Definition
DAgger (Dataset Aggregation) addresses the distribution shift problem in behavior cloning. In standard BC, the learner only sees states from the expert distribution, but during deployment it encounters novel states due to its own imperfect actions. DAgger iteratively runs the current policy, collects new states, asks the expert to label them with correct actions, and retrains. This progressively covers the states the learner actually visits. Variants include SafeDAgger (with safety constraints) and HG-DAgger (with human-gated interventions).
Why It Matters for Robot Teams
Understanding dagger is essential for teams building real-world robot systems. Whether you are collecting demonstration data, training policies in simulation, or deploying in production, this concept directly affects your workflow and system design.