Datasets:
image stringlengths 24 24 | frame int64 0 17.4k | type stringclasses 9
values | question stringclasses 85
values | answer stringlengths 1 54 | answer_point listlengths 2 2 ⌀ | answer_format stringclasses 1
value | in_seconds float64 0 8.2 ⌀ |
|---|---|---|---|---|---|---|---|
frames/clip_00/00000.jpg | 0 | count | How many instruments are working in the view? | 2 | null | null | null |
frames/clip_00/00000.jpg | 0 | action | What is the fenestrated grasper doing? | probe | null | null | null |
frames/clip_00/00000.jpg | 0 | jaw_state | Are the jaws of the fenestrated grasper open or closed? | closed | null | null | null |
frames/clip_00/00000.jpg | 0 | energy | Is energy being applied by the fenestrated grasper? | No | null | null | null |
frames/clip_00/00000.jpg | 0 | target | What is the fenestrated grasper acting on? | yellow fatty tissue | null | null | null |
frames/clip_00/00000.jpg | 0 | point | Point to the tip of the fenestrated grasper. | (527, 366) | [
527,
366
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00000.jpg | 0 | next_action | What will the fenestrated grasper do next? | withdraw | null | null | 1 |
frames/clip_00/00000.jpg | 0 | action | What is the dark rod doing? | enter | null | null | null |
frames/clip_00/00000.jpg | 0 | energy | Is energy being applied by the dark rod? | No | null | null | null |
frames/clip_00/00000.jpg | 0 | point | Point to the tip of the dark rod. | (597, 380) | [
597,
380
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00000.jpg | 0 | next_action | What will the dark rod do next? | sweep | null | null | 1.2 |
frames/clip_00/00000.jpg | 0 | tissue_interaction | Which instrument is working on the pale adherent material? | None | null | null | null |
frames/clip_00/00000.jpg | 0 | tissue_interaction | Which instrument is working on the yellow fatty tissue? | the fenestrated grasper | null | null | null |
frames/clip_00/00005.jpg | 5 | count | How many instruments are working in the view? | 2 | null | null | null |
frames/clip_00/00005.jpg | 5 | action | What is the fenestrated grasper doing? | probe | null | null | null |
frames/clip_00/00005.jpg | 5 | jaw_state | Are the jaws of the fenestrated grasper open or closed? | closed | null | null | null |
frames/clip_00/00005.jpg | 5 | energy | Is energy being applied by the fenestrated grasper? | No | null | null | null |
frames/clip_00/00005.jpg | 5 | target | What is the fenestrated grasper acting on? | yellow fatty tissue | null | null | null |
frames/clip_00/00005.jpg | 5 | point | Point to the tip of the fenestrated grasper. | (534, 250) | [
534,
250
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00005.jpg | 5 | next_action | What will the fenestrated grasper do next? | withdraw | null | null | 0.8 |
frames/clip_00/00005.jpg | 5 | action | What is the dark rod doing? | enter | null | null | null |
frames/clip_00/00005.jpg | 5 | energy | Is energy being applied by the dark rod? | No | null | null | null |
frames/clip_00/00005.jpg | 5 | point | Point to the tip of the dark rod. | (604, 352) | [
604,
352
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00005.jpg | 5 | next_action | What will the dark rod do next? | sweep | null | null | 1 |
frames/clip_00/00005.jpg | 5 | tissue_interaction | Which instrument is working on the pale adherent material? | None | null | null | null |
frames/clip_00/00005.jpg | 5 | tissue_interaction | Which instrument is working on the yellow fatty tissue? | the fenestrated grasper | null | null | null |
frames/clip_00/00010.jpg | 10 | count | How many instruments are working in the view? | 2 | null | null | null |
frames/clip_00/00010.jpg | 10 | action | What is the fenestrated grasper doing? | probe | null | null | null |
frames/clip_00/00010.jpg | 10 | jaw_state | Are the jaws of the fenestrated grasper open or closed? | closed | null | null | null |
frames/clip_00/00010.jpg | 10 | energy | Is energy being applied by the fenestrated grasper? | No | null | null | null |
frames/clip_00/00010.jpg | 10 | target | What is the fenestrated grasper acting on? | yellow fatty tissue | null | null | null |
frames/clip_00/00010.jpg | 10 | point | Point to the tip of the fenestrated grasper. | (528, 268) | [
528,
268
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00010.jpg | 10 | next_action | What will the fenestrated grasper do next? | withdraw | null | null | 0.6 |
