Data flywheels aim for continuous improvement by extracting insights from deployment data. Robot deployments generate massive high‑dimensional rollouts that are hard to interpret and to base data‑collection decisions on. To address this, the paper introduces behavior graphs, which compress policy rollouts into a single interpretable directed graph. The construction pipeline consists of: (1) learning to detect similar behavior modes across rollouts; (2) temporally linking these modes; (3) attaching outcome information to the nodes. In a user study on three real‑world tasks, participants chose data‑collection strategies from rollouts. Sixteen of eighteen preferred behavior graphs over video, reporting a 40.3% reduction in perceived temporal demand and a 37.6% drop in effort while reaching comparable conclusions. Beyond interpretability, behavior graphs serve as progress predictors: across 43 datasets they improve rank correlation with true temporal order by 22.9% and cut mean absolute error by 53.5% relative to the strongest baseline. As the fleet of deployed robots expands, behavior graphs impose structure on high‑dimensional rollouts, making deployment data easier to reason about.
Review: Behavior graphs transform complex robot rollout sequences into an intuitive graph via unsupervised learning, markedly lowering analysis overhead and boosting progress prediction accuracy, offering strong practical value.