Pith. sign in

REVIEW 3 major objections 3 minor 42 references

An imitation-learning policy autonomously steers a soft robotic guidewire to a target aneurysm in unseen vascular geometries with 83% success, matching expert clinicians.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 10:26 UTC pith:R56VHZFO

load-bearing objection A credible step for imitation learning in soft endovascular navigation, but the 83% generalization claim is anchored to three hand-picked mazes and no code — worth a referee, not a clean pass. the 3 major comments →

arxiv 2510.09497 v2 pith:R56VHZFO submitted 2025-10-10 cs.RO cs.AI

Toward Autonomous Soft Robotic Endovascular Navigation via Imitation Learning

Noah Barnes , Ji Woong Kim , Lingyun Di , Hannah Qu , Anuruddha Bhattacharjee , Miroslaw Janowski , Dheeraj Gandhi , Bailey Felix
show 6 more authors
Shaopeng Jiang Olivia Young Mark Fuge Ryan D. Sochol Jeremy D. Brown Axel Krieger
This is my paper
classification cs.RO cs.AI
keywords soft robot navigationimitation learningendovascular surgeryaction chunkinggoal conditioningautonomous guidewireaneurysm targetingfluoroscopy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that a transformer-based imitation learning model can teach a soft robotic guidewire to navigate to an aneurysm target in unseen vascular geometries, using only 647 demonstrations collected on a benchtop setup. The model achieved an 83% success rate on completely novel geometries, outperforming diffusion-policy, MLP, and centerline-following baselines, and statistically matching two expert clinicians. The authors argue this is a first step toward autonomous endovascular navigation, preserving the core difficulties of soft-robot control, visual occlusion, and geometric variation in a simplified 2D environment. If this approach transfers to clinical-scale robots, it could reduce procedural times and radiation exposure for patients and operators.

Core claim

The central claim is that an end-to-end imitation learning policy, built on an action-chunking transformer with goal-conditioned distance feature maps, relative motor actions, and learned contrast injection, can generalize to entirely new vascular geometries. On three novel geometries not seen in training, the policy reached the aneurysm in 83% of trials (15/18), compared to 22% for a diffusion policy, 28% for an MLP, and 50% for a centerline-following controller. The same policy reached the target in a patient-derived geometry with 75% success. Ablations demonstrate that each component—recovery demonstrations, learned contrast, distance-based goal map, and relative actions—is necessary for

What carries the argument

The key machinery is a BERT-style transformer encoder–decoder that receives a live simulated fluoroscopy image and a goal-conditioned feature map encoding the x- and y-distance to the aneurysm at every vessel pixel. The policy outputs a chunk of relative motor-position commands (bend and translate) plus a binary contrast-injection flag. The relative action representation handles actuator hysteresis and slippage, while the learned contrast scheduling lets the policy decide when to reveal the otherwise invisible vessels.

Load-bearing premise

The load-bearing premise is that the large-scale 2D benchtop maze, with a 5 mm pneumatic actuator and a camera-based fluoroscopy simulation, preserves the core challenges of real endovascular navigation, so that a policy trained here will transfer to sub-millimeter clinical guidewires in 3D patient anatomy.

