Turning Compute Into Data for Physical AI With NVIDIA Cosmos
An NVIDIA presenter explains how Cosmos uses compute-generated data to train and evaluate physical AI systems.
In this talk
- 0:01
Compute as a source of data
Physical AI must handle a world that is vast and unpredictable. The presenter argues that real-world data cannot scale to meet that need, so computation can provide training data.
- 0:14
How Cosmos combines inputs and models
Cosmos is described as an open model for physical AI, built with a Mixture-of-Transformers design. It processes pixels, actions, sound, and language, then combines reasoning and planning with video generation.
- 0:37
Uses for perception and simulation
Developers can adapt Cosmos for different embodiments and tasks. It can interpret scenes and identify relevant events, generate controllable synthetic video, and simulate plausible outcomes from actions to support policy training and evaluation.
- 1:17
From world model to action model
Further post-training turns Cosmos into a system that perceives, reasons, plans, and generates actions across embodiments and environments. The presenter frames compute-generated data as both a training resource and a way to teach physical AI.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.