This document discusses sim-to-real transfer in deep reinforcement learning. Deep RL can train robots by overcoming data inefficiency and collection costs through potentially infinite simulated data. However, there is performance degradation when transferring policies learned in simulation to the real world due to differences between the environments. Common methods for sim-to-real transfer include domain randomization, domain adaptation, and introducing disturbances to simulations to minimize mismatches with reality. Challenges include determining effective randomizations and unifying feature spaces between domains.