windjj/datumwell-handwriting

Datumwell Handwriting

Follow the phrase. Inspect the stroke. Read the joint target.

A research release of 200 synthetic writing episodes inspired by the “Write hello world” task in Artificial Analysis AA-Robotics v0.1. It packages synchronized visual observations and joint targets in LeRobot v3 for researchers inspecting a possible VLM/VLA training-data representation for tabletop writing.

This version re-renders existing torque-driven MuJoCo executions in the approved AA-inspired appearance scene: cream tabletop, brown tray, taped paper, black arm, and paneled room. It adds a third camera. This is not official Artificial Analysis data, not an exact reconstruction, and has no validated training benefit or AA benchmark score.

Illustrative camera snapshot from the synthetic episode at about 90 seconds, separate from the decision excerpt below. This is not AA footage or a policy input for earlier timestamps.

Mixed rights, no blanket open-data license. UJI handwriting retains its CC BY 4.0 attribution; the loader code is MIT licensed. Project-generated annotations and packaging are identified separately from third-party robot-rendered appearance. No applicable rendered-output distribution/reuse license has been established for the NexArm assets, and no manufacturer authorization is claimed. Public availability does not itself grant rights over those assets. See RIGHTS.md. No upstream endorsement is implied.

Version: 0.3.0-aa-style-rerender. This is a new rendering of 200 existing executions, not 200 additional robot executions or an exact AA-Robotics reconstruction.

Contents

Included	Count
Simulated writing episodes	200
Uniform numeric rows at 50 Hz	1,232,056
Native camera timestamps	74,068
Re-rendered native images across three views	222,204
640 × 480 H.264 observation videos	600
Distinct composed phrases	30
Distinct base handwriting layouts	84
Anonymous source writers	51

Each Parquet row is one 50 Hz state/command sample. Native observation capture is approximately 2.994 Hz (one image per 167 simulation steps at 2 ms). The 50 fps videos causally repeat the latest native image. They do not contain 50 independent observations per second. image_source_time_s, image_source_index, and observation.image_age_s expose this relationship; age can reach approximately 0.332 s.

The original source had 10 pilot executions; 190 additional executions expanded it to 200. Re-rendering preserves all numeric Parquet files byte for byte. Augmented handwriting is not independent human handwriting: 84 base layouts are composed from isolated recorded character strokes, with deterministic layout, rotation, translation, size and timing variations.

Download the release

Get the LeRobot dataset → · Project and synchronized example

The complete package is approximately 1.86 GB. To inspect one numeric episode without downloading every video:

from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download(
    repo_id="windjj/datumwell-handwriting", repo_type="dataset",
    revision="v0.3.0-aa-style-rerender",
    allow_patterns=["README.md", "RIGHTS.md", "meta/**", "examples/**",
                    "data/chunk-000/file-000.parquet"]
))

Omit allow_patterns to download the complete package, or add the desired videos/observation.images.*/chunk-000/file-000.mp4 paths for that episode. Raw NexArm model files and AA source footage are excluded. Read the mixed-rights record before reuse.

Inspect locally

Install pyarrow for lightweight numeric inspection:

from pathlib import Path
import pyarrow.parquet as pq

root = Path("/path/to/nexarm-handwriting-aa-200/resolve/main")
episode = pq.read_table(root / "data/chunk-000/file-000.parquet")
row = episode.slice(100, 1).to_pylist()[0]
print(row["observation.state"])
print(row["action"])
print(row["image_source_time_s"], row["observation.image_age_s"])

examples/loader.py lists the three corresponding video paths. examples/decision.json shows one current observation with a separate action label. action.schema.json, meta/info.json, and episode-index.json define the action, feature layout and episode identities. Numeric loading does not decode or automatically download video.

The LeRobot dataset ID on the local viewer is local/nexarm-handwriting-aa-200. Point the viewer at the same read-only local data server used for the previous build. No HF upload is required for local inspection.

Portable observation media is stored in the standard LeRobot MP4 paths. Native JPEGs are retained separately in the local rendering workspace; native paths in provenance frame records are local audit references, not files included in this dataset directory.

Record semantics

Field	Meaning
observation.state	Five actual simulated joint positions, rad
observation.velocity	Five actual joint velocities, rad/s
observation.tcp	Actual pen-tip position, metres in robot-base frame
observation.images.tabletop	Overhead camera video
observation.images.oblique	First oblique camera video
observation.images.right	Second oblique camera video
action	Five absolute planned reference joint positions, rad
action.velocity	Planned reference derivative, rad/s
observation.normal_force	Synthetic simulator pen/paper normal contact force, N
observation.contact, observation.valid_ink	Contact and ink validity diagnostics
source_time_s, source_task_time_s	Original physical and task clock values
image_source_index, image_source_time_s	Causal native camera-frame mapping
observation.image_age_s	Current row time minus latest native camera time
timestamp, frame_index, episode_index, task_index	Standard LeRobot indexing

The five arm joints are the command dimensions; the gripper is fixed and the pen is preheld rigidly. There is no learned grasping stage or pressure signal. Actual states are not the action labels. Actions are the command targets that drove the original gravity-compensated PD controller. Commands are evaluated from the original cubic Hermite plan; actual states were resampled by linear interpolation, with discrete contact/force diagnostics held causally. Initial settling was omitted and less than 20 ms trimmed at the end, as in the previous export.

