, the same architecture could be applied to other pivot-constrained microsurgical tasks, such as subretinal injection or membrane peeling, where the entry point is also variable.
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

5 major / 7 minor

Summary. The paper proposes RCM-ACT, an imitation-learning framework for autonomous intraocular foreign-body ring grasping and placement with an eye-surgical robot. The method combines a transformer-based action-chunking policy (ACT) with a one-time-per-episode 'dynamic RCM calibration' that realigns kinematic observations to a global coordinate system using three fiducial points, plus an exponential action-buffer during inference. Training uses stereo microscope images and 5D actuator-level proprioception from 30 expert demonstrations in an artificial eye model. The paper reports deployment and offline evaluation results, claiming a mean 3-D grasp deviation of 0.686 mm and a full-task success rate that appears as 11/20 in the abstract, 5/10 in Table I, and 3/5 in the text. The central claim is that the RCM calibration is what enables the improvement over baseline ACT.

Significance. If the claimed results were fully supported, RCM-ACT would be a notable step toward autonomous microsurgical manipulation: it addresses a real problem (RCM point variation under uncalibrated microscopy), uses a sensible state representation (actuator-level displacements), and demonstrates a complete hardware deployment with a challenging submillimeter ring-placement task. The paper also includes a useful baseline comparison and a plausible integration of temporal ensembling into ACT. However, the current evidence does not isolate the contribution of the proposed calibration method, the reported success counts are mutually inconsistent, and the statistical support is too thin to substantiate the headline submillimeter-accuracy claim. These issues are load-bearing for the paper's novelty and reliability, though they are addressable with additional experiments and reporting.

major comments (5)
  1. [Table I / Section F] The central benefit attributed to RCM calibration is confounded. The ablation 'RCM-ACT w/o Resample' includes the calibration and the same architecture but omits only the exponential action buffer. On every evaluation metric it is worse than baseline ACT: MSE 0.022512 vs 0.019876, grasp deviation 1.386 mm vs 1.182 mm, latency 94.25 vs 48.50 frames, and deployment success 0/5 vs 0/5. The full RCM-ACT differs from ACT by both calibration and resampling, so the improvement cannot be credited to calibration. An ACT+resampling condition is needed to separate the contribution of the proposed calibration from the temporal ensembling. Without it, the paper's central novelty is not supported.
  2. [Abstract vs Table I vs Section F] The full-task success rate is reported inconsistently: abstract says 11/20, Table I says 5/10, and Section F text says 'three out of five trials.' These cannot all be correct. The number of deployment trials must be stated precisely, and all sections should agree. With only 5 or 10 deployment trials, the success rate has high uncertainty; a Wilson interval or per-trial breakdown would be needed to support the headline claim.
  3. [Section F / Table I] All quantitative claims lack error bars, confidence intervals, or significance tests. MSE, grasp deviation, and latency are single numbers with no variance, and the deployment comparison is based on five trials. The parameter-sensitivity discussion in Section G likewise reports normalized values without error bars or number of repetitions. For a paper whose main quantitative claim is 'submillimeter accuracy,' this is insufficient statistical support.
  4. [Section III-B / Equation (2)] The calibration description contains a geometric impossibility: 'three fixed points that are not coplanar' cannot exist, since any three points define a plane. The intended condition is presumably 'non-collinear.' More importantly, the method assumes that one rigid transform computed from three fiducial points at episode start corrects for RCM drift for the entire episode; no sensitivity analysis or independent validation of the calibration accuracy is provided. If the fiducial points are poorly localized or the RCM drifts after calibration, the realigned data remain corrupted, and the policy inherits that error.
  5. [Section F / Equation (6)] The 'grasp deviation' metric compares the autonomous gripper position with the expert's gripper position at the moment of grasp, rather than with the actual ring or a task-defined target. A policy could have low deviation while still failing to grasp the ring, and conversely a successful grasp could have high deviation if the expert chose a slightly different approach. The claimed 0.686 mm accuracy should be complemented by a direct measure of positioning error relative to the ring (e.g., distance from gripper to ring center at contact) and by the task-success rate.
minor comments (7)
  1. [Title / Header] The arXiv title is 'RCM-ACT...' but the manuscript title in the text is 'AutoRing...'. The names should be consistent throughout.
  2. [Section III-B / Equation (3)] The use of d_t in Equation (2) and then p' = R_t^T (p - d_t) is unclear: is d_t the translation vector in the current frame or the global frame? Please define the transform direction explicitly. Also, the symbol t is overloaded as both timestep and translation vector in Algorithm 2.
  3. [Section III-D / Equation (4)] Equation (4) defines the weighted action as a finite sum with weights m^k (m=0.8, N=3), while Algorithm 2 says w_i = exp(-m i) without normalization. Please clarify the exact weighting scheme and whether the weights are normalized.
  4. [Section III-E / Training] Training details are ambiguous: '32-sequence batches (8 timesteps each)' is unclear, and 'five independent test sets (30 episodes each)' implies 150 episodes, which conflicts with the 30 total demonstration episodes. Please specify how data were split.
  5. [Section III-A] The task definition introduces p_Rblack[t] and p_Rorange[t] as positions of the rings, but these quantities are not used in the policy input or loss. Either remove them or explain their role.
  6. [Table I] The row 'RCM-ACT w.Encoder' has dashes for deployment results and grasp deviation. The text explains that the encoder variant was not deployed, but the table should state 'N/A' or include a footnote.
  7. [Figure 3] The parameter-sensitivity discussion in Section G refers to Figure 3, but the figure is not reproduced in the provided text and no numerical values are given. Please include the actual plot with error bars and axis labels.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: RCM calibration is a preprocessing registration, evaluation uses held-out expert trajectories, and the main issues are missing ablation and inconsistent reporting, not circularity.

