Datumwell / Synthetic robot data / v0.3.0
Handwriting,
with the motion
attached.
Inspired by the writing task in Artificial Analysis AA-Robotics v0.1. Explore 200 synthetic episodes connecting what the simulated arm saw, where the pen moved, and the planned joint command at that moment.
LeRobot data for inspecting a VLM/VLA training representation. Not official AA data or an exact reconstruction. No validated training benefit or AA benchmark score.
01 / Inspect an episode
Write “hello world”.
Local 3D simulation rendering. Videos have a 50 Hz clock and repeat the latest native camera image; they do not provide 50 independent observations each second. Camera changes keep the same playback time.
02 / Understand the representation
Observation, command,
and provenance.
Each row pairs a measured simulated joint state with the planned joint reference that drove the original controller. The latest camera image is held causally. Source record indices, image age, and camera calibration remain inspectable.
- Action
- 5 absolute joint targets · radians
- Camera media
- 640 × 480 H.264 · 3 views · 600 videos
- Numeric clock
- 50 Hz · 1,232,056 rows
- Handwriting source
- 84 composed base layouts · 51 anonymous writers
- Episode split
- 107 train / 21 validation / 72 test
03 / Assess the evidence
Clear limits.
Inspectable checks.
What changed in this version?
The cream table, brown tray, taped paper, black arm and paneled room follow the approved AA-inspired appearance sample. Three cameras replay the same saved simulation states. Numeric trajectory and action rows are unchanged; these are not 200 new executions.
How close is this to the AA setup?
Appearance is approximate. Paper remains a miniature 136.5 × 96.2 mm plane, the pen is preheld, and tray/tape/room objects are visual-only geometry. AA dimensions, camera calibration, contact parameters and scoring were not reconstructed.
Are these acceleration-compliant demonstrations?
No. Planned reference bounds passed the original audit, but actual simulated joint acceleration exceeds the configured 1.5 rad/s² threshold. This known issue remains. The data has no hardware-safety or sim-to-real validation.
Was every phrase judged legible?
Prior OCR matched 166 of 200 standalone final-ink diagnostics exactly; 34 are flagged. That is an auxiliary diagnostic, not human approval or a dense OCR audit of every new camera observation. Separately, OCR of the final new overhead observations matched 170 of 200 phrases; 30 are flagged. Both reports remain auxiliary.
What may enter a training observation?
Use current/past images and measured states with the task text. Keep action labels, planned target diagnostics, planned pen-down scheduling, final images and future feedback out of policy inputs. The package includes an explicit feature policy.
What are the reuse terms?
UJI handwriting retains CC BY 4.0 attribution, and the example loader is MIT licensed. Generated annotations and packaging are distinguished from the NexArm-rendered appearance. No applicable third-party rendered-output license has been established; no manufacturer authorization or blanket composite license is claimed. Raw model assets and AA reference footage are excluded.
Release validation passed: 200 episodes, 600 videos, byte-identical numeric labels and causal frame alignment. New overhead-camera OCR matched 170/200 phrases; 30 remain flagged.
Inspect the quality report →Sources and attribution
Handwriting: Prat et al., UJI Pen Characters v2, UCI DOI 10.24432/C5FG8S (CC BY 4.0). Isolated character strokes were composed, transformed, retimed and simulated. Robot geometry: Hiwonder/NexArm, pinned source commit retained in the package. Task and appearance inspiration: AA-Robotics v0.1. No upstream endorsement is implied.
Also from Datumwell: EmbodiedTrace — real observations and recorded-command action labels →