Train a robot in the real world: it learns to handle its specific environment but fails in new ones. Train in simulation: it generalizes better but the simulation gap — the mismatch between simulated and real physics — degrades performance. The standard tradeoff.
Choi et al. break it by scaling simulation diversity. Using generative 3D world models and a language-driven scene designer, they create hundreds of interactive environments — not hand-crafted, but generated. Simulation success rises from 9.7% to 79.8%. Real-world transfer improves from 21.7% to 75%.
The key variable is not simulation fidelity but simulation diversity. Each individual simulated scene is imperfect. But the diversity across scenes teaches the robot features that transfer to the real world — because the features that are consistent across many imperfect simulations are the features that are also present in reality. What varies across simulated scenes is noise. What persists is signal.
The correlation between scene diversity and zero-shot generalization is direct. More diverse training environments produce better transfer without any real-world fine-tuning. The robot never needs to see the real world during training — it just needs to see enough different fake worlds that the invariant structure of real-world manipulation emerges from the variation.
The structural insight: the sim-to-real gap is not a property of any single simulation's inaccuracy. It is a property of training on too few simulations. Diversity is the bridge that fidelity cannot build. One perfect simulation teaches the robot about one world. A thousand imperfect simulations teach it about the world.