Physical AI Explained: When Artificial Intelligence Acts in the Real World

Physical AI is artificial intelligence that perceives, reasons about and acts in the physical world through machines such as robots, self-driving cars and drones. It combines sensors, AI models that understand space and physics, and control systems. In 2026 its most mature form is the driverless taxi; general-purpose robots are still early.

This guide explains how physical AI works — from vision-language-action models and world models to simulation — who is building it, where it's already used and where it isn't, what the data shows, and what it means for the UAE, where driverless taxis now carry paying passengers in Dubai and Abu Dhabi.

Key takeaways

• Physical AI is AI with a body. Robots, cars, drones and cameras that sense, decide and act — not just humanoids.
• Robotaxis are the proof point. Waymo gave more than 400,000 paid rides a week in early 2026; Apollo Go delivered 3.2 million fully driverless rides in the first quarter of 2026.
• Data is the bottleneck. Physical-interaction data is far harder to collect than internet text, so developers combine simulation, world models, teleoperation and paid human video.
• Hands are still hard. Google DeepMind says multi-finger dexterous manipulation remains challenging even for Gemini Robotics 2.
• Dubai is scaling carefully. Commercial robotaxis started in March 2026; by August, 48 were in active operation completing about 50 trips a day, towards a target of 25% of all trips autonomous by 2030.

What is physical AI?

Physical AI is the branch of AI that operates in the real world: it takes in sensor data, builds an understanding of space, objects and physics, decides what to do and moves a machine to do it.

Definition

Physical AI — AI that lets autonomous systems such as cameras, robots and self-driving cars perceive, understand, reason and perform or orchestrate complex actions in the physical world, extending generative AI with an understanding of three-dimensional space and physics. (NVIDIA; Deloitte defines it similarly as AI that lets machines autonomously perceive, understand, reason about and interact with the physical world in real time.)

NVIDIA's Jensen Huang popularised the term, and NVIDIA's framing has escalated fast: "The ChatGPT moment for robotics is coming" in January 2025, "is here" in January 2026, and "Physical AI has arrived" in March 2026. Strip away the marketing and the category is broad: warehouse robots, industrial arms with AI vision, delivery robots, drones, driverless cars and humanoids all count.

	Generative AI	Physical AI
Output	Text, images, code	Movement and actions in the world
Training data	Abundant internet data	Scarce sensor, robot and simulation data
Speed required	Seconds are fine	Real-time control, milliseconds matter
Cost of a mistake	A wrong answer	A broken object, a crash or an injury
How it's proven	Benchmarks and user feedback	Hours, miles and incidents in the real world

How does physical AI work?

Physical AI runs a continuous loop: sense the world, model what's happening, decide on an action, move, and learn from the result — with much of the learning done in simulation before a machine touches anything real.

The physical AI loop

1. 01Perceive
   • Cameras
   • Lidar
   • Touch and force

   Sensors turn the world into data

2. 02Model the world
   • 3D space
   • Physics
   • Prediction

   A world model predicts what happens next

3. 03Decide
   • Instructions
   • Plans
   • Safety rules

   A policy model chooses the next action

4. 04Act
   • Motors
   • Grippers
   • Steering

   Control systems execute in milliseconds

5. 05Learn
   • Simulation
   • Real-world data
   • Human demonstrations

   Every run improves the next

Three ideas do most of the work.

Vision-language-action models. A vision-language-action (VLA) model takes camera images and a plain-language instruction and outputs robot actions. Google DeepMind's RT-2 introduced the idea in 2023 by writing robot actions as text tokens; Gemini Robotics, launched in March 2025, adds physical actions as a new output of a Gemini model. These are the "robot foundation models" behind NVIDIA's GR00T, Physical Intelligence's π models and Figure's Helix.

World models. A world model simulates how an environment behaves so a system can predict outcomes and practise. Google DeepMind's Genie 3 generates interactive worlds at 24 frames a second in 720p, staying consistent for a few minutes; NVIDIA's Cosmos models are built to generate training data for robots and cars; Waymo built its own world model on Genie 3 for driving simulation.

Simulation and sim-to-real. Robots learn much of their behaviour in simulation, through millions of trial-and-error attempts, then transfer it to reality. The standard trick, domain randomisation, varies a simulator's lighting, textures and physics so widely that the real world looks like just another variation.

Where does physical AI get its training data?

From four sources, each with trade-offs — and none yet at the scale of internet text.

