NVIDIA Releases Open-Weight Cosmos 3 World Models for Autonomous Driving and Robotics Simulation

NVIDIA has released four new open-weight checkpoints for its Cosmos 3 world model family, targeting interactive world simulation for autonomous driving and robotics. Based on Cosmos 3 Nano, the post-trained models are released under the OpenMDW1.1 license for commercial and non-commercial use. The models address multi-view driving data synthesis, synchronized camera and LiDAR streaming, simulator-to-RGB transfer, and action-conditioned robot video prediction.

Key takeaways

• Four models cover driving simulation (Transfer-Auto, Sim-Auto), simulator visual rendering (Sim-Transfer), and bimanual robot action prediction (Sim-Bimanual).
• Transfer-Auto supports up to 11 camera views and joint RGB-LiDAR generation, while Sim-Auto distills it for fast streaming rollouts with driving controls.
• Sim-Transfer converts physics simulator spatial controls into RGB observations, and Sim-Bimanual predicts environmental changes directly from two-arm robot actions across embodiments like YAM and AgiBot.
• The models use a Mixture-of-Transformers architecture with temporal-causal attention and a four-stage training recipe, including distillation and self-forcing to prevent compounding errors.