My long-term goal is to create robust artificial agents that can meaningfully interact with their environments and I am mainly interested in the field of Deep RL. Applying tabula rasa RL on complex real-world problems is extremely challenging and sample inefficient. One research direction I find interesting is to develop better techniques for incorporating prior information into RL algorithms, which could be in the form of imperfect demonstrations, differentiable physics, priors acquired via meta-learning etc. Another obstacle to applying RL algorithms to real-world tasks is the lack of suitable reward functions. I am also fascinated by the research direction of improving off-policy RL. In general, I believe that deep RL can be improved by inheriting ideas from supervised learning.