What would settle it

Run the trained policy on a small-scale (sub-millimeter) soft robotic guidewire in a 3D, patient-derived vascular phantom with pulsatile flow and deformable walls under real fluoroscopy. If the success rate drops from 83% to near the no-goal ablation level (around 33%), the claim that the benchtop setup preserves the core difficulties would be refuted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the 83% success generalizes beyond the benchtop, autonomous soft guidewire navigation could reduce the need for trial-and-error by clinicians and shorten procedures.
  • The learned contrast-injection policy shows that an agent can manage its own intermittent visual feedback, a principle extendable to other procedures where continuous imaging is harmful or unavailable.
  • The superior performance of dense goal-distance maps over binary goal images suggests that rich goal encoding is critical for generalization to new geometries.
  • The DAgger-style recovery data collection proved essential for out-of-distribution performance, offering a practical recipe for imitation learning in surgical robotics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach may transfer to other endovascular tasks (e.g., thrombectomy) by redefining the goal feature map for a target region or point, though the action space may need extension.
  • The 2D single-bifurcation simplification likely understates the difficulty of 3D tortuous anatomy; a model trained on the benchtop maze may require substantial fine-tuning on 3D phantoms or clinical data.
  • The relative action representation, by mitigating hysteresis, could be a broadly useful design choice for imitation learning on other soft-robot platforms with imprecise actuation.
  • The learned contrast-sensing policy resembles an active-sensing strategy; a similar sensing-action coupling could be applied in other domains where sensors are expensive, toxic, or intermittent.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. This paper presents an imitation learning framework for autonomous navigation of a 5-mm-scale soft robotic guidewire in 2D-printed bifurcated vessel mazes under simulated fluoroscopy. The policy, a transformer-based action-chunking model with goal-conditioned feature maps, relative motor actions, and learned contrast-injection decisions, is trained on 647 demonstrations across 36 mazes and evaluated on three rearranged and three novel mazes. The authors report an 83% success rate (15/18) on novel geometries, outperforming diffusion, MLP, and centerline baselines and roughly matching two clinicians. Ablations examine recovery data, contrast prediction, goal representation, and relative actions. The paper also includes DAgger-style data collection and a failure-mode analysis.

Significance. If the results hold, this would be a useful step toward end-to-end learning for soft endovascular tools, demonstrating that imitation learning with action chunking can handle partial observability and underactuated dynamics on physical hardware. The inclusion of a physical bench-top platform, clinician comparison, ablations, and a DAgger procedure are strengths. However, the generality of the central claim is currently limited by evaluation protocol issues, so the significance is conditional on a revised analysis.

major comments (3)
  1. [III-B, IV (Tables I-II)] The three novel and three rearranged test mazes are 'chosen' but no selection rule (randomization, stratification, or difficulty criterion) is reported. With only 18 trials per condition, the 83% estimate has a wide binomial CI (~59–96% at 95%), and no significance tests are provided for the comparisons in Tables I and II (e.g., ours vs. centerline 50%, or vs. 50% recovery 61%). Because every ablation and baseline is measured on these same three mazes, all relative conclusions are conditional on this non-random sample. The authors should report the maze-selection protocol, per-maze results, and confidence intervals or significance tests, or otherwise justify representativeness.
  2. [IV, 'Chunk size determination'] The chunk size K=50 was selected by evaluating on the rearranged test geometries, and the same rearranged set is then used to report the 89% success rate and to compare ablations. This constitutes hyperparameter tuning on a test set, making the rearranged figures optimistic. The novel-geometry numbers are less affected but still use the selected K. Please either use a validation set (or nested procedure) or explicitly acknowledge that the rearranged performance is a selected maximum.
  3. [Abstract vs. Sections IV-V] The abstract claims 'we extend the policy to achieve 75% success on an unseen patient-derived geometry,' but no such experiment appears in the provided full text. This is a missing result that supports an advertised contribution; please add the experimental setup, results, or remove the claim.
minor comments (3)
  1. [III-C, Eq. (4)] The notation is unclear: if ĉ_s is a binary flag, the max is already binary and no sigmoid is needed; presumably ĉ_s is a logit. Please define the variable as the raw model output before thresholding.
  2. [Fig. 5 and 6 captions] The caption 'Each individual bar represents 18 trials' does not match the box/violin-style plots and should describe what the distributions show.
  3. [General] The project website is cited, but the manuscript does not state whether code, data, or trained model weights will be released; for reproducibility, please add an availability statement.

Circularity Check

1 steps flagged

Chunk-size selection on the rearranged test set makes that column partly selected; the headline novel-geometry claim is independent.

specific steps
  1. fitted input called prediction [Section IV (Experiments and Results), 'Chunk size determination'; Table I]
    "We trained a model with chunk sizes of 25 (1 sec.), 50 (2 sec.), 75 (3 sec.), 100 (4 sec.). We then evaluated the model on the rearranged vascular geometries. Among these models, the chunk size of two seconds proved to be the best model with a success rate of 89% compared to 28%, 61%, and 78% for chunk sizes of 1, 3, and 4 seconds, respectively."

