Data Flywheel

A virtuous cycle where a deployed robot policy generates new interaction data, which is used to improve the policy, which generates better data, and so on. The concept is central to real-world robot learning systems — each deployment cycle adds to the training dataset. Self-improvement loops, active learning, and fleet learning are mechanisms that drive the data flywheel.