AI Is Moving From Screens Into the Physical World
For much of the modern artificial intelligence boom, the most important AI systems existed almost entirely inside computers. Large language models generated text, image models created pictures, video systems produced clips, recommendation engines predicted what people might watch, and software agents increasingly interacted with digital applications. The next phase of the technology is moving beyond the screen.
Technology companies are now investing heavily in systems that can perceive the physical environment, reason about what they see, plan actions, and control machines. This emerging field is commonly described as physical AI, embodied AI, or embodied intelligence. Instead of asking an AI system only to generate an answer, researchers are attempting to make AI systems capable of turning perception and reasoning into physical action.
The shift is already visible across robotics. Google DeepMind has developed Gemini Robotics models designed to connect multimodal reasoning with robot control. NVIDIA is building a broader physical-AI ecosystem involving simulation, synthetic data, robotics models, and deployment hardware. Figure is scaling compute and training data for humanoid robotics, while Boston Dynamics and Hyundai are preparing humanoid robots for industrial environments. These developments indicate that the competition around AI is expanding from software and data centers toward machines that can physically interact with the world.
That transition is technically difficult. A software AI can generate an incorrect sentence and produce another response. A physical AI system controlling a robot arm, vehicle, factory machine, or humanoid has to account for gravity, friction, collisions, moving objects, human behavior, uncertainty, battery limits, sensor errors, and the consequences of every action.
The central question is therefore not simply whether AI can control robots. It is whether increasingly general AI models can acquire enough physical understanding to operate safely and reliably in environments that were never completely programmed in advance.
What Is Physical Intelligence?
Physical intelligence refers broadly to AI systems that can perceive and reason about the physical world and then take actions within it. A physical-intelligence system may combine cameras, microphones, tactile sensors, motion sensors, language models, vision models, planning systems, and motor-control software.
A conventional industrial robot may perform a fixed sequence thousands of times. It can move an object from one location to another with extremely high precision when the environment remains predictable. The difficulty begins when the object changes shape, the lighting changes, another object blocks the path, or a human worker enters the workspace.
Physical AI attempts to give robots greater adaptability. Instead of programming every possible situation manually, developers train models to recognize objects, understand instructions, estimate spatial relationships, predict possible outcomes, and select appropriate actions.
This approach is closely related to the development of foundation models in software AI. A foundation model can be trained on enormous datasets and then adapted to many downstream tasks. Researchers are attempting to create analogous foundation models for robots, where the training data includes images, video, demonstrations, simulated environments, robot trajectories, language instructions, and sensor information.
The goal is not necessarily to create one humanoid robot that performs every task immediately. The broader objective is to develop an intelligence layer that can transfer skills across different environments and robot bodies.
Why Tech Companies Are Interested in Robots
One reason technology companies are moving into robotics is that software AI has reached a point where perception, language understanding, planning, and multimodal reasoning can be combined in increasingly sophisticated ways.
A robot needs several of these capabilities simultaneously. It must identify objects, understand a command, locate relevant objects, determine how they can be manipulated, choose a sequence of actions, monitor the result, and adapt when something goes wrong.
Modern AI models provide components that can potentially support each of these functions. Vision-language models can interpret images and instructions. Reasoning models can break complex tasks into smaller steps. Vision-language-action models can translate observations and instructions into robot actions.
Google DeepMind's Gemini Robotics 2, introduced in July 2026, illustrates this direction. Google describes the system as a vision-language-action model capable of controlling humanoid and other robotic systems, while its embodied-reasoning model is designed to understand physical environments and plan multi-step tasks. Google also introduced an on-device version optimized for local robotic operation.
This architecture is important because physical intelligence requires more than a chatbot connected to a motor controller. The system has to connect perception, reasoning, action, and feedback in a continuous loop.
The Robot Becomes an AI Interface
One way to understand the transition is to think of the robot as a physical interface for an AI model.
A smartphone allows software to interact with users through a screen and touchscreen. A robot can interact with the physical world through cameras, microphones, arms, hands, wheels, legs, grippers, and other mechanisms.
