yeeeiii111/wuji-writing

Wuji Writing Bundle

A paired human demonstration ↔ robot teleoperation dataset for the task of brush-writing digits 0–9 on grid paper. Built to train human-to-robot transfer policies (e.g. VAM-style action diffusion with human reference video).

Both subsets are in LeRobot v2.1 format and share the same 10 tasks (task_index ∈ {0..9} corresponds to the written digit). They live as separate sub-directories under the repo root because the camera layout and resolution differ.

.
├── teleop/   1 LeRobot v2.1 dataset — 487 ep, robot bimanual, 4 cameras, 480×640
├── ego_ref/  1 LeRobot v2.1 dataset — 896 ep, human egocentric, 1 head camera, 1280×720
└── README.md

Sample counts per task

task_index	digit	teleop ep	ego_ref ep
0	zero	80	90
1	one	80	90
2	two	48	89
3	three	40	90
4	four	47	89
5	five	35	89
6	six	31	89
7	seven	45	90
8	eight	29	90
9	nine	52	90
total		487	896

teleop/

• Format: LeRobot v2.1, fps 30
• Episodes: 487 (continuous episode_index 0..486, sorted by digit)
• Frames: 438,986
• Robot: bimanual Tianji arms + Wuji dexterous hands (54-D state/action)
• Cameras (480×640, av1+yuv420p):
  • observation.images.stereo_left
  • observation.images.stereo_right
  • observation.images.cam_left_wrist
  • observation.images.cam_right_wrist
• State / Action: 54-D float32. Layout: left arm joint 7 + right arm joint 7 + left hand joint 20 + right hand joint 20.
• Task prompt template: "the robot writes digit {word}" (e.g. "the robot writes digit four")

Data quality note: digits 0/1/2 come from the original March 2026 capture; digits 3–9 come from the May 2026 re-capture (which fixed a black-frame flickering bug present in the original 3–9 recordings — verified at 0% black frames after ffmpeg blackdetect scan). source_provenance.jsonl in meta/ records which source subset each merged episode came from.

ego_ref/

• Format: LeRobot v2.1, fps 30
• Episodes: 896 (continuous, 89–90 per digit)
• Frames: 357,427
• Camera: observation.images.head (1280×720, h264+yuv420p) — egocentric head-mounted view of a human writing with a brush on the same grid paper
• State / Action: 54-D zero-vector placeholders (this subset is reference video only; no robot action is recorded for the human demonstrations).
• Task prompt template: "the human writes digit {word}"

This subset is intended to be used as paired reference video when training a robot policy on teleop/; pair by task_index. Each teleop episode can be conditioned on a randomly chosen ego_ref episode of the same task.

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

teleop = LeRobotDataset("USER/wuji-writing-bundle", root="./data", subset="teleop")
ego    = LeRobotDataset("USER/wuji-writing-bundle", root="./data", subset="ego_ref")

Or read meta directly:

import json
info = json.load(open("teleop/meta/info.json"))
episodes = [json.loads(l) for l in open("teleop/meta/episodes.jsonl")]

Provenance

file	meaning
<subset>/meta/info.json	LeRobot info (features, fps, totals, codec)
<subset>/meta/episodes.jsonl	per-episode entry (episode_index, tasks, length)
<subset>/meta/tasks.jsonl	the 10 task strings (one per digit)
<subset>/meta/source_provenance.jsonl	merged-episode → source subset/episode mapping
<subset>/data/chunk-000/episode_NNNNNN.parquet	per-frame state, action, timestamp, indices
<subset>/videos/chunk-000/<cam_key>/episode_NNNNNN.mp4	per-camera video

Limitations

• teleop episodes are imbalanced across digits (29–80 per digit); upsample if needed.
• ego_ref state/action columns are zero-filled placeholders — only the head video carries information.
• teleop and ego_ref episodes are not 1-to-1 aligned in time; they are independent recordings of the same task.
• No language description beyond the digit-template prompt; if richer instructions are needed, use a VLM to re-caption the head videos.