HumynexSurg-1: A Curated Expert Liposuction Dataset
Rhea Huang, David L. Matlock, Laurence Reich
cs.RO · cs.LG
Quick verdict
Porcine liposuction dataset with expert narration, synced force/pressure/video and compiled labels; well built, but its PoC policy loses to trivial baselines.
Summary
HumynexSurg-1, from Humynex Robotics, is a commercially licensed expert demonstration dataset for liposuction: a single master surgeon operates on porcine abdominal wall across three sessions (14 episodes, 42,738 frames, 35.6 minutes at 20 Hz) while narrating every decision. The corpus synchronizes suction pressure (100 Hz), six-axis hand force/torque (~467 Hz), top-down RGB-D video, side video and lavalier audio; WhisperX word-aligned narration is compiled by deterministic rules (lipo v0) into nine frame-level label fields, tagging 95% of 356 utterances and 69.4% of frames. The action channel is a 2-D tracked tool-hand trajectory in the side video, not a measured pose, packaged in LeRobot/GR00T format with per-session QA reports. As a proof of concept, NVIDIA GR00T N1.7 fine-tunes in 36 minutes on one H100; held-out MAE on session 3 is 11.20±0.07 px, dropping to 10.67 px with extra views, and adding session 2 cuts error 18% (13.56→11.15 px). Limitations are substantial: per Appendix A the policy loses to zero-motion and linear-AR baselines on both held-out and in-sample frames; the session-scaling evidence rests on two additions and one test session; data are licensee-only; the tracker method is undisclosed; the force channel is unvalidated; tissue is porcine, not human.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 2
● low 3
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- HumynexSurg-1 fills a verifiable gap: across Open-H's 780-hour, 119-dataset surgical corpus only one dataset carries synchronized force and none covers aesthetic surgery (§1).
- Narrated surgeon speech compiles deterministically into nine frame-level label fields, tagging 95% of 356 utterances and 69.4% of 42,738 frames (§4.3, Table 6).
- NVIDIA GR00T N1.7 fine-tunes on the corpus with no custom code in 36 minutes on one H100 and, per the authors, 'learns the recorded sessions' (§6.1-6.2).
- Scaling probes show session diversity rather than episode count drives gains: adding session 2 lowers held-out error 18% (13.56 to 11.15 px) and adding one session-3 episode lowers error from 10.02 to 9.27 px (§6.3).
- Model-estimated channels (2-D hand track) can later be upgraded to measured sensors (6-DoF pose, validated force, ultrasound, tactile) without changing the data format (Table 2, §7).
Strengths
Verifiable gap, honestly positioned
§1 grounds the contribution in a checkable fact: across Open-H's 780-hour, 119-dataset surgical collection exactly one dataset carries synchronized force and none covers aesthetic surgery; liposuction's out-of-sight cannula makes force/narration the modality a camera-only corpus lacks.
Auditable labeling and QA discipline
§4.3/Table 6 compile WhisperX-aligned narration via deterministic rules into nine frame-level fields (95% of 356 utterances, 69.4% of frames, ~1 tagged maneuver/15 s); §5 ships per-session timing/drift/coherence reports and per-frame validity flags, with tracker observed-rates of 90/78/82% disclosed per session (Table 7).
Baselines and full training curves are published
Pre-registered evaluation and actionable next-capture design
Weaknesses
The PoC policy never beats trivial baselines, and the main text hides it
Scaling conclusions from n=2 on a single test session
Closed artifact, proprietary core method
2-D pixel action channel without metric grounding
Internal consistency loose ends
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
HumynexSurg-1 (arXiv 2609.23885, v0, September 2026) is a commercial, licensee-gated demonstration dataset for liposuction: one expert surgeon — a company co-founder — operating on porcine abdominal wall across three sessions (14 episodes, 42,738 frames, 35.6 minutes at 20 Hz), with synchronized suction pressure (100 Hz), six-axis hand force/torque (~467 Hz), top-down RGB-D, side video and narrated audio. Four central claims: (1) it fills a verifiable gap — across Open-H's 780-hour, 119-dataset surgical collection exactly one dataset carries synchronized force and none covers aesthetic surgery (§1); (2) narration compiles deterministically into a nine-field frame-level label schema covering 95% of 356 utterances and 69.4% of frames (§4.3, Table 6); (3) GR00T N1.7 fine-tunes with no custom code (36 minutes, one H100) and 'learns the recorded sessions' (§6.2); (4) every added session lowers unseen-session error, so the next capture should buy session diversity (§6.3, §7).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Is the HumynexSurg-1 dataset publicly available?
No. The paper states the dataset, its per-session quality reports and its tooling are available only to licensees under agreement; no code or data download is offered, and the hand-tracker method is disclosed only to licensees.
What exactly does HumynexSurg-1 contain?
Three sessions (14 episodes, 42,738 frames, 35.6 minutes at 20 Hz) of a single expert surgeon performing liposuction on porcine abdominal tissue, with synchronized suction pressure (100 Hz), six-axis hand force/torque (~467 Hz), top-down RGB-D video, side video, and word-aligned narration compiled into nine frame-level label fields covering 95% of 356 utterances.
Did the fine-tuned robot policy beat the baselines?
No. In Appendix A (Table 10) the GR00T N1.7 policy scores 11.20 px MAE on the held-out session versus 7.70 px for simply predicting no motion and 7.55 px for a linear autoregressive probe; it is also worse than those baselines on training data. The paper's value claims concern data structure and scaling trends, not current policy quality.
Why was porcine tissue used instead of human tissue?
The authors state porcine abdominal wall is commercially available, ethically simple, and mechanically a mammalian subcutaneous fat system that can be worked at full force; a cadaveric-tissue phase is planned to ground anatomical fidelity, with the current force and pressure profiles as the transfer baseline.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-7 are caption-only in the extracted text; visual claims (tracker overlay quality, scaling-curve shapes, force/hand spectra) cannot be checked.
- All quantitative claims rest on artifacts that are licensee-only; nothing in the paper could be independently verified.
- The plausible explanation for the 44,639-vs-42,738 frame discrepancy (all session frames vs in-episode frames) is inferred; the text does not define it.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.