frames/clip_00/00010.jpg | 10 | action | What is the dark rod doing? | enter | null | null | null |
frames/clip_00/00010.jpg | 10 | energy | Is energy being applied by the dark rod? | No | null | null | null |
frames/clip_00/00010.jpg | 10 | point | Point to the tip of the dark rod. | (598, 314) | [
598,
314
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00010.jpg | 10 | next_action | What will the dark rod do next? | sweep | null | null | 0.8 |
frames/clip_00/00010.jpg | 10 | tissue_interaction | Which instrument is working on the pale adherent material? | None | null | null | null |
frames/clip_00/00010.jpg | 10 | tissue_interaction | Which instrument is working on the yellow fatty tissue? | the fenestrated grasper | null | null | null |
frames/clip_00/00015.jpg | 15 | count | How many instruments are working in the view? | 2 | null | null | null |
frames/clip_00/00015.jpg | 15 | action | What is the fenestrated grasper doing? | probe | null | null | null |
frames/clip_00/00015.jpg | 15 | jaw_state | Are the jaws of the fenestrated grasper open or closed? | closed | null | null | null |
frames/clip_00/00015.jpg | 15 | energy | Is energy being applied by the fenestrated grasper? | No | null | null | null |
frames/clip_00/00015.jpg | 15 | target | What is the fenestrated grasper acting on? | yellow fatty tissue | null | null | null |
frames/clip_00/00015.jpg | 15 | point | Point to the tip of the fenestrated grasper. | (500, 256) | [
500,
256
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00015.jpg | 15 | next_action | What will the fenestrated grasper do next? | withdraw | null | null | 0.4 |
frames/clip_00/00015.jpg | 15 | action | What is the dark rod doing? | enter | null | null | null |
frames/clip_00/00015.jpg | 15 | energy | Is energy being applied by the dark rod? | No | null | null | null |
frames/clip_00/00015.jpg | 15 | point | Point to the tip of the dark rod. | (602, 280) | [
602,
280
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00015.jpg | 15 | next_action | What will the dark rod do next? | sweep | null | null | 0.6 |
frames/clip_00/00015.jpg | 15 | tissue_interaction | Which instrument is working on the pale adherent material? | None | null | null | null |
frames/clip_00/00015.jpg | 15 | tissue_interaction | Which instrument is working on the yellow fatty tissue? | the fenestrated grasper | null | null | null |
frames/clip_00/00020.jpg | 20 | count | How many instruments are working in the view? | 2 | null | null | null |
frames/clip_00/00020.jpg | 20 | action | What is the fenestrated grasper doing? | probe | null | null | null |
frames/clip_00/00020.jpg | 20 | jaw_state | Are the jaws of the fenestrated grasper open or closed? | closed | null | null | null |
frames/clip_00/00020.jpg | 20 | energy | Is energy being applied by the fenestrated grasper? | No | null | null | null |
frames/clip_00/00020.jpg | 20 | target | What is the fenestrated grasper acting on? | yellow fatty tissue | null | null | null |
frames/clip_00/00020.jpg | 20 | point | Point to the tip of the fenestrated grasper. | (484, 265) | [
484,
265
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00020.jpg | 20 | next_action | What will the fenestrated grasper do next? | withdraw | null | null | 0.2 |
frames/clip_00/00020.jpg | 20 | action | What is the dark rod doing? | enter | null | null | null |
frames/clip_00/00020.jpg | 20 | energy | Is energy being applied by the dark rod? | No | null | null | null |
frames/clip_00/00020.jpg | 20 | point | Point to the tip of the dark rod. | (603, 254) | [
603,
254
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00020.jpg | 20 | next_action | What will the dark rod do next? | sweep | null | null | 0.4 |
frames/clip_00/00020.jpg | 20 | tissue_interaction | Which instrument is working on the pale adherent material? | None | null | null | null |
frames/clip_00/00020.jpg | 20 | tissue_interaction | Which instrument is working on the yellow fatty tissue? | the fenestrated grasper | null | null | null |
frames/clip_00/00025.jpg | 25 | count | How many instruments are working in the view? | 2 | null | null | null |
frames/clip_00/00025.jpg | 25 | action | What is the fenestrated grasper doing? | withdraw | null | null | null |
frames/clip_00/00025.jpg | 25 | jaw_state | Are the jaws of the fenestrated grasper open or closed? | closed | null | null | null |
