DROID-Raw

Dataset Description

DROID-Raw is the DROID data that the Video-Point Model (VPM) DROID action policy is trained on (paper · project page · code). It is uploaded exactly as the training run reads it. Nothing is re-encoded, re-sampled or filtered beyond what training does. Three camera views, the raw 8D joint-position actions and states, the language instruction and the non-idle segments come together for every episode.

• 🎥 3 synchronized views per episode: exterior camera 1 (left), exterior camera 2 (right) and the wrist camera, all at 180×320 and 15 fps, stored as JPEG frames
• 🦾 raw 8D actions and states per frame: 7 Franka joint positions (rad) + gripper (0 = open … 1 = closed), not normalized
• 📝 one language instruction per episode
• ✂️ non-idle frame ranges per episode; training windows are sampled only inside them
• 🔢 50,430 episodes / 13.4 M frames per view (248 h) are used for training; the joint tables cover all 95,658 episodes of DROID 1.0.1

Three training episodes (rows), mid-episode frame of each view: exterior 1 (left), exterior 2 (right), wrist.

Relation to the original DROID release

The data comes from the LeRobot conversion of DROID 1.0.1 (lerobot/droid_1.0.1):

1. Frames: each LeRobot chunk video (videos/<camera>/chunk-000/file-NNN.mp4, the left image of each ZED camera, 180×320) is decoded once and stored as one JPEG per frame (quality 90) in a .framecache directory, so a frame can be read without video decoding. Only chunk files that hold episodes used for training are included (604 of them).
2. Actions / states: the LeRobot action and observation.state columns are copied unchanged into two row-aligned float32 tables (joint_actions.npy, joint_states.npy), one row per frame of every episode.
3. Episode index: episodes.npz maps every episode to its rows in those tables and to its first frame in each camera's frame cache, and stores the non-idle ranges and the keep flags described below.

What training uses

An episode is used when all of the following hold. tools/read_episode.py (DroidRaw.usable_rows()) applies the same rules and returns exactly the 50,430 episodes.

filter	episodes left
DROID 1.0.1 (LeRobot)	95,658
keep: successful, not blacklisted (episodes whose upstream video is missing or corrupt), has a non-idle range	55,618
… and a non-empty language instruction	55,618
… and frames cached for all three cameras	50,442
… and at least one non-idle range of ≥ 17 frames (current frame + 16-step action chunk)	50,430

The non-idle ranges come from the OpenPI-style joint-velocity idle detector, with DROID's success and language-present filters applied on top. A training sample is a window whose current frame lies inside one non-idle range, with the 16 future action steps inside the same range. That gives 62,675 ranges and 10,976,733 possible windows.

For reference, our training recipe reads a sample as follows:

• all three views in the fixed order exterior 1, exterior 2, wrist, resized to 192×320 (random-area crop and color jitter as augmentation);
• 16 future actions plus 0, 4 or 8 clean past frames and actions, with the current joint state as a prefix row;
• raw joint actions and states (no normalization).

The gripper is fed as 1 − gripper (1 = open … 0 = closed, the convention of the released VPM-DROID model). The files keep DROID's raw convention (0 = open … 1 = closed).

Dataset Structure

The layout mirrors the one our training loader reads: a LeRobot-style lerobot/droid_1.0.1/videos/ tree next to the episode/action cache oxe_policy_cache/droid/.

