Cosmos3-Nano-Sim-
Bimanual
Provide camera or robot actions to generate the next visual observation.
Model CardInteractive World Simulation with Cosmos for Physical AI
Worlds that respond to control, streamed in real time.
One post-training recipe.
Simulate an interactive world.
The same post-training recipe turns Cosmos 3 into real-time, streaming interactive world models. Use the resulting models to simulate camera and robot actions, generate RGB observations from spatial controls, or produce synchronized driving views, one video chunk at a time.
Cosmos3-Nano-Sim-Bimanual
Simulate the world from actions.
Predict visual dynamics from camera navigation and robot actions, using generated history to stream the next observation as new controls arrive.
Frame 1 / 901
Input Camera / 3D
Input Camera / 3D
Input Camera / 3D
Input Camera / 3D
Frame 1
Four generated robot rollouts with synchronized 3D input actions. Select an example to enlarge it and read its input caption.
Higher video quality and reference-frame PSNR than DreamDojo 2B and 14B on AgiBot-World and AgiBot-GEAR.
Cosmos3-Nano-Sim-Transfer
Turn spatial controls into visual worlds.
Turn depth, segmentation, edges, or blurred video into detailed RGB observations, using language to control appearance while the simulator maintains physical state.
Depth-conditioned RGB
Spatial guidance sets the robot's motion and scene layout. The model generates a new appearance from language.
- Appearance direction
- Scandinavian greige cabinetry, sandstone, and a blue toy mouse; Tuscan gray cabinetry, white Corian, and a green toy steamroller.
- Generated rollout
- Two examples · 20 seconds · 30 FPS
Segmentation-conditioned RGB
Semantic regions guide the scene's structure. Language changes the materials, colors, and surroundings of the generated observation.
- Appearance direction
- Mustard farmhouse cabinets, emerald granite, and a lavender unicorn; carved mahogany cabinets, light ash wood, and a green wheeled turtle.
- Generated rollout
- Two examples · 20 seconds · 30 FPS
Two examples per control type, with simulator controls on the left and generated RGB on the right. Language guides each appearance.
Stronger spatial control across all four evaluated modalities, including 59.2% higher edge F1 and 26.5% lower depth error than the Cosmos 3 GA reference.
Cosmos3-Nano-Sim-Auto
Generate a shared world across views.
Generate synchronized multi-view RGB for autonomous driving, using language, spatial controls, and available sensor context to condition each new video chunk.
Loading…
Click a camera or LiDAR to enlarge. Drag LiDAR to orbit.
← → Step
Explore four driving scenes with Transfer Auto and Sim Auto, eleven synchronized cameras, WSM overlays, and interactive LiDAR.
Start simulating with Cosmos.
Choose the model that matches your actions, spatial controls, or camera views.
Cosmos3-Nano-Sim-
Cosmos3-Nano-Sim-
Transfer
Provide spatial controls and an appearance prompt to generate detailed RGB video.
Model CardCosmos3-Nano-Sim-
Auto
Provide driving scene controls to generate synchronized video across camera views.
Model CardExplore the technical report and use the models for interactive world simulation.