Open world models and robotics

NVIDIA opens four Cosmos 3 checkpoints for interactive robotics and driving simulation

Cosmos3-Nano variants generate multi-view driving observations, simulator-conditioned video and robotic-action forecasts; licenses and physical fidelity still require scrutiny.

Open Model Weights published16 Nov 2026
Primary sourceNVIDIA — Cosmos 3 interactive world simulation technical blog
Source published2026-11-05
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original conceptual artwork of a luminous autonomous vehicle path, layered camera panels and cyan LiDAR scanning rings inside a three-dimensional simulated environment.
Original conceptual illustration of multi-sensor driving and robot world simulation, not footage from an NVIDIA demonstration or actual robotic performance. AI-generated editorial illustration; not a documentary photograph.

NVIDIA published four new Cosmos 3 Nano open-weight checkpoints in its November 5 technical release, aiming to make interactive world simulation more accessible to physical-AI researchers. The systems address different inputs and outputs rather than acting as interchangeable general-purpose language models. NVIDIA identifies each as a post-trained example of Cosmos 3 Nano with 16 billion trainable parameters. Their practical value lies in generating visual observations conditioned on traffic scenes, simulator states or robot actions, while the accuracy of those simulated consequences still needs to be established for each application. Primary source ↗

Cosmos3-Nano-Transfer-Auto is designed for offline synthetic driving data and can generate synchronized views across up to eleven cameras, including fisheye configurations, with joint RGB and LiDAR options. Cosmos3-Nano-Sim-Auto moves toward sequential interaction by streaming observations as the driving-control inputs change. Researchers could use these outputs to explore robustness across weather, lighting and traffic conditions, but a visually plausible rendered road scene is not evidence of physically correct vehicle behavior.

The remaining two variants address different forms of controllability. Cosmos3-Nano-Sim-Transfer turns simulator controls and text conditioning into rendered RGB observations, supporting workflows built around an existing physics simulation. Cosmos3-Nano-Sim-Bimanual predicts scenes following two-arm robot or camera actions, with a specific action representation documented by NVIDIA. Combining camera motion, action timing and observed state is a useful research direction, although error accumulation over long sequences and real-world policy transfer are major unresolved evaluation questions.

NVIDIA lists the releases under OpenMDW 1.1, which permits commercial and non-commercial activity subject to license conditions rather than carrying the unconditional implications sometimes associated with permissive software licenses. The published deployment recipe uses Linux, NVIDIA GPU infrastructure and supporting containers; a demonstrated GB200 environment should not be taken as a universal minimum specification. Open Model Weights treats this as four identifiable source-published checkpoint variants, not proof of real-robot performance or independent world-model fidelity.

OMW REGISTRY WATCH

What this changes for the evidence layer

These four Cosmos3-Nano post-trained checkpoints are not a general claim that simulated actions transfer faithfully to hardware. Verify each exact repository revision, OpenMDW 1.1 licensing terms and runtime dependencies before any curated model update or hardware deployment.

PRIMARY SOURCE

NVIDIA — Cosmos 3 interactive world simulation technical blog

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