Dimios45/molmo-fold-pink-towel-dagger

Pink Towel Folding — HITL / DAgger

30 real-robot towel-folding episodes collected with openpi-control on a YAM bimanual robot, combining policy rollouts with human corrections through a Meta Quest 3S VR interface. Every saved frame identifies whether the policy or human operator supplied the action.

The task instruction is “fold the towel”. Collection took place on September 13, 2026. This release contains the 30 saved episodes from a session initially configured for 50 attempts; collection was stopped after episode 30.

Dataset at a glance

Property	Value
Format	LeRobot v3.0, Parquet + MP4
Episodes	30 (zero-based indices 0–29)
Recorded frames	142,422
Human-controlled frames	20,916 (14.69%)
Policy-controlled frames	121,506 (85.31%)
Nominal recording rate	30 FPS
Nominal recorded duration	79.12 minutes, computed as frames / FPS
Cameras	Top, left wrist, right wrist
Image dimensions	848 × 480 RGB per camera
State / action dimensions	14 each
Operator-labeled outcomes	15 successes, 15 failures
Human takeovers	49, from the session manifest
Aborted episodes retained	1: attempt 22 / episode index 21
Split	train, all 30 episodes

Collection protocol

The policy drives the robot until the operator takes over through VR. During human intervention, policy queries are paused. The recorded action is the command from the active controller, and intervention marks its source. Autonomous portions and unsuccessful attempts are retained alongside human corrections.

Session settings: policy speed factor 0.7, action chunk size 30, episode time limit 180 seconds. The operator labels each attempt as successful or unsuccessful. These are outcomes of mixed human/policy rollouts, not an autonomous policy success benchmark. The exact policy checkpoint and training provenance are not recorded in the dataset metadata; the repository name alone does not establish them.

Schema

Feature	Type / shape	Meaning
observation.state	float32, (14,)	Measured robot state
action	float32, (14,)	Command issued by the active controller
intervention	float32, (1,)	1.0: human control; 0.0: policy control
observation.images.top	video, (480, 848, 3)	Overhead camera
observation.images.left_wrist	video, (480, 848, 3)	Left wrist camera
observation.images.right_wrist	video, (480, 848, 3)	Right wrist camera
timestamp	float32, (1,)	Frame timestamp
frame_index, episode_index, index, task_index	int64, (1,)	Frame, episode, global, and task indices

State and action ordering: left joints 1–6, left gripper, right joints 1–6, right gripper. Task text is stored in meta/tasks.parquet and referenced by task_index.

Files and labels

• data/: recorded state, action, intervention, and indexing columns.
• videos/: three camera streams in chunked MP4 files.
• meta/info.json: feature schema, FPS, paths, and dataset totals.
• meta/stats.json: feature statistics.
• meta/episodes/: episode lengths and references into data/video chunks.
• meta/tasks.parquet: task instructions.
• openpi_control_hitl.json: per-attempt prompts, success labels, saved/aborted flags, and runtime takeover statistics.

Join outcome labels to frame data using episode_index. attempt is one-based; episode_index is zero-based. Runtime counters in the HITL manifest are not identical to persisted frame counts: the manifest reports 20,917 human control ticks, while Parquet contains 20,916 human frames. Use the Parquet intervention column for training masks and saved-frame statistics. Nominal video duration is not wall-clock collection time.

Download and use

from huggingface_hub import snapshot_download

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

Read action/state data and select human corrections without decoding video:

from pathlib import Path
import numpy as np
import pyarrow.parquet as pq

frames = pq.read_table(str(Path(root) / "data"))
is_human = np.asarray(frames["intervention"].to_pylist()).reshape(-1) == 1.0
human_frames = frames.filter(is_human)

For image-based training, use a LeRobot v3-compatible loader that resolves episode video offsets. For human-correction imitation learning, apply the intervention == 1.0 mask to the training loss while retaining the temporal context needed by your model. Policy frames are policy-generated actions, not expert labels. Filter outcomes or aborted attempts explicitly if required by your experiment; they have not been removed from this release.

Scope and limitations

This is a small, single-task collection with one robot setup and one prompt. It contains both successful and unsuccessful trajectories, including one saved aborted attempt. There is no held-out validation or test split. Split by episode for evaluation to avoid overlap between neighboring frames. Outcomes are operator judgments, not independently audited labels.

No dataset license has been specified by the uploader in this release.