Comparison of Different Domain Randomization Methods for Policy Transfer in Reinforcement Learning

摘要

Deep reinforcement learning algorithms have made great progress in the field of control with the help of many high-efficiency simulation environments. However, due to the difference in state distribution and dynamics, these algorithms trained in the simulation cannot be effectively applied to the real world. The ability to reduce the impact of the Sim2Real gap is critical for transferring policy from the simulation to the real world. Although there are many methods for studying the Sim2Real problem, it is difficult to evaluate the performance of different algorithms due to the different evaluation platforms and evaluation metrics. In this paper, we construct a uniform robot navigation scenario, and revisit the ability of the popular domain randomization methods to transfer the policies from the simulation to the real world under the dynamics gaps. With the analysis of the performence in the simulation environment and the real world, we provide some recommendations of the domain randomization methods and hope to make these methods more efficient to use.