Datasets:
Download README.md from galafinger/galafinger-manipulation-sample: direct link, hf CLI and curl.
- Browser
- Download file 12 kB
-
https://huggingface.co/datasets/galafinger/galafinger-manipulation-sample/resolve/main/README.md
- Command line
-
hf download hf://datasets/galafinger/galafinger-manipulation-sample/README.md
-
curl -L -o README.md https://huggingface.co/datasets/galafinger/galafinger-manipulation-sample/resolve/main/README.md
license: cc-by-4.0
task_categories:
- robotics
tags:
- LeRobot
- tactile
- bimanual
- human-demonstration
- data-glove
- egocentric-video
- stereo
- manipulation
- tactile-sensing
- force
- imitation-learning
- robot-learning
configs:
- config_name: default
data_files: data/*/*.parquet
pretty_name: Galafinger Manipulation Sample
size_categories: 10K<n<100K
Tactile gloves + egocentric stereo video: 37 bimanual manipulation demos
A person does six everyday two-handed tasks at a table. A head-mounted stereo camera records what they see; a tactile glove on each hand records how hard every fingertip presses. Each video frame has one matching row of glove data, and every episode is cut into named sub-steps ("reach", "grasp", "squeeze hard", "release").
Use it to see what fingertip force adds to egocentric video, to test a data loader on LeRobot v2.1, or to judge our recording quality before asking for more. It is a small sample from Galafinger Labs; larger and custom collections (your tasks, objects, scenes) are available on request: hello@galafinger.com
Episode 5, head camera (left eye): squeeze the soft toy gently, then medium, then hard. The three force levels look alike on video and separate clearly in the fingertip force signal.
At a glance
| What | 37 episodes · 6 tasks · 24,617 frames at 30 fps · 13.7 min |
| Video | Orbbec EGO head-mounted stereo camera, left + right eye, 1600×900 |
| Touch | 2 × TouchTronix V3 gloves: per hand 5 fingertips × 4×4 pressure cells, fingertip normal force (N) |
| Hand motion | per hand 5 finger-bend values, 15 estimated joint angles, glove orientation (quaternion) |
| Labels | task instruction per episode · 445 time-ranged sub-step labels · automatic grasp phase per hand |
| Sync | camera and gloves on one host clock, measured 0.1–0.2 s |
| Format | LeRobot v2.1 (parquet + mp4), meta/modality.json for NVIDIA Isaac GR00T |
| Operator / scene | one operator, one room, one table |
| QA | 37 accepted · 0 rejected; every exported video frame machine-checked |
| License | CC BY 4.0 |
Tasks
| # | Instruction | Episodes | Duration |
|---|---|---|---|
| 0 | Pick the stones of two chosen colors from the mixed box and sort them onto two plates by color | 5 | 3 min 07 s |
| 1 | Squeeze the soft toy gently, then medium, then hard, and release | 10 | 3 min 34 s |
| 2 | Pick up the vials from the table one by one and place them upright into the slots of the tray | 5 | 1 min 23 s |
| 3 | Hold the cucumber on the cutting board and slice it into rounds with the knife | 5 | 1 min 46 s |
| 4 | Take the glasses out of the case, wipe both lenses with the cloth, and put the glasses back into the case | 6 | 2 min 13 s |
| 5 | Spread the T-shirt flat on the table and fold it into a neat rectangle | 6 | 1 min 39 s |
Episodes 30 (glasses) and 36 (T-shirt) contain a mistake and its correction, labeled as separate sub-steps.
Hardware
- Camera: Orbbec EGO head-mounted stereo camera (left + right eye), 1600×900, 30 fps, looking down at the hands and table.
- Gloves: 2 × TouchTronix V3 (firmware V01.12): 5 fingertip 4×4 pressure arrays (per-pixel kPa calibration, 22.5 mm² pixels), 5 finger-bend sensors, on-glove orientation; USB, 125 Hz each.
- Recording software: TQ Studio 1.4.0 — both gloves on one host clock; each glove clock aligned by minimum-latency offset.
