Complete guide
Physical AI
The intelligence that acts in the physical world. Not software that thinks — hardware that does.
Last updated: July 2026
Definition
What Physical AI means
Physical AI refers to AI systems embedded in physical hardware that can perceive the real world through sensors, reason about it using on-device or cloud AI, and act on it through actuators — motors, grippers, speakers, displays.
The term was coined in mainstream use by NVIDIA CEO Jensen Huang at CES 2024, though the underlying concept — embodied AI or situated cognition — has existed in robotics research for decades.
The difference from software AI is not just hardware. Physical AI must make decisions in real time with incomplete sensor data, under physical constraints, with consequences that cannot be undone with Ctrl+Z.
Physical AI vs Software AI
Categories
Six categories of Physical AI
Humanoid Robots
Bipedal robots designed for human environments. Tesla Optimus, Figure 03, 1X NEO, Agility Digit, Unitree G1, Boston Dynamics Atlas.
16,000 units deployed globally (2025)Robot Dogs
Quadruped robots for industrial inspection, security, and research. Boston Dynamics Spot, Unitree Go2, ANYbotics ANYmal.
From $1,600 (Unitree Go2 Air)AI Wearables
Glasses, rings, watches, and patches that bring persistent AI to the body. Meta Ray-Ban, Samsung Galaxy Ring, Humane AI Pin.
Mainstream and available nowCompanion Robots
Social robots for care settings, education, and home presence. Emotionally responsive, designed for human-robot relationships.
Growing 35% YoYSmart Home AI
AI embedded in home infrastructure: security cameras with scene understanding, robot vacuums with mapping, AI doorbells.
Already in millions of homesAutonomous Vehicles
Self-driving cars and trucks as the largest-scale Physical AI deployment. Waymo, Tesla FSD, Zoox, Aurora.
Waymo: 150,000+ paid rides/weekMarket context
Why Physical AI matters now
Three forces converged around 2023-2024: large language models became capable enough to power reasoning in robots, transformer-based vision models made sensor interpretation reliable, and hardware costs dropped far enough to make sub-$20K robots plausible.
The result: every major technology company — Tesla, Google, Amazon, Microsoft, Meta — plus dozens of well-funded startups are now racing to ship physical products.
16,000 humanoid robot units were deployed globally in 2025. China accounts for roughly 80% of that volume. Western companies lead on published performance benchmarks but lag on production scale.
Technical overview
How Physical AI works
Perceive
Sensors gather data about the physical world. Cameras, LiDAR, IMUs, tactile sensors, microphones. The output is high-dimensional, noisy, and continuous.
Reason
On-device AI (or cloud with acceptable latency) processes sensor data. Vision models classify objects. Language models interpret instructions. Policy networks plan actions.
Act
Actuators execute the plan. Motors move joints. Grippers close. Wheels turn. Each action changes the physical state and generates new sensor data for the next perception cycle.
NVIDIA's Physical AI stack
NVIDIA has positioned itself as the infrastructure layer for Physical AI, mirroring its role in software AI. The stack has three layers:
Physically accurate robot simulation for generating synthetic training data at scale. Runs in Omniverse.
General Robot 00 Technology — a multi-modal foundation model for robot learning, open-sourced in 2025.
Physical world simulator that generates realistic video of robot actions in novel environments.
FAQ
Common questions
What is Physical AI?
Physical AI refers to AI systems embedded in physical hardware that can perceive the real world through sensors, reason about it using on-device or cloud AI, and act on it through actuators like motors and grippers. Examples include humanoid robots, robot dogs, AI wearables, and autonomous vehicles.
Who coined the term Physical AI?
NVIDIA CEO Jensen Huang popularised the term in his CES 2024 keynote, framing it as the next major wave of AI after software-only systems. The underlying concept — embodied AI or situated cognition — has existed in robotics research for decades.
How is Physical AI different from software AI?
Software AI processes information and outputs text or code. Physical AI must additionally perceive the physical world through sensors, make real-time decisions under physical constraints, and execute actions through motors or other actuators — all while managing hardware reliability, latency, and safety.
What are examples of Physical AI?
Current deployed examples: Tesla Optimus (internal data collection), Figure 02 at BMW Spartanburg (90,000+ parts handled), Agility Digit at GXO warehouses (100,000+ totes moved, OSHA-certified), Unitree G1 ($16K, shipping now), and Meta Ray-Ban AI glasses (mainstream wearable).
When will Physical AI affect consumers?
AI wearables are already mainstream. Robot dogs start at $1,600. Consumer humanoid robots are expected in the $10-20K range by 2028-2030. Goldman Sachs projects a $38B humanoid market by 2035; Morgan Stanley puts the broader Physical AI ecosystem at $5 trillion by 2050.
Hub
Explore the Physical AI hub
Humanoid Robots
Every major player, specs, prices, and deployment status.
Robot Dogs
Quadruped robots for industrial inspection, security, and research.
AI Wearables
Glasses, rings, and sensors that bring AI to your body.
Companion Robots
Social robots designed for care, education, and daily presence.
Technology
Sensors, actuators, on-device inference, and the software stack.
Market & Investment
Goldman Sachs, Morgan Stanley, and BofA forecasts in context.
Applications
Manufacturing, healthcare, logistics, and home use cases.
Regulation
Safety standards, liability frameworks, and policy landscape.
Kin
Your personal entity
Physical AI acts in the world. Kin remembers it. A persistent memory layer that tracks your preferences, context, and history — so every physical AI device you use knows who you are.
Learn about KinLore
Field knowledge, captured
Physical AI needs to know how skilled work is actually done. Lore captures that knowledge from experienced workers — video and voice, on the phones and bodycams they already carry.
Learn about Lore