IMERSE-HF/Phantom_dataset_for_cotraining

Phantom Dataset for Co-training

This dataset contains episode-based image recordings and robot-state logs organized into two task collections: needle_task and chole_task. Each episode stores JPEG images from multiple named streams and an ee_csv.csv file containing timestamps, poses, joint values, and other recorded signals. The collection is intended to support research on imitation learning and co-training across tasks.

• Repository: IMERSE-HF/Phantom_dataset_for_cotraining
• Maintainer: Xinhao Chen; contact: xchen254@jhu.edu
• Related paper / project / training code: SurgVIL: Scaling Surgical Robot Imitation Learning with Open-source Surgical Videos
• Release version: V1

Dataset contents

Collection	Episode folders	JPEG images	CSV files	Size (GB)
needle_task	1,016	249,529	1,016	20.92
chole_task	713	105,061	713	11.63
Total	1,729	354,590	1,729	32.55

Episode counts refer to the stored episode folders. Image counts refer to JPEG files, not synchronized timesteps or independent demonstrations. Dataset size is the sum of file sizes in the two task folders. Root metadata files are excluded.

Collection	Task description
needle_task	Needle task recordings. Pick up a suturing needle in different poses with Large needle driver.
chole_task	Chole task recordings. Go to the cutting pose and ends before closing the Curved scissors.

Folder structure

Episode folders sit directly under their task folder. For example:

Phantom_dataset_for_cotraining/
├── README.md
├── .gitattributes
├── needle_task/
│   ├── 110521-547002/
│   │   ├── ee_csv.csv
│   │   ├── endo_psm1/
│   │   │   └── frame1766246721514829874_psm1.jpg
│   │   ├── left_img_dir/
│   │   │   └── frame1766246721520370681_left.jpg
│   │   └── right_img_dir/
│   │       └── frame1766246721521398671_right.jpg
│   └── ...
└── chole_task/
    ├── 101620-823564/
    │   ├── ee_csv.csv
    │   ├── endo_psm1/
    │   │   └── 1790345780824902840_psm1.jpg
    │   ├── left_img_dir/
    │   │   └── 1790345780819142181_left.jpg
    │   └── right_img_dir/
    │       └── 1790345780819279530_right.jpg
    └── ...

All 1,729 episode folders in this revision contain ee_csv.csv and images under endo_psm1, left_img_dir, and right_img_dir. There are no endo_psm2 image folders in this revision.

Images

Folder	Filename suffix	Dimensions observed in sampled files
endo_psm1	_psm1.jpg	640 × 480 pixels
left_img_dir	_left.jpg	960 × 540 pixels
right_img_dir	_right.jpg	960 × 540 pixels

The endo_psm1 contains the wrist camera view, while the left_img_dir and right_img_dir are the paired stereo endoscopic camera views.

Robot-state CSV files

Every episode has an ee_csv.csv file. Read its header by name.

Column or naming pattern	Observed contents / documentation needed
timestamp	Long integer timestamp values.
camera_source	Present in some episodes; observed values include left, right, psm1, and psm2. Other sampled needle CSVs omit this column.
psm1_pose.*, psm2_pose.*	Position components x, y, z and orientation components x, y, z, w.
psm1_sp.*, psm2_sp.*	Additional position and orientation fields. These are commanded setpoints.
psm1_jaw, psm2_jaw, *_jaw_sp	Jaw-related scalar values.
*_rcm_pose.*	Position and orientation of RCM point.
psm1_js[i], psm2_js[i], *_set_js[i]	Six joint angles in the PSM.
ecm_pose.*, ecm_rcm_pose.*, ecm_js[i], ecm_set_js[i]	ECM pose.
suj1_*, suj2_*, suj3_*, suj_ecm_*	Kinematics labels for SUJ.
psm3_js[i], psm3_set_js[i]	May present in some sampled needle episodes. Not useful in this dataset.

CSV rows may contain blank fields, which means these parts are not meaningful in that dataset episodes.

