DROID

DROID (Distributed Robot Interaction Dataset) is a large-scale, in-the-wild robot manipulation dataset developed by a collaboration led by Alexander Khazatsky, Karl Pertsch, Sergey Levine, Chelsea Finn, and over 80 co-authors from multiple institutions. The dataset contains 350 hours of diverse task demonstrations collected across 50+ environments. DROID features a standardized hardware and software setup deployed across multiple laboratories, enabling consistent data collection at scale. The dataset includes multi-camera observations (exterior and wrist cameras), robot states, actions, and natural language task instructions. Key features of DROID include: 87,000+ demonstration trajectories, diverse real-world environments (homes, offices, workshops, kitchens), tasks ranging from simple pick-and-place to complex multi-step manipulation, and standardized data format compatible with training VLA models. The DROID dataset provides a standard benchmark for evaluating generalization in robot manipulation — including scene, object, and lighting variations. The dataset is hosted on Google Research's cloud storage and is accessible via TensorFlow Datasets.