A new model shows how robots can learn from vast amounts of human video
Dyna-2 demonstrates that scaling first-person recordings of everyday human tasks produces predictable gains in robot performance across embodiments without robot data in pre-training.