    The 89% rearranged-geometry success rate is the criterion used to select the chunk size; the same three mazes then appear in Table I as the model's reported 'Rearranged geometries' result (89%, 16/18). Thus the rearranged column is not an independent hold-out prediction; it is the value that the hyperparameter choice was optimized against. This is a test-set selection artifact rather than a definitional equivalence, and it does not affect the separate 83% novel-geometry claim, because the novel mazes were held out from chunk-size selection. Still, presenting the selected model's rearranged performance as a result overstates the evidence for that specific comparison.

full rationale

This is an empirical imitation-learning systems paper without a formal derivation whose conclusion is equivalent to its assumptions. The headline 83% success is measured on three novel geometries that were excluded from both training and hyperparameter selection; the DAgger recovery data, while generated from the policy's own failures, is a standard training-data augmentation and does not make the test result circular. Goal conditioning, relative actions, and contrast-injection outputs are inputs or loss terms, not quantities defined by the success metric. Baseline and clinician comparisons are external. The only genuine selection-circularity is the chunk-size choice: it was selected by success on the rearranged set, and the same set is then reported as a result (89% in Table I). This affects only that secondary column; the novel-geometry generalization claim remains independent. Self-citations to prior actuator, teleoperation, and SRT work are descriptive and not load-bearing for the paper's main claim. The lack of a described random-selection rule for the three novel mazes is an internal-validity concern but not a circularity. Overall, the central claim does not reduce to a fit or to a self-citation chain, so the score is 2.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced; the contributions are algorithmic. The main free parameters are hyperparameters, one of which (chunk size) was selected using the rearranged test set. The load-bearing assumptions are the proxy validity of the 2D large-scale bench setup and the simulated fluoroscopy for real endovascular navigation.

free parameters (5)
  • Chunk size K = 50 timesteps (2 sec)
    Chosen by evaluating models with chunk sizes 25/50/75/100 on the rearranged test set; the 50-step model gave 89% vs 28-78% for others. This is a hyperparameter selected on a test set.
  • Contrast positive-class loss weight = 7
    Set 'to reflect the average frequency of contrast injections observed in the demonstration data'; the actual frequency is not reported.
  • Total loss weighting for BCE term = 0.5
    Chosen so that l_BCE reaches the same order of magnitude as l_L1 after a few epochs; hand-tuned.
  • Goal feature normalizing constant c = scales max possible distance to 1
    Hand-chosen to normalize the distance feature maps to the image's spatial dimensions; not fitted to data.
  • Trial success/timeout threshold = 60 seconds
    Trials are considered successful if the tip enters the aneurysm within 60 seconds; one branch set was also excluded when mean demonstration time exceeded 60 seconds.
axioms (4)
  • domain assumption A single bifurcation projected to 2D is representative of real endovascular navigation.
    Sec. I states 'navigation can be reduced to traversing a single bifurcation at a time, which often lies roughly in a plane after appropriate C-arm positioning.' If false, the bench-top result does not transfer to clinical anatomy.
  • domain assumption Large-scale 3D-printed pneumatic robot dynamics are representative of future small-scale clinical guidewires.
    Sec. I: 'larger-scale prototypes are currently more mechanically robust and easier to track... before extending to the small-scale.' The claim of generalizable soft-robot navigation depends on this transfer.
  • domain assumption Camera-based simulated fluoroscopy approximates real X-ray fluoroscopy with contrast dynamics.
    Sec. III-B describes translating overhead camera images into simulated fluoroscopy via color thresholding and BFS contrast fill; real X-ray physics, vessel motion, and tissue background are not modeled.
  • domain assumption Demonstrations from non-expert experimenters are sufficient to learn a clinically meaningful policy.
    The paper states in the Discussion that future work should collect demonstrations from an expert neurointerventional surgeon, acknowledging that the current demonstrations may not reflect optimal clinical behavior.

pith-pipeline@v1.3.0-alltime-deepseek · 13232 in / 11114 out tokens · 111884 ms · 2026-08-04T10:26:04.272121+00:00 · methodology

