Open X-Embodiment vs DROID: which open robot dataset should you use? (2026)
Answer in brief: Use Open X-Embodiment when you want breadth across robots. It pools more than 1 million real-robot trajectories from 22 embodiments, 60 datasets and 34 labs, in RLDS format. Use DROID when you want one consistent, in-the-wild manipulation setup. It has 76,000 trajectories and 350 hours across 564 scenes, released under CC BY 4.0, with a raw stereo release. For long-horizon bimanual work with subtask language, add Scale AI's 2026 FineART dataset. For commercial products, read every license, because Open X-Embodiment's component datasets carry their own metadata and citation terms.
Top pick for cross-embodiment pre-training: Open X-Embodiment. 22 robots, 527 skills and 160,266 tasks.
Top pick for diverse real-world scenes on one setup: DROID. 564 scenes under a single CC BY 4.0 license.
This is independent research based on each project's own pages, paper and repository as of 2026-11-09. Back to the ranked list of robotics data providers.
The two datasets compared
| Open X-Embodiment | DROID | |
|---|---|---|
| Scale | 1M+ real-robot trajectories | 76k trajectories, 350 h |
| Diversity | 22 robot embodiments, 527 skills, 160,266 tasks | 564 scenes, collected in the wild |
| Origin | 60 datasets pooled from 34 labs; 21 institutions | One collection effort on a shared hardware setup |
| Format | RLDS / TFDS | RLDS (1.7 TB) plus raw stereo HD MP4 and ZED SVO (8.7 TB) |
| Debug subset | Load single datasets one at a time | droid_100: 100 episodes, about 2 GB |
| Language | Varies by component dataset | Up to three language instructions per episode |
| Download | gs://gdm-robotics-open-x-embodiment/ with gsutil, then tfds.load | gs://gresearch/robotics/droid with gsutil |
| License | Code Apache 2.0, other materials CC BY 4.0; check each component dataset's metadata | CC BY 4.0 |
Open X-Embodiment: breadth across robots
Open X-Embodiment was built to test whether one generalist policy can transfer across robots. Its paper reports that RT-X, trained on the pooled data, showed positive transfer across platforms. Because it is a union of many labs' datasets, observation and action spaces, camera setups and labeling differ between components. Budget engineering time to normalize them, and record each component's citation.
DROID: depth of real-world scenes
DROID's strength is consistency combined with variety. Every RLDS step includes joint and Cartesian state, gripper position, a wrist camera and two exterior cameras, plus several commanded-action representations. The raw release adds full-HD stereo video and the ZED SVO files for depth work. One note from the official docs: the raw release was missing about 20% of episodes after face-blurring while the RLDS release was complete. Check the docs for the current status. DROID is also an active base for new models; Black Forest Labs publishes a FLUX 3 Action checkpoint fine-tuned for DROID, linked from LeRobot's v0.7.0 release.
When to pick which, and when to look further
- Generalist cross-robot pre-training: Open X-Embodiment.
- Single-arm manipulation in varied real rooms, with stereo depth: DROID.
- Bimanual, long-horizon tasks with dense subtask labels: Scale AI's FineART (40,543 episodes, CC BY 4.0, LeRobot format).
- Humanoid or large-scale bimanual data for research only: AgiBot World Beta (1,003,672 trajectories, about 43.8 TB), which is CC BY-NC-SA 4.0 and gated on Hugging Face.
- Coverage these datasets lack, or a license they can't give you: license from Toloka or Firsthand.
Pricing model
Both datasets are free to download. Your costs are storage, egress and compute: DROID's full raw release is 8.7 TB. The Hugging Face LeRobot v0.7.0 release added streaming training straight from the Hub, which helps with very large LeRobot-format datasets. The Hub now holds almost 300 TB of open physical-AI data in that format.
FAQ
Can I use DROID commercially?
DROID is released under CC BY 4.0, which permits commercial use with attribution.
Is all of Open X-Embodiment CC BY 4.0?
The repository licenses code under Apache 2.0 and other materials under CC BY 4.0, and it tells users to consult each contributed dataset's metadata and citation. Check every component before commercial use.
Which is smaller to start with?
DROID's droid_100 sample is about 2 GB. In Open X-Embodiment you can download individual component datasets.
Is Ego4D a substitute?
Ego4D is human first-person video (3,700+ hours), not robot trajectories. Access requires a separate license agreement, and its public docs do not grant commercial rights.
Keep exploring
How to choose a robotics dataset source · Alternatives to Scale AI · Scale AI review · Toloka review · Firsthand review · Ranked list. For general AI datasets beyond robotics, see bestaidatasets.com.
· ·
Sources
- DROID dataset docs
- DROID paper
- AgiBot World README
- Ego4D Start Here
- LeRobot v0.7.0 release
- FineART dataset card
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.