At SIGGRAPH, the throughline across NVIDIA's 21 papers is a shift from worlds that look real to worlds that behave real, in real time — 3D-grounded, physics-governed, and steerable by the creator. The same systems that generate a virtual scene can train and drive things in the physical one, whether that's a character, a robot, or a factory twin. Papers, code, and models are free. https://vist.ly/5b26a
The transition from AI systems that generate realistic representations to systems that can understand and influence real-world environments represents a significant institutional shift. As AI becomes increasingly embedded in physical and operational systems, institutions will require stronger capabilities for evaluation, oversight and accountability to ensure that technological advancement translates into trusted and responsible outcomes.
That demo uses gaussian splats, in my humble opinion, gaussian splat are a waste of raster power, becasue the application need to render millions just to approximate sharp edges. I am waiting for a true voxel rendering capable of rendering trillion of voxels per second. That would allow superior physics simulations and also effects like subsurface scatering and global illumination would be more precise and accurate from a lightnig physics point of view.
The future of AI is not only about generating content, but about understanding the real world. Digital twins and physical AI will redefine how we design, build, and innovate.
Nice shots they look real clean 👍
Great direction. The line that stuck with me is that the same system generating a virtual scene can also train something in the physical world. That makes me curious about the edge cases. A reconstruction only really knows what the camera saw. So when part of the scene is unusual, say a damaged car, and the camera never captured the damaged side, does the model fill it back in as a normal car from what it has learned? For most scenes that's harmless, but if that scene is later used to train a robot or a vehicle, the unusual case is often the one that matters most.