| --- |
| pretty_name: HandWM |
| license: other |
| license_name: mixed |
| tags: |
| - hands |
| - egocentric-video |
| - imu |
| - mano |
| - world-model |
| size_categories: |
| - 100K<n<1M |
| viewer: false |
| --- |
| |
| # HandWM |
|
|
| Everything needed to run the HandWM code (hand motion → simulated glove IMU) and the P3-WAM + IMU |
| re-implementation on EgoDex and ARCTIC: the datasets, MANO, intermediate assets and both code trees. |
| `organize.sh` turns a download into the directory layout the code expects; see [ORGANIZE.md](ORGANIZE.md). |
|
|
| ```bash |
| hf download RunyiY/HandWM --repo-type dataset --local-dir handwm_download |
| bash handwm_download/organize.sh handwm_download /path/to/root |
| ``` |
|
|
| ## Contents |
|
|
| | Path | Size | What | |
| |---|---|---| |
| | `data/egodex/<part>/*.tar` | 2.11 TB | EgoDex, all 338,234 episodes (`.hdf5` poses + `.mp4` video), ~10 GB tar shards; members are paths relative to `raw/egodex` | |
| | `data/arctic/images_zips/sXX/*.zip` | 696 GB | ARCTIC images, 339 sequences, 10 subjects (s03 is the official test subject, no public annotations) | |
| | `data/arctic/arctic_unpacked.tar` | 0.8 GB | ARCTIC `raw_seqs` (MANO, SMPL-X, object, egocentric camera; 301 sequences), `meta`, `splits_json`, `mano_v1_2` | |
| | `data/mano/MANO_{LEFT,RIGHT}.pkl` | 10 MB | MANO v1.2 hand models, chumpy-free copies loadable with `smplx` | |
| | `data/p3wam/egodex_mano.tar` | 23 GB | MANO fits of 162,836 EgoDex episodes (P3-WAM hand targets) | |
| | `data/p3wam/codec_v2/codec.pt` | 42 MB | frozen action codec (P3-WAM) | |
| | `data/p3wam/imu-stage-low/new_params.pt` | 2.2 GB | IMU warm-up weights for the low-noise expert (`train_wam.py --init-new`) | |
| | `code/handswm`, `code/p3wam_imu` | <0.2 GB | code, configs, tests, docs, evaluation results | |
| | `manifest/` | | `egodex_index.csv.gz` (file → shard), `egodex_fitted_files.txt` (P3-WAM training subset), `files.csv.gz` (sha256 of every payload file) | |
| | `data/*/receipts/` | | download receipts of the original archives | |
|
|
| Not included: the Wan2.2-Fun-A14B-Control backbone (Apache-2.0, `organize.sh` downloads it from |
| `alibaba-pai/Wan2.2-Fun-A14B-Control`), simulated IMU and canonical motion (regenerated by the code), |
| training checkpoints. |
|
|
| ## Licenses |
|
|
| Each component keeps its upstream license. All of them are **non-commercial**. |
|
|
| | Component | License | Source | |
| |---|---|---| |
| | EgoDex | CC BY-NC-ND 4.0, files unchanged (repackaged into tar archives only) | Apple, https://github.com/apple-aiml-research/ml-egodex | |
| | ARCTIC (images, annotations, meta) | ARCTIC Data & Software license, non-commercial scientific research | https://arctic.is.tue.mpg.de | |
| | MANO models | MANO license, non-commercial scientific research | https://mano.is.tue.mpg.de | |
| | EgoDex MANO fits | derived from EgoDex and MANO; their terms apply | this work | |
| | Code, codec and IMU weights | research use | this work | |
|
|
| Before using ARCTIC or MANO, register and accept their licenses on the sites above. |
|
|
| ## References |
|
|
| - EgoDex: R. Hoque et al., *EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video*, 2025. |
| - ARCTIC: Z. Fan et al., *ARCTIC: A Dataset for Dexterous Bimanual Hand-Object Manipulation*, CVPR 2023. |
| - MANO: J. Romero, D. Tzionas, M. J. Black, *Embodied Hands: Modeling and Capturing Hands and Bodies Together*, SIGGRAPH Asia 2017. |
| - P3-WAM: https://p3-wam.pages.dev/ |
|
|