0 comments
read the original abstract

In endovascular surgery, endovascular interventionists push a thin tube called a catheter, guided by a thin wire to a treatment site inside the patient's blood vessels to treat various conditions such as blood clots, aneurysms, and malformations. Robotic guidewires can enhance maneuverability but are difficult to model and control. Autonomous soft robotic guidewire navigation has the potential to overcome these challenges, increasing the precision and safety of endovascular navigation. As a first step, we establish a large-scale, 2D-projected environment for autonomous navigation. In other surgical domains, end-to-end imitation learning has shown promising results. Thus, we develop a transformer-based imitation learning framework with goal conditioning, relative action outputs, and automatic contrast dye injections to enable generalizable soft robot navigation in an aneurysm targeting task. We train the policy on 36 different modular bifurcated geometries, generating 647 total demonstrations under simulated fluoroscopy, and evaluate it on three previously unseen vascular geometries. The policy reaches the aneurysm with a success rate of 83% on the unseen geometries, outperforming several baselines. In addition, ablation and baseline studies evaluate the effectiveness of each design and data collection choice. Lastly, we extend the policy to achieve 75% success on an unseen patient-derived geometry. Project website: https://softrobotnavigation.github.io/

Figures

Figures reproduced from arXiv: 2510.09497 by Anuruddha Bhattacharjee, Axel Krieger, Bailey Felix, Dheeraj Gandhi, Hannah Qu, Jeremy D. Brown, Ji Woong Kim, Lingyun Di, Mark Fuge, Miroslaw Janowski, Noah Barnes, Olivia Young, Ryan D. Sochol, Shaopeng Jiang.

Figure 1
Figure 1. Figure 1: (Top) Commercial guidewire and microcatheter for neurovascular intervention next to a small-scale soft robotic guidewire (under development) and our tool. (Bottom) Illustration of a soft robotic guidewire inside the vessels in the Circle of Willis (neurovascular structure). Here, we deploy a 3d-printed soft robotic guidewire in a 2D projection of various vascular geometries. the severe under-actuation and … view at source ↗
Figure 3
Figure 3. Figure 3: First, we reserve a certain set of bifurcations and branches for the [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: To control the soft robotic guidewire, a user inputs force commands [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Proposed architecture for autonomous navigation. The static goal [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Distance to the goal aneurysm boundary at the end of the trial for each [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Representative successful rollout of our policy on a rearranged test [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Distance to the goal aneurysm boundary at the end of the trial for [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Representative examples of each failure mode, as defined by how the [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 5
Figure 5. Figure 5: Despite the similar results, it seems that the clinicians [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 10 linked inside Pith

  1. [1]

    Robotic endovascular surgery: current and future practice,

    L. Cruddas, G. Martin, and C. Riga, “Robotic endovascular surgery: current and future practice,”Seminars in V ascular Surgery, vol. 34, no. 4, pp. 233–240, 2021. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S0895796721000594

    2021

  2. [2]

    Au- tonomous navigation for robot-assisted intraluminal and endovascular procedures: A systematic review,

    A. Pore, Z. Li, D. Dall’Alba, A. Hernansanz, E. De Momi, A. Menciassi, A. Casals Gelp ´ı, J. Dankelman, P. Fiorini, and E. V . Poorten, “Au- tonomous navigation for robot-assisted intraluminal and endovascular procedures: A systematic review,”IEEE Transactions on Robotics, vol. 39, no. 4, pp. 2529–2548, 2023

    2023

  3. [3]

    Endovascular microrobotics for neurointervention,

    D. R. Santiago-Dieppa and J. Friend, “Endovascular microrobotics for neurointervention,”Annual Review of Control, Robotics, and Au- tonomous Systems, vol. 7, 2024

    2024

  4. [4]

    The critical tech- nologies of vascular interventional robotic catheterization: A review,

    B. Ren, Y . Zhao, J. Zhang, H. Li, K. Li, and J. Zhang, “The critical tech- nologies of vascular interventional robotic catheterization: A review,” IEEE Sensors Journal, vol. 23, no. 24, pp. 30 051–30 069, 2023

    2023

  5. [5]

    Soft robotic steerable microcatheter for the endovascular treatment of cerebral disorders,

    T. Gopesh, J. H. Wen, D. Santiago-Dieppa, B. Yan, J. S. Pannell, A. Kha- lessi, A. Norbash, and J. Friend, “Soft robotic steerable microcatheter for the endovascular treatment of cerebral disorders,”Science Robotics, vol. 6, no. 57, p. eabf0601, 2021

    2021

  6. [6]

    An approach for 3d mi- croprinting soft robotic surgical tools at 1.5 french length scales for endovascular interventions,

    B. Felix, O. Young, J. Andreou, N. Portwood, K. Barvenik, N. Barnes, C. R. Weiss, C. Bailey, D. Gandhi, M. Janowski, J. D. Brown, E. Tubaldi, M. Fuge, A. Krieger, and R. D. Sochol, “An approach for 3d mi- croprinting soft robotic surgical tools at 1.5 french length scales for endovascular interventions,” inThe IEEE 7th International Conference on Soft Rob...