Data format
| Feature | Shape | Description |
|---|---|---|
observation.state |
58 | per hand (left, then right): finger flex (5), joint angles (15), wrist quaternion (4), fingertip normal force (5) |
action |
30 | joint angles of both hands at the next frame (rad): left 15, right 15 |
observation.tactile.tips_left |
80 | left fingertip pressure, 5 fingers × 16 taxels, 0–255 (255 ≈ 356 kPa) |
observation.tactile.tips_adc_left |
80 | left raw 12-bit ADC per taxel (uncalibrated) |
observation.tactile.tips_right |
80 | right fingertip pressure, 5 fingers × 16 taxels, 0–255 (255 ≈ 356 kPa) |
observation.tactile.tips_adc_right |
80 | right raw 12-bit ADC per taxel (uncalibrated) |
observation.images.head |
video 1600x900 | camera stream |
observation.images.head_r |
video 1600x900 | camera stream |
task_index |
1 | index into meta/tasks.jsonl (language instruction) |
subtask_index |
1 | sub-step annotation: index into meta/subtasks.jsonl (time-ranged sub-step label), -1 = none; ranges in meta/annotations.jsonl |
phase_index_left |
1 | left-hand grasp phase: 0 idle, 1 grasp, 2 hold, 3 adjust, 4 release |
phase_index_right |
1 | right-hand grasp phase: 0 idle, 1 grasp, 2 hold, 3 adjust, 4 release |
Collection protocol
- One episode = one complete execution of the task card by an operator wearing both gloves.
- Gloves are zeroed (tare) with hands free at connection; bend sensors calibrated per wearer (open hand / fist).
- Every episode passes automatic checks (frame rate, gaps, dead or stuck taxels) and a human accept/reject review; by default only accepted episodes are exported.
- Grasp phases are detected automatically from fingertip force and may be corrected by the operator in the annotation view; sub-step labels (
subtask_index,meta/annotations.jsonl) in this release were drafted with AI assistance from the video and the force signals together, checked against force edges, and are subject to operator review; treat boundaries as accurate to about 0.25–0.5 s.
Repository layout
data/chunk-000/episode_000000.parquet … episode_000036.parquet glove streams + labels, one row per video frame (30 Hz)
videos/chunk-000/observation.images.head/episode_0000NN.mp4 head camera, left eye, 1600x900, 30 fps
videos/chunk-000/observation.images.head_r/episode_0000NN.mp4 head camera, right eye
meta/info.json feature names, shapes, fps
meta/tasks.jsonl task instructions
meta/episodes.jsonl episode -> task, length
meta/subtasks.jsonl sub-step label vocabulary
meta/annotations.jsonl sub-step time ranges per episode
meta/modality.json modality map (GR00T)
meta/stats.json, meta/episodes_stats.jsonl normalization statistics
Episodes by task: 0–4 stones, 5–14 squeeze, 15–19 vials, 20–24 cucumber, 25–30 glasses, 31–36 T-shirt.
Limitations
- Small. 37 episodes, 13.7 minutes, one operator, one room, one table. Useful for inspecting the format and signal quality, not for training a policy on its own.
- Sync is software-based. Camera and gloves share a host clock; measured alignment is 0.1–0.2 s. Fine for grasp/hold/release timing, not for contact-onset events at millisecond scale.
- Normal force only. No shear force, no palm sensing, no slip signal.
- No hand position. The gloves give finger flex and orientation. Joint angles are estimated from one flex sensor per finger; there is no wrist or fingertip position in space, no depth, no camera IMU, no camera calibration in this release.
- Sensor rest level. Some fingertip sensors read 1–2 N with nothing in the hand; gentle contacts sit close to that level.
- Labels. Sub-step labels were drafted with AI assistance from video plus force and are accurate to about 0.25–0.5 s; they have not all been reviewed by a second person. Grasp-phase columns are automatic.
- Repaired frames. 12 corrupted video frames (one each in 12 videos) were replaced by the preceding frame; frame counts are unchanged.
- Splits. No predefined split. Hold out whole episodes; with a single operator and scene, results will not show generalization to new people or places.
