acceptodds
Under review as a conference paper at ICLR 2027

Robotic Optical Coherence Tomography-Guided Geometric Modeling of Soft Tissue: 3D Representations for Data-Driven Closed-Loop Shape Control

Abstract

In medical robotics, data-driven soft-tissue shape control requires deforming tissue toward desired morphologies through sequential physical interactions. This challenge is particularly important in microscale manipulation, where localized deformation is difficult to observe, represent, and control, while physics-based models can be computationally expensive and difficult to adapt to realistic environments. Optical coherence tomography (OCT) provides high-resolution volumetric perception of tissue deformation, but target-conditioned control requires effective representations of both observed and desired tissue states. In this work, we introduce an OCT-guided data-driven framework for tissue shape control, learned from user-guided demonstrations and executed with closed-loop OCT feedback. We study full-volume, tissue-masked sparse-volume, and voxelized point-cloud representations, with or without residual conditioning, using deterministic (Explicit) and generative diffusion (Diffusion) policies. Across 12 configurations and 576 phantom rollouts on 3-DoF depth and 6-DoF directional tasks, OCT-derived 3D representations support both Explicit and Diffusion policies, with performance depending on representation and residual conditioning. Residual conditioning reduces directional Shape RMSE from 0.748 to 0.555 mm for full-volume Diffusion, while voxelized point-cloud Explicit with residuals achieves 0.534 mm for depth and 0.443 mm for directional control, with comparable accuracy at intermediate unseen targets and consistent performance across deployments. We further evaluate the framework in 60 porcine rollouts, extending to biological tissue with final geometric errors comparable to human demonstrations. These results establish volumetric OCT as current and goal representations for learned closed-loop tissue shape control across deterministic and generative action models.

Primary area:
applications to robotics, autonomy, planning
Keywords:
medical robotics, optical coherence tomography, deformable tissue manipulation, imitation learning, goal-conditioned control, 3D representations, diffusion policies

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zY4aOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents a data-driven, goal-conditioned robotic manipulation framework for closed-loop soft-tissue shape control guided by volumetric Optical Coherence Tomography (OCT) feedback. The authors introduce a parameterized inverse-warping target generation pipeline that constructs specimen-specific 3D target morphologies directly from pre-deformed OCT scans, and systematically compare three 3D representations, dense full-intensity volumes, tissue-masked sparse volumes, and geometry-only voxelized point clouds, paired with representation-specific current-to-target residual conditioning across both deterministic (Explicit) and generative (Diffusion) action-chunking policies. The proposed system is thoroughly evaluated across 576 physical silicone phantom rollouts (spanning 3-DoF depth and 6-DoF directional indentation tasks, including unseen intermediate targets) and 60 ex vivo porcine tissue rollouts.

Critical strengths
  1. This paper presents a well-posed new problem with substantial physical evaluation. Treating volumetric OCT tissue morphology as the controlled state (rather than a disturbance during navigation) is a genuine reframing. It is backed by 151 demonstrations and 576 closed-loop phantom rollouts across 12 configurations plus 60 ex vivo rollouts, with seen and unseen targets. Real-robot evidence at this scale is rare in medical robotics, and the phantom conclusions survive scrutiny.

  2. The work has representation-specific residual conditioning that demonstrably matters. The two-channel sparse residual and the voxelwise full-volume residual recover target-dependent directional control that the unconditioned variants lack (e.g., full-volume Explicit 0.746 → 0.488 mm, 14/27 → 27/27 improved). This is the paper's clearest mechanistic finding.

Critical weaknesses
  1. The ex vivo claim is not supported by the primary metric. Symmetric Shape RMSE and HD95 worsen relative to the initial state in 100% of porcine depth rollouts for both learned policies and the human operators (e.g., 0.592 → 1.220 mm). The 3 mm Gaussian target does not model the tooltip footprint or volume-conserving tissue, so the target is unreachable. Framing these results as "comparable to human demonstrations" conceals that doing nothing would have scored better on the headline metric.

  2. Statistical and architectural confounds limit the representation conclusions. All configurations use a single training seed. Reported SDs reflect three deployment sweeps of one checkpoint (df = 2), and the top six depth configurations overlap heavily. Representations differ in token capacity (64 spatial tokens vs. one pooled token), pooling (mean vs. max), and encoder sharing, and the shared heuristic support mask both constructs the sparse/point-cloud inputs and defines the evaluation ground truth. The ex vivo study tests only two configurations that vary representation and decoder simultaneously, so the observed ranking reversal cannot be attributed.