If an AI system can understand an instruction such as “move these boxes to the storage area,” physical intelligence requires the system to determine which boxes the user means, identify the storage area, understand obstacles, select a safe path, grasp the boxes, transport them, and verify that the task has been completed.
Every step introduces uncertainty. A box could be heavier than expected. Another worker could walk into the path. A shelf could be partially blocked. The robot could lose its grip. The lighting could change.
Therefore, physical AI is fundamentally an interaction problem. The model is not generating a static output. It is continuously observing the consequences of its previous actions and deciding what to do next.
Why Humanoid Robots Are Receiving So Much Attention
Humanoid robots are attracting particular interest because much of the human environment has already been designed around the human body.
Buildings contain stairs, doors, shelves, workbenches, tools, vehicles, and equipment designed for people. A robot with a human-like body could theoretically operate within these environments without requiring every facility to be redesigned.
Two arms can provide flexibility for manipulation. Legs can allow access to environments designed around human height. A human-like torso can carry batteries, computing equipment, and sensors.
However, humanoid design also introduces substantial engineering challenges. Walking is difficult. Maintaining balance while carrying an object is difficult. Coordinating two arms and multiple fingers is difficult. Operating safely around people is even more challenging.
The current commercial market remains very small compared with conventional robotics. Reuters reported in September 2026 that approximately 7,000 humanoid robots were sold globally during 2025 for industrial and professional service applications. The International Federation of Robotics has only recently begun separately tracking humanoid robot sales, reflecting how early this market remains.
That small installed base is important. The technology is progressing rapidly, but humanoid robots are still far from being a universal replacement for conventional automation.
Google DeepMind Is Treating Robotics as a Foundation-Model Problem
Google DeepMind's robotics strategy demonstrates how software AI research is being connected to physical machines.
Gemini Robotics 2 is designed around the idea that multimodal models can provide robots with more general capabilities. Google says the model can convert visual and language information into motor-control actions and can operate humanoid and bi-arm robots.
The company also describes an embodied-reasoning model that can plan multi-step activities in physical environments and coordinate multiple robots.
An especially important component is on-device operation. Running intelligence locally can reduce dependence on a constant cloud connection, decrease latency, and potentially improve privacy. For a robot moving around humans, waiting for a remote server to respond to every movement may not be practical.
Local inference also creates another engineering constraint: the robot has limited power and computing resources. The model must therefore become efficient enough to operate within the physical limitations of the machine.
NVIDIA Is Building the Infrastructure Behind Physical AI
NVIDIA's role in physical AI is broader than supplying processors. The company is building an ecosystem around simulation, synthetic data, robot models, training infrastructure, and deployment.
In March 2026, NVIDIA announced its Physical AI Data Factory Blueprint, a reference architecture intended to support large-scale data processing, synthetic-data generation, reinforcement learning, and evaluation for robotics, autonomous vehicles, and vision-based AI systems.
This addresses one of the largest problems in robotics: data.
Internet-scale AI models can learn from enormous collections of text, images, audio, and video. Robots cannot simply download a massive collection of real-world physical interactions. Physical training requires machines to manipulate objects and encounter different environments.
Collecting all that data physically is expensive and slow. Simulation can therefore become a major source of training data.
In a simulated environment, developers can generate millions of variations of lighting, objects, surfaces, obstacles, trajectories, and failures. Robots can learn from those simulations before being tested in the real world.
The Simulation Advantage
Simulation gives physical AI researchers something that traditional robotics development lacks: scale.
Imagine training a robot to pick up an object. In the real world, a robot may need thousands of attempts to learn the correct grip. Each attempt consumes time and hardware resources, and a failed attempt can damage equipment.
In simulation, developers can create thousands or millions of virtual environments. The robot can attempt different grips, arm trajectories, and approaches at high speed.
Researchers can also deliberately create difficult situations. Objects can be placed at unusual angles. Lighting can change. Floors can become slippery. Objects can be partially hidden.
The resulting synthetic data can then be combined with real-world demonstrations to train a model.
But simulation has a major limitation: the simulated world is not the real world. Small differences in friction, material properties, sensor noise, motor behavior, and physical dynamics can cause a policy that works perfectly in simulation to fail on an actual robot.
