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

It has been historically very challenging to collect large scale datasets for robotics. In this talk, we'll describe our journey in creating the Open X-Embodiment dataset, an initial and on-going attempt at pooling together robotic datasets from the academic community in an unified format. So far, the dataset contains 1M+ robot trajectories from 34 institutions, spanning 22 robot embodiments and more than 300 scenes. We’ll also describe the RT-X family of models, a series of large cross-embodiment policies we trained on the data that demonstrate strong transfer results. We’ll end with a discussion of future plans for continually expanding the Open X-Embodiment dataset, and enabling access to RT-X models.