Open X-Embodiment
Aggregate of 60 existing robot datasets from 34 labs, converted to a common RLDS format and spanning 22 embodiments. Most cross-embodiment VLAs pretrain on a subset of it (the "OXE magic soup").
Google DeepMind and 21 partner institutions · Oct 2023 · RobotReport an error
1MEpisodes
–Hours
–Size
527Tasks
Models trained on it11
π0.5Physical IntelligenceVLARoboLab 28.0%
π0Physical IntelligenceVLARoboLab 5.0%
GR00T N1.5NVIDIAVLA
OpenVLAStanford / UC Berkeley et al.VLA
RLDX-1RLWRLDVLA
π0-FASTPhysical IntelligenceVLARoboLab 15.5%
ABot-M0.5AMAP CV LabWorld-action
GalaxeaVLA (G0)Galaxea AIVLA
MolmoAct2Ai2VLASO-101 23.8%
EO-1Shanghai AI LabVLA
FLOWERKITVLA