RT-X / RT-1-X
Open X-Embodiment robot models. Language-conditioned manipulation across 22 robot types.
Overview
RT-X provides trained robot models (RT-1-X) from the Open X-Embodiment initiative. Models accept RGB images and task strings, output 7-DOF gripper actions. JAX and TensorFlow checkpoints available. Apache 2.0.
Architecture & Data
- 1M+ episodes, 22 robot types
- RGB + task string → 7D actions (x, y, z, roll, pitch, yaw, gripper)
- JAX and TensorFlow checkpoints
- 33 contributing institutions
Official Links
- robotics-transformer-x.github.io — Project site
- github.com/google-deepmind/open_x_embodiment — Dataset & models
Citation
See the project site for BibTeX and paper references.