NVIDIA AI Updates: June 4, 2026

1. NVIDIA Research Debuts GraspGen-X, LCDrive, and NitroGen at CVPR

NVIDIA. NVIDIA Research presented three new foundation models at CVPR aimed at robotics, autonomous driving, and agent training. GraspGen-X is a grasping model trained on 2 billion simulated grasps that adapts zero-shot to any robotic gripper, LCDrive replaces text-based reasoning with compressed latent representations to cut computational tokens by roughly 50% for autonomous vehicles, and NitroGen is a gameplay model built on the GR00T architecture and trained across more than 1,000 games and 40,000 hours of interaction. NVIDIA reports NitroGen improves performance by up to 52% over prior state-of-the-art methods in low-data scenarios and is available open-source on GitHub and Hugging Face. Source

2. NVIDIA Releases Agent Skills and Alpamayo 2 Super for Physical AI Research

NVIDIA. NVIDIA introduced new agent skills spanning autonomous vehicles, robotics, and vision AI to accelerate physical AI research, all built on the Cosmos 3 omnimodel. The release includes Neural Reconstruction skills that turn fleet-captured data into editable 3D scenes, agent-ready simulation workflows for NVIDIA Isaac Sim and Isaac Lab, a Cosmos-H-Surgical-Simulator for healthcare robotics, and Metropolis Defect Image Generation skills for vision AI. NVIDIA also launched Alpamayo 2 Super, an open 32-billion-parameter reasoning vision language action model for Level 4 autonomous driving development, with the skills available on GitHub and through NVIDIA Brev. Source

3. NVIDIA NemoClaw Powers Autonomous AI Engineers for Industrial Software

NVIDIA. NVIDIA announced NemoClaw, an open blueprint for building specialized, long-running agents with a secure runtime and frontier models, including a model router, NeMo libraries for customization, and the open-source NVIDIA OpenShell runtime that governs agent access to files, networks, and tools. The blueprint targets autonomous AI engineers that compress industrial engineering work from weeks to hours by automating CAD, meshing, simulation setup, debugging, and report generation. Cadence, Dassault Systemes, Siemens, and Synopsys are adopting NemoClaw, alongside startups such as Flexcompute, Neural Concept, nTop, PhysicsX, and SimScale building specialized agents for applications from photonic to aircraft design. Source