Source	How it works	Strength	Limit
Simulation	Robots practise in virtual worlds and digital twins	Cheap, safe, unlimited repetitions	The gap between simulation and reality
Teleoperation	People operate robots remotely to demonstrate tasks	High-quality, task-specific data	Slow and expensive to scale
Human video	Models learn from videos of people doing tasks	Enormous potential supply	Human bodies and robot bodies differ
Pooled robot data	Labs share data across robot types	Breadth across tasks and hardware	Still small: Open X-Embodiment pooled 22 robot types from 21 institutions

The scramble for data is now a business. Figure says it has paid $15 million to people who contribute video through its Index app, and reports that pre-training on that data raised its Helix model's success in unseen homes from 9% to 56%.

Who is building physical AI?

A platform layer (chips, simulators, open models), a model layer (robot foundation models and world models) and an application layer (robots and vehicles) are forming — with a handful of companies in each.

Company	Role	What it offers in 2026
NVIDIA	Platform	Cosmos world models, GR00T robot models, Isaac simulation, Jetson Thor robot computers (developer kit from $3,499)
Google DeepMind	Models	Gemini Robotics 2 (whole-body control; early-access partners), Genie 3 world model
Physical Intelligence	Models	π0 and π0.5 research models that cleaned kitchens and bedrooms in homes they'd never seen
Figure AI	Robots and models	Helix models for its humanoids; Index data programme
Skild AI	Models	"Robot brain" software; raised $1.4 billion in January 2026, led by SoftBank
Waymo, Baidu Apollo Go, WeRide, Pony.ai	Autonomous vehicles	Commercial driverless ride-hailing in the US, China and the Gulf
Amazon	Operations	More than 1 million warehouse robots coordinated by an AI model

For the humanoid side of the story — who's deploying robots in warehouses and factories, and at what cost — see humanoid robots: where they will be used first.

Where is physical AI used today?

In order of maturity: driverless ride-hailing, warehouse robotics, industrial inspection and handling, then homes and general-purpose robots.

• Driverless taxis — production. Waymo reported more than 400,000 rides a week across six US metro areas in February 2026 and 15 million rides in 2025. Baidu's Apollo Go delivered 3.2 million fully driverless rides in the first quarter of 2026.
• Warehouses — production. Amazon deployed its millionth robot in 2025 and says a new AI foundation model coordinating the fleet improves travel efficiency by 10%.
• Factories — pilots and early deployments. AI vision and mobile robots are common; humanoids are in pilots at carmakers and logistics firms.
• Homes — research. Physical Intelligence's π0.5 cleaned kitchens and bedrooms in entirely new homes in research tests; Figure says Helix 2.5 worked in 30 Bay Area homes it had never collected data in. Neither is a product.
• Buildings and cities — emerging. Digital twins double as simulators for robots and vehicles; construction uses 3D printing and robotic assembly.

What does the latest data show?

What the data shows

• Driverless miles (through June 2026): 271.3 million rider-only miles; compared with human drivers in its operating cities, Waymo reports 95% fewer crashes causing serious injury or worse. — Waymo
• Driverless rides (Q1 2026): 3.2 million fully driverless rides for Apollo Go, with weekly rides peaking above 350,000 in March. — Baidu
• Investment (first half of 2026): $47.4 billion of venture funding across 521 deals, using a broad definition that includes robotics, autonomous vehicles and drones. — Crunchbase
• Developers (vendor figure, January 2026): 2 million robotics developers using NVIDIA's stack. — NVIDIA
• Warehouse scale (July 2025): more than 1 million Amazon robots in operation. — Amazon

What this means

The money and the models are real, but the proof is uneven. Driving works at scale because it has a narrow task, huge data and years of miles; manipulation in messy rooms doesn't yet. Judge physical AI claims by miles, hours and intervention rates in the real world, not by demo videos.

What are the limits of physical AI?

The main limits are data, dexterity, the long tail of rare situations, safety certification and compute on the machine itself.

• Data scarcity. An April 2026 Nature Machine Intelligence editorial notes that data capturing physical interactions is much harder and more time-consuming to obtain than internet data.
• Dexterity. Google DeepMind reports medium-to-high success for whole-body and gripper tasks with Gemini Robotics 2, but says multi-finger dexterous manipulation remains challenging.
• World-model horizons. Genie 3 stays consistent for a few minutes — useful for training, not yet a faithful simulator of a working day.
• The long tail. Rare events — an unusual obstacle, a child running into the road, a slipping load — dominate safety risk and are the hardest to train for.
• Certification. There's no published safety standard yet for general-purpose humanoid robots; cars are regulated city by city.
• Compute and power. Robots must run models on board in real time, which is why purpose-built robot computers such as NVIDIA's Jetson Thor exist, and battery life limits how long they can work.

Common misconception

"Simulation has solved robotics' data problem." Simulation helps enormously, but the gap between simulated and real physics remains, which is why companies still pay for teleoperation and human video. Figure alone has paid contributors $15 million for training video.

