For most of its history, artificial intelligence has lived in software. It reads documents, generates text, classifies images, and answers questions. What it has not done is act on the world. It cannot pick up a component from a production line, navigate a warehouse at 2 am, or respond to a hazard it has never seen before without being told exactly what to look for.
That is changing. Physical AI refers to AI systems that do more than process information. They perceive, reason, and act within the physical world around them.
This shift, from AI as a software tool to AI as a presence in the real world, is one of the biggest developments in the technology’s history. For organisations that run physical environments, it is happening now.
What Is Physical AI?
Physical AI refers to the class of AI systems designed to operate in and interact with the physical world. Traditional AI works with digital inputs and produces digital outputs.
Physical AI is different. It takes in sensory data from cameras, lidar, microphones, and other instruments. It reasons about what that data means. Then it takes physical action in response.
In one sentence: physical AI is artificial intelligence that can see, understand, and act in real-world environments, not just in software.
Physical AI is different from the generative AI systems that have dominated public attention since 2022. Generative AI produces content, including text, images, code, and audio. Physical AI produces action.
A generative AI model can describe how to navigate a warehouse. A physical AI system actually navigates it. Generative AI produces information. Physical AI produces action in the real world.
The term has become more common, partly because Jensen Huang declared at CES 2025 that “the ChatGPT moment for robotics is here.” The comment framed physical AI as the next big shift after the large language model era. The global physical AI market was worth about $5 billion in 2025. It is projected to reach between $83 billion and $119 billion by 2033, a compound annual growth rate of about 32 to 36 percent.
Bring a new AI vision application to life.
What Physical AI Refers To: The Core Components
Physical AI relies on a combination of capabilities that let a machine act intelligently in the physical world. To understand how it works, look at the three layers every system depends on.
Perception: Computer Vision and Sensor Fusion
The foundation of any physical AI system is the ability to perceive its environment. Computer vision is the dominant perception technology. It accounts for about 42 to 45 percent of the physical AI technology market. It gives machines the ability to see and interpret visual data from cameras in real time, including:
- Detecting objects
- Understanding spatial relationships
- Recognising people
- Tracking movement
Computer vision in physical AI goes well beyond the object classification tasks that defined the field for most of the past decade. A physical AI system must understand depth, motion, and context at the same time. It needs to reason about what a scene means and what is likely to happen next, not just label what is present. This is why the shift from narrow-trained computer vision models to large vision models matters: general visual understanding, not task-specific detection, is what a physical environment actually requires.
Reasoning and Decision-Making
Perception alone is not enough. A physical AI system must decide what to do with what it perceives. This is where reasoning capabilities, increasingly powered by large language models, vision language models, and multimodal foundation models, become critical. The system must interpret the situation it finds itself in, identify the right response, plan a sequence of actions, and carry them out in real time.
For complex tasks, this reasoning layer must handle uncertainty, missing information, and new situations no one planned for. This is a key challenge that separates physical AI from older rule-based automation, which could only respond to situations it had been explicitly programmed for.
Action: Control Systems and Actuation
The third layer is physical action. Control systems translate the decisions made by the reasoning layer into physical movements, whether that means steering an autonomous vehicle, closing the fingers of a robotic arm, adjusting the gait of a humanoid robot, or triggering an alert in an operational workflow. The design and precision of these control systems determine whether the physical AI system can operate reliably and safely alongside human workers.
How Physical AI Works: Training in Simulated and Real-World Environments
Training physical AI is one of the hardest challenges in the field. A language model can be trained on text scraped from the internet. A physical AI system cannot. It must learn from physical interaction, which is slow, costly, and sometimes dangerous to repeat at scale.
The dominant approaches to training physical AI are:
- Reinforcement learning in simulation: The AI system learns by trial and error in a simulated environment. It gets rewards for success and penalties for failure. Running millions of trials in simulation saves time that real-world testing would not allow.
- Imitation learning from human demonstration: The AI system learns by watching and copying human behaviour. A human operator demonstrates the task, and the system learns to repeat it. This approach works well for complex tasks where it is hard to define a reward.
- Foundation model transfer: Large pre-trained models already hold broad world knowledge. These models are fine-tuned on physical task data, using their existing reasoning ability instead of training entirely from scratch. This greatly reduces the data needed to train physical AI for specific domains.
Reinforcement learning is still the main approach for physical AI in environments where the rules can be defined clearly and simulated reliably. This includes autonomous vehicle navigation, robotic arm manipulation, and logistics routing. The trial-and-error nature of reinforcement learning means the system can develop strategies that human designers would not have anticipated. These strategies are often faster, more reliable, or more energy-efficient than human-designed alternatives.
