---
license: mit
task_categories:
  - robotics
  - video-classification
tags:
  - robotics
  - world-model
  - video-prediction
  - imitation-learning
  - depth-estimation
  - affordance
  - robocasa
size_categories:
  - 1K<n<10K
---

# DeVA — RoboCasa (processed)

Processed **RoboCasa** data for [DeVA](https://github.com/Mq-Zhang1/deva), a robot
video-world model with affordance + depth physical guidance. Derived from
[robocasa/robocasa](https://github.com/robocasa/robocasa) —
**please cite the original benchmark as well**.

| | |
|---|---|
| Episodes | 1,199 |
| Frames | 316,995 |
| Video | 448 × 448, 16 fps (3 views on a 2×2 grid) |
| Views | `agentview_left`, `agentview_right`, `eye_in_hand` |
| Action dim | 7 (absolute) |
| Proprio state | not included |
| Size | 56 GB |

## Download

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

## Layout

```
deva_robocasa/
├── 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_6], ...]}  len == video frames
├── data_mapping.json   # episode -> source task
├── norm_stats.json     # action normalization statistics
├── 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` — the
bottom-right cell is empty:

```
[[agentview_left, agentview_right],
 [eye_in_hand,    null           ]]
```

## 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, 224, 224]` | `[T, 224, 224]` |
| 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**. No percentile clip is applied: the range is set by the
farthest pixel, so agentview content occupies roughly the lower half of `[0, 1]` and
never saturates.

**Affordance source** — simulator oracle contacts rendered as Gaussian heatmaps.

## Citation

If you use this data, please cite DeVA and RoboCasa:

```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}
}

@inproceedings{robocasa2024,
  title     = {RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots},
  author    = {Nasiriany, Soroush and Maddukuri, Abhiram and Zhang, Lance and
               Parikh, Adeet and Lo, Aaron and Joshi, Abhishek and
               Mandlekar, Ajay and Zhu, Yuke},
  booktitle = {Robotics: Science and Systems (RSS)},
  year      = {2024}
}
```