Kosmic Worlds V1
Completed training run
Our first video world model: a 1.6B-parameter transformer trained on 3,600 hours of video. A public demo and technical appendix document the run.
About Kosmic Worlds
World models for
physical systems.
We develop action-conditioned world models for robotics and autonomous driving.
Our focus is long-horizon simulation and physical grounding: generating future observations that reflect the actions taken in a scene.
Explore our models
Our work
Models, data,
and validation.
We build the models and the data tools needed to train them. V1 established our video-modeling stack. V2 is being developed around explicit actions, longer rollouts, and physical consistency.
Explore V1
Kosmic Worlds V2
In development
Our next-generation world action model. Planned variants V2D for driving and V2R for robotics target actuator-level action conditioning, synthetic policy training data, and closed-loop simulation.
Kosmic PACE
Public demonstration
Our internal control-extraction tool supports action labeling for driving data. The public demonstration compares its estimates against recorded vehicle controls.
Founder
Shaan Vats
Founder, Kosmic Worlds
Building the models and infrastructure behind Kosmic Worlds.
Shaan leads Kosmic Worlds’ work on video world models and action-conditioned simulation, bringing together model development, training systems, and physically grounded action data.
Get in touchDesign partnerships
Built with teams
shipping autonomous systems.
6 design partners
Our design partners span electric and autonomous freight, sensing, building inspection, autonomous drones, robot foundation models, and legged robotics. Their applications help inform the scenarios and control requirements we work toward.
Meet our design partnersResearch and partnerships
Work with Kosmic Worlds.
Talk with us about world models, simulation, and training data for robotics or autonomous driving.