This problem is known as the sim-to-real gap. Closing that gap is one of the central technical challenges in physical AI.
Why Robot Data Could Become the Next AI Data Advantage
Data has always been a major competitive factor in machine learning. Physical AI makes data collection even more complicated because useful data is tied to actions.
A video showing a person opening a drawer contains information about perception and motion. But a robot also needs information about the forces involved, the position of its joints, the grip applied to the handle, and what happened after the drawer moved.
This creates a new category of training data: action data.
Figure announced in September 2026 that it was building a large humanoid training dataset and said its partnership with Nscale would provide access to up to 100,000 NVIDIA Vera Rubin GPUs, with an initial compute commitment of $3.5 billion and an intention to scale beyond $6 billion.
Figure's announcement demonstrates the relationship between robot data and AI compute. Training a general-purpose physical model can require enormous amounts of computational infrastructure.
As a result, the physical-AI race is not only a robot manufacturing race. It is simultaneously a data race and a compute race.
Why One Video Could Become Valuable Training Data
One emerging approach is to teach robots through demonstrations rather than programming each task manually.
NVIDIA highlighted Skild AI's S1 robotic foundation model in September 2026. The system is designed to learn previously unseen long-horizon tasks from a single video demonstration, using in-context learning rather than requiring task-specific weight updates for every new task.
If such approaches become reliable, the economics of robot programming could change substantially.
Instead of an engineer manually programming a robot for every variation of a task, a worker could demonstrate the task and the AI system could infer the required sequence of actions.
However, demonstration-based learning still requires careful validation. A human demonstration may contain movements that are safe for a human but unsafe for a robot. The robot also needs to understand differences in body structure, reach, strength, balance, and available tools.
Factories Are the First Major Testing Ground
Industrial environments are attractive for physical AI because they are more structured than homes or public spaces.
A factory can restrict the robot's operating area, define safety zones, standardize objects, and provide controlled workflows. These constraints make it easier to test advanced robotics than in an unpredictable household.
Automotive manufacturing is one of the most important early markets. Robots already perform welding, painting, assembly, material movement, and inspection. Physical AI could extend automation to tasks that currently require more flexible human manipulation.
Boston Dynamics has opened a humanoid-robot testing facility at a Hyundai Motor Group plant in Georgia, where Atlas robots are being trained for automotive manufacturing tasks. This represents an effort to move humanoid robotics from demonstrations into actual industrial workflows.
The significance is not that humanoid robots have already replaced large numbers of factory workers. They have not. The significance is that major industrial companies are beginning to create environments in which general-purpose robot intelligence can be tested against real production requirements.
Why Warehouses Could Be Another Major Market
Warehouses contain thousands of repetitive but variable physical tasks. Workers may need to pick products, move containers, sort packages, unload shipments, scan items, and transport materials.
Some of these tasks are already highly automated. Autonomous mobile robots can move shelves and containers, while robotic arms can perform specialized picking and sorting.
Physical AI could potentially make these systems more adaptable. A robot might learn to manipulate a wider range of objects rather than requiring separate programming for every product.
The challenge is reliability. A warehouse robot must operate for long periods without creating dangerous situations or damaging products. A small error repeated thousands of times can become a major operational problem.
For this reason, the first commercial applications are likely to focus on environments where the task can be measured clearly and the operating conditions can be controlled.
The Home Is a Much Harder Robotics Problem
A household may appear simpler than a factory, but from a robotics perspective it is extraordinarily complicated.
Homes contain thousands of objects with different shapes, textures, weights, and materials. Objects are frequently left in unexpected places. Furniture changes position. Lighting changes throughout the day. Children and pets move unpredictably.
A robot that can reliably fold one type of towel in a laboratory is not necessarily capable of cleaning a normal household.
This is why the transition from industrial physical AI to general household robotics could take considerably longer than the transition from laboratory demonstrations to controlled factories.
Household robots must also satisfy extremely high safety requirements because they operate directly around people in confined spaces.
Autonomous Vehicles Are Another Form of Physical AI
Physical intelligence is not limited to humanoid robots.
