US What physical AI means
Physical AI is artificial intelligence that perceives the physical world through sensors, works out what is happening and what to do next, and then acts through a machine. The machine might be a robot arm, an autonomous mobile robot, a humanoid, a self-driving vehicle or a network of fixed cameras that controls a site.
The contrast is with the generative AI most people know. A chatbot or image model is trained on text and pictures from the internet and its output stays on a screen. A physical AI model has to understand space, motion, friction, gravity and cause and effect, because its output is a movement that can hit something. NVIDIA has used the term widely since 2024, sometimes as “generative physical AI”, for models that bring the reasoning of large models to machines that move.
In practice the term covers a whole way of building autonomous machines: foundation models trained in the data centre, skills learned and tested in simulation, and a compact computer on the machine that runs the result in real time.
How physical AI works
Every physical AI system runs the same loop many times a second. The loop is short, and each step has to finish before the world moves on, which is why most of it runs on the machine rather than in the cloud. Combining several sensor types in the first two steps is known as sensor fusion.
-
//
Sense. Cameras, depth and stereo cameras, lidar, radar, an IMU and joint encoders stream raw data into the computer.
-
//
Perceive. Neural networks detect objects, estimate depth and pose, and build a map or 3D picture of the scene.
-
//
Reason and plan. A policy or planning model decides what to do next. Newer systems use large multimodal models that can also follow a spoken or written instruction.
-
//
Act. The decision becomes motor commands, steering angles or gripper forces, sent over CAN, EtherCAT or another real-time link to the drives.
The three computers behind physical AI
NVIDIA describes physical AI development as work split across three kinds of computer. The split is useful because each stage has very different hardware needs.
| Stage | What it does | Typical hardware | Where it sits |
|---|---|---|---|
| Training | Trains or fine-tunes foundation models and robot policies on real and synthetic data | Multi-GPU servers | Data centre or cloud |
| Simulation | Builds a physically accurate copy of the site, generates synthetic data and tests policies before deployment | RTX GPU servers and workstations running Omniverse and Isaac Sim | Lab or data centre |
| Runtime | Runs perception and the policy in milliseconds, on the power the machine can supply | NVIDIA Jetson Thor or Jetson AGX Orin | Inside the robot or vehicle |
The simulation stage depends on a digital twin of the factory, warehouse or road network. Training and simulation usually share the same class of hardware, such as our data centre GPUs in a server, while the runtime computer is a different design entirely: small, efficient and built to survive vibration and heat.
Why simulation and synthetic data matter
A physical AI model needs far more varied experience than anyone can safely collect with real machines. Crashing a robot into a shelf a thousand times to teach it not to is slow and expensive, and rare events such as a person stepping out from behind a pallet are hard to capture at all.
Simulation solves part of that. Engineers randomise lighting, textures, object positions and sensor noise, a technique called domain randomisation, so the model does not overfit to one clean scene. World foundation models such as NVIDIA Cosmos can then generate further realistic variations from a smaller set of real recordings. Robots learn skills in simulation through trial and error or by copying human demonstrations, and the policy is checked in simulation before it reaches the real machine.
The remaining difference between simulated and real behavior is called the sim-to-real gap. Closing it usually means a final round of tuning on real data collected from the deployed robots.
What physical AI needs from the computer on the machine
The runtime computer is where most industrial projects spend their engineering effort, because it has to meet several limits at once.
-
//
Compute per watt. Large multimodal models and several perception networks run side by side. NVIDIA quotes Jetson AGX Thor at up to 2,070 FP4 TFLOPS with up to 128 GB of memory, within a configurable 40 to 130 W power range.
-
//
Sensor bandwidth. A mobile robot may carry eight or more cameras. GMSL cameras keep latency low over long cable runs, and 10 or 25 GbE ports carry lidar and network cameras.
-
//
Real-time control links. CAN FD and isolated I/O connect the AI computer to motor drives and safety controllers.
-
//
Environment. Shock, vibration, dust, wide temperature ranges and vehicle power all apply, so fanless and sealed designs are common.
Quick answers on physical AI
Our Jetson Thor computers include robot computers with up to 16 GMSL camera inputs and CAN FD, and our comparison of Jetson Thor and AGX Orin explains when the extra compute is worth it.
Physical AI: common questions
Is physical AI the same as generative AI?
No, but it builds on it. Generative AI produces text, images or code. Physical AI uses similar large models, trained with added spatial and physical data, to produce actions for a machine.
Is physical AI the same as embodied AI?
Close, with a different emphasis. Embodied AI is the older research term for intelligence that lives in a body and learns through it. Physical AI is the newer industry term for the full workflow of training, simulating and deploying those machines.
What is an example of physical AI?
A warehouse robot that picks mixed items. It sees each item with cameras, decides how to grasp it, and adjusts its grip as it lifts. Self-driving vehicles, inspection drones and humanoid robots are other examples.
Which companies are working on physical AI?
Chip makers, robot makers and AI labs. NVIDIA supplies much of the computing and simulation software. Robot makers and research labs, including Google DeepMind and Physical Intelligence, publish robot foundation models.
Does physical AI need a connection to work?
Not for real-time control. The machine runs its models locally. A connection is used to upload data, monitor the fleet and install updated models.
We build physical AI systems on NVIDIA Jetson Orin and Jetson Thor as standard, with the cameras, storage, wireless and JetPack image fitted and tested before they ship. Tell us about the robot or vehicle, its sensors and its power supply, and we will specify the computer to match.