DROID-Raw/
├── README.md
├── assets/example.png
├── oxe_policy_cache/droid/
│   ├── meta.json                 # fps, camera keys / sizes, joint-table layout and q01/q99 ranges
│   ├── tasks.json                # language instruction of every episode (list, episode row order)
│   ├── episodes.npz              # episode index, see below
│   ├── joint_actions.npy         # (27,630,375, 8) float32
│   └── joint_states.npy          # (27,630,375, 8) float32
├── lerobot/droid_1.0.1/videos/
│   ├── observation.images.exterior_1_left/chunk-000/file-NNN.framecache/   # 231 files
│   ├── observation.images.exterior_2_left/chunk-000/file-NNN.framecache/   # 230 files
│   └── observation.images.wrist_left/chunk-000/file-NNN.framecache/        # 143 files
│       ├── frames.bin            # concatenated JPEG bytes, one image per frame
│       ├── index.npy             # (N, 2) int64: byte offset, byte length of frame i
│       └── meta.json             # fps, num_frames, JPEG quality, source video path (LeRobot-relative)
└── tools/
    ├── frame_cache.py            # frame-cache reader (numpy + Pillow)
    └── read_episode.py           # episode loader + usable-episode filter + 3-view video export

Data Format

episodes.npz

One entry per episode (95,658 rows, row i = LeRobot episode_index i):

key	shape	content
episode_index	(E,) int64	LeRobot episode index
row_start, length	(E,) int64	the episode's rows in joint_actions.npy / joint_states.npy are row_start … row_start + length − 1; row row_start + k is frame k
dataset_from_index	(E,) int64	global LeRobot frame index of the episode's first frame
keep	(E,) bool	successful, not blacklisted, has a non-idle range
cam_ok	(E, 3) bool	frames of camera j are cached for this episode
cam_chunk, cam_file	(E, 3) int32	frame cache holding the episode: videos/<camera j>/chunk-{cam_chunk:03d}/file-{cam_file:03d}.framecache
cam_base	(E, 3) int64	frame index of the episode's first frame inside that frame cache
range_ep, range_start, range_end	(R,) int64	non-idle ranges: episode row and [start, end) episode-relative frames

The camera order j = 0, 1, 2 is meta.json → camera_keys: exterior 1 (left), exterior 2 (right), wrist.

Actions and states

file	shape	content
joint_actions.npy	(N, 8) float32	LeRobot action: commanded absolute joint positions joint_0 … joint_6 (rad) + gripper (0 = open … 1 = closed)
joint_states.npy	(N, 8) float32	LeRobot observation.state: measured joint positions + gripper position, same layout

meta.json → joint_action lists the per-dimension 1st/99th percentiles of both tables.

Frame caches

Frame i of a .framecache is the i-th frame of the original concatenated chunk video. Episode frame k of camera j is frame cam_base[ep, j] + k of the episode's frame cache. All frames are 180×320 RGB JPEGs (quality 90).

Usage

hf download rooty2020/DROID-Raw --repo-type dataset --local-dir DROID-Raw      # ~984 GB
cd DROID-Raw
python tools/read_episode.py --root . --list 5                  # first usable episodes: row, episode index, length, instruction
python tools/read_episode.py --root . --row 42 --out ep42.mp4   # 3 views side by side (needs imageio[ffmpeg])

import sys; sys.path.insert(0, "DROID-Raw/tools")
from read_episode import DroidRaw

ds = DroidRaw("DROID-Raw")
rows = ds.usable_rows()                      # the 50,430 training episodes
r = rows[0]
actions, states = ds.joints(r)               # (L, 8) each, frame-aligned
instruction = ds.tasks[r]
ranges = ds.nonidle_ranges(r)                # [(start, end), ...]
frames = [ds.frame(r, j, k=0) for j in range(3)]   # exterior 1, exterior 2, wrist; (180, 320, 3) uint8 each

The frame caches can also be read without the helper: index.npy[i] gives the byte offset and length of frame i in frames.bin, and those bytes decode as a JPEG.

