Ninkofu/openarm-cube-pick-place-sim-20260908

OpenArm Cube Pick and Place — Simulation

A robot-learning dataset contributed by Ninkofu. The author-owned contributions are released under CC BY 4.0; third-party rights are described below.

Task and collection

Instruction: Pick up the cube and place it in the tray.

53 selected successful episodes, 13,593 frames at 30 Hz; OpenArm v2 bimanual robot, PICO/WebXR teleoperation, simulation. Four camera views and both arms are retained. The active task label is the instruction shown above. Earlier collection or editing labels, where different, are not training labels for this export.

Data representation

LeRobot v3.0 layout: data/ contains Parquet samples, videos/ contains MP4 camera streams, and meta/ contains episode, task, split and provenance metadata. Videos have no audio streams.

• action: 16 absolute target joint positions in radians, including the gripper joint position.
• observation.state: 16 observed joint positions in radians.
• Order for both vectors: right joint 1–7, right gripper, left joint 1–7, left gripper. Exact names are in meta/info.json.
• timestamp: seconds relative to the episode; frame_index resets per episode. index is the global row index.
• task_index: 0, mapped to the instruction in meta/tasks.parquet.
• success: 1 throughout these selected demonstrations. It is the episode annotation, not a measured policy success rate.
• Camera images are RGB; all published streams use a 30 Hz timeline.

Camera	Width × height
head_left	640 × 480
head_right	640 × 480
wrist_left	960 × 600
wrist_right	960 × 600

Source sampling and conversion

The following source rates were measured across all 53 retained source episodes. Each pooled rate is 1e9 / median(consecutive timestamp differences in ns) without joining episode boundaries. The range describes per-episode median rates, not timing-jitter bounds. Camera timestamps come from source image filenames; arm timestamps come from the recorded Parquet streams.

Source stream	Pooled measured Hz	Per-episode median Hz range
action/arms/left	84.893	70.588–94.732
action/arms/right	84.892	70.588–94.732
obs/arms/left	47.609	37.821–119.360
obs/arms/right	81.157	63.298–123.600
cameras/head_left	29.374	29.299–29.462
cameras/head_right	29.374	29.299–29.462
cameras/wrist_left	29.374	29.299–29.462
cameras/wrist_right	29.374	29.299–29.462

These measured source rates replace stale/incomplete declarations in meta/source_metadata.yaml. Original declarations remain explicitly marked as superseded in its publication metadata. The converted dataset is uniformly 30 Hz. Conversion used OpenArmDataset 0.4.0 with smoothing disabled; this release does not resample or modify the converted samples or videos.

Experimental splits and statistics

Partition	Episodes
train	43
validation	5
test	5

The exact converted episode indices are in meta/episode_split.json; the three lists are disjoint and cover every episode. meta/stats.json retains the existing statistics computed using training episodes only.

The Hub viewer's train storage split contains the complete Parquet table, including held-out episodes. It is not the experimental training partition. Select episodes from meta/episode_split.json when training or evaluating; do not train on the entire viewer split.

Local loading

From this dataset directory, using a LeRobot installation that supports v3.0:

import json
from pathlib import Path
from lerobot.datasets.lerobot_dataset import LeRobotDataset

root = Path.cwd()
split = json.loads((root / "meta/episode_split.json").read_text())
dataset = LeRobotDataset(
    repo_id="local/openarm-cube-pick-place-sim-20260908",
    root=root,
    episodes=split["train"],
    video_backend="pyav",
)

meta/preparation.json uses paths relative to this dataset for output_root and split_path. Archive references beginning with data/raw/, data/edited/, outputs/ or source-assets/ are provenance identifiers from the source project, not files bundled here. Absolute usernames and machine paths have been removed. Original source fingerprints identify the unchanged source archive; refreshed metadata checksums identify the files in this export. The original project's validator binds absolute local paths and should not be applied directly to this portable export; the native LeRobot loader is the intended reader.

Quality and limitations

Task labels, episode/frame counts, finite state/action values and experimental split coverage were checked. Images were sampled from every episode (start, middle and end of the left head view), supplemented with other camera views; this was not an exhaustive inspection of every frame. No faces or private documents were observed in those samples. Small, success-selected demonstrations in a narrow setup do not establish multi-task language grounding or generalization. Sensor metadata marks calibration as unverified; nominal optical parameters must not be treated as measured calibration.

License, rights and attribution

Copyright 2026 Ninkofu, to the extent copyright subsists in the contributor-owned material. The contributor-owned recordings, annotations, dataset compilation and documentation are made available under CC BY 4.0; see LICENSE. Attribute Ninkofu and this dataset repository, retain the notices and identify modifications. The contributor authorized public release on 2026-10-07. No trademark, endorsement, publicity or other third-party rights are licensed by this grant.

Robot geometry provenance includes Enactic OpenArm v2 and CAD components identified upstream as CERN-OHL-S-2.0. That is a component notice, not an assigned license for this dataset. CAD/mesh source files are not included. Public upstream references and component license notices in meta/source_metadata.yaml have been retained.

Branded bottle labels, where visible, are part of the recorded or rendered scene; no affiliation or endorsement is implied. Please cite Ninkofu, the dataset title and its Hugging Face repository URL. Questions can be raised in the repository Discussions.