Rigid-body ready Embodied / Robotics CC BY Open

DROID In-the-Wild Robot Manipulation Dataset

Provider：Stanford, UC Berkeley and TRI with 13 institutions (DROID Dataset Team)

DROID provides 76k in-the-wild teleoperated trajectories / 350 hours over 564 scenes and 86 tasks on unified hardware in RLDS form, with CC-BY 4.0 data licensing.

Get it from the source →

At a glance

Readiness
L1 · Rigid-body ready
Rationale
Manipulation L1: 76k real teleoperated trajectories are unified into RLDS/TFDS form with three stereo RGB-D cameras, 7-DoF joint state, end-effector pose and language instructions, directly usable for diffusion-policy or VLA training; but the assets are real-world video and action logs rather than simulation-loadable 3D models, hence L1 as a real behavior data layer, consistent with how BridgeData V2 and RLBench are graded.
Scale
76k real teleoperated trajectories / 350 hours across 564 scenes and 86 tasks, collected by 50 data collectors over 12 months (Two task counts circulate: the project site and the v1 paper abstract say 86 tasks, while the v2 abstract of arXiv 2403.12945 says 84; neither is reconciled here. Institution counts differ as well, with the site stating 13 institutions on one identical hardware setup and the official repository README stating 18 DROID robot platforms were assembled during collection. Releases also come in several sizes, including a 1.7TB RLDS build, 8.7TB raw stereo and 5.6TB raw non-stereo data. Checked 2026-10-26.)
License
CC BY — The data is licensed CC-BY 4.0, based on the CC BY 4.0 label on the arXiv 2403.12945 paper page and the cc-by-4.0 dataset card of the official Hugging Face mirror KarlP/droid. The project site and download page do not spell out a license string, and the companion policy-learning repository droid_policy_learning is MIT, which covers code only and must not be read as the data license. Many third-party LeRobot conversions list apache-2.0, but that is a conversion-script default rather than an upstream data term.
Access
Open
Formats
other
Runtimes
other
Physics properties
Collision geometryJoint structureSensor attributes
Domain
Other
Published
2026-10-26
Updated
2026-10-26

Why it is listed

DROID takes the opposite design route to large aggregation projects: instead of pooling whatever datasets exist, it fixed one hardware rig and shipped it to every collection site. The platform is a Franka Panda 7-DoF arm, two adjustable ZED 2 stereo cameras, a wrist-mounted ZED Mini stereo camera and an Oculus Quest 2 headset for teleoperation, all mounted on a portable height-adjustable desk so scenes could be changed quickly. That standardization yields a dataset with a uniform observation shape, a uniform action convention and a uniform language annotation format, which is why it drops into policy-learning pipelines with far less preprocessing than heterogeneous corpora.

Readiness assessment

Episodes are distributed in the RLDS/TFDS layout already used by Open X-Embodiment, so the same loader path covers both. Every episode carries multi-view RGB-D imagery, proprioceptive state, end-effector pose, gripper width and a natural-language instruction; a 100-episode DROID-100 sample is also mirrored for quick experiments. With 564 real scenes spanning offices, kitchens, labs and homes, the dataset is unusually long-tailed in task description, which makes it a standard pre-training corpus for models such as OpenVLA and pi-zero.

Practical notes

The data license is CC-BY 4.0, while the policy-learning code repository is MIT and covers code only; several community LeRobot conversions list apache-2.0 as a default and do not reflect the upstream term. Camera calibration for about 36k episodes was improved in an April 2025 update, and a December 2024 update added three natural-language annotations for 95 percent of the roughly 75k successful episodes, so it is worth pinning the release you use. The raw stereo and non-stereo dumps are an order of magnitude larger than the RLDS build, so download sizing matters on shared clusters.

Disclaimer: license and usage terms follow the official source — verify them yourself before use. This site indexes and links only; it does not host or mirror any asset file.

License and usage terms are those of the official source — verify before use. No asset file is hosted or mirrored on this site.