Dimios45/bimanual-fold-pink-towel-vr

Bimanual Pink Towel Folding — VR Demonstrations

30 bimanual towel-folding episodes collected on a YAM robot using the Quest VR teleoperation kit integrated into openpi collect. Every episode has the exact task instruction fold the towel.

Dataset summary

Property	Value
Collection date	September 16, 2026
Format	LeRobot v3.0
Robot	yam_bimanual
Saved episodes	30, indexed 0–29
Saved frames	43,814
Nominal recording rate	30 FPS
Nominal saved duration	24.34 minutes (frames / FPS)
Task	fold the towel
Views	Top, left wrist, right wrist
Resolution	848 × 480 RGB per view
State and action size	14 values each
Split	train, all 30 episodes

Collection

A human operator controlled both arms through VR: grip to clutch an arm and trigger to control its gripper. Right B started an episode or discarded and restarted the current take; left Y saved it. The session stopped after 30 saves and parked and de-energized both arms.

The terminal summary reported 29 discarded takes and 162 stalled ticks. Discarded takes do not count toward the 30 saved episodes. These counters describe the collection session; they are not per-episode quality labels, and the stalled-tick count is not a count of corrupted saved frames. This release contains human teleoperation demonstrations, with no policy/human intervention column or per-episode success labels.

Features

Feature	Type / shape	Description
observation.state	float32, (14,)	Measured robot joint and gripper state
action	float32, (14,)	Recorded robot command
observation.images.top	video, (480, 848, 3)	Top view
observation.images.left_wrist	video, (480, 848, 3)	Left wrist view
observation.images.right_wrist	video, (480, 848, 3)	Right wrist view
timestamp	float32, (1,)	Frame timestamp within the episode
frame_index, episode_index, index, task_index	int64, (1,)	Frame, episode, global, and task indices

State and action ordering is left joints 1–6, left gripper, right joints 1–6, right gripper. Joint positions are in radians. Gripper values use the dataset convention: 0.0 is fully open and 1.0 is fully closed.

The prompt is represented natively by task_index = 0 on every frame, mapped to fold the towel in meta/tasks.parquet. Each episode's task list also contains fold the towel.

Download

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="Dimios45/bimanual-fold-pink-towel-vr",
    repo_type="dataset",
)

Use a LeRobot v3-compatible loader to resolve the camera video offsets. The repository contains data/ for frame Parquet files, videos/ for the three camera streams, and meta/ for the feature schema, statistics, task mapping, and episode offsets.

Validation and scope

Before upload, all 30 episode task lists and every frame's task index were checked. Stored frame counts, episode lengths, contiguous frame indices, and referenced video paths were validated. This is structural validation; individual demonstrations have not all been visually audited for task success or precise camera/action alignment.

This is a small, single-task collection. No held-out evaluation split is provided; split by episode to avoid overlap between neighboring frames. No dataset license has been specified by the uploader.