Base model · In development
Kosmic Worlds V2
General world modeling with action-conditioned generation, temporal context, and physical grounding as core research priorities.
World Action Model · In development
Kosmic Worlds V2
Our next-generation world action model, designed for action-conditioned generation over longer horizons.
Building on V1’s training and inference stack, with physically grounded controls, longer context, and a hybrid generation architecture.
Development priorities
What changes in V2.
- 01
Physically grounded actions
Condition generation on actuator-level actions. V2D is planned around steering wheel angle, throttle, and brake inputs. V2R is planned around joint angles, velocities, trajectories, and torques.
Action conditioning - 02
Longer temporal context
Train on substantially longer sequences and develop a persistent representation of the world, so earlier scene information remains available during extended rollouts.
Long-horizon consistency - 03
Autoregression with flow matching
Combine causal autoregressive generation with a flow-matching objective for visual refinement. This addresses the blurring and accumulated prediction errors observed in V1.
Visual quality - 04
Stronger data preparation
Improve filtering before tokenization, increase native training resolution, and prepare action-aligned video data for domain-specific training.
Training data
These are development goals. V1 results are reported separately on the V1 page and in its technical appendix.
Action-conditioned generation
Observations and actions,
modeled together.
A world action model is intended to predict how a scene changes under a supplied action sequence. This supports rollouts directed by a robot or driving policy.
Past frames + scene descriptionActuator-level actionssteering, throttle, brake · joint commands
Model in developmentKosmic Worlds V2
Future frames
Policy training data
How V2 is conditioned
Model family
A base model.
Domain-specific training.
V2 is being developed as the shared foundation. Fine-tuned variants will adapt that foundation to the observations and control inputs of a particular domain.
Driving variant · Planned
Kosmic Worlds V2D
A driving-specific fine-tune intended for generating synthetic training data for E2E driving policies and closed-loop simulation.
Robotics variant · Planned
Kosmic Worlds V2R
A robotics-specific fine-tune intended for action-conditioned rollouts, synthetic policy training data, and closed-loop simulation, using joint angles, velocities, trajectories, and torques.
Scalable action labeling
PACE supports
action conditioning.
Kosmic PACE is our internal physics-based control-extraction tool. It estimates steering and pedal inputs from egocentric driving video to help prepare action labels for model training.
The public demonstration compares these estimates with recorded vehicle controls.
Research and partnerships
Work with us on V2.
For technical discussions, design partnerships, and robotics or autonomous-driving applications, contact our team.
contact@kosmicworlds.com