Open Source RoboticsRobots, software & learning

DROID (Distributed Robot Interaction Dataset)

Software

Large-scale in-the-wild robot manipulation dataset and data-collection platform distributed in RLDS (1.7 TB) and raw stereo HD video (8.7 TB) formats via Google Cloud Storage. [2] [4]

Installation & first result →Official websiteSource codeGetting started

For
Robot learning researchers; Manipulation dataset engineers; Teams building multi-camera teleoperation rigs [2] [4]
Hardware
DROID robot platform (18 standardized setups documented in hardware setup guide), ZED stereo cameras (SVO recordings… [2] [4]
Source rights
Some licenses are unresolved. Review licenses
Unknowns
operating system not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
License
not fully assessed for software, datasets, documentation · Details by part

Requirements · Research details and dates

On this page

Overview

DROID (Distributed Robot Interaction Dataset) distributes large-scale, in-the-wild robot manipulation episodes on Google Cloud Storage alongside the data-collection platform repository (droid-dataset/droid) used to set up 18 collection rigs. It targets visuomotor policy training, and teams adopting it need to separate three practical issues up front: the storage jump between RLDS and raw stereo HD formats, an apparent action-dimension discrepancy in the documented RLDS schema, and the absence of an explicit license file in the main repository and dataset guide. [1] [2] [4]

Is this for you?

Workflows
  • Interactive dataset visualization and Colab inspection
  • RLDS download via gsutil (droid_100 or full droid)
  • Raw stereo HD / non-stereo HD download via gsutil
  • Hardware assembly, camera calibration, and teleoperated data collection

[2] [4]

Known limitations
  • droid-dataset/droid has no root LICENSE file and pyproject.toml omits a license declaration; neither the repository nor the inspected dataset guide states a dataset-wide reuse license.
  • Upstream documentation notes that ~20% of raw episodes were missed during face-blurring and copying to the raw bucket (while RLDS is reported complete).
  • Full dataset downloads require substantial storage (1.7 TB for RLDS; 5.6 TB non-stereo raw; 8.7 TB stereo HD raw) and gsutil access.

[1] [3] [4]

What you need

Languages / runtimes
  • Python

[3]

Dependencies
  • gsutil (for GCS dataset download)
  • Python >=3.7
  • mujoco==2.3.2
  • dm-control==1.0.5
  • opencv-python==4.6.0.66
  • protobuf==3.20.1
  • h5py
  • open3d
  • zerorpc

[3] [4]

Host computer

Local storage of 2 GB for droid_100 debugging, 1.7 TB for full RLDS, 5.6 TB for non-stereo raw HD, or 8.7 TB for full stereo HD raw data.

[4]

Compute for the workload
  • 2 GB disk for droid_100 RLDS debug subset
  • 1.7 TB disk for full RLDS dataset
  • 5.6 TB disk for raw non-stereo HD video
  • 8.7 TB disk for full raw stereo HD + SVO dataset

[4]

Skills
  • Python
  • Google Cloud gsutil
  • RLDS / TensorFlow Datasets
  • HDF5 / multi-camera video handling

Inferred from gsutil download commands, RLDS schema, and pyproject.toml dependencies.

[3] [4] · Editorial assessment

Installation & first result

Open the upstream getting-started guide [4]

  • Install gsutil and start with the 2 GB droid_100 RLDS sample (gs://gresearch/robotics/droid_100) before downloading the 1.7 TB full RLDS or 5.6–8.7 TB raw buckets.
  • Inspect the RLDS observation/action_dict schema (180x320 RGB images, 7-DoF joint/gripper states, 6x joint velocity + 1x gripper position action) and verify alignment with your policy loader.
  • For policy training, use the separate MIT-licensed droid_policy_learning repository rather than the data-collection repo.

[2] [4] [5]

Installation

Download RLDS or raw subsets via gsutil from gs://gresearch/robotics/; install the data-collection package from droid-dataset/droid (Python >=3.7) or use droid_policy_learning for PyTorch/JAX training loaders.

[2] [3] [4]

Hardware, models & data

Supported hardware
  • DROID robot platform (18 standardized setups documented in hardware setup guide)
  • ZED stereo cameras (SVO recordings in raw data)

[2] [4]

Interfaces
  • Google Cloud Storage (gsutil)
  • RLDS / TFDS episode schema
  • HDF5 trajectory files (trajectory.h5)
  • MP4 and ZED SVO camera recordings

[4]

Inputs and outputs

RLDS episodes contain 180x320 uint8 RGB views (wrist_image_left, exterior_image_1_left, exterior_image_2_left), proprioceptive states (joint_position, cartesian_position, gripper_position), three language_instruction strings, and 7-D action vectors (6x joint velocities + 1x gripper position) alongside full action_dict commands.

