NVIDIA Research unveiled advances in physical AI at CVPR 2026, including new agent training frameworks for autonomous vehicles, robotic grasping, and vision-based navigation. The techniques combine GPU-accelerated sim-to-real training with foundation models to reduce engineering overhead in autonomous systems.
This underscores NVIDIA's shift from inference acceleration toward end-to-end agent training infrastructure—a critical capability for enterprises deploying autonomous fleets or robotics where simulation and data flywheel cost determines competitive advantage.