The robots do not have to be humanoid robots but he wants the bots to go into a very, very messy room and clean it up and prepare a meal for two and set the table. There would be a prepared clean room, the meal cooked and candles lit.

James Fan talks about having ten of thousands of randomized environments to get enough training data. The massive simulation enables years of simulated training to enable one shot first time successful real world performance of tasks. This is for walking, shooting a basketball or spinning a pencil.

1.5 million parameters is all that is needed to capture the sophistication of the human body for humanoid robot training.

Nvidia predicts that everything that moves will be autonomous.