Statistics

Episodes (training set)	50,430 (of 95,658 in DROID 1.0.1)
Frames (training set)	13,401,413 per view (248.2 h at 15 fps); 11,979,533 inside non-idle ranges
Episode length (training set)	50–2,449 frames (median 220, mean 265.7)
Non-idle ranges / training windows	62,675 / 10,976,733
Unique instructions (training set)	35,203
Views	exterior 1, exterior 2, wrist; 180×320 JPEG (q90), 15 fps
Frames stored	21.1 M / 21.3 M / 21.5 M (exterior 1 / exterior 2 / wrist), whole chunk files
Action / state	8D raw joint positions + gripper, 27,630,375 rows (all episodes)
Robot	Franka Panda + Robotiq 2F-85
Size	984 GB (frame caches 982 GB, episode/action cache 1.8 GB)

Original DROID Dataset

DROID-Raw is derived from DROID, a large in-the-wild robot manipulation dataset collected with Franka Panda robots across 13 institutions, through its LeRobot conversion lerobot/droid_1.0.1:

• Paper: DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset
• Website: https://droid-dataset.github.io/
• License: CC BY 4.0

Citation

If you use this dataset, please cite the original DROID paper and the VPM paper.

@article{vpm2026,
  title   = {Learning the 4D World through Point Trajectory Forecasting},
  author  = {Jeon, Byungwoo and Hong, Sunghwan and Kim, Dohyeon and Lee, Jonghoon and Park, Jungwoo and Lee, Inhee and
             Baek, Doosan and Kim, Jinil and Lim, Taehyun and Joo, Hanbyul and Pollefeys, Marc and Xie, Saining and
             Shin, Jinwoo},
  journal = {arXiv preprint arXiv:2610.06940},
  year    = {2026}
}

@article{khazatsky2024droid,
  title   = {DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset},
  author  = {Alexander Khazatsky and Karl Pertsch and Suraj Nair and Ashwin Balakrishna and Sudeep Dasari and Siddharth Karamcheti and Soroush Nasiriany and Mohan Kumar Srirama and Lawrence Yunliang Chen and Kirsty Ellis and Peter David Fagan and Joey Hejna and Masha Itkina and Marion Lepert and Yecheng Jason Ma and Patrick Tree Miller and Jimmy Wu and Suneel Belkhale and Shivin Dass and Huy Ha and Arhan Jain and Abraham Lee and Youngwoon Lee and Marius Memmel and Sungjae Park and Ilija Radosavovic and Kaiyuan Wang and Albert Zhan and Kevin Black and Cheng Chi and Kyle Beltran Hatch and Shan Lin and Jingpei Lu and Jean Mercat and Abdul Rehman and Pannag R Sanketi and Archit Sharma and Cody Simpson and Quan Vuong and Homer Rich Walke and Blake Wulfe and Ted Xiao and Jonathan Heewon Yang and Arefeh Yavary and Tony Z. Zhao and Christopher Agia and Rohan Baijal and Mateo Guaman Castro and Daphne Chen and Qiuyu Chen and Trinity Chung and Jaimyn Drake and Ethan Paul Foster and Jensen Gao and Vitor Guizilini and David Antonio Herrera and Minho Heo and Kyle Hsu and Jiaheng Hu and Muhammad Zubair Irshad and Donovon Jackson and Charlotte Le and Yunshuang Li and Kevin Lin and Roy Lin and Zehan Ma and Abhiram Maddukuri and Suvir Mirchandani and Daniel Morton and Tony Nguyen and Abigail O'Neill and Rosario Scalise and Derick Seale and Victor Son and Stephen Tian and Emi Tran and Andrew E. Wang and Yilin Wu and Annie Xie and Jingyun Yang and Patrick Yin and Yunchu Zhang and Osbert Bastani and Glen Berseth and Jeannette Bohg and Ken Goldberg and Abhinav Gupta and Abhishek Gupta and Dinesh Jayaraman and Joseph J Lim and Jitendra Malik and Roberto Martín-Martín and Subramanian Ramamoorthy and Dorsa Sadigh and Shuran Song and Jiajun Wu and Michael C. Yip and Yuke Zhu and Thomas Kollar and Sergey Levine and Chelsea Finn},
  journal = {arXiv preprint arXiv:2403.12945},
  year    = {2024}
}

License

DROID-Raw is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, the license of the original DROID dataset.