DROID: A Large-Scale In-the-Wild Robot Manipulation Dataset

DROID is arguably the largest and most diverse open-source robot manipulation dataset available to date. Released by researchers at Stanford and UC Berkeley for RSS 2024, it aggregates over 350 hours of teleoperated robot demonstrations collected across 22 different robot platforms operating in 71 distinct real-world households and laboratories. Each trajectory includes high-resolution stereo RGB video, proprioceptive joint states, end-effector poses, gripper forces, and text task descriptions. The dataset is intentionally designed for diversity rather than scale on any single task. It covers a wide variety of everyday manipulation primitives: opening drawers, picking fruits, pouring liquids, tying knots, and repositioning household objects. By collecting across many different robot morphologies (single-arm, dual-arm, mobile manipulators, gripper types) and many different environments, DROID provides the kind of breadth needed to train generalist policies rather than narrow task-specific ones. The authors benchmark several imitation learning methods on DROID, showing that policies pretrained on the full dataset and then fine-tuned on a small target task outperform policies trained from scratch. The dataset includes standardized train/validation splits, preprocessed observation sequences, and an evaluation server. A PyTorch dataloader and dataset viewer are provided, making it straightforward to integrate DROID into existing training pipelines. DROID fills a critical gap in robot learning: while simulation datasets can scale to billions of samples, they lack the visual and physical diversity of the real world. DROID provides a real-world counterpart at a scale large enough to train meaningful generalist models. It has become a standard benchmark for imitation learning and VLA pretraining, alongside Open X-Embodiment. For researchers building the next generation of generalist robot policies, DROID is an essential starting point.