New RL method handles diverse human operator data.
problem Handling data from multiple, unknown policies.
method Domain-Invariant Model-based Offline RL (DIMORL) with Risk Extrapolation (REx).
result Models trained with REx generalize better across different demonstrators.