RoboNet

RoboNet is an open database for sharing robotic experience, proposed by researchers from UC Berkeley, Stanford University, University of Pennsylvania, and CMU (Sudeep Dasari, Frederik Ebert, Stephen Tian, Suraj Nair, Bernadette Bucher, Karl Schmeckpeper, Siddharth Singh, Sergey Levine, Chelsea Finn), published at CoRL 2019. The dataset contains 15 million video frames collected from 7 distinct robot platforms including Franka, Kuka, Sawyer, Baxter, and other common research platforms across 4 institutions. It was one of the first large-scale efforts to unify robotic manipulation data across different hardware platforms into a common format. RoboNet addresses a fundamental tension in robot learning: how to learn generalizable controllers without collecting impractically large amounts of data per experiment. By pooling data across robots, the dataset enables cross-embodiment transfer learning — a model pretrained on RoboNet and fine-tuned on a held-out robot can outperform training from scratch using 4x-20x more task-specific data. The paper evaluated two learning paradigms on RoboNet: visual foresight using forward video prediction models, and supervised inverse models. Experiments tested generalization across new objects, tasks, scenes, camera viewpoints, grippers, and entirely new robots. RoboNet was foundational in demonstrating that multi-robot, multi-institution data sharing can dramatically improve generalization in robotic manipulation. It inspired subsequent large-scale datasets like Open X-Embodiment and established the paradigm of cross-embodiment robot learning that underpins modern VLA models.