---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- droid
- openpi
- federated-learning
---

# droid8 (openpi format)

38,979 episodes of [DROID](https://droid-dataset.github.io/) for an 8-client federated-learning study:
each client is **one DROID building** (the 8 buildings with the most usable episodes) with **all** its usable episodes
(successful, a real language instruction, >= 2 s). 9,601,563 frames at 15 Hz, 25,686
distinct instructions.

| client | building | episodes | episode_index range |
|---|---|---|---|
| c0 | Gates | 8,863 | 0–8862 |
| c1 | TRI | 5,665 | 8863–14527 |
| c2 | BWW | 5,051 | 14528–19578 |
| c3 | KAIST | 4,699 | 19579–24277 |
| c4 | BAIR | 3,976 | 24278–28253 |
| c5 | Autolab | 3,981 | 28254–32234 |
| c6 | 403 Fairmont | 3,493 | 32235–35727 |
| c7 | RAIL | 3,251 | 35728–38978 |

## Download

LeRobot v2.1 stores one file per episode per camera (~117k files), so the episodes ship as **39 tar shards**
(`shards/droid8_chunk-XXX.tar`, uncompressed: data + videos of 1,000 episodes each). Unpack them into the dataset root:

```python
import glob, os, tarfile
from huggingface_hub import snapshot_download
d = os.path.join(os.environ["HF_LEROBOT_HOME"], "DecentVLA/droid8_openpi")
snapshot_download("DecentVLA/droid8_openpi", repo_type="dataset", local_dir=d, max_workers=8)
for t in sorted(glob.glob(f"{d}/shards/*.tar")):
    with tarfile.open(t) as tf:
        tf.extractall(d, filter="data")
    os.remove(t)
```

## Contents

* **Source:** [`lerobot/droid_1.0.1`](https://huggingface.co/datasets/lerobot/droid_1.0.1) (DROID 1.0.1, Apache-2.0).
  Per-client source episode indices: `meta/decent_vla_clients.json`. Note: in that release the data files' `episode_index`
  column is not unique (42,644 values name two different episodes in two files); episodes here were taken by DROID's own
  episode table (global `index` in [dataset_from_index, dataset_to_index) of each episode's data file) and re-indexed 0..N-1.
* **Format (after unpacking):** LeRobot v2.1 as read by openpi's pinned lerobot. Videos `exterior_image_1_left`
  (← exterior_1_left) and `wrist_image_left` are 180x320 AV1. `joint_position` (7) and `gripper_position` (1); `actions` (8) =
  ABSOLUTE joint-position targets + gripper position; `task` = `language_instruction`.
* **Norm stats:** `assets/droid8/norm_stats.json` (openpi format), for the joint-position recipe. State = joints + gripper;
  actions = 15-step chunks with the 7 joints as deltas to the current state.
* **Made by:** decent-vla `scripts/droid_to_openpi.py` (commit `69c379f`, tag `droid8-v1`). Recipe:
  `adapters/openpi/_droid.py` (`train_config: pi05_droid_jointpos`). Runbook: `examples/droid/README_droid8.md`.

Please cite DROID (Khazatsky et al., "DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset", RSS 2024).