Instead of making a simulator accurate, domain randomization makes it varied: masses, friction coefficients, motor latencies, textures, lighting, and camera poses are all resampled every episode, and a single policy must succeed across the whole distribution. If the randomization is broad enough, the real world is effectively just another sample, and the policy transfers without ever training on real data.
The technique powered the landmark sim-to-real results in dexterous manipulation and locomotion. Its limitation is dimensionality: randomization covers low-dimensional gaps like friction and latency well, but cannot enumerate open-ended variation like every object in a kitchen, which is a core reason simulation-trained policies conquered legged locomotion before general manipulation.