Feature learning is very fundamental in DL. For RL, we can make better decisions if we can extract features strongly relevant to the problem domain. In the meta-learning, we may consider learning the feature is also an important part to improve learning efficiency. (Rather just focus on calculating the rewards) So in controlling a robot, we can use camera images as input, derive features from it and then apply RL on those features. This process usually involves a separate learning step to learn what features are. This is a very active research area. I have an article related to it but not release yet. I hope I will resume publishing the remaining RL articles when time is allowed.