Physical AI Is Coming: When AI Stops Generating Content And Starts Acting On The World
At CES 2026, Nvidia unveiled new open AI models designed to help machines perceive, understand and act on the physical world, marking a significant shift toward physical AI that performs real actions rather than just generating content.
Since the advent of AI, technology has focused primarily on generating information - writing documents, creating images, answering questions. The next generation, however, represents a fundamental shift: physical AI, which performs real actions in the physical world rather than just describing them. Unlike traditional AI, physical AI has no safety net - once a robotic arm moves, there's no 'undo' button.
This transition creates real consequences as AI systems move beyond digital outputs to physical actions. Most companies are still focused on agentic AI and governance for generative AI, while technology has already moved to this third phase, where the real risks lie.
Vision-Language-Action Models
Physical AI operates on vision-language-action (VLA) models, physical-world counterparts to the language models most people already know. Traditional factory robots followed pre-programmed rules and failed when something was out of place, while VLA-based robots can adapt to new situations, understand plain-language instructions, and complete multistep tasks with minimal human input.
This technology is no longer confined to labs. Platforms like NVIDIA Isaac and Omniverse allow robots to train themselves across millions of virtual scenarios before touching the real world. Manufacturers can now build and test robots in digital workspaces first, compressing deployment timelines from years to months.
The 'simulate first' approach allows companies to run multiple tests and fail in simulation before any mistake with real consequences occurs, making the technology more practical and reliable.
Real-World Applications
Amazon has deployed over 1 million robots across its fulfillment network, coordinated by a generative AI model called DeepFleet. This system has improved fleet travel time by 10%, with roughly 75% of Amazon's global deliveries now robot-assisted. Warehouse automation analysis shows Amazon's robotics deployment delivering a 50% to 75% increase in items picked per hour.
BMW Group's Leipzig plant is testing a humanoid robot named AEON, developed by Hexagon Robotics. The robot began performing production tasks in June 2026, with applications in high-voltage battery assembly and component manufacturing.
Amazon illustrates how AI can improve coordination across a large, established robotic operation. BMW illustrates how manufacturers are exploring new ways to integrate adaptable robots into existing production workflows.
Leadership Challenges
The biggest mistake leaders make when evaluating physical AI is treating it like new software - focusing on ROI and deployment speed while treating safety and governance as afterthoughts. This approach is backward because a robot that misjudges a person's position on a factory floor can cause real injuries, unlike a chatbot giving a wrong answer.
Physical AI creates both opportunities and responsibilities. Factories can become more flexible, infrastructure can monitor itself, and robots can work in previously too-variable environments. But risks increase as mistakes become physical, security becomes physical, and the boundary between software and systems engineering disappears.
As AI begins to shape the physical world around us, the challenge becomes learning to trust what machines do when they're standing next to us, fundamentally changing how we interact with and regulate autonomous systems.
Related
