DYNA introduces Dyna-2 world-action model trained on 1M hours.
It reveals surprising scaling laws, enables cross-embodiment transfer, and showcases new capabilities like language following and exceptional robustness, achieving 87% quality and throughput in new environments.