What is physical AI?

A rigorous definition

Physical AI refers to artificial intelligence systems that can perceive their surroundings, reason about what they observe, and act in the real world: industrial robots, humanoids, autonomous vehicles, drones, agricultural machines. Where conventional AI manipulates data behind a screen, physical AI drives a body. It turns raw sensor streams (cameras, lidar, force sensors, microphones) into decisions, and decisions into motions that physically change the environment.

Three building blocks define it: an AI model trained to understand the real world, typically a multimodal foundation model; embedded compute that runs that model in real time, close to the sensors; and a control layer that guarantees precision and safety. The term took hold in 2025-2026, championed in particular by NVIDIA, to describe the convergence of robotics, edge AI and large models: intelligence leaving the screen and entering machines.

How it differs from generative AI

Generative AI produces content: text, images, code. When it gets something wrong, you regenerate. Physical AI acts: a mistake can break a part, halt a production line or injure someone. That difference in consequences reshapes the entire system design.

In practice, physical AI has to cope with a continuous, noisy world: shifting light, deformable objects, friction, imperfect sensors. It must meet hard real-time constraints (control loops often run at several hundred hertz) and run locally, because a round trip to the cloud is unacceptable in the middle of a motion. Its training data is also scarce: there is no web-scale corpus of robot motions, which is why the field leans so heavily on simulation and human demonstrations.

This distinction was not built overnight: our history of AI in robotics pillar traces the successive waves, from programmed automation to learned perception and then to generalist models, that shaped what counts as physical AI today.

The perception, reasoning, action loop

Every physical AI system runs the same fundamental loop, continuously:

1. Perception: sensor data (cameras, lidar, inertial units, force sensors) is fused into a representation of the scene: where the objects, obstacles and people are.
2. Reasoning: the system interprets the scene, plans and decides. This is where the newest models come in, notably VLA (vision-language-action) models and world models, which predict the consequences of an action before executing it.
3. Action: decisions are translated into motor commands, with fast control loops correcting the trajectory at every instant.

The loop closes immediately: every action changes the environment, which perception observes again right away. In practice, several loops coexist at different frequencies: safety reflexes in milliseconds, motion adjustments in hundredths of a second, planning in seconds. That hierarchy is what lets a robot be both reactive and deliberative.

Concrete examples in the field

Physical AI is not a lab concept: it is already deployed, at very different levels of maturity.

• Humanoids in factories: pilots are under way at several carmakers and logistics operators, on tightly scoped tasks such as moving totes or tending machines. At CES 2026, Boston Dynamics unveiled the production version of its Atlas humanoid, paired with Google DeepMind models.
• Autonomous mobile robots (AMRs): the most mature category. Entire fleets navigate warehouses for picking and internal transport, with well-documented returns on investment.
• Autonomous vehicles: robotaxis operate commercially in several US and Chinese cities. This is physical AI at its largest scale today.
• Inspection drones: autonomous inspection of energy infrastructure, construction sites and industrial plants, with automatic anomaly detection. Quadrupeds such as those from ANYbotics play the same role on the ground.
• Agricultural machines: targeted spraying, precision weeding and autonomous tractors cut inputs and physical strain.

To compare these machines side by side (size, payload, price), see our humanoid robots comparator.

Why 2026 is the tipping point

The shift happened between 2025 and 2026, driven by three converging forces.

The models changed in kind. VLA models and world models let a single brain generalize to tasks that were never explicitly programmed. At CES 2026, physical AI dominated the show: humanoids, home robots, edge compute modules. NVIDIA CEO Jensen Huang declared there that the "ChatGPT moment" for physical AI was nearly here, having described it a year earlier as merely around the corner.

Capital is pouring in. According to Crunchbase, robotics startups raised $13.8 billion in 2025, up from $7.8 billion in 2024, and Figure was valued at $39 billion after its September 2025 round.

Adoption is under way. Deloitte's State of AI in the Enterprise 2026 survey (3,235 senior leaders across 24 countries) found that 58% of companies already use physical AI at least to a limited extent, a share expected to reach 80% within two years. Asia Pacific leads at 71% adoption, with Europe at 56% today but projected to reach 81% within two years. The report also notes that physical AI adoption is climbing more slowly than software agents (expected to jump from 23% to 74% over the same period): costs, development cycles and safety requirements explain the gap. The window to position yourself is now. Find these players and their funding rounds in our companies map of physical AI.

What physical AI still cannot do

Credibility demands honesty: the physical AI of 2026 is still limited on several fronts.

• Brittle generalization: a robot that performs well under demo conditions can fail when the lighting changes, an object is unfamiliar or the space is cluttered. Robustness outside the training distribution remains the central problem.
• Insufficient reliability: a production line demands uptime figures that few physical AI systems achieve today without human supervision, or even standby teleoperation.
• Fine dexterity: cables, fabrics, deformable objects, dense bins: manipulations that are trivial for a human operator are still research topics.
• Speed: on many tasks, today's humanoids remain markedly slower than an experienced operator.
• Total cost: the robot's price tag is only a fraction of the full cost. Integration, safety engineering, certification, maintenance and change management often weigh more.

These limits do not doom the technology: they define the roadmap. The deployments that succeed in 2026 are the ones that pick bounded, measurable use cases matched to the actual state of the art, not to the demo videos.

Keep reading

• See how these building blocks come together in a real machine: our physical AI stack pillar details sensors, models and actuators.
• Understand the model that drives the decision loop: our guide to VLA models explains how they work.
• Compare humanoid robots already deployed in our robots comparator.