full rationale

The derivation chain is: define the task; collect demonstrations; compute an episode-level rigid transform from three fiducial correspondences via Eq. (2); realign all kinematic data with Eq. (3); train an ACT policy on the realigned data; at inference realign observations before predicting action chunks and temporally ensembling them. No step defines the output in terms of the target result or fits a parameter to the evaluation metric. The RCM calibration rotation is solved from fiducial point pairs, not from grasp outcomes or task success. The evaluation metrics compare predicted actions against expert trajectories on a held-out set, and the paper explicitly states 'The expert trajectory is used strictly for evaluation and is not available to the policy during inference,' so the comparison is not circular. There are no load-bearing self-citations: the overlapping-author citations (Zhou et al. [22], [23]) appear only in related-work context and are not used to justify the RCM calibration or the ACT modification. The paper's substantive weaknesses lie elsewhere: the 'RCM-ACT w/o Resample' ablation still includes calibration yet is worse than baseline ACT on every metric, so the contribution of calibration is not isolated by an ACT+resampling condition; and the reported success counts are inconsistent (3/5 in text, 5/10 in Table I, 11/20 in the abstract). These are attribution and reporting problems, not circularity. Accordingly, no circular step can be quoted or reduced to an equation-level identity, and the score is 0.

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

The paper introduces no new physical entities. Its main intellectual content is an algorithmic preprocessing step (RCM calibration) plus a standard ACT policy. The free parameters are training and filtering choices, not physically fitted constants. The key domain assumptions are that a one-time rigid-frame correction fully captures RCM variation and that the plastic-eye ring task is a meaningful proxy for intraocular surgery.

