Physical AI is artificial intelligence that perceives, decides and acts in the physical world, through robots, machines, vehicles, sensors and the systems that control them. Unlike a chatbot, which answers with text, a Physical AI system changes something real: it adjusts a valve, reroutes a forklift, stops a line or dispatches a crew.
Updated October 2026
What changed in 2026
- Open robot and world models shipped at CES. In January 2026 NVIDIA released Cosmos Reason 2, a vision-language model for physical AI, Cosmos Predict 2.5 for synthetic video worlds, and the GR00T N1.6 robot foundation model (NVIDIA).
- Industrial robot makers adopted the stack. At GTC in March 2026 NVIDIA announced Cosmos 3 and GR00T N1.7 (with GR00T N2 due by year-end), with ABB Robotics, FANUC, KUKA, YASKAWA and Universal Robots among the partners (NVIDIA).
- Edge compute for robots got cheaper. NVIDIA's Jetson T4000 module (64 GB, up to 1,200 FP4 TFLOPS) went on order in January 2026 with JetPack 7.1 (JetsonHacks).
- Research moved to embodied AI. CVPR 2026 drew a record 16,092 submissions, with embodied AI, robotics and agentic AI as leading themes (CVPR).
- Regulation got dates. EU AI Act high-risk duties for AI in regulated products such as machinery now apply from 2 August 2028; transparency duties started on 2 August 2026 (Gibson Dunn).
Two things made Physical AI practical in the last two years. Models now understand video, sensor streams and language together, so they can reason about what is happening on a factory floor or a city street. And AI agents can now plan multi-step actions and call the systems that control equipment.
The hard part is no longer whether the AI can act. It is whether it should act, within what limits, and with what record of what happened. That is what this guide focuses on.
How Physical AI works
- Sense: sensors and cameras stream data about machines, sites and people.
- Reason: an AI agent reasons over that data and a digital twin of the asset.
- Check: before acting, it checks the limits people have set, and asks for approval where rules require it.
- Act and record: it acts through equipment or people and records what it did. The results flow back as new data, and the loop starts again.
Physical AI vs related ideas
| Term | What it is | How it relates to Physical AI |
|---|---|---|
| Robotics | Machines that move and manipulate | Robots are one kind of body Physical AI can control; it also runs fixed equipment, fleets and buildings |
| Embodied AI | AI that learns through a body in an environment | The research side; Physical AI is the broader deployment in real operations |
| Digital twin | A live virtual model of an asset or process | Gives Physical AI a safe place to test actions and a current picture of the asset (digital twins guide) |
| Cyber-physical systems | Computers tightly coupled with physical processes, such as PLCs and SCADA | The control layer Physical AI acts through |
| Edge AI | AI that runs close to the device | Where most Physical AI decisions run, for speed and privacy (edge AI) |
| Computer vision | AI that understands images and video | One of Physical AI’s main senses (computer vision solutions) |
Physical AI use cases by industry
| Industry | Physical AI in practice |
|---|---|
| Manufacturing | Vision agents spot defects and hold affected batches; maintenance agents act on vibration and temperature data before a failure |
| Energy and utilities | Agents balance loads, flag transformer stress, and schedule crews from sensor and weather data |
| Logistics and warehousing | Agents route robots and forklifts, re-plan picking when stock moves, and track yard congestion |
| Public safety and emergency services | Agents triage incidents from cameras and calls, suggest dispatch, and track response times |
| Healthcare | Agents monitor equipment and patient flow, and alert staff to bottlenecks |
| Smart buildings and cities | Agents tune heating, cooling and lighting, and manage traffic signals and parking |
| Retail | Agents watch shelves and stock, and run cashierless checkout |
Have a use case in mind? See how it would run on NexaStack →
Running Physical AI safely
When AI acts on equipment, a wrong answer is no longer just a bad paragraph. It can damage a machine, stop a site or put people at risk. Production Physical AI needs five controls:
- Identity: every agent and device is known and authenticated.
- Limits: clear rules for what an agent may do on its own, and what needs a person’s approval.
- Simulation first: risky actions are tested on a digital twin before they touch the real asset.
- Local execution: time-critical decisions run at the edge, so a network outage doesn’t leave equipment uncontrolled.
- A full record: what the agent saw, what it decided, who approved it and what happened next, available for audit and incident review.
How NexaStack fits
NexaStack is a platform for running and orchestrating AI agents on physical operations. It connects agents to cameras, sensors, digital twins and control systems, and runs them on your own infrastructure, from the edge to your cloud. It enforces the limits and approvals your team sets, and records every action. Teams use it to move from one pilot to many sites without building the plumbing again each time.
Where it's heading: 2026 → 2030
| When | What it looks like |
|---|---|
| Around 2020 | Robots ran fixed programs in fenced cells. AI mostly flagged problems on dashboards; a person made the change. |
| 2026 (now) | Open world models and robot foundation models are in early commercial use. Narrow Physical AI runs in inspection, maintenance, warehouses and building control, with AI agents acting under human approval. |
| By 2028 (expected) | EU high-risk rules apply to AI in machinery and other regulated products from 2 August 2028, so audit trails and human oversight become a condition of sale in the EU (EU Digital Omnibus). |
| By 2030 (expected) | Gartner predicts 80% of frontline workers at international companies will be assisted by physical AI systems such as robots, drones and self-driving vehicles (Gartner). |
Frequently asked questions
What is a Physical AI example?
A camera on a production line spots a defect. An agent checks the batch record, holds the affected units and opens a quality ticket, all within seconds, and logs every step.
Is Physical AI the same as robotics?
No. Robotics is about machines that move. Physical AI is the intelligence that decides what any physical system should do, whether that’s a robot, a pump, a vehicle fleet or a building.
Who coined the term “Physical AI”?
No single person owns it. NVIDIA and its CEO Jensen Huang popularised it from 2024 onward; the idea draws on earlier work in embodied AI and cyber-physical systems.
How far are we from Physical AI?
Narrow Physical AI is in production today in inspection, predictive maintenance, warehouse robotics and building control. General-purpose robots that handle any task are still in early deployment.
Does Physical AI need the cloud?
Not for every decision. Time-critical actions run at the edge; the cloud is used for training, coordination across sites and reporting.
What are the latest Physical AI trends in 2026?
Open world models and robot foundation models (NVIDIA Cosmos, GR00T), cheaper edge compute for robots, industrial robot makers building on the same stack, and governance moving to the front as EU AI Act deadlines firm up.
See Physical AI on your operations
Tell us one asset or site you want AI agents to run safely. Book a 30-minute Physical AI readiness session →
