alexzms's picture
Update README.md
ae64413 verified
|
Raw
History Blame Contribute Delete
1.95 kB
---
license: apache-2.0
language:
- en
---
# FastvideoWorldModel-MC
This dataset contains short Minecraft (MineDojo) episodes. Each episode is a folder with a video and optional JSON action annotations.
## File layout
```text
<dataset_root>/
episode_<timestamp_utc>_w<worker_id>_<episode_index>_<run_name>/
manifest.json
video.mp4
mg/
mg_actions.json
data.npz # optional (advanced users)
```
## `video.mp4`
- H.264 (libx264)
- FPS and resolution are recorded in `manifest.json` (`fps`, `image_size_hw` where `image_size_hw = [H, W]`)
- The video has **96 frames**
## `mg/mg_actions.json`
Per-frame action annotations (one entry per timestep).
Format:
- Top-level: JSON object (dict)
- Keys: `"0"` ... `"T-1"` (string timestep)
- Value per timestep:
- `action`: `List[int]` (raw MineDojo MultiDiscrete action vector)
- `move_action`: `str` in `{ "", "W", "A", "S", "D", "WA", "WD", "SA", "SD" }`
- `view_action`: `str` in `{ "", "up", "down", "left", "right", "up_left", "up_right", "down_left", "down_right" }`
Notes:
- `view_action` is derived from the per-frame yaw/pitch delta.
- Convention: `dyaw > 0 => right`; `dpitch > 0 => down` (common Minecraft convention).
## Advanced (Pro): `data.npz` (optional)
Some episodes also include `data.npz` (created with `numpy.savez_compressed`). It contains **per-frame aligned** numpy arrays.
Keys (all aligned to the video frames; typically `T = 96`):
- `actions`: `int64`, shape `(T, D)`
- raw MineDojo `MultiDiscrete` action vector per frame
- `D` matches `len(manifest["action_nvec"])`
- `positions`: `float32` (sometimes `float64`), shape `(T, 3)`
- Minecraft world position `(x, y, z)`
- `yaws`: `float32` (sometimes `float64`), shape `(T,)`
- camera yaw in **degrees**
- `pitches`: `float32` (sometimes `float64`), shape `(T,)`
- camera pitch in **degrees**
If you only need video + `mg_actions.json`, you can ignore `data.npz`.