free parameters (4)
  • Chunk size k = 90
    Selected via the parameter sensitivity analysis in Fig. 3 to balance latency, prediction accuracy, and grasp deviation; not derived from first principles.
  • Exponential action buffer weight m and window N = m = 0.8, N = 3
    Used in Eq. 4 to smooth predictions and mitigate temporal misalignment; the exact values are chosen without a principled derivation.
  • Motion scaling factor = 1/30
    Applied during expert demonstrations to map controller movements to instrument motion; this directly shapes the kinematic data the policy learns from.
  • KL regularization weight beta = 0.5
    Training hyperparameter in the ACT-style loss L = L_reconst + beta * L_reg; affects how much latent noise is allowed.
assumptions (5)
  • standard math A rigid 3D rotation and translation can be recovered uniquely from three point correspondences.
    Invoked in Section III-B when solving Eq. 2 for R_t and d_t; requires the three fixed points to be non-collinear, not the stated 'non-coplanar'.
  • domain assumption RCM point variation is a fixed rigid-frame change that remains constant within a single episode.
    Section III-B says calibration is performed once per episode, so any mid-episode drift in the entry point is not corrected.
  • domain assumption The eye phantom and ring task are a valid surrogate for intraocular foreign body removal.
    Experiments use a plastic eye model and rings, not biological tissue; clinical transferability is untested.
  • domain assumption Stereo visual data plus 5D instrument kinematics contain enough information to learn the task without explicit depth sensing.
    The policy input is stereo images and proprioception; the method does not use depth or force feedback.
  • domain assumption The 30 expert demonstrations are sufficiently consistent and near-optimal for behavioral cloning.
    Imitation learning performance is bounded by demonstration quality, and the paper does not analyze inter-demonstration variance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RCM-ACT: Imitation Learning with Dynamic RCM Calibration for Autonomous Intraocular Foreign Body Removal." pith.science (2026). https://pith.science/paper/HNLMEH55

@misc{pith2026250819191,
  author       = {Pith},
  title        = {Pith review of: RCM-ACT: Imitation Learning with Dynamic RCM Calibration for Autonomous Intraocular Foreign Body Removal},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HNLMEH55}},
  note         = {Machine review of arXiv:2508.19191}
}
read the original abstract

Intraocular foreign body removal demands millimeter-level precision in confined intraocular spaces, yet existing robotic systems predominantly rely on manual teleoperation with steep learning curves. To address the challenges of autonomous manipulation, particularly kinematic uncertainties from variable motion scaling and Remote Center of Motion (RCM) point variation, we propose RCM-ACT, an imitation learning framework for autonomous intraocular foreign body ring manipulation. Our approach integrates RCM dynamic calibration to resolve coordinate system inconsistencies caused by intraocular instrument variation and introduces the RCM-ACT architecture, which combines action chunking transformers with episode-level kinematic realignment. Trained solely on stereo visual data and instrument kinematics from expert demonstrations in an artificial eye model, RCM-ACT successfully completes ring grasping and positioning tasks without explicit depth sensing. Experimental validation demonstrates the successful implementation of end-to-end autonomy under uncalibrated microscopy conditions, achieving a mean 3-D Euclidean grasp deviation of 0.686 mm and 11/20 full-task successes. The results provide a viable framework for developing intelligent eye surgical systems capable of complex intraocular procedures.

Figures

Figures reproduced from arXiv: 2508.19191 by the authors.

Figure 1
Figure 1. Overview of RCM-ACT Framework: Instrument Actuation with Data Acquisition and Multimodal Training [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Diagram of the Deployment Process for the Eye Surgery Robot [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Impact of the Chunk Size achieving near-real-time responsiveness at intermediate con￾figurations. Notably, while smaller chunk sizes (<90) yield the lowest normalized grasping deviation—critical for pre￾cise ring positioning in our ring task—their latency remains suboptimal for dynamic robotic operations. Our chosen con￾figuration (chunk size 90) demonstrates an optimal balance, achieving the highest normalized pred… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 29 canonical work pages

  1. [1]

    Removal of intraocular foreign bodies,

    W. J. Wirostko, S. Bhatia, W. F. Mieler, and C. M. McCabe, “Removal of intraocular foreign bodies,” in Vitreoretinal Surgical Techniques, Second Edition. Routledge, 2019, pp. 492–504

    work page 2019

  2. [2]