frames/clip_00/00025.jpg | 25 | energy | Is energy being applied by the fenestrated grasper? | No | null | null | null |
frames/clip_00/00025.jpg | 25 | point | Point to the tip of the fenestrated grasper. | (470, 376) | [
470,
376
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00025.jpg | 25 | next_action | What will the fenestrated grasper do next? | probe | null | null | 6 |
frames/clip_00/00025.jpg | 25 | action | What is the dark rod doing? | enter | null | null | null |
frames/clip_00/00025.jpg | 25 | energy | Is energy being applied by the dark rod? | No | null | null | null |
frames/clip_00/00025.jpg | 25 | point | Point to the tip of the dark rod. | (583, 277) | [
583,
277
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00025.jpg | 25 | next_action | What will the dark rod do next? | sweep | null | null | 0.2 |
frames/clip_00/00025.jpg | 25 | tissue_interaction | Which instrument is working on the pale adherent material? | None | null | null | null |
frames/clip_00/00025.jpg | 25 | tissue_interaction | Which instrument is working on the yellow fatty tissue? | None | null | null | null |
frames/clip_00/00030.jpg | 30 | count | How many instruments are working in the view? | 2 | null | null | null |
frames/clip_00/00030.jpg | 30 | action | What is the fenestrated grasper doing? | withdraw | null | null | null |
frames/clip_00/00030.jpg | 30 | jaw_state | Are the jaws of the fenestrated grasper open or closed? | closed | null | null | null |
frames/clip_00/00030.jpg | 30 | energy | Is energy being applied by the fenestrated grasper? | No | null | null | null |
frames/clip_00/00030.jpg | 30 | point | Point to the tip of the fenestrated grasper. | (486, 470) | [
486,
470
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00030.jpg | 30 | next_action | What will the fenestrated grasper do next? | probe | null | null | 5.8 |
frames/clip_00/00030.jpg | 30 | action | What is the dark rod doing? | sweep | null | null | null |
frames/clip_00/00030.jpg | 30 | energy | Is energy being applied by the dark rod? | No | null | null | null |
frames/clip_00/00030.jpg | 30 | point | Point to the tip of the dark rod. | (575, 320) | [
575,
320
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00030.jpg | 30 | next_action | What will the dark rod do next? | probe | null | null | 1.2 |
frames/clip_00/00030.jpg | 30 | tissue_interaction | Which instrument is working on the pale adherent material? | None | null | null | null |
frames/clip_00/00030.jpg | 30 | tissue_interaction | Which instrument is working on the yellow fatty tissue? | None | null | null | null |
frames/clip_00/00035.jpg | 35 | count | How many instruments are working in the view? | 2 | null | null | null |
frames/clip_00/00035.jpg | 35 | action | What is the fenestrated grasper doing? | withdraw | null | null | null |
frames/clip_00/00035.jpg | 35 | jaw_state | Are the jaws of the fenestrated grasper open or closed? | closed | null | null | null |
frames/clip_00/00035.jpg | 35 | energy | Is energy being applied by the fenestrated grasper? | No | null | null | null |
frames/clip_00/00035.jpg | 35 | point | Point to the tip of the fenestrated grasper. | (482, 535) | [
482,
535
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00035.jpg | 35 | next_action | What will the fenestrated grasper do next? | probe | null | null | 5.6 |
frames/clip_00/00035.jpg | 35 | action | What is the dark rod doing? | sweep | null | null | null |
frames/clip_00/00035.jpg | 35 | energy | Is energy being applied by the dark rod? | No | null | null | null |
frames/clip_00/00035.jpg | 35 | point | Point to the tip of the dark rod. | (561, 361) | [
561,
361
] | (x, y) in 720x576 pixels | null |
frames/clip_00/00035.jpg | 35 | next_action | What will the dark rod do next? | probe | null | null | 1 |
frames/clip_00/00035.jpg | 35 | tissue_interaction | Which instrument is working on the pale adherent material? | None | null | null | null |
Laparoscopic Surgical Video for Robot Learning — Annotated Sample
This dataset is a fully annotated, physician-reviewed laparoscopic surgery video (11 min 57 s, 17,370 frames) prepared for training surgical AI and robot-learning models. Every frame carries SAM 3.1 instrument masks, 2.5D tool-tip motion and scene graphs; every 10-second clip carries action labels, task graphs and surgeon-style instructions, exported as VLA, VQA and reasoning training sets.
