Rigid-body ready Embodied / Robotics Mixed (check per asset) Open

Open X-Embodiment Cross-Embodiment Robot Dataset

Provider：Google DeepMind with 21/34 partner institutions (Open X-Embodiment Collaboration)

Open X-Embodiment unifies 60 source datasets into RLDS form with 1M+ real trajectories, 22 embodiments and 527 skills under a mixed Apache-2.0 / CC-BY licensing.

Get it from the source →

At a glance

Readiness
L1 · Rigid-body ready
Rationale
Manipulation L1: 60 source datasets are unified into RLDS/TFDS form providing more than one million real robot trajectories, each with a 7-DoF end-effector action, multi-view images and a language instruction, directly usable for policy 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
More than 1 million real robot trajectories across 22 embodiments, 60 source datasets, 527 skills and 160,266 tasks (The project site states 1M+ trajectories across 22 embodiments pooled from 60 existing datasets collected by 34 robotics labs; the paper abstract describes a collaboration between 21 institutions. The 527-skill and 160,266-task figures agree across both. The 21-institution and 34-lab counts are two different framings and are not reconciled here. Checked 2026-10-26.)
License
Mixed (check per asset) — The aggregation repository licenses in two layers: all software under Apache License 2.0 and all other materials under CC-BY 4.0, per the official README. Each of the 60 source datasets keeps its own original license, some of which are research-use only, so the per-dataset spreadsheet must be checked before using any subset; the aggregate protocol cannot be assumed to apply.
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

Open X-Embodiment is the largest open collection of real robot demonstrations assembled to date. Instead of releasing a single lab's recordings, the collaboration pooled 60 existing datasets from 34 labs across 22 robot embodiments, covering 527 skills and 160,266 task instances, and converted everything into the common RLDS episode format loadable through TensorFlow Datasets.

Readiness assessment

Every trajectory is expressed in one action space: a 7-dimensional end-effector vector of x, y, z, roll, pitch, yaw and gripper opening, with dimensions a particular robot never exercises set to zero during training. Observations carry multi-view images, proprioceptive state and a natural-language instruction, so the data can be consumed directly by policy learning pipelines such as OpenVLA, pi-zero, Octo and GR00T N1.

Practical notes

Licensing has two layers: the repository code is Apache 2.0 and the other materials are CC-BY 4.0, but each of the 60 source datasets keeps its own original license, some restricted to research use, so the official dataset spreadsheet must be checked per subset. Trajectory counts are also unevenly distributed, with a few large platforms contributing the majority of episodes, which is worth accounting for when mixing sources.

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.