Autonomous vehicles are also physical AI systems. They perceive the environment through cameras, radar, lidar, and other sensors, interpret traffic conditions, predict the behavior of other road users, plan trajectories, and control steering, acceleration, and braking.
NVIDIA describes robotaxis as an early commercial application of physical AI because autonomous vehicles require AI to interact with a complex physical environment.
The underlying technical problem is similar to robotics. The AI must continuously perceive the world, predict what might happen next, make decisions, execute actions, and monitor the consequences.
The difference is that the consequences of an incorrect decision can be much more serious. This makes safety validation a central component of autonomous physical AI.
Why Physical AI Is Harder Than Chatbots
A chatbot operates primarily in an information space. The consequences of a wrong answer can be significant, but the model usually does not directly exert physical force.
A robot does.
When a robot moves an arm, it generates force. When it drives a vehicle, it creates momentum. When it lifts an object, it affects the object's physical state. When it operates machinery, it can create hazards.
This means physical AI must understand not only semantic meaning but also physical consequences.
The system needs some representation of geometry, depth, movement, force, object permanence, spatial relationships, and uncertainty.
Researchers therefore increasingly combine language and vision models with specialized robotics systems, physics simulation, motion planners, control algorithms, and real-time safety mechanisms.
Reasoning Is Not Enough Without Motor Control
A powerful language model may be able to describe how to pick up a cup. That does not mean it can physically pick up the cup.
The robot needs to know where the cup is, estimate its orientation, determine an appropriate grasp, move its arm through a safe trajectory, apply the correct force, and adjust if the cup moves.
This distinction explains the importance of vision-language-action models. The model must connect high-level concepts to low-level physical actions.
Google's Gemini Robotics architecture is an example of this direction. The model is designed to connect visual and linguistic understanding with motor control rather than stopping at text generation.
Physical AI Needs Continuous Feedback
Traditional software can often execute a predefined sequence because the software environment is deterministic enough to support it.
Robots operate in a world filled with uncertainty.
A robot may expect an object to be at a specific location and discover that it has moved. It may expect a door to be open and find it closed. It may attempt to grasp an object and discover that the object is heavier or more slippery than expected.
Physical AI therefore needs a feedback loop.
Observe. Reason. Act. Observe again.
This cycle may happen many times per second for certain control tasks, while higher-level planning may occur more slowly.
Combining these time scales is one of the central architectural challenges in robotics.
Safety Becomes a Core AI Problem
As AI systems gain control over physical machines, safety can no longer be treated as an optional feature added after model development.
A physical AI system should have boundaries around what it can do, where it can operate, how much force it can apply, and when it must stop.
Recent research has also demonstrated why this matters. A September 2026 report from Robocurve described experiments in which several AI models controlling robot arms were tested on potentially harmful physical tasks. The results raised concerns about whether general-purpose AI systems can reliably distinguish unsafe physical instructions without specialized safeguards.
Such experiments do not demonstrate that robots are generally unsafe in commercial environments. They do demonstrate that connecting general-purpose models to physical actuators introduces a safety problem that is qualitatively different from ordinary chatbot interaction.
Future physical-AI systems are therefore likely to require layered safety mechanisms: model-level constraints, action-level validation, hardware limits, environmental sensors, emergency stops, human supervision, and independent testing.
Why the Physical AI Market Could Be Extremely Large
The economic opportunity comes from the enormous number of physical tasks performed around the world.
Factories move materials. Warehouses handle products. Hospitals transport equipment. Farms harvest crops. Construction sites move heavy objects. Retail stores restock shelves. Hotels clean rooms. Logistics companies sort packages.
Software automation has transformed information work because digital information can be processed cheaply at enormous scale. Physical work remains constrained by machines, humans, transportation, and physical environments.
If AI can make robots more flexible, the potential market extends far beyond traditional industrial robotics.
But market size should not be confused with near-term adoption. Physical deployment requires hardware manufacturing, maintenance, batteries, sensors, actuators, safety certification, insurance, workforce integration, and reliable economics.
Why the Robot Hardware Problem Is Still Huge
AI models can be updated through software. Robots cannot.
