Pretrains robot policies on large-scale web video to acquire dynamics and physics understanding instead of static images or VLMs. Yields faster training, better generalization, and superior dexterous manipulation results in real-world tasks.