    2024

  7. [7]

    Toxicity of mri and ct contrast agents,

    K. M. Hasebroock and N. J. Serkova, “Toxicity of mri and ct contrast agents,”Expert opinion on drug metabolism & toxicology, vol. 5, no. 4, pp. 403–416, 2009

    2009

  8. [8]

    Model-less feedback control of con- tinuum manipulators in constrained environments,

    M. C. Yip and D. B. Camarillo, “Model-less feedback control of con- tinuum manipulators in constrained environments,”IEEE Transactions on Robotics, vol. 30, no. 4, pp. 880–889, 2014

    2014

  9. [9]

    Cosserat rod-based dynamic modeling of a hybrid-actuated soft robot for robot-assisted cardiac ablation,

    M. Roshanfar, J. Dargahi, and A. Hooshiar, “Cosserat rod-based dynamic modeling of a hybrid-actuated soft robot for robot-assisted cardiac ablation,”Actuators, vol. 13, no. 1, 2024

    2024

  10. [10]

    Vision-based online learning kinematic control for soft robots using local gaussian process regression,

    G. Fang, X. Wang, K. Wang, K.-H. Lee, J. D. L. Ho, H.-C. Fu, D. K. C. Fu, and K.-W. Kwok, “Vision-based online learning kinematic control for soft robots using local gaussian process regression,”IEEE Robotics and Automation Letters, vol. 4, no. 2, pp. 1194–1201, 2019

    2019

  11. [11]

    Deep- learning-based compliant motion control of a pneumatically-driven robotic catheter,

    D. Wu, X. T. Ha, Y . Zhang, M. Ourak, G. Borghesan, K. Niu, F. Trauzettel, J. Dankelman, A. Menciassi, and E. V . Poorten, “Deep- learning-based compliant motion control of a pneumatically-driven robotic catheter,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 8853–8860, 2022

    2022

  12. [12]

    Data-driven methods applied to soft robot modeling and control: A review,

    Z. Chen, F. Renda, A. L. Gall, L. Mocellin, M. Bernabei, T. Dangel, G. Ciuti, M. Cianchetti, and C. Stefanini, “Data-driven methods applied to soft robot modeling and control: A review,”IEEE Transactions on Automation Science and Engineering, pp. 1–16, 2024

    2024

  13. [13]

    Automated motion control of the coast robotic guidewire under fluoroscopic guidance,

    S. R. Ravigopal, T. A. Brumfiel, and J. P. Desai, “Automated motion control of the coast robotic guidewire under fluoroscopic guidance,” in 2021 International Symposium on Medical Robotics (ISMR), 2021, pp. 1–7

    2021

  14. [14]

    Autonomous robotic intracardiac catheter navigation using haptic vision,

    G. Fagogenis, M. Mencattelli, Z. Machaidze, B. Rosa, K. Price, F. Wu, V . Weixler, M. Saeed, J. E. Mayer, and P. E. Dupont, “Autonomous robotic intracardiac catheter navigation using haptic vision,”Science Robotics, vol. 4, no. 29, p. eaaw1977, 2019

    2019

  15. [15]

    Vascular centerline- guided autonomous navigation methods for robot-lead endovascular interventions,

    N. Li, Y . Wang, H. Cheng, H. Zhao, and H. Ding, “Vascular centerline- guided autonomous navigation methods for robot-lead endovascular interventions,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 11 578–11 584

    2024

  16. [16]

    Benchmarking reinforcement learning algorithms for autonomous mechanical thrombectomy,