Collection and preparation

• Robot: dVRK-Si
• Phantom: Suture Pad and liver&gallbladder phantom
• Operator experience: People has experience in teleoperating da Vinci robot but not medical students

Download and use

Install the Hub client and an image reader:

python -m pip install -U huggingface_hub pillow

Download the full dataset, keeping both task folders:

hf download IMERSE-HF/Phantom_dataset_for_cotraining --repo-type dataset --local-dir phantom_dataset

For a smaller first download, retrieve one episode at the documented revision:

from pathlib import Path
import csv
from huggingface_hub import snapshot_download
from PIL import Image

root = snapshot_download(
    repo_id="IMERSE-HF/Phantom_dataset_for_cotraining",
    repo_type="dataset",
    revision="cdc2928568687eb88ea15d4b824fb4b0053a40d0",
    allow_patterns=["needle_task/110521-547002/**"],
    local_dir="phantom_dataset",
)

episode = Path(root) / "needle_task" / "110521-547002"
with (episode / "ee_csv.csv").open(encoding="utf-8-sig", newline="") as file:
    reader = csv.DictReader(file)
    print("Columns:", reader.fieldnames)
    first_row = next(reader)
    print("First timestamp:", first_row["timestamp"])  # kept as a string

image_path = next((episode / "endo_psm1").glob("*.jpg"))
with Image.open(image_path) as image:
    print("Example image:", image.size, image.mode)

The Hub download guide describes revision pinning and file filters.

Training and evaluation

The repository has task folders rather than explicit train/validation/test folders. The automatically generated Hub viewer currently presents images with directory-derived labels such as endo_psm1 and left_img_dir. Those labels identify image streams; they are not task success labels or curated co-training targets. The viewer's train split should not be interpreted as a documented episode-level benchmark split.

License and citation

The repository's existing metadata declares MIT.

If you use this dataset, please cite: Chen, X., Chen, J., Nelson, N., Goldenberg, A., Haworth, J., Huver, S. D., & Krieger, A. (2026). SurgVIL: Scaling Surgical Robot Imitation Learning with Open-source Surgical Videos. arXiv preprint arXiv:2608.16058.

@article{chen2026surgvil,
  title={SurgVIL: Scaling Surgical Robot Imitation Learning with Open-source Surgical Videos},
  author={Chen, Xinhao and Chen, JuoTung and Nelson, Nigel and Goldenberg, Antony and Haworth, Jesse and Huver, Sean D and Krieger, Axel},
  journal={arXiv preprint arXiv:2608.16058},
  year={2026}
}

Action Label Pipeline

This section describes the labeling method in SurgVIL. The paper uses two sources of supervision: kinematics recorded directly from the dVRK-Si for in-house phantom demonstrations, and approximate kinematics estimated from monocular surgical videos. The video estimation pipeline described below applies to the video dataset.

Video-to-action labeling

1. Segment the tool. A lightweight U-Net variant with a MobileNetV2 encoder segments the surgical tool and needle from endoscopic images. It is trained using 60 surgical videos annotated with SAM2, with a weighted combination of multi-class cross-entropy and foreground Dice losses. Tool masks provide evidence for pose estimation. For the optional needle-pose input, curve fitting further processes the needle mask to remove the attached suture thread and estimate the needle center and local tangent.
2. Estimate the tool pose by differentiable rendering. Given a monocular video, the pipeline uses a trained differentiable rendering policy to optimize the pose of a renderable tool model so its appearance matches each observed frame. The estimated pose is represented as $T_t \in SE(3)$. The pipeline supports the large needle driver and is extended to curved scissors using a custom CAD model.
3. Apply image, mask, and temporal losses. The objective combines SSIM image similarity, Dice mask overlap, and a temporal consistency penalty. The temporal term promotes a smooth pose trajectory.
4. Correct systematic depth bias. An affine calibration maps estimated depth to ground-truth depth: $z_{gt} \approx a z_{est} + b$. The coefficients are fitted once on synthetic MuJoCo training data and then kept fixed. The appendix also states that pose-estimation training data are generated synthetically in MuJoCo.
5. Convert calibrated poses into relative actions. The final calibrated pose sequence is converted into labels by computing frame-to-frame relative motion. These estimated labels provide weak supervision for policy training.

The pose-estimation loss is:

