A Tutorial on Physical AI from a Multimedia Perspective: Part II Cosmos Foundation Model

Physical AI aims to enable systems to perceive, reason, and act in the unstructured physical world, yet it faces bottlenecks of scarce real-world data and high deployment risks. Traditional hand-engineered simulators provide controllable environments yet struggle with scalable coverage of diverse scenes and long-tail scenarios. Cosmos addresses these challenges as a data-driven world foundation model platform, unifying generation- verification loop to form a closed-loop data flywheel for Physical AI development. This tutorial systematically introduces Cosmos's multi-modal design, core components, and practical deployment, offering a comprehensive guide for leveraging learned world models in physical system training and evaluation.