    Clinical outcomes and characterization of intraocular foreign body injuries from a canadian centre: a 20-year retrospective study and literature review,

    A. Al-Ani, M. S. Bondok, K. Madjedi, S. Kherani, and A. Kherani, “Clinical outcomes and characterization of intraocular foreign body injuries from a canadian centre: a 20-year retrospective study and literature review,” Canadian journal of ophthalmology , 2024

    work page 2024

  3. [3]

    Clinical and imaging characteristics, outcomes and prognostic factors of intraocular foreign bodies extracted by vitrectomy,

    X. Liu, Q. Bai, and X. Song, “Clinical and imaging characteristics, outcomes and prognostic factors of intraocular foreign bodies extracted by vitrectomy,” Scientific Reports, vol. 13, no. 1, p. 14136, 2023

    work page 2023

  4. [4]

    Intraocular foreign body: diagnostic protocols and treatment strategies in ocular trauma patients,

    H. C. Jung, S. Y . Lee, C. K. Yoon, U. C. Park, J. W. Heo, and E. K. Lee, “Intraocular foreign body: diagnostic protocols and treatment strategies in ocular trauma patients,” Journal of clinical medicine , vol. 10, no. 9, p. 1861, 2021

    work page 2021

  5. [5]

    Overcoming the impact of physiologic tremors in ophthalmology,

    G. Singh, W. W. J. Jie, M. T. Sun, R. Casson, D. Selva, and W. Chan, “Overcoming the impact of physiologic tremors in ophthalmology,” Graefe’s Archive for Clinical and Experimental Ophthalmology , vol. 260, no. 12, pp. 3723–3736, 2022

    work page 2022

  6. [6]

    Robotics Applications in Neurology: A Review of Recent Advancements and Future Directions

    R. Retnaningsih, A. Budiyono, R. Ismail, D. Tugasworo, R. Danuaji, S. Syahrul, and H. Gunawan, “Robotics applications in neurology: a review of recent advancements and future directions,” arXiv preprint arXiv:2312.06956, 2023

    work page Pith review arXiv 2023

  7. [7]

    Robot-assisted eye surgery: a systematic review of effectiveness, safety, and practicality in clinical settings,

    A. J. Thirunavukarasu, M. L. Hu, W. P. Foster, K. Xue, J. Cehajic- Kapetanovic, and R. E. MacLaren, “Robot-assisted eye surgery: a systematic review of effectiveness, safety, and practicality in clinical settings,” Translational Vision Science & Technology , vol. 13, no. 6, pp. 20–20, 2024

    work page 2024

  8. [8]

    Real-time operator takeover for visuomotor diffusion policy training,

    N. Ingelhag, J. Munkeby, M. C. Welle, M. Moletta, and D. Kragic, “Real-time operator takeover for visuomotor diffusion policy training,” arXiv preprint arXiv:2502.02308 , 2025

    arXiv 2025

Show all 34 references
  1. [9]

    Deri-igp: Learning to manipulate rigid objects using deformable linear objects via iterative grasp-pull,

    Z. Wang and A. H. Qureshi, “Deri-igp: Learning to manipulate rigid objects using deformable linear objects via iterative grasp-pull,” IEEE Robotics and Automation Letters , 2025

    2025

  2. [10]

    Multi-critic reinforcement learning for garment handling: Addressing unpredictability in temporal-phase continuous contact tasks,

    Y . Zhang, D. Chen, W. He, A. E. P. Barcel ´o, J. V . S. Luces, and Y . Hirata, “Multi-critic reinforcement learning for garment handling: Addressing unpredictability in temporal-phase continuous contact tasks,” IEEE Transactions on Automation Science and Engineering , 2025

    2025

  3. [11]

    Autonomous teleoperated robotic arm based on imitation learning using instance segmentation and haptics information,

