Embodied AI Breakthrough: One-Shot Robot Learning and World Models

GEN-1.5 achieves one-shot in-context learning for robot manipulation from a single 3-12 second demo, no retraining (59% one-shot, 83% with 10 steps), analogous to GPT-3 moment for robotics. This follows other embodied AI advances: Dyna's DYNA-2 World-Action Model achieving human-to-robot scaling law (1.55x task completion over VLA, published in Nature), GEN-0/EgoScale arguing coverage over volume, and HiPHI dataset for humanoid control. The field is moving toward generalist physical agents that learn from minimal data.