Evaluating Domain Randomization in Deep Reinforcement Learning Locomotion Tasks

Abstract

Domain randomization in the context of Reinforcement learning (RL) involves training RL agents with randomized environmental properties or parameters to improve the generalization capabilities of the resulting agents. Although domain randomization has been favorably studied in the literature, it has been studied in terms of varying the operational characters of the associated systems or physical dynamics rather than their environmental characteristics. This is counter-intuitive as it is unrealistic to alter the mechanical dynamics of a system in operation. Furthermore, most works were based on cherry-picked environments within different classes of RL tasks. Therefore, in this work, we investigated domain randomization by varying only the properties or parameters of the environment rather than varying the mechanical dynamics of the featured systems. Furthermore, the analysis conducted was based on all six RL locomotion tasks. In terms of training the RL agents, we employed two proven RL algorithms (SAC and TD3) and evaluated the generalization capabilities of the resulting agents on several train–test scenarios that involve both in-distribution and out-distribution evaluations as well as scenarios applicable in the real world. The results demonstrate that, although domain randomization favors generalization, some tasks only require randomization from low-dimensional distributions while others require randomization from high-dimensional randomization. Hence the question of what level of randomization is optimal for any given task becomes very important.