Large language models can train on the internet. Robots cannot. Physical AI must learn from scarce real-world interactions involving vision, touch, motion, and physics. Collecting that data is slow and expensive.