Open X-Embodiment (OXE)

Aggregated RLDS dataset of 1M+ real-robot trajectories across 22 embodiments from 60 datasets; basis for RT-X and OpenVLA.

Dataset specifications
Year2023
Episodes1,000,000
Embodimentsfranka_panda, kuka_iiwa, multi, sawyer, ur5, xarm, widowx
Modalitiesdepth, ee_pose, language, proprioception, rgb
Task categoriesmanipulation
LicenseCC-BY-4.0
Accessopen — commercial use permitted
MaintainerOpen X-Embodiment Collaboration (led by Google DeepMind)

Open X-Embodiment is a large-scale, standardized aggregation of robot learning datasets assembled by a collaboration of 21 institutions, pooling 60 existing datasets from 34 labs covering 22 robot embodiments and over 1 million real robot trajectories demonstrating 527 skills (160,266 tasks). All datasets are converted to a common RLDS (Reinforcement Learning Datasets) episode format loadable via TensorFlow Datasets, and the corpus was used to train the cross-embodiment RT-1-X and RT-2-X generalist policies. It is widely used to train cross-embodiment models such as OpenVLA.