UniAction: Unified Action Conditioning for World Models
Paper • 2610.32977 • Published
schema string | num_samples int64 | streams list |
|---|---|---|
uniaction-training-sources | 40,000 | [
{
"name": "hand",
"source_root": ".",
"metadata": "metadata/hand.csv",
"samples": 10000
},
{
"name": "force_balls",
"source_root": ".",
"metadata": "metadata/force_balls.csv",
"samples": 5000
},
{
"name": "force_dominos",
"source_root": ".",
"metadata": "metadata/... |
UniAction-Train
Paper · Model · Evaluation set
40,000 samples covering force, hand pose, robot pose and camera control, with 10,000 samples per task. The dataset includes repeated samples to balance the tasks. Source videos are retained; the training loader samples and preprocesses 81 frames at 832 × 480.
The included training-sources.json is used by uniaction-train and uniaction-cache encode.
Sources and licenses
| Task | Source | License |
|---|---|---|
| Hand pose | EgoDex | CC BY-NC-ND |
| Force | Goal Force | MIT |
| Robot pose | DROID | CC BY 4.0 |
| Camera control | SolarWM | Source terms; see LICENSE |
Video sampling, resizing and action alignment are performed by the training loader. Source materials retain their original licenses.
- Downloads last month
- 1