README.md

---
license: mit
task_categories:
  - robotics
  - video-classification
tags:
  - robotics
  - world-model
  - video-prediction
  - imitation-learning
  - depth-estimation
  - affordance
  - bimanual
  - real-robot
size_categories:
  - n<1K
---

# DeVA — YAM (processed)

Real bimanual-robot data collected on the [YAM](https://www.i2rt.com) platform and
processed for [DeVA](https://github.com/Mq-Zhang1/deva), a robot video-world model
with affordance + depth physical guidance.

| | |
|---|---|
| Episodes | 90 |
| Frames | 30,749 |
| Video | 480 × 1920, 30 fps (3 views tiled side by side) |
| Views | `left_wrist`, `head`, `right_wrist` |
| Action dim | 14 (bimanual, delta actions at train time) |
| Proprio state | included (`state.json`) |
| Size | 19 GB |

## Download

```bash
hf download mengqz9/deva_yam --repo-type dataset --local-dir datasets/deva_yam
```

## Layout

```
deva_yam/
├── dataset_info.json   # view names, tile grid, affordance/depth encodings
├── videos/             # <episode>.mp4   tiled multi-view RGB
├── metas/              # <episode>.txt   language instruction
├── t5_xxl/             # <episode>.pickle  precomputed T5-XXL embedding [n_tok, 1024]
├── action.json         # {episode: [[a_0..a_13], ...]}  len == video frames
├── state.json          # {episode: [[s_0..s_13], ...]}  proprioceptive state
├── norm_stats.json     # action normalization statistics (quantile)
├── affordance/         # <episode>.npz   physical guidance (optional at train time)
└── depth/              # <episode>.npz   physical guidance (optional at train time)
```

`videos/` tiles the views on the grid given by `dataset_info.json::tile`:

```
[[left_wrist, head, right_wrist]]
```

The gripper channels of `action.json` are **binarized** (open/closed), which is
intentional for this dataset.

## Affordance / depth `.npz`

Both are keyed by **canonical view name** (never by tile position), one array per view:

| | affordance | depth |
|---|---|---|
| key | `<view>` | `<view>` |
| dtype | `float16` | `float16` |
| shape | `[T, 128, 128]` | `[T, 480, 640]` |
| range | raw, un-normalized | `[0, 1]` |
| note | loader resizes to 128×128, then per-frame max-normalizes | per-view per-clip min-max applied at export |

`T` equals the episode's video frame count, so aux arrays index by absolute frame.
The validity mask is derived from the tile grid (`null` cells are invalid) and is not stored.

**Depth source** — DepthAnything (ViT-L) disparity, converted as `depth = 1 / disparity`;
**higher value = farther**. Clipped at export (percentile not recorded); ~5% of pixels
saturate at 1.0.

**Affordance source** — [UAD](https://unsup-affordance.github.io/) pseudo-labels rendered as
Gaussian heatmaps. Note that the real-robot affordance footprint is noticeably broader
than the simulator oracle labels used for LIBERO / RoboCasa.

## Citation

If you use this data, please cite DeVA:

```bibtex
@article{zhang2026deva,
  title   = {{DeVA}: Decoupled Video-Action Model with physical guidance for robot policy learning},
  author  = {Zhang, Mengqi and Khose, Sahil and Kareer, Simar and
             Song, Yuchen and Jain, Unnat and Hoffman, Judy},
  journal = {arXiv preprint arXiv:2607.24159},
  year    = {2026}
}
```