Questions for authors
  1. Can the authors report a top-surface-only error (before the ~300-voxel interior fill) to separate surface control from lower-boundary attenuation artifacts, and a target-agnostic baseline (constant 3.5 mm indentation) for the directional task?
  2. Why was w = 3 mm chosen for porcine tissue, and would re-evaluating against a width calibrated to the empirical tooltip footprint change the sign of the ex vivo result?
  3. What is the end-to-end control-cycle time (OCT acquisition, mask extraction, inference, robot motion) for each representation?
  4. What causes the ~15× latency increase for the sparse residual model (35 → 518 ms): implementation or algorithmic?
  5. Can multi-seed offline validation MSE and target-paired statistical tests be provided to substantiate the comparative claims?
Reviewer_acMxOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies OCT-guided robotic soft-tissue shape control using policies learned from demonstrations. Target morphologies are constructed by applying parameterized deformations to measured OCT scans. The framework compares full volumes, tissue-masked sparse volumes, and voxelized point clouds, with optional current–target residual conditioning and either deterministic or diffusion-based action prediction. The robot executes incremental tool motions and acquires new OCT observations for feedback. Evaluation includes physical phantom experiments, intermediate target parameters absent from training, and separately trained policies on ex vivo porcine tissue. I have the following comments.

Critical strengths

1、A complete physical system for OCT-guided tissue shape control. The framework connects target construction, volumetric observation, learned action prediction, and robotic execution within a closed loop. Representing both current and desired tissue morphology in the same OCT space provides a concrete way to learn shape-directed manipulation from demonstrations. The integration and physical implementation constitute a useful systems contribution.

2、A systematic experimental study with informative comparisons. The 12 configurations cover three representations, two action models, and residual conditioning across 576 phantom rollouts. Evaluation includes intermediate targets absent from training and repeated deployments after repositioning. The four point-cloud configurations reduce Shape RMSE in 187 of 192 trials, while the comparisons identify where residual conditioning helps and where it does not. These results provide useful guidance for designing OCT-based manipulation policies and support acceptance within the tested setting.

Critical weaknesses

1、The ex vivo results do not yet establish reliable target-shape control. Table 13 shows that Shape RMSE increases in all 30 depth rollouts, from 0.650 to 1.102 mm for Full-volume Diffusion and from 0.592 to 1.220 mm for Point-cloud Explicit. Only 12 of 30 directional rollouts improve this metric. Comparable human performance does not resolve the issue, since human depth demonstrations also increase the same error (Table 10). The directed metrics show partial alignment, but the paper should clarify whether the remaining mismatch arises from physically unattainable targets, controller limitations, or the evaluation objective. This limits the biological-tissue conclusion, although the phantom results remain a substantive contribution.

Questions for authors

1、Are the parameterized porcine targets physically attainable with the available tool and contact configuration? How should the consistent increase in depth Shape RMSE be interpreted relative to the intended control objective?

2、Are results available for a policy that receives the initial OCT scan, target, and pose history but no updated OCT observations? How much does online visual feedback contribute to the final shape accuracy?

3、How many independent porcine specimens were used, and were demonstration collection and evaluation separated by specimen? This would help distinguish interpolation within familiar tissue conditions from transfer to new tissue.

Reviewer_VgLzOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents substantial physical experimentation and results using optical coherence tomography (OCT).

The study compares three input pipelines:

  1. Dense OCT intensity volumes.
  2. Tissue-masked sparse intensity volumes.
  3. Voxelized point clouds sampled from the tissue support.

The strongest supported finding is that compact geometric representations enable useful control within the tested phantom target families. My main concerns are that the biological-tissue experiments generally worsen the stated primary shape metric, the representation comparisons contain substantial architectural confounds, and the contribution of live volumetric feedback is not isolated from simpler alternatives.

Critical strengths
  1. The study goes beyond offline action prediction and provides substantial evaluation on physical hardware.
  2. The reporting is informative and also includes the negative results. It also reports a useful result favoring a relatively simple policy.
  3. Data separation is reasonable and has clearly defined interpolation.
Critical weaknesses
  1. The biological tissue results do not demonstrate successful overall shape matching. Table 13 reports the following ex vivo outcomes, and no depth-control policy improves the primary objective.
  2. The twelve configurations provide a useful comparison of practical implementations. However, performance differences may therefore arise from resolution, sample count, encoder architecture, weight sharing, pooling, or capacity. Representation and residual effects are not isolated.
  3. The paper acknowledges its principal predecessors, and the main contribution is the integration and systems contribution. However, algorithmic novelty is limited, and the method needs more insight to be more general.
Questions for authors
  1. Why do all human and policy porcine depth trajectories worsen symmetric Shape RMSE?
  2. How much performance remains without live OCT feedback?
  3. What are the measured full-cycle feedback rate and near-goal stopping behavior?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.