File size: 3,668 Bytes
7fae406 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 | ---
license: other
license_name: mixed-cc-by-and-public-domain
license_link: https://creativecommons.org/licenses/by/4.0/
task_categories: [other]
tags: [robotics, motion, expressive-robots, animacy, reachy-mini, autonomous-lamp, vla, talking-head, speech]
pretty_name: animacy human motion (large harvest)
size_categories: [n<1K]
---
# squaredcuber/animacy-human-motion-large
Canonical human conversational motion (`animacy.human.v1`, 30 Hz) captured at scale with
[animacy](https://github.com/Hcoder10/animacy) — the open interaction layer for expressive robots.
Every clip is a directory `clips/<shard>/<name>/` holding
* `motion.parquet` — 28 channels: head 6-DoF, gaze, brows, eyes, mouth, torso, a puppet arm chain,
`speaking` / validity flags (schema: `docs/CANONICAL.md` in the repo);
* `audio.opus` — the clip's audio, mono, Opus 32 kb/s from the 16 kHz capture track, on the same
clock as `motion.t` (`ffmpeg -i audio.opus -ar 16000 -ac 1 audio.wav` restores the 16 kHz wav that
`animacy.schema.HumanClip.load` reads);
* `meta.json` — source url, **license and the machine-readable license evidence**, capture settings,
neutral pose, validity stats, and a `harvest` block (source family, channel, speaker key, language
guess, chunk offset, prescreen result).
`index.json` is one row per clip (all shards). `manifests/<shard>.json` is the same for one commit.
**26 clips, 2.9 hours (2.7 face-valid hours), 2 speaker keys, 2 languages.**
## License policy (enforced in code, not remembered per file)
Only material whose *machine-readable* license metadata says public domain / CC0 / CC-BY is
captured; anything ND, NC, SA, or with missing license metadata is refused
(`scripts/fetch_sources.py::classify_license`, applied by `scripts/harvest/fetch.py`). Evidence per source:
| source | evidence recorded in `meta.json.license_evidence` |
|---|---|
| YouTube (CC search) | info-json `license` == "Creative Commons Attribution license (reuse allowed)" (CC-BY 3.0), channel id, upload date |
| U.S. Government channels | allowlisted official agency channel + channel id match; basis 17 U.S.C. § 105 (works of the U.S. Government). VOA clips carry a caveat about third-party newswire material |
| Wikimedia Commons | `videoinfo.extmetadata` LicenseShortName / License / LicenseUrl / Copyrighted=False |
| archive.org | item metadata `licenseurl` / `rights` |
Attribution: every clip's `meta.json` has `source_url`, `artist`/channel, and the license label;
CC-BY clips must be credited to that channel/author when redistributed. This dataset is a derived
work (motion parameters + downsampled audio), not the source video.
## Quality gate
Chunks of <= 600 s; prescreen (>= 60% of 16 sampled frames show exactly one face);
kept if `face_valid >= 60%` and `face_valid * duration >= 60 s`. No channel contributes
more than 5% of the kept seconds (fetch-time cap). Speaker keys over-merge (a channel = one
speaker), so a per-speaker cap at training time is conservative.
## Counts
| source | hours |
|---|---|
| usgov | 2.9 |
| license | hours |
|---|---|
| Public Domain | 2.9 |
| language (guess) | hours |
|---|---|
| en | 2.5 |
| en-US | 0.5 |
| series | hours |
|---|---|
| briefing | 1.9 |
| usgov_other | 0.7 |
| address | 0.3 |
## Provenance and reproducibility
Harvested by `scripts/harvest/` (crawl -> fetch -> workers -> push) on a CPU-only workstation; the
`harvest` block in each `meta.json` names the worker and chunk. Full pipeline description:
`docs/HARVEST.md`. Nothing here was captured from a source whose license metadata was missing.
|