    F. Moosa, H. Robertshaw, L. Karstensen, T. C. Booth, and A. Grana- dos, “Benchmarking reinforcement learning algorithms for autonomous mechanical thrombectomy,”International Journal of Computer Assisted Radiology and Surgery, pp. 1–8, 2025

    2025

  17. [17]

    Recurrent neural networks for generalization towards the vessel geometry in autonomous endovascular guidewire navigation in the aortic arch,

    L. Karstensen, J. Ritter, J. Hatzl, F. Ernst, J. Langej ¨urgen, C. Uhl, and F. Mathis-Ullrich, “Recurrent neural networks for generalization towards the vessel geometry in autonomous endovascular guidewire navigation in the aortic arch,”International Journal of Computer Assisted Radiology and Surgery, vol. 18, no. 9, pp. 1735–1744, 2023

    2023

  18. [18]

    Autonomous guidewire navigation in dynamic environments,

    V . Scarponi, F. Lecomte, M. Duprez, F. Nageotte, and S. Cotin, “Autonomous guidewire navigation in dynamic environments,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 6931–6938

    2024

  19. [19]

    Learning-based autonomous navigation, benchmark environments and simulation framework for endovascular interventions,

    L. Karstensen, H. Robertshaw, J. Hatzl, B. Jackson, J. Langej ¨urgen, K. Breininger, C. Uhl, S. M. H. Sadati, T. Booth, C. Bergeles, and F. Mathis-Ullrich, “Learning-based autonomous navigation, benchmark environments and simulation framework for endovascular interventions,”

  20. [20]

    Sim2real learning with domain randomization for autonomous guidewire navigation in robotic-assisted endovascular procedures,

    T. Yao, H. Wang, B. Lu, J. Ge, Z. Pei, M. Kowarschik, L. Sun, L. Seneviratne, and P. Qi, “Sim2real learning with domain randomization for autonomous guidewire navigation in robotic-assisted endovascular procedures,”IEEE Transactions on Automation Science and Engineer- ing, vol. 22, pp. 13 842–13 854, 2025

    2025

  21. [21]

    Collaborative robot-assisted endovascular catheterization with generative adversarial imitation learning,

    W. Chi, G. Dagnino, T. M. Y . Kwok, A. Nguyen, D. Kundrat, M. E. M. K. Abdelaziz, C. Riga, C. Bicknell, and G.-Z. Yang, “Collaborative robot-assisted endovascular catheterization with generative adversarial imitation learning,” in2020 IEEE International Conference on Robotics and Automation (ICRA), 2020, pp. 2414–2420

    2020

  22. [22]

    Surgical gan: Towards real-time path planning for passive flexible tools in endovascular surgeries,

    Y . Zhao, Y . Wang, J. Zhang, X. Liu, Y . Li, S. Guo, X. Yang, and S. Hong, “Surgical gan: Towards real-time path planning for passive flexible tools in endovascular surgeries,”Neurocomputing, vol. 500, pp. 567–580, 2022

    2022

  23. [23]

    Imitation learning for path planning in cardiac percutaneous interven- tions,

    A. Peloso, R. Damiano, X. Zhang, A. Bicchi, E. V otta, and E. De Momi, “Imitation learning for path planning in cardiac percutaneous interven- tions,”IEEE Transactions on Biomedical Engineering, 2025

    2025

  24. [24]

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

    J. W. B. Kim, J.-T. Chen, P. Hansen, L. X. Shi, A. Goldenberg, S. Schmidgall, P. M. Scheikl, A. Deguet, B. M. White, D. R. Tsai, R. J. Cha, J. Jopling, C. Finn, and A. Krieger, “Srt-h: A hierarchical framework for autonomous surgery via language- conditioned imitation learning,”Science Robotics, vol. 10, no. 104, p. eadt5254, 2025. [Online]. Available: ht...

    doi:10.1126/scirobotics.adt5254 2025

  25. [25]

    Rt-sketch: Goal-conditioned imitation learning from hand-drawn sketches,

    P. Sundaresan, Q. Vuong, J. Gu, P. Xu, T. Xiao, S. Kirmani, T. Yu, M. Stark, A. Jain, K. Hausman, D. Sadigh, J. Bohg, and S. Schaal, “Rt-sketch: Goal-conditioned imitation learning from hand-drawn sketches,” inProceedings of The 8th Conference on Robot Learning, ser. Proceedings of Machine Learning Research, P. Agrawal, O. Kroemer, and W. Burgard, Eds., v...

