Tactile Video Pretrain — non-commercial sibling (yxma/tactile-video-pretrain-nc)

Companion to yxma/tactile-video-pretrain — aggregates GelSight-style tactile-video sources whose licenses do not allow commercial use and so cannot mix into the MIT main repo.

Currently included: Baihu-VTouch (sample subset of 4 tasks across 2 robot platforms).

Format

Each episode lives at:

videos/baihu_vtouch/<task>/<robot>/<episode_id>/
├── bimanual.mp4          ← compact preview (cams on top, 4 tactile bottom)
├── tactile_left_l.mp4    ← 240×240 H.264 CRF 18, all frames
├── tactile_left_r.mp4
├── tactile_right_l.mp4
├── tactile_right_r.mp4
├── cam_hand_left.mp4     ← 848×480 H.264 CRF 18, all frames
├── cam_hand_right.mp4
├── cam_head.mp4          ← 640×480
├── head_ir_left.mp4      ← 848×480 grayscale H.264 CRF 18
├── head_ir_right.mp4
├── head_depth.mkv        ← FFV1 lossless gray16le (16-bit precision)
├── joints.parquet        ← state + action arrays
├── metadata.json
└── params/               ← camera intrinsics + extrinsics

original.h5 not mirrored to HF. The full Baihu-VTouch H5s (the upstream source-of-truth, ~970 MB / episode) are kept on the maintainer's disk but not pushed to HF — HF's LFS rate-limit would take weeks for ~700 GB. To get the H5s, fetch them directly from the official ModelScope page.

Task previews (live!)

We currently have processed episodes from 4 Baihu-VTouch tasks across 2 robot platforms.

1. Waste Recycling — D-WHEEL (wheeled humanoid)

273 episodes processed.

Full 10-sample montage →

2. Stack Plates — D-WHEEL

84 episodes processed.

Full 10-sample montage →

3. Ruler into Pouch — PIKA (handheld smart terminal)

261 episodes processed.

Full 10-sample montage →

4. Router Cable — PIKA

140 episodes processed.

Full 10-sample montage →

Bimanual layout (per-episode preview)

Each episode's bimanual.mp4 lays out the bimanual sensors as:

+-------------------+-------------------+
|  LEFT hand cam    |  RIGHT hand cam   |   ← scene RGB from each wrist
+--------+----------+--------+----------+
| L-fing | L-fing   | R-fing | R-fing   |   ← 4 tactile sensors
| left   | right    | left   | right    |
+--------+----------+--------+----------+

The wide top row shows each wrist-camera's view of the workspace; the four cells below are the 4 individual GelSight-style sensors (2 per finger, one for each side of the gel).

Schema (upstream H5)

Each upstream Baihu-VTouch episode is one HDF5 file with the following layout (preserved in this repo's compressed outputs):

<task>/<robot_config>/<episode_id>.h5
├── cameras/{head,hand_left,hand_right}/color/data     # (T,) JPEG bytes
├── cameras/head/{depth,ir_left,ir_right}/data         # (T,) bytes/array
├── tactile/{hand_left,hand_right}/{left,right}/data   # (T,) PNG bytes  ← gel images
├── joints/state/{arm,effector,robot,waist}/...        # (T, K) float32
├── joints/action/...
├── parameters/camera/{head,hand_left,hand_right}.json
├── metadata.json
└── timestamp                                          # (T,) int64

Each Baihu-VTouch episode runs ~400–1500 frames at 30 Hz (a few seconds to ~50 s of motion).

Robot platforms

robot	description	tasks in this repo
D-WHEEL	Wheeled humanoid (Qinglong-class)	Waste Recycling, Stack Plates
PIKA	Handheld dual-gripper smart terminal	Ruler into Pouch, Router Cable

Both expose the same 4-tactile + 3-camera + 16-bit depth + 2-IR + joint-state schema, so cross-platform pretraining is straightforward.

Quick start

from huggingface_hub import snapshot_download
import glob, os, pyarrow.parquet as pq

root = snapshot_download("yxma/tactile-video-pretrain-nc", repo_type="dataset")
episodes = glob.glob(os.path.join(root, "videos/baihu_vtouch/*/*/*"))
print(f"{len(episodes)} episodes available locally")

ep = episodes[0]
tactile_left = os.path.join(ep, "tactile_left_l.mp4")
depth        = os.path.join(ep, "head_depth.mkv")     # FFV1 16-bit
joints       = pq.read_table(os.path.join(ep, "joints.parquet"))

What's still being processed

The full Baihu-VTouch corpus is 17 TB across 29 tasks. We're downloading additional tasks in the background (smallest-first), processing each into the format above, and auto-pushing every 15 min. Check back later for more.

License

CC-BY-NC-SA-4.0 — inherited from upstream Baihu-VTouch.

• Attribution required — cite VTouch++ (Hua et al., 2026) below.
• Non-commercial only — no commercial use of derivative works.
• Share-alike — derivatives must be licensed under CC-BY-NC-SA-4.0.

Citation

@article{hua2026vtouch,
  title  = {VTouch++: A Multimodal Dataset with Vision-Based Tactile Enhancement for Bimanual Manipulation},
  author = {Hua, Qianxi and Li, Xinyue and Yan, Zheng and Li, Yang and Zhang, Chi and Li, Yongyao and Liu, Yufei},
  journal = {arXiv preprint arXiv:2604.20444},
  year   = {2026}
}