livsynrobotics/LivUMI-Ego-Both-Sample

UMI-Ego Dataset (Subset 34)

A multi-task egocentric manipulation dataset collected with the LiveGo platform and stored in the LeRobot dataset format (codebase version v2.1). Each subdirectory corresponds to one household manipulation skill with synchronized multi-camera video and proprioceptive / UMI tracking signals.

Overview

Property	Value
Format	LeRobot v2.1
Robot platform	livego
Sampling rate	30 Hz
Total tasks	9
Total episodes	101
Total frames	55,624
Approx. duration	~31 min (at 30 FPS)
On-disk size	~2.1 GB
Train split	All episodes (train: 0:N per task)

Tasks

Each task is a self-contained LeRobot dataset under its own folder:

Folder	Description	Episodes	Frames	Size
clean_body_fat_scale	Clean a body-fat scale	8	3,516	~140 MB
clean_desk	Clean a desk surface	9	3,132	~146 MB
clean_toliet_table	Clean a toilet-area table (folder name retains original spelling)	22	5,450	~159 MB
fold_cloth	Fold a piece of cloth	11	12,291	~448 MB
open_bottle_cup	Open a bottle or cup	12	5,663	~255 MB
put_battery_into_box	Place a battery into a box	10	3,817	~181 MB
put_iphone_to_charge	Put an iPhone on a charger	10	3,460	~110 MB
put_toy_into_cupboard	Put a toy into a cupboard	10	7,507	~279 MB
remove_the_pills_from_the_board	Remove pills from a board	9	10,788	~470 MB

Natural-language task labels in meta/tasks.jsonl are currently placeholders ("task example"). Use the folder names above as the canonical task identifiers, or replace the labels in tasks.jsonl / episodes.jsonl with your own descriptions.

Directory Layout

Each task folder follows the standard LeRobot layout:

<task_name>/
├── meta/
│   ├── info.json              # Dataset schema, feature specs, paths, splits
│   ├── tasks.jsonl            # Task index → language instruction
│   ├── episodes.jsonl         # Per-episode metadata (length, task labels)
│   └── episodes_stats.jsonl   # Per-episode feature statistics
├── data/
│   └── chunk-000/
│       └── episode_XXXXXX.parquet   # Tabular state & indices (no embedded video)
└── videos/
    └── chunk-000/
        └── <video_key>/
            └── episode_XXXXXX.mp4   # One MP4 per camera stream per episode

Path templates (from meta/info.json):

• Tabular data: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Videos: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4

Sensors & Observations

Cameras (5 streams per episode)

Feature key	Resolution (H×W)	Notes
observation.images.cam_top.oak	800 × 1280	Top-mounted OAK camera
observation.images.cam_left_fisheye.fisheye	480 × 640	Left fisheye
observation.images.cam_right_fisheye.fisheye	480 × 640	Right fisheye
observation.images.cam_left_depth	480 × 640	Left depth (stored as RGB video)
observation.images.cam_right_depth	480 × 640	Right depth (stored as RGB video)

Videos are stored as MP4 files; frame-level references live in the Parquet episodes.

Robot / UMI state (per timestep)

For each of three end-effector groups — arm.left, arm.right, and arm.top — the Parquet files include:

Feature suffix	Shape	Description
end_effector_pose	6	End-effector pose
end_effector_value	1	Scalar gripper / actuator value
umi_accel	3	Accelerometer (ax, ay, az)
umi_gyro	3	Gyroscope (gx, gy, gz)
umi_mag	3	Magnetometer (mx, my, mz)
umi_raw	7	Raw UMI packet fields
umi_pose_m_rpy_deg	6	Pose, meters + roll-pitch-yaw (degrees)
umi_pose_m_zyx_rad	6	Pose, meters + ZYX Euler (radians)
umi_flags	5	Protocol / status flags
umi_reserved	3	Reserved bytes
umi_timestamp	1	Unix timestamp (float64)

Note: In several tasks, arm.top UMI channels are zero-filled while arm.left and arm.right carry active UMI streams. Check episodes_stats.jsonl or your own loaders if you rely on top-arm UMI data.

Episode indexing fields

Field	Description
timestamp	Time within episode (seconds, 0-based)
frame_index	Frame index within episode
episode_index	Episode ID
index	Global frame index across the dataset
task_index	Index into tasks.jsonl

This release contains observations only (images + state). There are no separate action.* features in meta/info.json.

Loading the Data

With LeRobot

If you have LeRobot installed:

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

# Load a single task
dataset = LeRobotDataset(
    repo_id="local/umi-ego-34-clean_desk",
    root="/path/to/umi-ego/34/clean_desk",
)

sample = dataset[0]
print(sample.keys())

Point root at any task subdirectory (e.g. clean_desk, fold_cloth). Adjust repo_id to a local identifier of your choice.

Manual access

import json
from pathlib import Path

task_root = Path("clean_desk")
info = json.loads((task_root / "meta/info.json").read_text())

episodes = [
    json.loads(line)
    for line in (task_root / "meta/episodes.jsonl").read_text().splitlines()
    if line.strip()
]

print(f"{len(episodes)} episodes, {info['total_frames']} frames @ {info['fps']} FPS")

Read data/chunk-000/episode_*.parquet for state time series and decode matching files under videos/chunk-000/<video_key>/.

Data Splits

Every task uses a single training split covering all episodes, e.g. "train": "0:9" for 9 episodes. No held-out validation or test split is provided in this subset.

Citation & License

If you publish work using this data, cite the UMI-Ego project and the LeRobot dataset tooling as appropriate. License terms are not bundled in this folder; refer to the parent umi-ego distribution or data provider for usage restrictions.

Known Issues

• clean_toliet_table: Directory name uses toliet instead of toilet.
• Task strings: meta/tasks.jsonl entries are placeholders; episode-level tasks fields mirror the same placeholder text.
• Top-arm UMI: arm.top IMU/pose fields may be all zeros depending on the recording setup.

File Checklist (per task)

Path	Purpose
meta/info.json	Feature schema, FPS, chunk size, path templates
meta/tasks.jsonl	Task vocabulary
meta/episodes.jsonl	Episode lengths and labels
meta/episodes_stats.jsonl	Min/max/mean/std per feature per episode
data/chunk-*/episode_*.parquet	Low-dimensional state & indices
videos/chunk-*/<camera>/episode_*.mp4	Synchronized camera recordings