Reinforcement Learning (RL) enables robots to acquire complex behaviors through trial-and-error interactions, placing simulators at the heart of scalable and efficient training pipelines. This study presents a qualitative comparison of simulators employed for robot training through RL, focusing on both sequential and parallel training architectures. While conventional simulators have advantages in terms of fidelity and sim-to-real, it is limited in terms of training time and generalization. On the other hand, modern simulators that allow parallel training offers significant improvements in terms of learning speed and data variety but also bring an advanced hardware cost. In this context, the study analyzes widely used simulators according to training architecture, training speed, parallel computation capacity, ROS compatibility, accessibility and scientific prevalence. Based on the analysis, a conceptual simulator selection tool was developed using a polychotomous key to assist researchers in identifying the most suitable simulation environment.