- 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
-
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.
-
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
-
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.
-
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
- 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?
- 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?
- What is the end-to-end control-cycle time (OCT acquisition, mask extraction, inference, robot motion) for each representation?
- What causes the ~15× latency increase for the sparse residual model (35 → 518 ms): implementation or algorithmic?
- Can multi-seed offline validation MSE and target-paired statistical tests be provided to substantiate the comparative claims?