Nate B Jones interviews Ming-Yu Liu, vice president of NVIDIA's Cosmos Lab, about world modelsA world model learns how an environment is structured and how it may change in response to actions. for robots and autonomous systems. Liu describes world understanding, simulationRobot simulation models a robot and its environment in software so behavior can be developed or evaluated before physical deployment. and action as related capabilities: a model may interpret a scene, generate a plausible environment for testing, or help produce physical control actionsPhysical AI combines learned perception and decision-making with systems that interact with the physical environment.. That is more specific than saying every world model simply issues robot commands.
The conversation covers model sizes for edge devices and data centers, scaling through more data and richer descriptions, and using physics engines to create synthetic training scenarios. Liu stresses that a generated scene approximating physical behavior is not proof the model understands physics. Cosmos can help screen driving or robot policiesA robot policy is the decision-making function that maps observations and goals to actions a robot should take in its environment. in simulation, but real-world validation remains important, especially for rare events and changing environments.
For home robots, Liu expects tasks with observable outcomes, such as folding laundry, to become verifiable sooner than taste-dependent cooking. Both speakers frame general household competence as an open challenge. They also discuss specialized agent tools and NVIDIA's preference for open models; future capabilities described near the end are expectations, not demonstrated household products.
Watch on YouTube
