Open X-Embodiment: Robotic Learning Datasets and RT-X Models

ABSTRACT

Large, diverse datasets have catalyzed breakthroughs in natural language and computer vision, yet robotics has struggled to build generalist models due to the fragmented nature of hardware platforms and data silos. We present the Open X-Embodiment dataset, a massive aggregation of robotic learning data encompassing over 1 million real-world trajectories from 22 distinct robot embodiments across 34 academic and industrial laboratories. Utilizing this unprecedented dataset, we train RT-X (Robotics Transformer X), a family of cross-embodiment foundation models. We demonstrate that co-training on diverse robotic data significantly enhances the performance and generalization capabilities of individual robot platforms, advancing the field toward truly universal robotic controllers.