Vision-language-action models can understand a robot’s instructions—but their vision-language backbones are trained to describe the world, not predict how it will change after the robot acts.
World Action Models start with a video world model instead, giving robot policies a stronger foundation for physical reasoning. Whether you’re comparing against a VLA or post-training a policy with your own robot data, this post explores how to:
- Learn from diverse interaction data and generalize to new motions and environments.
- Build on NVIDIA Cosmos 3, an open omni-model foundation with native support for action, video, image, audio, and text.
- Generate robot actions alongside an optional imagined video of the predicted outcome.
- Deploy Cosmos3-Nano-Policy-DROID on a workstation or Cosmos3-Edge-Policy-DROID on-device.
In an evaluation using the same training recipe, data, and compute, initializing from the Cosmos 3 omni checkpoint increased RoboLab success from 28.1% to 36.8%.
Choose how you want to get started:
📖 Read the Technical Blog
🤗 Explore the Cosmos 3 Model Collection
🧑💻 Get Models and Code on GitHub
📺 Tune in to the Cosmos Labs Livestream on August 13
Building a policy with your own robot data? Share your embodiment, dataset, VLA baseline, results, or current blocker with the community in the Cosmos Discord channel.