Last updated: 17 November 2026 · Physician-reviewed · Licence: Apache-2.0 · Maintainer: hkust-aaron
▶ Watch first: showreel.mp4 — masks, tool-tip trails, live action labels and instructions drawn on the video.
At a glance
| Item | Value |
|---|---|
| Video | 1 laparoscopic recording, 720 × 576, 25 fps; 694.8 s of surgical field |
| Instrument masks | 17,370 frames, 222 tracks (SAM 3.1), masks on 96.6 % of frames |
| Tool kinematics | tip, shaft angle, speed, relative depth and camera motion for 31,313 instrument-frames |
| Action labels | 771 segments across all 70 clips; 20-verb vocabulary; jaw state, energy, target, instruction |
| Task structure | 178 subtasks with preconditions and effects; 58 decision points with outcomes |
| Tissue masks | 118 target-tissue tracks in 68 clips; 79 accepted, 14,150 reliable frame masks |
| Training sets | 38,166 VQA pairs · 3,474 VLA steps · 17,370 scene graphs · 7,348 action tokens · 128 reasoning · 623 next-action |
| Clinical review | all layers reviewed and approved by the contributing physician (17 Nov 2026) |
| Pipeline | 20 reproducible scripts; SHA-256 for every file |
Quick start
from datasets import load_dataset
REPO = "hkust-aaron/laparoscopic-robotic-vision-sample"
vqa = load_dataset(REPO, "vqa", split="train") # 38,166 grounded question–answer pairs
vla = load_dataset(REPO, "vla_steps", split="train") # image + state + action chunk + instruction
print(vqa[0]["question"], "->", vqa[0]["answer"])
Other subsets: reasoning, next_action, action_tokens, scene_graphs, clip_captions. Records reference images by path (frames/clip_XX/<frame>.jpg); fetch them with:
from huggingface_hub import snapshot_download
snapshot_download(REPO, repo_type="dataset", allow_patterns=["annotated/layer6_robot_learning/frames/*"])
What the dataset contains
Each layer builds on the previous one, from what is in view to what the instruments do and why.
| Layer | Path | Contents | Coverage |
|---|---|---|---|
| 1 Quality | annotated/layer1_quality/ |
per-frame sharpness, brightness, glare, motion; 683 of 694 s usable | whole video |
| 2 Timeline | annotated/layer2_timeline/ |
12 scene segments, 6 events | whole video |
| 3 Captions | annotated/layer3_clip_captions/ |
70 ten-second clips with English captions (video–text) | whole video |
| Keyframes | annotated/gold_subset/ |
81 frames, 200 instances, COCO format | 81 frames |
| Dense tracks | annotated/dense_tracks/ |
instance masks on every frame (PNG, pixel = track id) | 17,370 frames |
| 4 Kinematics | annotated/layer4_tool_kinematics/ |
tip, shaft entry, angle, speed, camera motion, camera-compensated motion, relative depth | 17,370 frames |
| 5 Actions | annotated/layer5_actions/ |
action segments, task graphs, decision points, tissue masks, timelines, showcase videos | whole video |
| 6 Robot learning | annotated/layer6_robot_learning/ |
VLA steps, VQA, reasoning, next-action, action tokens, scene graphs, frames | whole video |
Example annotation
A single action segment from clip 43 (peeling adherent material):
{"track_id": 102, "start_frame": 10813, "end_frame": 10870, "verb": "peel",
"target": "pale adherent material [202] off the pink surface [201]",
"jaw_state": "closed", "energy_visible": false,
"instruction": "Keep pulling to the left so the material comes off the surface."}
And the action token for one 0.2 s step of an instrument: <dx+2><dy-1><dz0><rot0><jaw:close><e0> — tip motion with camera motion removed, depth change, rotation, jaw command, energy.
How it was built
- Instrument segmentation — AI-drafted keyframe polygons refined with Meta SAM 3.1, then propagated through every 10 s clip (text prompt
"metal tool"plus keyframe clicks). - Kinematics — tool tip and shaft from each mask; laparoscope motion from background optical flow (RANSAC similarity); relative depth from Depth Anything V2 Small.
- Actions and reasoning — every clip labelled from contact sheets (0.2–0.4 s sampling) with a fixed verb vocabulary, symbolic pre/effects and decision points; validated automatically for vocabulary, frame ranges and overlaps.
- Target tissue — SAM 3.1 tracks from clicks placed just ahead of the acting instrument's tip; per-frame quality rules flag masks that drift onto instruments or spread over the field.
- Exports — scene graphs, tokens and VLA / VQA / reasoning / next-action records generated from the layers above.
