easyminnn/exumi-lerobot

exUMI (LeRobot v2.0)

1,452 UMI-gripper episodes across 8 manipulation tasks with an image-type tactile 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	1,452
Frames	218,527 (20 fps)
Tasks	8 — insert_pen, open_bottle, peg_in_hole, pick_brocolli, pick_carrot, pick_cube, …
Video	ego_view 224×224 · tactile_left 680×460
Action	7-D per-step EE delta + gripper (FreeTacMan layout)
Tactile type	image — usable with a frozen image tactile tokenizer

Notes

• Source runs at 59.94 fps; this conversion applies stride 3 to reach 20 fps, matching the exUMI task config (dataset_frequeny 59.94, obs_down_sample_steps 3).
• Tactile is stored as a video stream (observation.images.tactile_left), not an array — it comes from the source tactile_combined RGB buffer.
• The source also ships raw 2K GoPro footage, which is not included here; only the down-scaled vision used for training was converted.

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 silicx/exUMI (https://huggingface.co/datasets/silicx/exUMI). 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.

License note. The upstream repository does not state a license. This conversion is redistributed for research use on the assumption that the authors intend open research access; if you are an author and would like it taken down or relicensed, please open a discussion on this repo.