Open X-Embodiment

Open X-Embodiment (OXE) unifies open-sourced robotic demonstration data from 60 datasets contributed by 34 research labs into one RLDS (Reinforcement Learning Datasets) episode format for easy downstream training. It spans 22 robot embodiments and over 1 million real robot trajectories, ranging from single-arm manipulators (e.g. Google RT-1 / fractal, Kuka, WidowX/bridge, Franka) to mobile navigation robots. Each dataset is stored as a sequence of episodes; each episode is a sequence of steps carrying an observation (camera image(s) plus proprioceptive robot state), an action (typically a 7-DoF end-effector command: x, y, z, roll, pitch, yaw, and gripper), a natural-language task instruction, a scalar reward, and RLDS boundary flags (is_first / is_last / is_terminal). Data was released by the Open X-Embodiment Collaboration alongside the RT-X models and is downloadable from public Google Cloud Storage buckets (gs://gresearch/robotics/) or via tensorflow_datasets; no login or terms acceptance is required. Repository code is Apache-2.0 and DeepMind-authored materials are CC-BY-4.0, but each constituent dataset retains its own license, so the aggregate license is mixed.