Open X-Embodiment

Open X-Embodiment (OXE) is a collaborative effort to pool robot-learning data from many labs into a single, standardized corpus. It aggregates 60+ existing robot datasets contributed by 21 institutions, yielding more than one million real robot trajectories spanning 22 distinct embodiments, from single-arm manipulators to bi-manual setups and mobile robots, and covering hundreds of skills, objects, and scenes.

What’s in it

Each trajectory records observations (typically one or more RGB camera streams, sometimes depth or wrist views) paired with robot actions and natural-language task instructions. To make heterogeneous robots trainable together, all data is converted to a common RLDS (Reinforcement Learning Datasets) format built on TensorFlow tfrecord files, which accommodates differing action spaces, control frequencies, and sensor configurations without discarding the original structure.

Tasks, license & access

OXE is used to train generalist manipulation policies, most notably the RT-1-X and RT-2-X models, which demonstrated positive transfer across robot types, and it underpins research on vision-language-action (VLA) models and cross-embodiment generalization. The collection was led by Google DeepMind with academic and industry partners; component datasets keep their original licenses (many permissive, such as Apache-2.0 or CC BY) and require appropriate citation, so users should check terms per source. Loader utilities and per-dataset readers are published so practitioners can mix subsets at chosen sampling ratios.

• Modality: robot trajectories (vision, proprioception, actions, language)
• Scale: 1M+ trajectories, 60+ datasets, 22 embodiments
• Tasks: imitation learning, generalist/VLA policies, cross-embodiment transfer
• License: per-source (mixed; citation required)
• Creator: Google DeepMind and a multi-lab consortium

By unifying fragmented robot data, Open X-Embodiment is a landmark resource for training foundation models that control physical robots.