Abstract: Online Reinforcement Learning requires access to the environment for trials-and-error. Offline Reinforcement Learning learns from existing logged trajectories (i.e. observational studies) but must either weaken the learning goals or making unrealistic assumptions. Are there any meaningful setting in between? The talk starts by discuss the statistical complexity and limitation of offline RL, then review a burgeoning problem of **low-adaptive exploration** which addresses these limitations by providing a sweet middle ground between offline and online RL. Somewhat surprisingly, we show that only O(log-log T) batches of umbarrassingly parallel access to the environment is needed for us to solve exploration with near-optimal sqrt(T) regret (up to log factors). We also discuss the influence of function approximation, two-player game and other settings such as pure and reward-free) exploration.