Kosmic Worlds
Our Mission
We are building world models that give robotics and autonomous-driving teams more ways to learn from experience, train policies, and test decisions in simulation.
Explore our modelsWhat we’re building toward
Simulation can expand the situations a system encounters during development. Our goal is to make those experiences useful for both learning and evaluation.
01 / Training
Generate varied experiences paired with physical actions, giving teams data for training driving policies and robot controllers.
Explore changes in environments, motion, and operating conditions without collecting every variation individually.
02 / Validation
Use closed-loop simulation to study how a policy’s actions change what happens next.
Revisit scenarios, vary conditions, and compare behavior to understand where a system succeeds and where it needs more work.
Research priorities
Our research connects generated observations with the controls, time horizons, and physical behavior that matter to autonomous systems.
Steering, throttle, and brake for driving. Joint commands for robotics. The aim is to condition a rollout on controls that a policy can produce and a machine can execute.
A useful simulation needs to retain scene information as a sequence unfolds. We focus on preserving geometry, objects, and state across longer rollouts.
The relationship between an action and its outcome needs to be credible. We work toward generated motion and observations that remain consistent with physical constraints.
Putting the mission into practice
V2 is in development. V2D and V2R are planned variants for the different controls and environments these systems use.
Explore V2 developmentDriving · V2D · Planned
Synthetic training data and closed-loop simulation for end-to-end driving policies, using steering wheel angle, throttle percentage, and brake percentage.
Robotics · V2R · Planned
Action-conditioned experiences and closed-loop simulation for robot policies, using joint angles, velocities, trajectories, and torques.
Research and partnerships
Help define the scenarios, controls, and evaluation needs that useful world models should support.