    K. Imai, Y . Takahashi, S. Tsuichihara, and M. Haruna, “Autonomous teleoperated robotic arm based on imitation learning using instance segmentation and haptics information,” Journal of Advanced Compu- tational Intelligence and Intelligent Informatics , vol. 29, no. 1, pp. 79–94, 2025

    2025

  4. [12]

    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,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 606–612

    2024

  5. [13]

    Motion planning and control of active robot in orthopedic surgery by cdmp-based imitation learning and constrained optimization,

    X. Jian, Y . Song, D. Liu, Y . Wang, X. Guo, B. Wu, and N. Zhang, “Motion planning and control of active robot in orthopedic surgery by cdmp-based imitation learning and constrained optimization,” IEEE Transactions on Automation Science and Engineering , 2025

    2025

  6. [14]

    Evalu- ation of (shared) autonomy in robot-assisted vitreoretinal surgery using a surgical model,

    M. M. Marinho, Y . Koyama, Y . Taniguchi, T. Yamanaka, F. Arai, S. Omata, K. Sugimoto, T. Ueta, K. Totsuka, T. Shiraya, et al., “Evalu- ation of (shared) autonomy in robot-assisted vitreoretinal surgery using a surgical model,” The International Journal of Medical Robotics and...

    2025

  7. [15]

    Towards motion compensation in autonomous robotic subretinal injections,

    D. Arikan, P. Zhang, M. Sommersperger, S. Dehghani, M. Esfandiari, R. H. Taylor, M. A. Nasseri, P. Gehlbach, N. Navab, and I. Iordachita, “Towards motion compensation in autonomous robotic subretinal injections,” arXiv preprint arXiv:2411.18521 , 2024

    2024 arXiv

  8. [16]

    To- wards autonomous eye surgery by combining deep imitation learning with optimal control,

    J. W. Kim, P. Zhang, P. Gehlbach, I. Iordachita, and M. Kobilarov, “To- wards autonomous eye surgery by combining deep imitation learning with optimal control,” in Conference on Robot Learning . PMLR, 2021, pp. 2347–2358

    2021

  9. [17]

    Intraoperative adaptive eye model based on instrument-integrated oct for robot- assisted vitreoretinal surgery,

    M. Briel, L. Haide, M. Hess, J. Schimmelpfennig, P. Matten, R. Peter, M. Hillenbrand, E. Tagliabue, and F. Mathis-Ullrich, “Intraoperative adaptive eye model based on instrument-integrated oct for robot- assisted vitreoretinal surgery,” International Journal of Computer Assist...

    2025

  10. [18]

    Robot-assisted intraocular surgery: development of the iriss and feasibility studies in an animal model,

    E. Rahimy, J. Wilson, T. Tsao, S. Schwartz, and J. Hubschman, “Robot-assisted intraocular surgery: development of the iriss and feasibility studies in an animal model,” Eye, vol. 27, no. 8, pp. 972– 978, 2013

    2013

  11. [19]

    Simultaneous online registration-independent stiffness identification and tip localization of surgical instruments in robot-assisted eye surgery,

    A. Ebrahimi, S. Sefati, P. Gehlbach, R. H. Taylor, and I. I. Iordachita, “Simultaneous online registration-independent stiffness identification and tip localization of surgical instruments in robot-assisted eye surgery,” IEEE Transactions on Robotics , vol. 39, no. 2, pp. 1373...

    2022

  12. [20]

    Optical coherence tomography-guided robotic ophthalmic microsurgery via reinforcement learning from demonstration,

    B. Keller, M. Draelos, K. Zhou, R. Qian, A. N. Kuo, G. Konidaris, K. Hauser, and J. A. Izatt, “Optical coherence tomography-guided robotic ophthalmic microsurgery via reinforcement learning from demonstration,” IEEE Transactions on Robotics , vol. 36, no. 4, pp. 1207–1218, 2020

    2020

  13. [21]

    Robot-assisted epiretinal membrane peeling: a prospective assessment of pre-and intra-operative times and of sur- geons’ subjective perceptions,

    F. Turgut, G. M. Somfai, F. M. Heussen, A. Eberle, M. D. de Smet, and M. D. Becker, “Robot-assisted epiretinal membrane peeling: a prospective assessment of pre-and intra-operative times and of sur- geons’ subjective perceptions,” Journal of clinical medicine , vol. 12, no. 8,...

