NVIDIA Cosmos Predict, Transfer and Reason

2025-03-18 · Publisher : NVIDIA · Domain : Robotics

At GTC 2025, NVIDIA organizes Cosmos into three families: Predict to generate future world states, Transfer to turn simulations into photoreal video, and Reason, an open physical reasoning model.

What changes

Cosmos Transfer turns segmentation, depth or lidar maps into photoreal video.
Cosmos Predict generates intermediate states between start and end images.
Cosmos Reason reasons about video in natural language with chain of thought.
1X, Agility, Figure AI and Skild AI are among early users.

### Background On March 18, 2025, at GTC, NVIDIA announced a major update to its Cosmos world foundation models, launched at CES in January. The platform is organized into three families with distinct roles, along with two synthetic data generation blueprints for robots and autonomous vehicles. ### Specifications - **Cosmos Transfer**: ingests structured video inputs (segmentation maps, depth maps, lidar scans) and outputs photoreal video, turning one simulation into many realistic variants. - **Cosmos Predict**: generates virtual world states from text, images or video; new versions predict intermediate frames between given start and end images. - **Cosmos Reason**: an open, customizable model with spatiotemporal awareness that uses chain-of-thought reasoning to describe in natural language what will happen (a person stepping into a crosswalk, a box falling from a shelf). - **Cited uses**: 1X uses Predict and Transfer to train NEO Gamma; Skild AI augments its data with Transfer. ### Why it matters NVIDIA now covers the whole physical AI data chain: generate, vary and annotate. Cosmos Reason later became the "slow brain" of subsequent GR00T models. ### Key takeaway GTC 2025 established Cosmos as a pillar of NVIDIA's robotics strategy, alongside GR00T and Isaac.

NVIDIA Cosmos Predict, Transfer and Reason

What changes

Links