Physical AI in Practice: Where It Is Already Deployed
Physical AI is not a future technology. It is already deployed at commercial scale across multiple sectors.
| Sector | Physical AI application | Examples |
|---|---|---|
| Automotive | Autonomous vehicle navigation, driver assistance | Tesla Full Self-Driving, Waymo robotaxi fleet |
| Manufacturing | Robotic arm assembly, quality inspection, and collaborative robots | Boston Dynamics Spot at Hyundai, ABB robotic systems |
| Logistics | Autonomous mobile robots, warehouse navigation, sortation | Agility Robotics Digit at Amazon warehouses |
| Healthcare | Surgical assistance, patient monitoring, and pharmacy automation | da Vinci surgical systems, autonomous dispensing |
| Construction | Site monitoring, equipment automation, safety compliance | Drone-based inspection, autonomous excavators |
| Operations/HSE | Real-time visual intelligence, agentic safety monitoring | VGI-powered query-based operational insight |
The clearest public illustration of how physical AI works is the autonomous vehicle. Driving a car autonomously requires every part of the physical AI stack at once:
- Computer vision to perceive lanes, obstacles, and pedestrians
- Reasoning to interpret complex, ambiguous traffic situations
- Control systems to translate decisions into steering, acceleration, and braking in real time
The same architecture that lets AI drive a car on a public road also lets it operate a robotic arm on a production line, or an autonomous mobile robot in a fulfilment warehouse.
The most visible physical AI systems are humanoid robots. They have moved from research curiosity to commercial deployment:
- Tesla’s Optimus Gen 3 began production in January 2026
- Boston Dynamics’ electric Atlas is deployed at Hyundai facilities
- Agility Robotics’ Digit operates in Amazon warehouses
These humanoid robots are the most recognisable form of physical AI. They are not the most common, though. The largest volumes of deployed physical AI are in less visible systems:
- Autonomous vehicle fleets operating on public roads
- Robotic arm units running production lines in electronics and automotive manufacturing
- Autonomous systems managing inventory and fulfilment in logistics facilities
Physical AI in Operations: The Role of Visual Intelligence
For most organisations, the cameras already installed across their facilities are the most accessible and valuable form of physical AI. This matters more for day-to-day operations than any humanoid robot or autonomous vehicle.
Computer vision is the largest technology segment in the physical AI market. This is because it is the easiest form of physical world perception to deploy widely. Every facility with a camera already generates a constant stream of physical world data. The real question is whether AI can extract intelligence from that data in real time, and whether that intelligence can drive action instead of just producing reports.
This is where physical AI converges with agentic computer vision. An agentic physical AI system perceives what is happening, reasons about the context, and works out the right response. Then it runs a complete workflow, often before a human has reviewed a single frame:
- Alerting the right team
- Updating the relevant system
- Documenting the incident
- Initiating corrective action
The physical environment generates the data. The AI system processes it and acts on it.
Visual General Intelligence (VGI) is the category viso.ai defined in a foundational white paper published in 2025. It applies physical AI principles to operational visual intelligence at enterprise scale. Traditional computer vision was trained for specific tasks in specific environments.
A VGI system works differently. It understands any physical environment, answers questions about it in natural language, and acts on what it finds, without needing model training or annotation. This is what it looks like when physical AI becomes operational rather than experimental.
Physical AI vs. Generative AI: Understanding the Difference
The relationship between physical AI and generative AI is often misunderstood. They are not competing categories. Generative AI is increasingly part of physical AI systems. The distinction is about what the AI ultimately does.
- Generative AI produces digital content in response to digital inputs. Its outputs are text, images, code, or audio. It operates in software environments.
- Physical AI produces physical actions in response to sensory inputs from real-world environments. Its outputs are movements, decisions, and operational consequences. It operates in the physical world.
- The convergence: Modern physical AI systems increasingly use generative AI models as their reasoning layer. A large language model interprets the situation. A vision language model processes the visual inputs. A control system acts.
This convergence makes the current moment different from earlier waves of robotics and automation. Earlier physical systems ran on fixed rules and could not adapt to situations outside their training. Physical AI systems built on foundation models can generalise.
With the right foundation model integration, this ability to generalise shows up across use cases:
- A robotic arm trained on assembly tasks can adapt to new components it has never handled before
- An autonomous vehicle trained in one city can navigate another
- An operational visual intelligence system in a manufacturing facility can be queried in plain language by an HSE manager who has never written a line of code