[4]

File formats
  • RLDS / TFDS
  • HDF5 (trajectory.h5)
  • MP4 (mono and stereo HD)
  • ZED SVO
  • JSON episode metadata

[4]

Project notes

Start with droid_100 before committing terabytes

The dataset documentation hosts four gsutil download routes under gs://gresearch/robotics/: [4]

  • droid_100 (2 GB): 100 RLDS episodes intended for debugging loaders before pulling the full corpus.
  • droid (1.7 TB): the full dataset in RLDS format with 180×320 RGB images.
  • droid_raw non-stereo HD (5.6 TB): gsutil rsync excluding *SVO* and *stereo*.mp4 files.
  • droid_raw full stereo HD (8.7 TB): full-HD stereo MP4 videos, raw ZED .svo camera files, trajectory.h5 low-dimensional logs, and metadata_*.json episode metadata.

Two caveats matter before choosing raw over RLDS. First, the dataset guide explicitly warns that roughly 20% of episodes were missed during face-blurring and copying to the raw bucket, whereas the RLDS version is reported as complete; this review did not download the bucket to verify whether that backfill has since finished. Second, droid-dataset/droid is the robot platform and collection codebase (pinning older packages such as mujoco==2.3.2, dm-control==1.0.5, opencv-python==4.6.0.66, and protobuf==3.20.1 in pyproject.toml), while PyTorch and JAX policy-training loaders live in the separate droid-dataset/droid_policy_learning repository. [2] [3] [4]

Check action conventions in the RLDS schema

Every RLDS step exposes three 180×320 uint8 RGB camera views (wrist_image_left, exterior_image_1_left, exterior_image_2_left), proprioceptive states (joint_position [7], cartesian_position [6], gripper_position [1]), and three natural-language instruction fields (language_instruction, language_instruction_2, language_instruction_3). The default 7-D action tensor is documented as [6x joint velocities, 1x gripper position], while action_dict additionally stores commanded Cartesian and joint positions and velocities—including a 7-D joint_velocity tensor and a 6-D cartesian_velocity tensor. That apparent discrepancy between six joint velocities in the action comment and seven dimensions in action_dict.joint_velocity is unresolved in the inspected guide; default action semantics were not verified by loading episodes here, so inspect the loader code in droid_policy_learning or confirm with the maintainers before interpreting the vector as robot commands. [2] [4]

What we checked

Primary-source metadata, README, pyproject.toml, dataset documentation, and companion policy-repo LICENSE inspection only. No GCS data was downloaded, no RLDS episodes were parsed, and no hardware or training code was executed.

Not yet established

  • Requirements: operating system

Updates

Upstream last updated
Sep 15, 2025 [1]
Latest release
Not established
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
droid-dataset/droid@33ae6a67274f36d2e29525b86f23a56616ef43a7 · combined hardware and software repository

Source files and licenses

High-value multi-scene manipulation dataset with documented GCS downloads and RLDS schema, included as partial because droid-dataset/droid lacks a root LICENSE file and the inspected dataset guide does not state an explicit dataset reuse license.

LayerAvailability / license
softwareavailable · License not fully assessed
droid-dataset/droid has no root LICENSE file and pyproject.toml omits a license field. The separate companion repo droid-dataset/droid_policy_learning declares MIT. [1] [3] [5]
datasetsavailable · License not fully assessed
Public GCS download paths (gs://gresearch/robotics/droid, droid_100, droid_raw) are documented, but neither droid-dataset/droid nor the inspected dataset guide states an explicit dataset license. [4]
documentationavailable · License not fully assessed
Public developer and dataset documentation available; no explicit documentation license notice found on the inspected pages. [2] [4]

Scoped to the DROID dataset and collection platform; no pretrained model checkpoints are hosted in droid-dataset/droid.

License gaps remain unresolved in primary sources

At pinned commit 33ae6a67274f36d2e29525b86f23a56616ef43a7 (2025-09-15), droid-dataset/droid has no root LICENSE file and pyproject.toml declares no license field. The inspected DROID dataset guide (the-droid-dataset) also omits an explicit dataset-wide license statement. By contrast, the companion droid-dataset/droid_policy_learning repository carries a root MIT LICENSE file. Do not assume that MIT on the policy repo—or third-party mirrors—grants blanket reuse terms for the collection repo or raw/RLDS episodes without verifying rights directly with the maintainers. [1] [3] [4] [5]

No GCS buckets were downloaded, no RLDS or HDF5 files were parsed, and no collection or training scripts were executed in this review. The concrete next step is pulling the 2 GB droid_100 split via gsutil, inspecting its schema against your policy loader, and clearing dataset licensing requirements before redistributing derived artifacts. [4]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history