license: mit
task_categories:
- robotics
tags:
- robotics
- manipulation
- droid
- lerobot
- imitation-learning
size_categories:
- 10B<n<100B
DROID 1.0.1 — native 720p, packed (LeRobot v3.0)
A native 1280×720 re-encode of the DROID robot-manipulation
dataset in LeRobot v3.0 format, with videos packed into ~500 MB files like the original
lerobot/droid_1.0.1. It reuses droid_1.0.1's
exact tabular data (states / actions / camera extrinsics / task instructions) and replaces the
low-resolution (320×180) video with full 1280×720 frames decoded from the original DROID raw MP4s
(left-eye mono per camera).
What's different from lerobot/droid_1.0.1
lerobot/droid_1.0.1 |
this dataset | |
|---|---|---|
| Video resolution | 320×180 | 1280×720 (native) |
| Codec | h264 | h264 (CRF 20) |
| GOP / keyframes | sparse | g=2 (dense — fast random-access decode) |
| Episodes | 95,658 | 72,112 (only those with public raw video) |
| Video packing | packed (~500 MB files) | packed (~500 MB files) |
| Tabular data | full | identical |
Everything else — feature schema, fps (15), robot state/action tensors, camera extrinsics, task
instructions, episode structure — matches droid_1.0.1. Each per-episode video has exactly length
frames (frame-exact with the tabular data).
Coverage & indexing
Only the 72,112 / 95,658 episodes with publicly available DROID raw video are included (the rest are
absent from the public raw bucket). Episodes are re-indexed 0…72,111; index_to_droid_1_0_1.json
maps each new index back to its original droid_1.0.1 episode index.
Usage
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("XinxuanLuM/droid_720p", video_backend="pyav")
frame = ds[0] # observation.images.{wrist_left,exterior_1_left,exterior_2_left} at 720×1280
A lower-resolution 640×360 packed variant is available at
XinxuanLuM/droid_360p.
Provenance & license
Derived from the DROID dataset (Khazatsky et al., 2024) and lerobot/droid_1.0.1. Please cite DROID and
respect its terms. Re-encode & repackaging only; no new annotations.