Domain Randomization

A training technique that randomizes simulator parameters (textures, lighting, friction, dynamics) so that policies generalize to the real world. Often combined with learned world models to expand the distribution of training scenarios beyond what real data alone can provide.