A model developer can release a new model globally after completing the necessary deployment process. A robot manufacturer must build physical machines, source components, assemble them, test them, ship them, install them, maintain them, and replace damaged components.
Actuators, motors, gearboxes, batteries, cameras, tactile sensors, processors, wiring, structural components, and mechanical joints all have failure modes.
This means the physical-AI industry requires two forms of scaling simultaneously: scaling intelligence and scaling manufacturing.
A highly capable model is not useful if there are only a few machines capable of running it.
The Compute Requirement Could Become Enormous
Training physical AI models may require substantial computing power because the models can learn from video, simulation, sensor data, robot trajectories, and large collections of demonstrations.
Figure's announced partnership for up to 100,000 NVIDIA GPUs illustrates how physical intelligence can become closely linked with hyperscale AI infrastructure.
This creates an unusual feedback loop. AI data centers train robot models. Robot models control physical machines. Physical machines generate new training data. That data returns to the data center and improves the next generation of models.
The result resembles a continuous learning cycle connecting cloud computing, simulation, robotics hardware, and real-world deployment.
Why Robots Could Generate a New Type of AI Data
Text and images tell an AI model what the world looks like. Robot interaction can tell the model what happens when an action is performed.
A robot can learn that pushing an object produces movement, that lifting an object requires force, that a soft object behaves differently from a rigid one, and that some actions produce unexpected outcomes.
This kind of interaction data can provide information about cause and effect.
That is particularly valuable for planning. A physical AI system must not only recognize what exists. It must predict what will happen if it acts.
This makes robotics an important potential source of embodied world knowledge.
The Competition Is Expanding Beyond Traditional AI Companies
The physical-AI race includes semiconductor companies, cloud providers, AI laboratories, robot manufacturers, automotive companies, logistics firms, and industrial manufacturers.
Google brings its Gemini models and DeepMind research. NVIDIA provides computing, simulation, robotics software, and physical-AI infrastructure. Automotive companies such as Hyundai have direct industrial environments where humanoid robots can be tested.
Specialized robotics companies such as Figure and Skild AI are focused on developing robot intelligence and hardware integration.
This creates a different competitive structure from the early generative-AI market. No single company necessarily controls every layer.
The stack can include chips, cloud infrastructure, simulation, foundation models, robot operating systems, sensors, actuators, robot hardware, industrial software, and deployment services.
Open Ecosystems Could Matter
Robotics is historically fragmented. Different manufacturers use different hardware, software interfaces, sensors, and control systems.
Open frameworks can make it easier for developers to transfer AI capabilities across robot platforms.
NVIDIA's Isaac ecosystem and ROS-based robotics development are examples of efforts to standardize parts of the development stack.
Google's robotics models also emphasize adaptation across different robot embodiments. If a model can transfer skills from one robot body to another, developers may not need to retrain an entire intelligence system whenever hardware changes.
That capability could become an important factor in the economics of physical AI.
What Happens to Human Workers?
The impact on employment will depend heavily on which tasks robots can perform reliably and economically.
In some cases, robots may automate specific repetitive or physically demanding tasks while humans continue to supervise operations. In other cases, companies may redesign entire workflows around robots.
Physical AI could also create new jobs in robot maintenance, deployment, data collection, safety engineering, fleet management, model evaluation, simulation, and robotics software.
The transition is unlikely to occur uniformly across industries. Tasks requiring high precision and repetitive movement may be easier to automate than tasks involving unpredictable environments, complex social interaction, or frequent physical adaptation.
The important distinction is between automating an occupation and automating individual tasks. A robot may perform some tasks previously handled by a worker without eliminating the entire occupation.
Why 2026 Is an Important Period for Physical AI
The year 2026 is becoming an important period because several parts of the physical-AI stack are developing simultaneously.
Foundation models are becoming more capable. Robotics models are increasingly multimodal. Simulation systems are becoming more sophisticated. AI accelerators are becoming more powerful. Robot manufacturers are increasing production ambitions. Industrial companies are beginning to test humanoid systems in real facilities.
The International Federation of Robotics' decision to begin separately tracking humanoid robot sales also reflects the growing attention around the category.
