NVIDIA AI Updates: October 30, 2026

1. NVIDIA Shows 12 Omniverse Simulation Projects Built by Frontier AI Agents, Adds Omniverse Agent Skills

NVIDIA published a roundup of 12 developer projects in which frontier AI agents, guided by natural-language instructions, built simulations on Omniverse libraries (ovrtx, ovstage, ovstream, ovphysx, ovui). Examples include a warehouse rebuilt from site scans as an editable OpenUSD model, a kitchen scene assembled from 63 SimReady assets plus 18 custom pieces, and a warehouse with 80 conveyors, 40 robot arms, and 800 boxes that held 30 fps at full HD while completing 640 box placements. Other projects cover cable routing with Isaac Lab and Newton, a YuMi robot converted from CAD to a SimReady asset, a Market Street driving test environment, and a Unitree G1 that cleared a simulated hurdle in 64 of 100 trials. NVIDIA also pointed developers to new Omniverse agent skills, an Omniverse Real-Time Viewer skill, an Onshape importer guide, and USD Content Agents. The post flags limits: the YuMi physics were not calibrated against the real robot, and the cable-routing controller used ground-truth simulation data rather than sensor input. Source

2. FP32 Emulation on Tensor Cores Speeds Up Machine Learning Interatomic Potentials by Up to 1.7x

NVIDIA detailed how two Tensor Core FP32 emulation methods accelerate EPFL’s Point Edge Transformer (PET) interatomic potential for molecular dynamics. On Blackwell, cuBLAS BF16x9 emulation (enabled with CUBLAS_EMULATE_SINGLE_PRECISION=1, no code changes) made large GEMMs 2.2 to 2.5x faster on a 17,280-atom system and cut full MD step time by up to 1.7x, with error below the model’s own run-to-run non-determinism. On Ampere, a custom TF32Linear layer backed by a CUTLASS TF32x3 kernel gave up to 1.5x faster MD steps and matched FP32 convergence in a structural relaxation test (218 to 222 steps versus 218), while one-pass TF32 failed to converge within 1,000 steps. Reference code is available via CUTLASS Example 27, the NVIDIA ALCHEMI Toolkit, and the UPET checkpoint repository. Source