Illustration brief the model wrote
A large worktable holds 22 differently shaped robot arms and grippers reaching toward a central pile of assorted objects. Glowing lines connect each robot to a single server rack labeled 'RT-X' in the background. No text, no faces.
Open X-Embodiment: Robotic Learning Datasets and RT-X Models
Aggregates 1M+ trajectories from 22 robots; cross-embodiment RT-X policies achieve positive transfer.
What changed
ModelBefore, robotic learning datasets were mostly single-robot or limited in diversity. This paper released a multi-robot dataset spanning 22 embodiments and showed that training a single RT-X model across all robots improves policy performance, enabling positive transfer between different robot morphologies.
So what, in plain English
ModelThink of it as a shared school library where each robot donates its own homework problems and solutions; now every robot can study from everyone else's work. This lets a robot learn a task it never saw before by drawing on experience from a completely different robot type.
Contribution to industry
ModelThis paper contributed the Open X-Embodiment dataset as a public benchmark and the RT-X model family as pre-trained starting points. Industry teams now use this corpus to bootstrap robot policies, evaluate cross-embodiment transfer, and inform data collection strategies across heterogeneous fleets.
Typed judgments
Jev★★★★★ robots · Physical AI 98%
Real hardware 75% · Open code / data 75% · Foundation-model based 36% · Survey / benchmark-only 3%
Probabilities from TypeSafe Jev over the title, venue and abstract; the filters on the atlas apply code-owned thresholds to these numbers.
Editor's notes
Summary. Assembles more than one million trajectories from 22 robot embodiments into a single dataset and shows RT-X policies trained across them gain positive transfer.
Industry. Gave industry a shared multi-robot corpus plus evidence that cross-embodiment data helps, shaping how teams now mix data across heterogeneous robot fleets.