What does physical AI mean for the UAE?

The UAE is deploying physical AI in public — driverless taxis carry paying passengers in Dubai and Abu Dhabi — while building research capacity in world models and robotics.

• The target. Dubai's Autonomous Transportation Strategy, adopted in 2016, aims for 25% of all transportation trips in Dubai to be autonomous by 2030, with AED 22 billion in annual savings. The RTA said in February 2024 that 9.4% of Dubai's transportation was already autonomous.
• Robotaxis in Dubai. Commercial operations began on 30 March 2026 in Umm Suqeim and Jumeirah, with Apollo Go vehicles on the Apollo Go app and WeRide vehicles on Uber. By August 2026 the service had driven more than 4 million kilometres and completed 7,613 passenger trips, with 144 vehicles in the fleet, 48 in active operation, about 50 trips a day and 97% customer satisfaction. Apollo Go's agreement with the RTA plans to scale from 50 vehicles to 1,000 over three years.
• Robotaxis in Abu Dhabi. Autonomous taxis expanded to Al Reem and Al Maryah islands in July 2025.
• Research. MBZUAI's Institute of Foundation Models released PAN, a world model for simulating how environments change, in November 2025; TII and NVIDIA opened a joint AI and robotics lab in Abu Dhabi in September 2025.
• Capital. Mubadala Capital was among the investors in Waymo's $16 billion round in February 2026.

Common misconception

"Dubai's 25% target means a quarter of cars will drive themselves by 2030." The target covers all transportation trips, not only cars, and the RTA already reported 9.4% in February 2024 — two years before commercial robotaxis launched. Robotaxis are one part of the plan, and in August 2026 they were completing about 50 trips a day.

How will physical AI affect real estate and construction?

Physical AI will reach property through construction, building operations and mobility — each with different timelines.

• Construction. Dubai's 3D Printing Strategy aims for 25% of buildings to be based on 3D printing technology by 2030 — a target, not a measure of progress. Trakhees issued Dubai's first 3D-printed building licence to Nakheel for Al Furjan Hills in December 2023.
• Building operations. Cleaning, inspection, security and delivery robots start in controlled back-of-house areas; buildings with digital twins, callable lifts and good connectivity are easier to automate. See smart buildings and digital twins in real estate.
• Mobility and location. Dubai's first robotaxi zones — Umm Suqeim and Jumeirah — are established residential neighbourhoods; as autonomous mobility spreads, access to it becomes another factor in how locations are compared.
• Logistics real estate. Warehouses are where robots already work at scale, and where humanoids are being piloted first.

How should businesses approach physical AI?

Start with proven categories, demand real-world metrics, and plan for safety and data from day one.

1. Buy mature categories first. Autonomous mobile robots, AI vision inspection and warehouse automation have track records; general-purpose humanoids don't yet.
2. Ask for real-world numbers. Hours worked, tasks completed, interventions per shift, failure rates — and whether a demo was teleoperated.
3. Write the safety case early. Where people and machines share space, who can stop the machine, and how incidents are reported.
4. Plan your data. Physical AI improves with operational data; agree who owns it before deployment.
5. Use digital twins. Simulating a site before deploying robots cuts risk and cost.
6. Stay regulation-aware. In the UAE, autonomous vehicles and robots in public spaces operate under authority approvals and defined zones.

What is likely to happen next?

Expect robotaxis to expand city by city, robot foundation models to improve quickly on grippers before hands, and the first safety standards for humanoid-type robots.

• Models. NVIDIA says GR00T N2, previewed in March, is slated for release by the end of 2026; Google DeepMind has opened its Gemini Robotics reasoning model to developers on Google AI Studio, while its action model stays with early-access partners.
• Mobility. Waymo planned to lay groundwork for ride-hailing in more than 20 additional cities in 2026, including Tokyo and London; Dubai and Abu Dhabi expand zones and fleets.
• Standards and rules. Safety standards for dynamically stable robots, and city-level rules for autonomous vehicles, will decide how fast deployments grow.
• Money. Record funding raises the stakes for real-world results in 2027.

The software side of autonomy — agents that act in digital systems — is covered in agentic AI explained, and the compute behind all of it in the AI infrastructure guide.

Final takeaway

Physical AI is AI that moves things: cars, robots, drones and machines that sense and act. It already works at scale where the task is narrow and the data is plentiful — driverless taxis and warehouse fleets — and it's advancing fast in labs on everything else. The constraints are data, dexterity, the long tail of rare events and certification. In the UAE, robotaxis in Dubai and Abu Dhabi make physical AI tangible today; for businesses, the right move is to deploy the mature categories now and pilot the rest with real-world metrics, not demo videos.