DROID: the most diverse robot manipulation dataset with 76,000 demonstrations
In one sentence Stanford, Berkeley, and CMU release DROID, the most diverse robot manipulation dataset ever collected: 76,000 demonstrations, 564 scenes, 86 tasks, 52 robot arms. It enables cross-embodiment generalization and is the reference for robot foundation models.
Training a robot capable of doing many things in many different environments requires data collected in many different situations. Until 2024, the problem was that existing robot datasets were too small, collected in a few laboratories with a few robots, in always-identical environments.
DROID solves this problem at industrial scale. Stanford, Berkeley, and CMU joined forces to collect 76,000 robot demonstrations across 564 different scenes, covering 86 different tasks, using 52 different robot arms. It is the most diverse robot dataset ever created.
Diversity is the key characteristic: data comes from laboratories, kitchens, offices, improvised environments. This variety allows training robots that work in new situations, not just those already seen during training.
DROID was released publicly with all the tools to use it. Any laboratory in the world can now train robot policies starting from a huge and diverse data corpus, instead of having to collect everything from scratch. It is the ImageNet equivalent for robotics.
Companies
Stanford, UC Berkeley, CMU
Tools
—
Tags
Sources