botdata1 / README.md
xijiazhao's picture
README: English version
920ca23 verified
|
Raw History Blame Contribute Delete
6.34 kB
metadata
language:
  - en
tags:
  - robotics
  - lip-sync
  - servo-control
  - real-images
size_categories:
  - 10K<n<100K
configs:
  - config_name: all
    default: true
    data_files:
      - split: train
        path: data/round*/train-*.parquet
  - config_name: round1
    data_files:
      - split: train
        path: data/round1/train-*.parquet
  - config_name: round2
    data_files:
      - split: train
        path: data/round2/train-*.parquet

botdata1: robot mouth grid data (commands ↔ real photos)

botdata1 is a full grid sweep of the 13 mouth-servo commands of one humanoid robot face, with one photo per command combination. It has 49,248 photos from two capture sessions:

Config Capture date Rows Notes
round1 2026-06-12 to 06-13 38,880 Full grid over 9 independent degrees of freedom
round2 2026-06-25 10,368 Wider command range (jaw up to 1.0, corners go negative); col1/col3 pinned at 0.33
all (default) — 49,248 Both rounds together

The command convention is the same in both rounds:

  • The same command gives the same physical mouth pose.
  • The left/right mirror rules are bit-identical.

So the two rounds can be pooled as one command → appearance dataset. No command appears in both rounds, and the rounds complement each other. Round 1 covers the upper-lip and lower-lip centre motors (col1/col3). Round 2 covers a larger jaw opening and wider lower-lip and corner ranges.

Images are stored as the original JPEG bytes, with no re-cropping, resizing or re-encoding.

1. Quick start

from datasets import load_dataset
ds = load_dataset("xijiazhao/botdata1", "all", split="train")  # or "round1" / "round2"

x = ds[0]
x["face_480"]   # PIL.Image 480×480, whole face
x["lip_256"]    # PIL.Image 256×256, mouth crop
x["commands"]   # 13 float32 servo commands
x["round"], x["idx"], x["global_idx"], x["timestamp"]

The commands are also available without datasets as raw/round{1,2}_commands.npy. Row k (0-based) belongs to the photo with idx = k + 1.

2. Fields

Field Meaning
round 1 or 2
idx Photo number within the round, 1-based. Original file names are {idx:05d}_face.jpg and {idx:05d}_lip.jpg
global_idx Index over both rounds. Round 1 is 0…38879; round 2 is 38880…49247 (= 38880 + idx − 1)
timestamp Capture time (local time), taken from the capture logs
commands 13 normalized servo commands. These are the commanded targets; there is no encoder feedback
face_480 480×480 whole-face photo
lip_256 Mouth crop: face_480[305:433, 180:308] (128×128) upscaled 2× to 256×256. Both rounds use the same box

3. The 13 command columns

There are 9 independent degrees of freedom: cols 0, 1, 2, 3, 4, 6, 8, 9 and 12. The other 4 columns are right-side mirrors derived from the left side:

  • col5 = 0.272 − col4
  • col7 = 0.944 − col6
  • col10 = 0.467 − col8
  • col11 = 1.0 − col9
col Name Motor Round 1 levels Round 2 levels
0 upperlip_center_fwd upper lip centre, forward 0 / .33 / .667 .33 / .67 / .9
1 upperlip_center_up upper lip centre, up 0 / .28 / .556 fixed .33
2 lowerlip_center_fwd lower lip centre, forward 0 / .35 / .694 0 / .35 / .69
3 lowerlip_center_up lower lip centre, up 0 / .30 / .611 fixed .33
4 upperlip_fwd_L upper lip left, forward 0 / .272 0 / .272
5 upperlip_fwd_R upper lip right, forward mirror mirror
6 lowerlip_up_L lower lip left, up 0 / .3 / .6 / .944 .3 / .6 / .94 / 1.2
7 lowerlip_up_R lower lip right, up mirror mirror
8 corner_frontback_L left corner, front/back 0 / .23 / .467 −1 / −.5 / 0 / .23
9 corner_updown_L left corner, up/down 0 / .25 / .5 / .75 / 1 −.6 to .9, step .3 (6 levels)
10 corner_frontback_R right corner, front/back mirror mirror
11 corner_updown_R right corner, up/down mirror mirror
12 jaw_fwd jaw (mouth opening) 0 / .23 / .46 / .694 0 / .2 / .4 / .6 / .8 / 1.0

Grid sizes:

  • Round 1: 3·3·3·3·2·4·3·5·4 = 38,880.
  • Round 2: 3·3·2·4·4·6·6 = 10,368.

Every photo has a unique command vector. The original motor notes are in raw/round{1,2}_motor_meta.json and raw/command_sample.md; they include the travel angles and normalization ranges. command_sample.md is in Chinese.

4. Mixing the two rounds

  • Commands can be pooled directly. The detailed check is in raw/round2_CONSISTENCY.md.
  • Pixels should not be mixed as-is. Round 2 was shot in a separate session, and the camera, lighting and wig all shifted:
    • The position offset is not a single global shift. It is about (+2.9, +4.8) px around the eyes but only about (+0.5, +0.1) px over the lower face.
    • The light direction also changed. For the same command, even the pose-invariant eye/brow region differs by about 8 MAE (0–255 grey levels).
  • Register round 2 before pixel-level use such as image retrieval, image comparison or training image models. Use a translation plus a low-frequency lighting gain. In our experiments this brought the lower-face gap between the rounds down to about one jaw grid step.
  • In round 2, jaw values of 0.8 and 1.0 really do open the mouth further; they are not clipped. The mouth travel of this robot head is narrow, though.

5. Files

data/round1/train-0000{0..5}-of-00006.parquet   # 38,880 rows
data/round2/train-0000{0,1}-of-00002.parquet    # 10,368 rows
raw/round1_commands.npy, raw/round2_commands.npy  # (N, 13) float32
raw/round1_motor_meta.json, raw/round2_motor_meta.json
raw/command_sample.md       # motor travel / normalization range / sample points
raw/round2_CONSISTENCY.md   # round 1 vs round 2 consistency check
raw/summary.json            # row counts and command ranges per round

6. Limitations

  • Commands are the targets that were sent to the servos, not measured positions.
  • All photos come from one robot head, one camera viewpoint and two capture sessions.
  • Mapping human faces onto this robot needs an extra retargeting step that is not part of this dataset.

Access / License

Publicly visible research dataset. No license has been assigned yet; contact the authors before redistribution or commercial use.