    2025

  26. [26]

    Learning by cheating,

    D. Chen, B. Zhou, V . Koltun, and P. Kr ¨ahenb¨uhl, “Learning by cheating,” 2019. [Online]. Available: https://arxiv.org/abs/1912.12294

    Pith/arXiv arXiv 2019

  27. [27]

    One-piece 3d-printed pneumatic catheter: Dual-segment design with integrated robotics control for endovascular interventions,

    S. Jiang, L. Di, N. Barnes, H. Qu, O. Young, J. D. Brown, R. Sochol, and A. Krieger, “One-piece 3d-printed pneumatic catheter: Dual-segment design with integrated robotics control for endovascular interventions,” in2024 IEEE 7th International Conference on Soft Robotics (RoboSoft), 2024, pp. 832–838

    2024

  28. [28]

    Bedside admittance control of a dual-segment soft robot for catheter-based interventions,

    N. Barnes, S. Jiang, L. Di, H. Qu, M. Janowski, C. I. Berul, A. Colton, O. Young, R. D. Sochol, J. D. Brown, , and A. Krieger, “Bedside admittance control of a dual-segment soft robot for catheter-based interventions,” in2024 46th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), 2024, in pres

    2024

  29. [29]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” 2015. [Online]. Available: https://arxiv.org/abs/1505.04597

    Pith/arXiv arXiv 2015

  30. [30]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. Girshick, P. Doll ´ar, and C. Feichtenhofer, “Sam 2: Segment anything in images and videos,”arXiv preprint arXiv:2408.00714, 2024. [Online]. Available: https://arxiv.org/abs/2408. 00714

    Pith/arXiv arXiv 2024

  31. [31]

    A reduction of imitation learning and structured prediction to no-regret online learning,

    S. Ross, G. J. Gordon, and J. A. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” 2011. [Online]. Available: https://arxiv.org/abs/1011.0686

    Pith/arXiv arXiv 2011

  32. [32]

    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,” 2024. [Online]. Available: https://arxiv.org/ abs/2407.12998

    Pith/arXiv arXiv 2024

  33. [33]

    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,” 2023. [Online]. Available: https://arxiv.org/abs/2304.13705

    Pith/arXiv arXiv 2023

  34. [34]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” 2019. [Online]. Available: https://arxiv.org/abs/1810.04805

    Pith/arXiv arXiv 2019

  35. [35]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” 2015. [Online]. Available: https://arxiv.org/abs/1512.03385

    Pith/arXiv arXiv 2015

  36. [36]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,” in Proceedings of Robotics: Science and Systems (RSS), 2023

    2023

  37. [37]

    Towards autonomous eye surgery by combining deep imitation learning with optimal control,

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

    2021

  38. [38]

    Movement primitive diffusion: Learning gentle robotic manipulation of deformable objects,

    P. M. Scheikl, N. Schreiber, C. Haas, N. Freymuth, G. Neumann, R. Lioutikov, and F. Mathis-Ullrich, “Movement primitive diffusion: Learning gentle robotic manipulation of deformable objects,”IEEE Robotics and Automation Letters, vol. 9, no. 6, p. 5338–5345, June

  39. [39]

    What matters in learning from offline human demonstrations for robot manipulation,

    A. Mandlekar, D. Xu, J. Wong, S. Nasiriany, C. Wang, R. Kulkarni, L. Fei-Fei, S. Savarese, Y . Zhu, and R. Mart ´ın-Mart´ın, “What matters in learning from offline human demonstrations for robot manipulation,”

  40. [40]

    Available: http://dx.doi.org/10.1109/LRA.2024.3382529

    [Online]. Available: http://dx.doi.org/10.1109/LRA.2024.3382529

    arXiv 2024

  41. [2021]

    Available: https://arxiv.org/abs/2108.03298

    [Online]. Available: https://arxiv.org/abs/2108.03298

    Pith/arXiv arXiv

  42. [2024]

    Available: https://arxiv.org/abs/2410.01956

    [Online]. Available: https://arxiv.org/abs/2410.01956

    Pith/arXiv arXiv