BridgeData V2

BridgeData V2 is the second version of the large-scale robot manipulation dataset from the RAIL Lab at UC Berkeley. It contains over 60,000 demonstration trajectories collected using a WidowX 6-DoF robot arm across hundreds of different tasks and visual settings. Each trajectory includes multi-view RGB images, joint positions, end-effector poses, actions, and free-form language task descriptions. The dataset is specifically designed to study generalization in robot learning — how policies trained on diverse data adapt to novel objects, backgrounds, and task configurations. BridgeData V2 significantly improved upon the original BridgeData by adding more scenes, more diverse tasks, and higher-quality demonstrations. It has become one of the most widely used training datasets for generalist Vision-Language-Action (VLA) models, including OpenVLA, and serves as a standard benchmark for evaluating generalist robot policies. The dataset is hosted on TensorFlow Datasets and Hugging Face Datasets, and is widely used in the open-source robotics community for training and evaluating imitation learning models and VLA architectures.