World Action Models (WAMs) are proposed as a successor to Vision-Language-Action (VLA) models for robot manipulation. While VLAs use vision-language model backbones that describe the world, WAMs use video world models that predict how the world evolves, giving them a physics prior. NVIDIA's Cosmos 3 is presented as an open foundation model for building WAMs, featuring a Mixture-of-Transformers architecture trained on 767M images, 348M videos, and 8M action samples. Two post-trained policy models are available: Cosmos3-Nano-Policy-DROID (16B) for workstation serving and Cosmos3-Edge-Policy-DROID (4B) for on-device deployment on Jetson hardware. Benchmarks show the omni-pretrained checkpoint improved RoboLab success rates from 28.1% to 36.8% over a base checkpoint, demonstrating the value of the physical prior. WAMs require less task-specific data, generalize better to unseen environments, and adapt faster to new robot embodiments.
