iris-kaist/origami-v3-merged

Robotic Origami Challenge — Unified LeRobot v3.0

A single, ready-to-train LeRobot v3.0 dataset of real-world bimanual dexterous paper-airplane folding (robot type north_ces, Sharpa Hands), consolidated from the per-season releases of the Robotic Origami Challenge.

Provenance. The upstream release ships as 46 separate per-season datasets, each a self-contained v3.0 tree whose episode/frame/file indices restart at 0, plus a redundant v2.1 copy. This repo merges all lerobot3.0 season trees into one unified, globally-indexed v3.0 dataset. The v2.1 copies are intentionally excluded (same trajectories, different encoding). No video re-encoding was performed. Original data © Sharpa — https://www.sharpa.com/ — released under CC-BY-4.0.

Summary

Format	LeRobot v3.0
Robot	north_ces — bimanual arms + Sharpa dexterous hands
Episodes	595
Frames	4,091,138
Duration	37.88 h @ 30 fps
Task	1 — "north ces task" (fold a paper airplane)
Source seasons	46
Cameras	6 synchronized streams
On disk	~459 GB

Episode length: 174.8 s min / 229.2 s mean / 418.0 s max.

Features

Key	Type	Shape	Notes
observation.state	float32	(65,)	left arm 7 + left hand 22 + right arm 7 + right hand 22 + motor 7
action	float32	(65,)	same joint layout
observation.state.joint_torque	float32	(65,)	joint torques
observation.state.tcp	float32	(24,)	left pose6+forces6, right pose6+forces6 — all-zero (never populated upstream)
observation.tactile	float32	(60,)	2 hands × 5 fingers × 6-axis (fx,fy,fz,tx,ty,tz)
observation.images.head_left	video	480×480
observation.images.head_right	video	480×480
observation.images.wrist_left	video	480×480
observation.images.wrist_right	video	480×480
observation.images.tactile_deform	video	480×1200	tactile deformation view
observation.images.tactile_raw	video	480×1600	raw tactile view

Known quirks.

• observation.state.tcp is exactly zero across all 4,091,138 frames. Verified against the upstream v2.1 copies (also zero where present) — this is an upstream recording gap, not a conversion loss.
• Video files are a mix of AV1 and H.264 (~907 AV1 / ~766 H.264). info.json labels them all av1; readers detect the real codec, so this is cosmetic. All files decode and every frame dimension is correct.

Usage

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("iris-kaist/origami-v3-merged")
print(ds.num_episodes, ds.num_frames)
print(ds[0].keys())

How it was merged

Per season, in global order: episodes renumbered 0→594 and frames 0→4,091,137; data-parquet index/episode_index columns rewritten; episode-metadata data/video pointers remapped; videos carried over verbatim (no re-encode); meta/stats.json recomputed as a count-weighted aggregate of the per-season stats. 15/16 structural integrity checks pass (the 16th is the cosmetic codec label above).

License & attribution

Released under CC-BY-4.0, following the upstream Sharpa release. If you use this data, please credit Sharpa and the Robotic Origami Challenge.