Initialize DROID shard publication
Browse files- README.md +22 -0
- publication.json +1 -0
README.md
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
viewer: false
|
| 3 |
+
---
|
| 4 |
+
|
| 5 |
+
# DROID episodes
|
| 6 |
+
|
| 7 |
+
Publication status: **uploading — do not use yet**.
|
| 8 |
+
|
| 9 |
+
33,966 episodes in 1,018 uncompressed TAR shards. Episodes are ordered lexicographically by folder name; no episode spans shards.
|
| 10 |
+
|
| 11 |
+
All original file contents and directory names are preserved, including RGB, depth, TCP, actions, observations, calibration and metadata. TAR ownership, permissions and timestamps are normalized. No image/video re-encoding is performed.
|
| 12 |
+
|
| 13 |
+
Use `episodes.jsonl` to map episode IDs to shards and filter language instructions, lab, status or frame count. `shards.jsonl` provides byte sizes and SHA-256 hashes. Each TAR extracts to original episode directories. A subset requires only its corresponding shards, but each selected shard is downloaded in full.
|
| 14 |
+
|
| 15 |
+
Download `droid_hf.py` and `requirements-droid-hf.txt` from this repository. With Python >= 3.10, install the requirements and run:
|
| 16 |
+
|
| 17 |
+
```bash
|
| 18 |
+
python droid_hf.py download --repo-id OWNER/DATASET --shards 0-2 --output-dir ./episodes
|
| 19 |
+
python droid_hf.py download --repo-id OWNER/DATASET --episodes-file ids.txt --output-dir ./episodes
|
| 20 |
+
```
|
| 21 |
+
|
| 22 |
+
The download command verifies and extracts one shard at a time. It removes verified local archives after extraction unless `--keep-archives` is set. Existing differing episode directories are never overwritten.
|
publication.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"complete":false,"episode_count":33966,"format_version":1,"plan_id":"1c4ea6c01cd1b4fbfda774efdb7a039cb8fd06325add28edfdfad12745d01b1e","shard_count":1018}
|