Yet the industry remains early. Thousands of humanoid robots sold in a year is tiny compared with the global population of industrial robots.
The next stage will therefore be defined less by demonstrations and more by measurable deployment: how many hours robots operate, how often they fail, how much human intervention they require, how much they cost per task, and whether they produce a measurable economic benefit.
The Real Test Will Be Reliability
A robot that performs a spectacular demonstration once is not necessarily commercially useful.
Industrial customers need machines that work repeatedly. They need predictable maintenance costs, uptime, safety procedures, replacement parts, and integration with existing systems.
For physical AI, reliability may therefore become a more important metric than benchmark performance alone.
A model that performs extremely well on a laboratory benchmark but frequently fails in an unpredictable factory may be less useful than a simpler system that performs the same task consistently for thousands of hours.
This is one reason real-world data and long-term deployment will become increasingly valuable.
Could Physical AI Eventually Become a General-Purpose Platform?
The long-term ambition is much larger than individual factory tasks.
If a physical foundation model can understand instructions, perceive environments, reason about physical objects, transfer skills between machines, and learn from demonstrations, it could potentially become a general-purpose intelligence layer for many types of robots.
One model could support warehouse robots, industrial arms, humanoids, autonomous vehicles, agricultural machines, and other physical systems.
That would resemble the transition from individual software applications toward general-purpose AI platforms.
Whether this vision becomes practical remains uncertain. Physical environments are extraordinarily diverse, and transferring intelligence between different robot bodies is difficult.
The Biggest Barrier May Be the Real World Itself
Software systems exist in environments that can often be standardized. Physical environments cannot.
Every factory has different machinery. Every warehouse has different layouts. Every household contains different objects. Every road contains different combinations of vehicles, weather, pedestrians, and infrastructure.
This variability is what makes physical intelligence so difficult.
AI researchers can build increasingly capable models, but those models must still interact with the physical limitations of motors, sensors, batteries, materials, and environments.
The race to physical intelligence is therefore ultimately a race to make AI robust against uncertainty.
What the Next Generation of Robots Could Look Like
The next generation of robots is likely to be more software-defined than earlier machines.
Instead of programming every behavior manually, developers may specify high-level objectives. The AI system could then interpret the objective, plan actions, execute them, monitor progress, and adapt when circumstances change.
Robots could also receive continuous software updates that improve their capabilities without changing the physical hardware.
A robot purchased for warehouse picking could potentially learn new manipulation skills later. A factory robot could receive updated models for new products. A humanoid could acquire new capabilities through demonstrations.
This would make robot hardware increasingly dependent on the quality of its AI software.
Conclusion
The movement of AI from software into robots represents a major expansion of the artificial-intelligence industry. Instead of limiting AI to text, images, code, and digital applications, technology companies are attempting to create systems that perceive the physical world, reason about it, and act within it.
Google DeepMind's Gemini Robotics models, NVIDIA's physical-AI infrastructure, Figure's large-scale robotics compute strategy, and the increasing use of humanoid robots in industrial testing all point toward the same broader direction: physical intelligence is becoming an important frontier for AI development.
The transition will not be immediate or simple. Robots must operate safely, reliably, and economically. They require hardware, batteries, sensors, actuators, simulation, training data, computing infrastructure, maintenance, and carefully controlled deployment environments.
The current scale of humanoid deployment also remains small compared with conventional industrial robotics. The technology is therefore better understood as an emerging platform rather than a mature replacement for human physical labor.
What makes the current moment significant is the convergence of several technologies that previously developed separately. Foundation models provide increasingly powerful perception and reasoning. Simulation can generate enormous quantities of training experience. Advanced chips provide the compute required for model training and inference. Robotics hardware provides a physical body. Together, these technologies are creating the foundations of physical AI.
If researchers can close the gap between digital intelligence and reliable physical action, robots could become a new interface through which AI interacts with the real world. The resulting competition would extend far beyond chatbots and software applications into factories, warehouses, transportation, healthcare, logistics, agriculture, and eventually homes.
The defining question is no longer simply whether AI can understand the world. It is whether AI can understand the world well enough to act in it safely.
