NVIDIA opens four Cosmos 3 Nano models for interactive world simulation

The release covers complementary jobs rather than a single do-everything model[1]:

• Cosmos3-Nano-Transfer-Auto generates synchronized multi-view driving data with varied weather, lighting and time of day, including up to 11 camera views and optional joint RGB-LiDAR.
• Cosmos3-Nano-Sim-Auto is the streaming, causally distilled variant that produces successive camera and LiDAR chunks for interactive driving rollouts.
• Cosmos3-Nano-Sim-Transfer turns simulator spatial controls (depth, segmentation, edges, blur) and a text prompt into photorealistic RGB for closed-loop policy work.
• Cosmos3-Nano-Sim-Bimanual predicts video from robot or camera actions for interactive two-arm simulation on YAM and AgiBot embodiments.

All four are post-trained examples of Cosmos 3 Nano[1].

NVIDIA argues robotics and autonomy need environments that stay consistent as scenes and viewpoints change. Transfer and Sim-Auto keep appearance and geometry aligned across cameras and LiDAR; Sim-Transfer leaves physics with the simulator while generating observations a policy can consume; Sim-Bimanual lets a policy or joystick drive the next visual frame from action[1]. Streaming generation uses a Mixture-of-Transformers backbone with temporal-causal attention and a four-stage recipe that ends in self-forcing so the student trains on its own generated history[1].

NVIDIA is shipping open skills, notebooks and post-training recipes so labs can adapt Transfer-Auto to their own sensor rig, distill a Sim-Auto streamer, or retarget Sim-Bimanual to another embodiment[1]. The company is explicit that these problems remain hard: the point of open checkpoints is to measure where the models support policy learning and where they fall short[1].

Closed world-model stacks are hard to compare and expensive to reproduce. Putting multi-view driving, simulator-to-RGB transfer and bimanual action-to-video under one open Cosmos 3 Nano family gives academic and industry labs a shared starting point for physical AI evaluation without rebuilding the foundation from scratch[1][2].