Internal consistency: steps labelled position, hold and grasp move at a median 54–95 px/s; transport, probe, counter_traction and peel at 190–325 px/s — hand-written labels and independently measured motion agree. All layers were reviewed and approved by the contributing physician on 17 November 2026 (CLINICAL_REVIEW.json); the pipeline's own needs_review, reliability and status fields remain available as filters. Full methods: annotated/DATACARD.md.
How it relates to other surgical datasets
Public surgical datasets usually provide one or two label types over many cases. This sample shows every layer at once on a single case, in training-ready formats.
| Dataset | Content | Labels |
|---|---|---|
| Cholec80 | 80 laparoscopic cholecystectomy videos | surgical phase, tool presence |
| CholecT50 | 50 laparoscopic cholecystectomy videos | instrument–verb–target action triplets |
| CholecSeg8k | 8,080 frames from Cholec80 | semantic segmentation |
| EndoVis robotic instrument segmentation (2017/2018) | da Vinci robotic surgery frames | instrument masks |
| JIGSAWS | benchtop da Vinci tasks (suturing, knot tying, needle passing) | robot kinematics, gestures, skill scores |
| This dataset | 1 laparoscopic recording, every frame | instance masks, 2.5D tool motion, camera motion, depth, action segments, task graphs, reasoning, scene graphs, VLA / VQA exports |
FAQ
What is this dataset for? Training and evaluating models that understand surgical video at the level a robot needs: where instruments are, how they move, what action they perform on which tissue, and what should happen next. It suits vision-language-action (VLA) pre-training, surgical VQA, action recognition, video–language models and surgical world models.
Can it be used to train a model that controls a surgical robot?
It provides the perception, action-semantics and planning layers of that stack. Policies that output robot commands additionally need synchronised robot kinematics; ROBOT_LEARNING_DATA_SPEC.md describes the full data stack and how video archives and robot demonstrations fit together.
How were the labels created?
With Meta SAM 3.1 for segmentation and tissue tracking, Depth Anything V2 for depth, optical flow for camera motion, and an AI annotator for actions, task graphs and reasoning. Every step is reproducible from annotated/scripts/. All layers were then reviewed and approved by the contributing physician (17 November 2026).
What kind of surgery is shown? Laparoscopic (minimally invasive) abdominal surgery with hand-held instruments: insulated graspers, a fenestrated grasper, a rod-shaped instrument and gauze; scenes include gauze tamponade, grasping, coagulation, blunt dissection and peeling of adherent material.
Has a clinician reviewed the labels?
Yes. All annotation layers were reviewed and approved by the contributing physician on 17 November 2026, with no corrections requested; the record is in CLINICAL_REVIEW.json.
Can I use it commercially? Yes. The video and labels are released under Apache-2.0; the contributor confirmed permission for public distribution and commercial AI training.
Is more data available? Yes. This sample comes from a collection of approximately 50,000 hours of surgical video; the same pipeline extends across it with physician-reviewed annotation and custom label vocabularies. Contact aaron@aaronxu.com.
Toward surgical robot foundation models
annotated/ROBOT_LEARNING_DATA_SPEC.md sets out the data a model that plans, acts and monitors in surgery needs — from large-scale video pre-training through dense perception and action semantics to synchronised robot demonstrations, corrections and outcomes.
Scale with us
The wider collection (~50,000 hours) supports pilot deliveries with physician-reviewed labels, procedure-specific ontologies and formats matched to your models. Availability, scope and licensing are agreed separately.
Contact: aaron@aaronxu.com — dataset inquiries, annotated pilots and collaboration.
Citation
@misc{laparoscopic_robot_learning_sample_2026,
title = {Laparoscopic Surgical Video for Robot Learning -- Annotated Sample},
author = {hkust-aaron},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/hkust-aaron/laparoscopic-robotic-vision-sample}},
note = {Instance masks, tool kinematics, action labels, task graphs and VLA/VQA exports for a laparoscopic surgery video}
}
Licence and tools
- Video and labels: Apache-2.0.
- Original recording:
videos/laparoscopic_sample_001.mp4(also at the repository root under its original file name); training cut:annotated/video/laparoscopic_sample_001_intracorporeal.mp4(surgical field only, frame-exact). - Models used: Meta SAM 3.1 (
sam3.1_multiplex.pt, SAM License) and Depth Anything V2 Small (Apache-2.0). No model weights are included.
- Downloads last month
- 154