$${\mathcal{L}}_t={\lambda }_{\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{m}}(1-\mathrm{S}\mathrm{S}\mathrm{I}\mathrm{M}(I_t^{\mathrm{r}\mathrm{e}\mathrm{n}},I_t^{\mathrm{o}\mathrm{b}\mathrm{s}}))+{\lambda }_{\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{e}}{\mathcal{L}}_{\mathrm{D}\mathrm{i}\mathrm{c}\mathrm{e}}(M_t^{\mathrm{r}\mathrm{e}\mathrm{n}},M_t^{\mathrm{o}\mathrm{b}\mathrm{s}})+{\lambda }_{\mathrm{t}\mathrm{e}\mathrm{m}\mathrm{p}}\rho ({\xi }_t-{\xi }_{t-1}),\backslash mathcal\{L\}\_t\; =\; \backslash lambda\_\{\backslash mathrm\{ssim\}\}\backslash left(1-\backslash mathrm\{SSIM\}(I\_t^\{\backslash mathrm\{ren\}\},I\_t^\{\backslash mathrm\{obs\}\})\backslash right)\; +\; \backslash lambda\_\{\backslash mathrm\{dice\}\}\backslash mathcal\{L\}\_\{\backslash mathrm\{Dice\}\}(M\_t^\{\backslash mathrm\{ren\}\},M\_t^\{\backslash mathrm\{obs\}\})\; +\; \backslash lambda\_\{\backslash mathrm\{temp\}\}\backslash rho(\backslash xi\_t-\backslash xi\_\{t-1\}),$$

where $I^{\mathrm{ren}}$ and $I^{\mathrm{obs}}$ are rendered and observed images, $M^{\mathrm{ren}}$ and $M^{\mathrm{obs}}$ are their masks, $\xi_t=\log(T_{t-1}^{-1}T_t)$, and $\rho(x)=\sqrt{\lVert x\rVert_2^2+\epsilon^2}$ is the Charbonnier penalty.

Labeling parameter	Value
SSIM weight, $\lambda_{\mathrm{ssim}}$	0.5
Dice weight, $\lambda_{\mathrm{dice}}$	1.0
Temporal consistency weight, $\lambda_{\mathrm{temp}}$	0.05
Charbonnier stability constant, $\epsilon$	1 × 10⁻⁶

These video-derived labels remain approximate. The following are pose estimation errors.

Instrument	Evaluation dataset	Position error	Orientation error
Large needle driver	SurgRIPE	6.3 ± 1.7 mm	4.12 ± 1.55°
Curved scissors	ImitateCholec	9.5 ± 2.8 mm	7.01 ± 2.43°

Policy Training Details

The details below are ACT and VLA models' training detail in SurgVIL.

ACT

Parameter	Needle task	Chole task
Model / initialization	ACT	ACT
Optimizer	AdamW	AdamW
Learning rate	5 × 10⁻⁴	5 × 10⁻⁴
Batch size	256	256
Adam β₁	0.9	0.9
Adam β₂	0.999	0.999
Adam ε	1 × 10⁻⁸	1 × 10⁻⁸
Weight decay	1 × 10⁻⁴	1 × 10⁻⁴
Learning-rate schedule	LambdaLR	LambdaLR
Warmup	750 steps	1,000 steps
Training steps	15,000	20,000
Hidden dimension	512	512
Feedforward dimension	3200	3200
Training chunk size	60	60
Evaluation action horizon	20	20
KL weight / β	10	10
Data FPS	30 Hz	30 Hz

π₀

Parameter	Needle task	Chole task
Model / initialization	π₀	π₀
Optimizer	AdamW	AdamW
Learning rate	1 × 10⁻⁴	1 × 10⁻⁴
Batch size	128	128
Adam β₁	0.9	0.9
Adam β₂	0.95	0.95
Adam ε	1 × 10⁻⁸	1 × 10⁻⁸
Weight decay	0.1	0.1
Learning-rate schedule	Cosine	Cosine
Warmup	1,000 steps	1,500 steps
Training steps	11,000	15,000
Full fine-tuning	True	True
Training chunk size	50	50
Evaluation action horizon	20	20
Data FPS	30 Hz	30 Hz

GR00T-H

Parameter	Needle task	Chole task
Model / initialization	GR00T-H	GR00T-H
Optimizer	AdamW	AdamW
Learning rate	2 × 10⁻⁵	2 × 10⁻⁵
Batch size	64	64
Weight decay	1 × 10⁻⁵	1 × 10⁻⁵
Learning-rate schedule	Cosine decay	Cosine decay
Warmup	5%	5%
Training steps	25,000	35,000
Backbone tuning	LLM backbone frozen	LLM backbone frozen
Action horizon	50	50
Action normalization	Mean/std	Mean/std
Random crop	95%	95%
Rotation augmentation	±5°	±5°
Coarse pixel dropout	0.5	0.5
Color jitter brightness	0.12	0.12
Color jitter contrast	0.15	0.15
Color jitter saturation	0.15	0.15
Color jitter hue	0.02	0.02
Multi-view dropout	0.25	0.25
Gradient clipping	1.0	1.0