Tactile Video Pretrain (yxma/tactile-video-pretrain)

~48 hours of GelSight-style tactile video, aggregated across 4 public sources. Each row is one demonstration / episode / scan / touch session. Two configs share one shared video folder:

• tactile_only — one tactile stream per row. Use for tactile-only self-supervised pretraining (video MAE / V-JEPA / contrastive / masked).
• tactile_rgb — paired (tactile, scene-RGB) videos per row. Use for cross-modal alignment (CLIP-style, contrastive, masked cross-modal).

Quick start

from datasets import load_dataset

# tactile-only stream
ds = load_dataset("yxma/tactile-video-pretrain", "tactile_only", split="train")
print(ds[0]["video"])           # repo-relative path
print(ds[0]["source"])          # "freetacman" | "gelslam_tracking" | ...

# paired tactile + RGB
ds2 = load_dataset("yxma/tactile-video-pretrain", "tactile_rgb", split="train")
print(ds2[0]["tactile_video"], ds2[0]["rgb_video"])

Resolve repo-relative paths to local files:

from huggingface_hub import snapshot_download
import os
root = snapshot_download("yxma/tactile-video-pretrain", repo_type="dataset")
local = os.path.join(root, ds[0]["video"])

Composition

source	clips	hours	tactile resolution	RGB pair?
freetacman	11,967	29.71	640×480 @ 30 FPS	✅ 11,965
touchandgo	140	14.66	640×480 @ 27 FPS	✅ 138
gelslam_recon	15	2.47	320×240 @ 25 FPS	–
gelslam_tracking	140	0.81	320×240 @ 25 FPS	–
TOTAL	12,262	47.65	mixed	12,103

Total frames	4,929,371
Mean clip length	14 s
Median clip length	8 s
Codec	H.264 (re-encoded for browser playback)

Preview clips

Short (≤ 8 s) MP4 previews of one randomly-picked clip per source.

freetacman tactile	freetacman tactile + scene
touchandgo tactile	touchandgo tactile + scene
gelslam_tracking	gelslam_recon (long surface scan)

Sample frames

Per-source 40-frame grids:

freetacman tactile	touchandgo tactile
gelslam_tracking tactile	gelslam_recon tactile

Paired tactile + scene RGB (16 random pairs):

Schema

tactile_only

column	type	description
clip_id	string	stable per-clip id, format depends on source
source	string	one of freetacman, gelslam_tracking, gelslam_recon, touchandgo
task	string	task name where applicable (FreeTacMan)
object_class	string	object class (GelSLAM)
demo_id	int32	demonstration / episode id
side	string	left/right (FreeTacMan); else null
n_frames	int32	number of video frames
fps	float32	source frame rate
duration_s	float32	n_frames / fps
width, height	int32	spatial resolution
has_pose	bool	true ⇒ GelSLAM tracking 6-DoF poses available
has_gradient	bool	true ⇒ GelSLAM tracking gradient maps available
has_contact_mask	bool	true ⇒ GelSLAM tracking contact masks available
video	string	repo-relative path to MP4
trajectory	nullable struct	per-frame state, schema depends on source

tactile_rgb

Same schema with video → tactile_video, plus rgb_video.

trajectory struct (nullable)

The struct columns differ per source — every field is optional. Convenient quick filter: row["trajectory"] is None ⇒ no per-frame state available.

field	sources
timestamp, TCP_pos_*, TCP_euler_*, quat_*, gripper_distance	freetacman
time_tactile, time_scene	touchandgo
(none)	gelslam_recon
6-DoF poses, contact masks, gradient maps	gelslam_tracking — stored alongside videos as .npy, not in parquet

Sources

See SOURCES.md for per-source details (license, sensor, processing recipe) and SHORTLIST.md for the full roadmap.

source	sensor	license	upstream
FreeTacMan	McTac-derived GelSight-style	MIT	OpenDriveLab/FreeTacMan
GelSLAM (tracking + reconstruction)	GelSight Mini (markerless)	MIT	joehjhuang/GelSLAM_dataset
Touch and Go	GelSight (markered) + ego RGB	CC-BY-4.0	project page

Investigated but not included

• AgiBot World — audited 7 spreadsheet-flagged "tactile gripper" tasks; 0 hours of tactile video in the public release. Details in SOURCES.md.

Pipeline (vs. upstream)

1. Rename / re-organize to a unified naming convention per source.
2. Re-encode AVI → MP4 (libx264, CRF 18–20) so browser playback works on HF preview. GelSLAM's FFV1-AVI files become MP4 here.
3. Hardlink where possible so on-disk usage stays low.
4. Wrap available per-frame state as a nullable trajectory parquet struct.
5. Add clip_id + source for stable filtering across configs.

Recommended uses

• Tactile-only SSL — video MAE / V-JEPA on tactile_only; ~30 h FreeTacMan + ~3 h GelSLAM clean Mini frames.
• Cross-modal pretrain — contrastive (tactile ↔ scene) on tactile_rgb.
• Geometry-aware SSL — has_pose/has_gradient/has_contact_mask filters surface the GelSLAM tracking rows for dense supervision.
• In-the-wild material reasoning — filter source = "touchandgo".
• Sensor-shift robustness — train markerless, test markered (Touch and Go).

License

MIT, inherited from all currently included upstream sources.

Citation

@article{wu2025freetacman,
  title   = {FreeTacMan: Robot-free visuo-tactile data collection system for contact-rich manipulation},
  author  = {Wu, Longyan and Yu, Checheng and Ren, Jieji and Chen, Li and Jiang, Yufei and Huang, Ran and Gu, Guoying and Li, Hongyang},
  journal = {IEEE International Conference on Robotics and Automation},
  year    = {2026}
}
@inproceedings{huang2024gelslam,
  title   = {{GelSLAM}: Long-Horizon Object Tracking with Vision-Based Tactile Sensors},
  author  = {Huang, Hung-Jui and ...},
  year    = {2024}
}
@inproceedings{yang2022touchgo,
  title   = {Touch and Go: Learning from Human-Collected Vision and Touch},
  author  = {Yang, Fengyu and ...},
  booktitle = {NeurIPS Datasets and Benchmarks Track},
  year    = {2022}
}