RoboBoom
NVIDIA Omniverse· November 3, 2026· 35:57

From Generating Pixels to Understanding Worlds

Efe Sviša, senior director of product for AI Labs at NVIDIA, moderates a GTC panel with Robin Rombach of Black Forest Labs, Zaeem Farbman of VLA-IX, and Nikolaus Enßlin of Visoma on how world models and embodied learning can help machines understand and act in the physical world.

In this talk

  1. 2:12

    What video can teach world models

    Robin Rombach argues that large video datasets can teach models about lighting, human movement, and other features of the physical world. He sees current errors as a sign that these systems need more scale and better ways to identify actions and events in video.

  2. 4:17

    Why manipulation needs practice

    Nikolaus Enßlin says perception and reasoning models can help robots understand scenes, while reinforcement learning already supports locomotion and whole-body control. Manipulation remains harder: demonstrations show a robot what a task looks like, but self-learning through physical attempts may be needed to build practical skill.

  3. 6:40

    Combining models, tools, and feedback

    Rombach says foundation models need interaction and ways to verify results, with feedback from simulation and real-world trials. Zaeem Farbman compares simulators to calculators for physical tasks and says scaling models alongside tools could improve their ability to predict and act.

  4. 25:21

    Separating skills from the body

    Enßlin says learned skills, such as using a screwdriver, should transfer across bodies, while the agent must learn how to operate its new embodiment. Rombach says both unified multimodal models and coordinated specialist models could be practical approaches, depending on deployment needs.

  5. 29:35

    A test for progress and Europe's role

    Enßlin proposes giving a robot an unfamiliar screwdriver and seeing whether it can learn to use it overnight. Farbman expects language and world models to converge, while Rombach points to Europe's specialized manufacturers as demanding customers that could drive systems to generalize across varied industrial setups.

Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.