Training inputs and leakage

Use task text, current/past images, and current measured state as policy inputs. Actual velocity, TCP, contact/force and image age can be optional inputs when compatible with the intended deployment. Keep action, action.velocity, future observations, final ink diagnostics, planned target TCP, planned pen-down scheduling, applied command diagnostics and provenance/bookkeeping out of policy observation inputs. Do not blindly feed all observation.* fields into a policy: some retained fields are diagnostics. See provenance/training-feature-policy.json.

The renderer uses saved joint states at the exact camera record indices. Only physical contact ink stamps at or before that timestamp enter the image. The new three-camera intrinsics/extrinsics are in provenance/calibration.json. They are synthetic ideal pinhole parameters, not AA camera estimates.

Splits and coverage

Split	Episode range	Count
Train	[0,107)	107
Validation	[107,128)	21
Test	[128,200)	72

Source writer IDs, base layouts, raw source samples and phrase groups were separated before augmentation; variants retain their parent split. Test phrases are held out. This small synthetic collection is not a performance benchmark or a claim of real robot generalization. Thirty phrases and 84 base layouts do not imply broad language or handwriting coverage. coverage.json records counts and caveats.

Quality and limits

• The prior physics audit reported 225,804 checks with zero failures under its stated criteria. It checked finite values, planned bounds, contacts, ink provenance, frame alignment, tracking and collisions. This appearance release does not rerun physics; numeric/source hashes establish continuity.
• The final release validation checks all 200 numeric files, all 600 video lengths/rates/dimensions, original numeric-file hashes, causal frame mappings, exact source record/FK alignment, updated camera calibration, split isolation and image/video correspondence. See quality/release-validation.json for the actual result; do not infer a pass from this card alone.
• Prior OCR of the standalone 2D actual-ink diagnostics matched 166/200 phrases exactly after normalization; 34 are flagged. These are not 200 human-approved legibility labels or dense OCR checks of the new camera images. The report is retained at quality/readability.json.
• Planned reference speed/acceleration bounds passed, but actual simulated joint acceleration exceeds the configured 1.5 rad/s² threshold. That known issue remains. Do not treat these as acceleration-compliant demonstrations, safety-certified actions, or hardware-ready trajectories.
• The paper remains a miniature 136.5 × 96.2 mm plane. Tray, tape and room are visual-only geometry; the tray is sunk into the existing table to preserve the original paper contact plane. Actual tray collision, friction and placement are not reconstructed.
• AA paper dimensions, tool mounting, lighting, camera calibration and material parameters were not measured. There is no simulation-to-real validation, model training result or robot success-rate claim.

Image statistics are estimated from 64 uniformly sampled output timestamps per episode after resizing to 160 × 120. Numeric statistics are exact and inherited from the byte-identical prior numeric data.

Provenance and rights

Handwriting: Prat, F., Castro, M., Llorens, D., Marzal, A., & Vilar, J. (2008), UJI Pen Characters, Version 2, UCI Machine Learning Repository, DOI 10.24432/C5FG8S, source dataset, CC BY 4.0 as marked by UCI. Changes include coordinate-scale normalization, glyph composition, smoothing/resampling, layout augmentation, synthetic timing/lifts, robot planning, simulation and rendering. Isolated characters are composed into phrases; these are not recordings of natural connected phrase writing. IAM-OnDB is not included.

Robot geometry: official Hiwonder/NexArm, pinned commit 0eb4fb8d806454fa1ff753f6a5ad305c63acceb4. Root README states educational/research availability; nexarm_description/package.xml still contains TODO: License declaration. No applicable explicit redistribution/render-output license has been established. Robot mesh, URDF/Xacro and model XML assets are excluded from this release package, but excluding them does not establish rights to release their rendered appearance.

Task/appearance inspiration: AA-Robotics article. Its video was inspected locally as a visual reference and is not included in the dataset or used as a training image. No official assets, logs, camera calibration or grading procedure were obtained from AA.

Our new room/tray/tape geometry and rendering scripts are separate from the official robot assets. Loader code terms are in CODE_LICENSE. NOTICE retains attribution; RIGHTS.md documents the unresolved publication gate. No permissive license is asserted for the entire composite release.

File SHA-256 fingerprints are generated in checksums.json after validation. Pin the eventual release commit/version when using the package; a candidate repository name is not proof of publication.

New camera readability check

Local macOS Vision OCR on the cropped final new overhead camera observations matched 170/200 target phrases after case/whitespace normalization; 30 are flagged. Cropping uses the calibrated paper bounds, without resizing or adding pixels. This check is separate from the prior 166/200 standalone ink diagnostic. OCR is auxiliary and does not certify human legibility. See quality/camera-readability.json and quality/episode-flags.json for per-episode filtering. Final-image OCR results are evaluation/provenance fields, not current policy inputs.