Physical AI

NVIDIA coined the term "Physical AI" at GTC 2025 to describe artificial intelligence that understands and interacts with the physical world. Unlike cloud-based AI confined to text and images, Physical AI encompasses foundation models like NVIDIA Isaac GR00T, Cosmos world models, diffusion policies, and Vision-Language-Action (VLA) models that give robots the ability to perceive, reason about, and act within real environments. This represents a fundamental shift from cloud AI to embodied AI—intelligence that lives in physical form.

Foundation models are proving to be the key enabler for general-purpose robotics. By training on massive datasets of robot interactions, simulated physics, and multimodal sensor data, these models allow robots to generalize across tasks they were never explicitly programmed for. A robot trained with Physical AI can pick up novel objects, navigate unfamiliar spaces, and adapt to unexpected situations—capabilities that were impossible with traditional rule-based programming. Companies like NVIDIA, Google DeepMind, Covariant, Physical Intelligence, and Skild AI are leading the charge in building these foundational systems.

The Physical AI market is projected to transform the entire robotics industry by enabling truly general-purpose robots. As simulation platforms generate synthetic training data at scale, and VLA models bridge the gap between language understanding and physical action, robots are becoming more capable, adaptable, and autonomous than ever before. This convergence of large language models, computer vision, and robotics is ushering in an era where intelligent machines can work alongside humans across manufacturing, logistics, healthcare, and everyday life.