Physical AI Is Coming: When AI Stops Generating Content And Starts Acting On The World

Harsh Verma is the Principal Software Engineer - AI/ML at Palo Alto Networks, an IEEE Senior Member & BCS Fellow.

Since the advent of AI, it has been all about generating information. Large language models were used to write documents, generate software, summarize research, create images and answer questions. The next generation is something else entirely. Physical AI is something different. It has no safety net. It is AI connected to cameras, sensors or robotic arms or legs. These systems don’t just describe an action, as AI is used to; they perform it.

At its CES 2026 keynote, Nvidia signaled a potential turning point for robotics, unveiling new open AI models designed to help machines perceive, understand and act on the physical world.​ This is a huge leap in what the future of AI looks like because there is no “undo” button on a robotic arm that has already moved.

A new phase is emerging: physical AI, which is the turning point because it creates real consequences. It moves a real object in a real room, which leaves a trail of actions behind it. I think most companies are still caught up with agentic AI and building governance that barely covers generative AI at this point, when technology has already moved to its third phase. That’s where the real risk is.​

From Outputs To Outcomes: The Technology That Makes This Possible

​Physical AI runs on the vision-language-action model, or VLA. This is a physical-world version of the language models most people already know. In a guide to VLA architecture, older robots followed pre-programmed rules, while VLA models are learning real behavior. With VLA, a robot can adapt to a new object or situation it wasn’t specifically trained on, understand a plain-language instruction and carry out a multistep task with very little human input along the way.