Privacy
The only person recorded is the operator, a member of Galafinger Labs, who consented to publication. The camera looks down at the table and hands. Videos were machine-scanned frame by frame for corruption and reviewed visually at one frame per second; no faces, screens or documents were found. A brand wordmark on the operator's clothing is visible in some frames. If you spot anything that should not be public, write to hello@galafinger.com.
Loading data
With LeRobot (a release that reads the v2.1 dataset format; not yet tested by us against the newest LeRobot):
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("galafinger/galafinger-manipulation-sample")
print(ds[0]["observation.tactile.tips_right"].shape)
Without LeRobot (download the repo, then read the files directly):
import json, numpy as np, pyarrow.parquet as pq
t = pq.read_table("data/chunk-000/episode_000005.parquet")
col = lambda name: np.array(t.column(name).to_pylist(), dtype=np.float32)
state = col("observation.state") # (N, 58), 30 Hz
names = json.load(open("meta/info.json"))["features"]["observation.state"]["names"]
ix = lambda prefix: [i for i, n in enumerate(names) if n.startswith(prefix)]
right_bend = state[:, ix("right_bend_")] # (N, 5) finger flex, 0 = open, 1 = fist
right_joints = state[:, ix("right_joint_")] # (N, 15) joint angles, rad (estimated from flex)
right_quat = state[:, ix("right_quat_")] # (N, 4) glove orientation (no position)
right_force = state[:, [names.index(f"right_{f}_fz") for f in
("thumb", "index", "middle", "ring", "little")]] # (N, 5) normal force, N
# Fingertip pressure maps: 5 fingers x 4x4 taxels, 0-255 (255 ~ 356 kPa)
right_tips = col("observation.tactile.tips_right").reshape(-1, 5, 4, 4)
left_tips = col("observation.tactile.tips_left").reshape(-1, 5, 4, 4)
# Sub-step labels: per-frame index, or time ranges per episode
subtask = np.array(t.column("subtask_index").to_pylist()).ravel() # -1 = unlabeled
labels = [json.loads(l)["subtask"] for l in open("meta/subtasks.jsonl")]
ranges = [json.loads(l) for l in open("meta/annotations.jsonl")]
ep5 = [(r["frame_0"], r["frame_1"], r["subtask"]) for r in ranges if r["episode_index"] == 5]
# Peak right-hand force inside each labeled sub-step of this episode
for f0, f1, name in ep5:
print(f"{name:28s} {f0:4d}-{f1:4d} peak {right_force[f0:f1].max():5.1f} N")
# Video frame k of videos/chunk-000/observation.images.head/episode_000005.mp4
# corresponds to row k of this table (30 fps; camera-glove sync 0.1-0.2 s).
Example output for episode 5 (squeeze at three force levels):
pick up the toy 80- 162 peak 1.8 N
squeeze gently 162- 189 peak 2.3 N
squeeze medium 303- 387 peak 3.3 N
squeeze hard 471- 578 peak 13.2 N
put the toy down 642- 717 peak 2.0 N
release 717- 747 peak 1.8 N
Not in this release: hand or wrist position in space (the gloves give finger flex and orientation only), shear force, palm pressure, camera IMU.
Citation
@misc{galafinger2026manipulationsample,
title = {Galafinger Manipulation Sample: bimanual tactile-glove demonstrations with egocentric stereo video},
author = {{Galafinger Labs}},
year = {2026},
howpublished = {Hugging Face},
url = {https://huggingface.co/datasets/galafinger/galafinger-manipulation-sample}
}
Using this data? We'd love to see what you build. Tag Galafinger Labs on LinkedIn or write to hello@galafinger.com. We share the best projects and give priority to requests from people who use the sample.
License and contact
CC BY 4.0: free to use, including commercially, with attribution (see Citation). For larger or custom collections (your tasks, objects, scenes, hours) contact Galafinger Labs: hello@galafinger.com · https://galafinger.com
Author: Galafinger Labs
Generated by TQ Studio on 2026-11-16.