One of the three computers, in four parts
Dr. Pascalis Trentsios (Digital Environments Leader, Boston Dynamics) framed NVIDIA Omniverse as the simulation computer in the three-computer picture from the GTC update: train, simulate, deploy. The easiest label is “simulation platform.” His working model has four parts.
OpenUSD is the scene format Omniverse depends on, which is why a USD file can meet CAD tools or Blender instead of living in one viewer. It can also be written out as text, which is why an agent can be prompted to build or edit a scene. Kit is what makes the product modular: viewport, stage, and properties are separate apps, and a missing tool can be added. He noted his slides were already behind the product - the renderer itself is now a component - so you can pair the Omniverse renderer with another simulator such as MuJoCo, or Omniverse physics with another renderer. RTX ray tracing is what makes cameras, lidars, depth cameras, and fisheye sensors believable enough to test a perception pipeline, and to mint synthetic data. Physics ships with the platform, but because the stack is modular you can bring another solver. He described Newton, the joint physics effort with Google DeepMind and Disney, as a place where different solvers (granular material, rigid bodies, and others) can sit under the same rendering.
Omniverse, Isaac Sim, and Isaac Lab are one stack with more parts
The naming, he said, is the confusing part. Base Omniverse can be the viewer and the rendering. Isaac Sim is that plus physics, ROS, and synthetic data. Isaac Lab adds training environments and the learning loop. There is no single product you must take whole: a team can assemble the components it needs. Python for reach, C++ for speed, and - in his current practice - agents writing most of that glue. He said he has not opened the Omniverse UI himself in months; he points an agent at the scene.
