WorldArchive/mono-india-workplace-lerobot

World Archive Mono — LeRobot Mirror

Robot-ready export of the Mono India Workplace sample. One episode per clip; frames subsampled to 15 fps at 480p.

Episodes	9 (one per workplace clip)
Frames	46,436
FPS	15
Tasks	75 unique verb–noun labels
Metadata source	mono-india-workplace-sample
Explorer	data-explorer Space
Collection	Physical AI India

Dataset Description

Each episode is a real Indian workplace manipulation clip (factory, catering, repair, craft). Per frame:

Feature	Shape	Description
observation.images.ego	video 480×852	RGB egocentric frame (H.264)
observation.state	126-d float	2 hands × 21 landmarks × (x, y, z)
action	126-d float	Hand-state proxy (human motion signal)
task	string	Verb–noun label from covering action segment

Full JSONL annotations, schema docs, and ~19 GB source MP4s: S3 pack or the metadata dataset.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("WorldArchive/mono-india-workplace-lerobot")
print(ds.num_episodes, ds.num_frames, ds.fps)
sample = ds[0]
# sample["observation.images.ego"], sample["observation.state"], sample["task"]

Browse clips interactively: WorldArchive/data-explorer

Limitations

• Hand action is a keypoint proxy, not joint torques or robot commands.
• 15 fps subsample from 30 fps source; fine-grained contact timing may be coarser than native video.
• Same geographic and license constraints as the source sample.

License

CC BY-NC 4.0 for evaluation. Commercial production training: shubham@worldarchive.co.

Citation

@dataset{worldarchive_mono_lerobot_2026,
  title        = {World Archive Mono LeRobot: India Workplace Egocentric Manipulation},
  author       = {World Archive / GGN},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/WorldArchive/mono-india-workplace-lerobot}}
}

Contact

Book a call · shubham@worldarchive.co · worldarchive.co