Datasets:

easyminnn
/

freetacman-lerobot

5,627 handheld visuo-tactile manipulation episodes (42 tasks) with a GelSight-style tactile pair and an egocentric RGB view, converted to LeRobot v2.0.

Part of a set of tactile manipulation datasets converted to a single common LeRobot v2.0 layout so they can be mixed in one training run.

At a glance


Episodes	5,627
Frames	1,513,182 (30 fps)
Tasks	42
Video streams	`ego_view`, `tactile_left`, `tactile_right` — all 640×480
State	8-D — EE position (3) + EE quaternion wxyz (4) + gripper (1)
Action	7-D — EE position delta (3) + rotation delta axis-angle (3) + gripper target (1)
Tactile type	image (GelSight-style), usable with a frozen image tactile tokenizer

Conversion

Source layout is one directory per task with {task}_{ep}_camera1/2/3.mp4 (tactile left, tactile right, ego RGB) plus {task}_{ep}_traj.csv, one CSV row per video frame.
Episode length is min(csv_rows, frames of each of the 3 videos), so a video is never shorter than its Parquet table; a few episodes carry 1–17 unused trailing frames.
Videos are copied unmodified from the source release (MPEG-4 Part 2, 640×480, 30 fps). They decode fine with FFmpeg-based readers (PyAV, torchcodec, OpenCV, decord) but will not play in Chromium-based viewers, which do not support that codec.
episodes.jsonl carries an extra outlier_steps list per episode: step indices whose action delta exceeds 20 mm or 0.12 rad in one 33 ms step (corrupt boundary rows in the source CSVs). 958 of 5,627 episodes have at least one. They are flagged, not removed.

Known gaps

episode_index is not contiguous (0…6228 for 5,627 episodes). The gaps are episodes the conversion could not produce, almost all because the source lacks the third camera: FragileCup (200), Stamp (200), TextureClassify (200), Write (199) and one SteakSaucing episode ship only Camera1/Camera2. One further episode (ScrewInTheBulb_70) was skipped for an unrelated reason. Index maths must not assume 0..N-1; iterate meta/episodes.jsonl instead.
meta/tasks.jsonl lists all 46 source tasks, but only 42 have episodes — task indices 8, 25, 32 and 45 are never referenced.
info.json omits the optional features, total_frames, total_videos, total_chunks and splits keys, and the Parquet tables have no global index column. Stock LeRobotDataset therefore will not open this as-is; GR00T-style loaders that read modality.json will.
No meta/stats.json — compute normalization statistics yourself.

Verification

Every file was checked before upload: all 22,512 files present and consistent with episodes.jsonl; all 16,881 MP4s structurally intact (no truncation, stsz/stts agree, duration matches frames/30); ~1,100 videos fully decoded with zero errors; all 5,627 Parquet tables read clean with zero NaN/Inf, monotonic timestamps at exactly frame_index/30, and row counts matching the metadata.

Format

GR00T-flavoured LeRobot v2.0: per-episode Parquet under data/chunk-{:03d}/ and per-episode MP4 under videos/chunk-{:03d}/{video_key}/, with meta/info.json, meta/episodes.jsonl, meta/tasks.jsonl and meta/modality.json.

from datasets import load_dataset          # metadata / parquet only
# or read directly:
import pyarrow.parquet as pq, json
info = json.load(open("meta/info.json"))
tbl  = pq.read_table("data/chunk-000/episode_000000.parquet")

meta/modality.json names the tactile stream and the state/action slot layout, so the dataset can be loaded by an Isaac-GR00T style loader without extra configuration.

Provenance & license

This is a format conversion of OpenDriveLab/FreeTacMan (https://huggingface.co/datasets/OpenDriveLab/FreeTacMan), released under the MIT license. All sensor data originates from the authors of that dataset; please cite their work. No frames were dropped or re-timed beyond what is listed under Conversion above.

Downloads last month: 588