Datasets:

allen-datumwell
/

datumwell-handwriting-sample

Inspect the full temporal sequence for each sample episode: measured simulated states, planned joint commands, three video views and source alignment metadata. This is a ten-episode inspection sample from an existing 200-episode collection, not ten rows cut from otherwise downloadable full data.

Request the full 200-episode collection or discuss a collaboration →
The website form submits directly to Datumwell. You can also email allen@datumwell.io. For full-repository access, include your Hugging Face username and intended use; access requests are reviewed manually. Larger or tailored collections are a discussion of future work, not a claim of additional ready inventory.

Resource	Availability
This public repository	Exactly 10 complete episodes, freely inspectable without access approval
Full 200 collection	Manual access approval; request through the project form and the HF access request
Tailored data collection	Discuss tasks, coverage, formats and feasibility with Datumwell

Important scope and rights

These are synthetic MuJoCo torque-driven writing sequences, derived from isolated UJI handwriting characters through composition, transformation and simulation. They are not 200 independent human robot demonstrations, official Artificial Analysis data, an exact AA-Robotics reconstruction, or evidence of validated training gains. The v0.3.0 appearance rerender reused existing executions.

Mixed rights; no blanket open-data or commercial license. UJI source handwriting retains CC BY 4.0 attribution; the example loader is MIT licensed. An applicable license for distribution/reuse of Hiwonder/NexArm-rendered appearance has not been established. Public availability or manual approval does not grant missing third-party rights. See RIGHTS.md, NOTICE, and CODE_LICENSE. No manufacturer or AA endorsement is claimed. Raw model assets and AA reference footage are excluded.

Load all ten complete episodes

from huggingface_hub import snapshot_download
root = snapshot_download(
    repo_id="windjj/datumwell-handwriting-sample",
    repo_type="dataset",
    revision="main",  # pin the returned commit for reproducible experiments
)

With LeRobot 0.4.x and a supported video backend:

from lerobot.datasets.lerobot_dataset import LeRobotDataset
sample = LeRobotDataset("windjj/datumwell-handwriting-sample", root=root,
                        video_backend="pyav")
first = sample[0]
print(first["observation.state"].shape)  # 5 joints
print(first["action"].shape)             # 5 absolute joint targets

The train/test source labels are retained as contiguous sample splits (train=0:6, test=6:10). No validation episodes happened to be selected. These small splits are for provenance/inspection, not a new benchmark or recommended training evaluation. The source 200 has107/21/72 episodes in train/validation/test.

Contents and timing

Numeric stream: 50Hz; observation.state/velocity and action/action.velocity are five-joint vectors (radians and radians/s). Action is the planned absolute joint reference used by the original controller, not measured motion.
Videos: three 640×480 H.264 views (tabletop, oblique, right), ten complete videos per camera.50fps playback causally repeats native camera observations captured about 2.994Hz. Video frames are not 50 independent new images per second.
timestamp and frame_index span the entire exported source episode. next.done is true only on its final numeric row.
source_time_s, source_record_before, image_source_time_s, image_source_index, and observation.image_age_s preserve original alignment. Native frame provenance is included for the selected episodes.
LeRobot episode_index, global index, task indices, episode offsets, file indices, splits, counts and aggregate statistics are rebuilt for this sample. State/action values and video bytes otherwise match the source release.

meta/info.json, stats.json, tasks.parquet, episodes/chunk-000/file-000.parquet
data/chunk-000/file-000.parquet ... file-009.parquet
videos/observation.images.{tabletop,oblique,right}/chunk-000/file-000.mp4 ... file-009.mp4
provenance/episodes/<selected-source-id>/frames.json
sample-manifest.json, episode-index.json, checksums.json, quality/

Selection and known limitations

The manifest records the source commit and all ten sourceepisode indices. The deterministic procedure takes one complete episode from each contiguous 20episode source-index bin. It fixes the canonical hello world demo in the first bin; for later bins it ranks episode IDs by SHA256(datumwell-public-sample-v1:<id>) and prefers different phrase/base/writer IDs. OCR outcomes were not selection inputs. This is a convenience inspection sample, not a statistically representative quality estimate.

Known source issues remain: actual simulated joint acceleration exceeds the configured 1.5 rad/s² bound; appearance, geometry scale, calibration and physics are approximate; no hardware-safety, sim-to-real, model-training benefit or AA score is validated. Some final ink/camera OCR results fail. Auxiliary OCR is not human approval or a dense audit of every observation.

This sample: final overhead-camera OCR matched 9/10 phrases exactly; prior standalone ink OCR matched 8/10. Per-episode flags are in quality/episode-flags.json. The source 200 reported 170/200 camera matches and 166/200 prior ink matches. These are distinct auxiliary diagnostics.

Keep current/past observations and measured states separate from labels and diagnostics: planned targets, planned pen-down scheduling, final images, future feedback and action labels must not leak into policy inputs. See training feature policy.

Attribution

Prat etal., UJI Pen Characters (Version2), UCI DOI10.24432/C5FG8S, source CC BY 4.0. Character strokes were composed, transformed, retimed and simulated. Geometry: Hiwonder/NexArm, unresolved rendered-output terms as detailed above. Task/appearance inspiration: AA-Robotics v0.1.

Downloads last month: -