Deep reinforcement learning methods require active data collection, where the model actively interacts with its environment. A data-driven paradigm for reinforcement learning will enable us to pre-train and deploy agents capable of sample-efficient learning in the real-world.