What domain randomization fundamentally trades off

In the domain randomization approach of Tobin et al. (2017), a policy or perception model is trained across simulated environments whose parameters (textures, lighting, object poses, and for dynamics randomization also mass, friction, and latency) are randomly sampled each episode. Which statement best characterizes the *core mechanism and trade-off* of domain randomization for sim-to-real transfer? Use $\theta \sim p(\theta)$ to denote the randomized simulation parameters.