Dataset Aggregation

Dataset aggregation combines examples from multiple sources or collection rounds to train models and policies with broader coverage. In imitation learning, DAgger-style methods repeatedly collect policy rollouts and add expert or human corrections, while shared control and uncertainty-guided review aim to make those corrections more useful and efficient. Aggregated labels and benchmark records can also support learning and evaluation, but pooling may hide differences among annotators, users, domains, or scoring rules. A key tension is balancing the robustness of larger combined datasets against information loss, distribution bias, and the cost of collecting reliable labels.