    2023

  14. [22]

    Towards robotic- assisted subretinal injection: A hybrid parallel–serial robot system design and preliminary evaluation,

    M. Zhou, Q. Yu, K. Huang, S. Mahov, A. Eslami, M. Maier, C. P. Lohmann, N. Navab, D. Zapp, A. Knoll, et al. , “Towards robotic- assisted subretinal injection: A hybrid parallel–serial robot system design and preliminary evaluation,” IEEE Transactions on Industrial Electronics,...

    2019

  15. [23]

    Needle detection and localisation for robot-assisted subretinal injection using deep learning,

    M. Zhou, X. Guo, M. Grimm, E. Lochner, Z. Jiang, A. Eslami, J. Ye, N. Navab, A. Knoll, and M. A. Nasseri, “Needle detection and localisation for robot-assisted subretinal injection using deep learning,” CAAI Transactions on Intelligence Technology , 2023

    2023

  16. [24]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929 , 2020

    2010 arXiv

  17. [25]

    Design of adaptive robot control system using recurrent neural network,

    S. Yildirim, “Design of adaptive robot control system using recurrent neural network,” Journal of Intelligent and Robotic Systems , vol. 44, pp. 247–261, 2005

    2005

  18. [26]

    Transformer-based deep imitation learning for dual-arm robot manipulation,

    H. Kim, Y . Ohmura, and Y . Kuniyoshi, “Transformer-based deep imitation learning for dual-arm robot manipulation,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2021, pp. 8965–8972

    2021

  19. [27]

    When transformer meets robotic grasping: Exploits context for efficient grasp detection,

    S. Wang, Z. Zhou, and Z. Kan, “When transformer meets robotic grasping: Exploits context for efficient grasp detection,” IEEE robotics and automation letters , vol. 7, no. 3, pp. 8170–8177, 2022

    2022

  20. [28]

    Gen2sim: Scaling up robot learning in simulation with generative models,

    P. Katara, Z. Xian, and K. Fragkiadaki, “Gen2sim: Scaling up robot learning in simulation with generative models,” in 2024 IEEE Interna- tional Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 6672–6679

    2024

  21. [29]

    Trajectory prediction with latent belief energy-based model,

    B. Pang, T. Zhao, X. Xie, and Y . N. Wu, “Trajectory prediction with latent belief energy-based model,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 11 814–11 824

    2021

  22. [30]

    Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning,

    Y . Wang, Y . Zhang, M. Huo, R. Tian, X. Zhang, Y . Xie, C. Xu, P. Ji, W. Zhan, M. Ding, et al. , “Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning,” arXiv preprint arXiv:2407.01531, 2024

    2024 arXiv

  23. [31]

    Semantically controllable augmentations for generalizable robot learning,

    Z. Chen, Z. Mandi, H. Bharadhwaj, M. Sharma, S. Song, A. Gupta, and V . Kumar, “Semantically controllable augmentations for generalizable robot learning,” The International Journal of Robotics Research , p. 02783649241273686, 2024

    2024

  24. [32]

    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,” arXiv preprint arXiv:2304.13705, 2023

    2023 arXiv

  25. [33]

    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,” arXiv preprint arXiv:2407.12998 , 2024

    2024 arXiv

  26. [34]

    A low cost surrogate eye model for corneal foreign body removal,

    J. Lin and M. T. Chua, “A low cost surrogate eye model for corneal foreign body removal,” BMC ophthalmology, vol. 20, pp. 1–7, 2020

    2020

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.