NVIDIA pushes Cosmos further for physical AI with Transfer 2.5, Predict 2.5, and Reason 2
Original: Autonomous vehicles, humanoid robots - even surgical ones — depend on massive, diverse, and physics-aware datasets for training. Better models generate better data to train even better downstream models. NVIDIA Cosmos is at the center. New model checkpoints are up to: Scale synthetic data generation with realism and diversity faster with Transfer 2.5; Improve long-tail scenario coverage using Predict 2.5; Faster and advanced vision reasoning with Reason 2. Read how Cosmos WFMs are setting the foundation for the next generation of robots and autonomous systems. View original →
What NVIDIA highlighted on X
On March 20, 2026, NVIDIA said new Cosmos checkpoints are aimed at one of the hardest problems in physical AI: creating enough diverse, physics-aware data and reasoning capability to train real robots and autonomous systems. The post grouped the update into three pieces: Cosmos Transfer 2.5, Cosmos Predict 2.5, and Cosmos Reason 2.
The framing is strategically important. NVIDIA is not talking only about faster model inference. It is positioning Cosmos as the infrastructure layer that links simulation, synthetic data generation, future-world prediction, and embodied reasoning. That makes the announcement relevant well beyond robotics research labs.
What the NVIDIA Technical Blog confirms
The linked NVIDIA Technical Blog says the next generation of humanoids, autonomous vehicles, and other physical AI systems depends on high-fidelity, physics-aware training data. The article then details how each Cosmos update is supposed to move that stack forward.
- Cosmos Transfer 2.5 is described as faster and more scalable data augmentation from simulation and 3D spatial inputs, helping teams create more photorealistic and varied scene data across lighting, environment, and composition changes.
- Cosmos Predict 2.5 focuses on future-world generation for sequences up to 30 seconds and, according to NVIDIA, can reach up to 10x higher accuracy after post-training on proprietary or domain-specific data. It also adds multiview outputs and custom camera layouts.
- Cosmos Reason 2 adds stronger spatiotemporal understanding, object localization in 2D and 3D, reasoning explanations, and long-context support up to 256K input tokens.
The article repeatedly connects these changes to physical AI workflows in which simulation data, perception, scenario generation, and downstream decision systems all need to reinforce each other. In NVIDIA’s view, better world models do not just interpret physical environments; they also help create the training data used to improve later-stage systems.
Why this matters
This update matters because physical AI has a data bottleneck that looks very different from web-scale language modeling. Robots and autonomous machines need reliable coverage of lighting changes, rare edge cases, motion dynamics, sensor layouts, and real-world physics. Collecting enough real data is slow and expensive, so synthetic generation and world modeling become core infrastructure rather than optional research extras.
NVIDIA is using Cosmos to occupy that infrastructure layer. If Transfer improves photoreal generation, Predict improves scenario rollouts, and Reason improves grounded multimodal interpretation, then Cosmos becomes a platform play for training and validating downstream robotics systems. That helps explain why the company is emphasizing world foundation models as central to the next generation of physical AI instead of treating them as a side project.
Sources: NVIDIA AI Developer X post · NVIDIA Technical Blog
Related Articles
NVIDIA announced an open Physical AI Data Factory Blueprint on March 16, 2026. The reference architecture is designed to unify data generation, augmentation, and evaluation for robotics, vision AI agents, and autonomous vehicles, with GitHub availability expected in April.
NVIDIA on March 16, 2026 introduced its Physical AI Data Factory Blueprint, an open reference architecture for generating, augmenting, and evaluating training data for robotics, vision AI agents, and autonomous vehicles. The company says the stack combines Cosmos models, coding agents, and cloud infrastructure from partners such as Microsoft Azure and Nebius to lower the cost and time of physical AI training at scale.
NVIDIA said on March 16, 2026 that ABB, FANUC, Figure, KUKA, Skild AI and other robotics players are building on Cosmos, Isaac and GR00T to move physical AI from simulation into production. The release also introduced Cosmos 3, Isaac Lab 3.0 in early access, and commercial early access for GR00T N1.7.
Comments (0)
No comments yet. Be the first to comment!