World Action Models and Robotics Breakthroughs

Dyna-2 world action model trained on 1M+ hours of human video achieves human-to-robot scaling law, 13-minute fine-tuning for complex tasks, 87% zero-shot deployment vs 46% for teleoperation-heavy approaches. Scaling ego-centric data works well, and all labs expected to switch to WAMs by EOY. SimWAM for autonomous driving achieves SOTA with lower latency. This signals near-term commercial opportunities in WAM infrastructure and ego-centric data pipelines, and validates video-based pretraining for robotics.