Open X-Embodiment Unites 60 Robot Datasets to Advance Robot Learning

November 16, 2026
Open X-Embodiment Unites 60 Robot Datasets to Advance Robot Learning
  • Open X-Embodiment brings together 60 datasets from 21 institutions, creating a shared robot-learning resource with more than one million real-world robot trajectories across 22 embodiments, spanning 527 skills and 160,266 tasks.

  • The collection is designed to support robot-policy pre-training and co-training, benchmarking in a common format, and testing RT-1-X on recorded episodes; its data is organized as robot-run episodes in RLDS format.

  • The project also released RT-1-X, trained on the pooled data, and RT-2-X, based on the RT-2 vision-language model; the authors report RT-1-X outperformed original methods by about 50% in small-data domains, while RT-2-X scored about three times better than RT-2 on emergent-skill evaluations, with results dependent on their evaluation setup.

  • The RT-1-X checkpoint takes a single RGB image from a workspace camera and a task description, then returns seven gripper-related values: x, y, z, roll, pitch, yaw, and gripper opening; depending on the dataset, these values may represent absolute positions, changes, or velocities.

  • There are important input and action-format limits: RT-1-X does not accept wrist- or in-hand-camera images or depth, and researchers should confirm predicted and ground-truth actions use the same value convention and consult dataset-specific documentation.

  • Researchers can run RT-1-X inference on recorded episodes with a TensorFlow Colab, or use the JAX checkpoint from the project’s Google Cloud Storage bucket with the repository’s Flax example.

  • To explore the data, users can use the repository’s Open X-Embodiment Datasets Colab, load datasets through TensorFlow Datasets using names in the dataset spreadsheet, or download them with gsutil if TensorFlow Datasets returns DatasetNotFoundError.

  • Users are asked to cite the Open X-Embodiment collaboration and each dataset they use; the dataset spreadsheet includes metadata and dataset-specific citations.

  • The repository notes that the project is not an official Google product; its software is licensed under Apache 2.0 and other materials under CC-BY 4.0.

Summary based on 1 source


Get a daily email with more AI stories

More Stories