Kosmic Worlds
Our models
We develop models that generate future observations from visual context and actions. V1 established our video-modeling stack. V2 builds toward longer, more controllable rollouts.
V1 · Completed training run
Our first video world model.
A causal spatiotemporal transformer trained on egocentric video. Generate from text, an image, or a video, with a learned latent-action pathway.
V2 · In development
Our next-generation World Action Model.
Development focuses on reliable action-conditioning, long-horizon simulation, and physical grounding. Planned variants include V2D for driving and V2R for robotics.
Development priorities
Continuous controls with meaning in the real world.
Scene information retained across a rollout.
Autoregressive generation with flow matching.
Research and partnerships
We work with robotics and autonomous-vehicle teams on the data and control requirements for world-model training.
Contact us