Physical AI is AI that perceives the real world through sensors, decides what to do, and then acts on it through a machine: a robot arm, a warehouse vehicle, a drone or an inspection camera that rejects a faulty part. Embodied AI is the closely related research term for an AI that has a body and learns by interacting with its surroundings. In practice the two phrases describe the same shift. AI is moving off the screen and into equipment that has to work safely in a factory, a hospital corridor or a shop floor.
That is also why a powerful GPU is only one part of the system. A chatbot that answers a second late is mildly annoying. A robot that reacts a second late can drop a part or hit a person. Physical AI needs sensors, real-time edge compute, simulation for training, and safety controls, and the GPU sits in the middle of all of them.
This post is part of our AI concepts series. For a related idea, see what a world model is, the type of model that lets machines predict how the physical world will change.
Physical AI vs embodied AI vs generative AI: what is the difference?
| Term | What it means | Typical output | Example |
|---|---|---|---|
| Generative AI | Models that create text, images, code or audio | Content on a screen | A chatbot drafting a quotation |
| Physical AI | AI that senses, reasons and acts in the physical world | A movement, a decision, a control signal | A robot picking mixed items from a bin |
| Embodied AI | Research term for AI with a body that learns through interaction | Learned behaviour | A humanoid learning to open doors in simulation, then in a lab |
| Edge AI | AI that runs on a device near the data, not in a cloud data centre | Local inference | A smart camera counting people without streaming video out |
Physical AI usually uses generative AI inside it. A robot may run a vision-language model to understand an instruction such as "put the blue box on the second shelf", then a control policy to move its arm. The difference is that the final output is an action, and actions have consequences.
How does a physical AI system work?
Every physical AI system runs the same loop: sense, think, act, and repeat many times a second.
- Sense. Cameras, depth sensors, lidar, microphones, force sensors and joint encoders produce a constant stream of data. A mobile robot may carry several cameras at once.
- Think. Perception models identify objects and people, a planning layer decides what to do, and increasingly a single foundation model handles language, vision and action together.
- Act. Motor controllers turn the decision into torque, speed and position, with safety limits that override the AI if something goes wrong.
The loop has a deadline. Perception and control that run in milliseconds can't wait for a round trip to a cloud server over a mobile network, which is why the compute has to sit on the machine itself.
Why don't robots just use a data-centre GPU?
Because a robot has a battery, a heat limit and no guaranteed internet connection. A data-centre GPU can draw several hundred watts on its own. A robot computer has to fit inside the chassis, survive vibration and run on a power budget measured in tens of watts. It also needs many camera inputs, industrial interfaces such as CAN bus, and predictable timing.
NVIDIA's answer for robots is the Jetson family. Its newest module, Jetson Thor, delivers up to 2,070 FP4 TFLOPS and 128 GB of LPDDR5X memory in a power envelope of 40 to 130 W, which NVIDIA says is 7.5 times the AI performance of Jetson AGX Orin at 3.5 times the energy efficiency. Its memory bandwidth is about 273 GB/s, and the developer kit includes four 25GbE ports for high-bandwidth sensors. The headline TFLOPS figure is a sparse FP4 number; dense FP8 is about a quarter of it, which is why the same chip can look very different depending on the model format (see FP4 vs FP8 vs INT8 vs INT4).
| Edge computer | Memory | Best fit |
|---|---|---|
| Jetson Orin Nano Super 8GB | 8 GB | Learning, single-camera vision, small robots |
| Jetson AGX Orin 32GB / 64GB | 32 or 64 GB | Multi-camera AMRs, inspection cells, existing Orin fleets |
| Jetson AGX Thor 128GB | 128 GB | Humanoids and robots running generative models on board |
| DGX Spark / GB10 desktop | 128 GB | Developing and fine-tuning models before deploying to Jetson |
| RTX PRO 6000 workstation or DGX Station | 96 GB VRAM and up | Simulation, synthetic data and training |
Source: NVIDIA Jetson Thor product page and developer blog (Aug 2025); SourceIT product listings. Compiled 11 Nov 2026.
Why does physical AI depend so much on simulation?
Collecting real-world robot data is slow, expensive and sometimes dangerous. You can't crash a forklift a thousand times to teach it what not to do. So most physical AI is trained and tested in simulation first, using digital twins of the factory or warehouse, then refined on real hardware.
World foundation models are speeding this up. NVIDIA launched Cosmos 3 at GTC Taipei on 1 June 2026 as an open model that can act as a vision-language model, a world model that predicts future states, or the backbone for action models that train robots. NVIDIA describes it as a mixture-of-transformers design, related to the mixture-of-experts approach used in large language models. Simulation and synthetic data generation are the heaviest part of the compute bill, and they run on workstations and servers, not on the robot.
A robot needs three computers, not one: a large system to train and simulate, a desktop to develop and test, and a small, power-efficient computer on the machine itself. Buying only the biggest GPU covers the first and leaves the robot with nothing.
Where is physical AI being used today?
- Manufacturing inspection. Vision models spot defects on a production line and trigger a reject. We cover the hardware in AI for manufacturing quality inspection.
- Logistics. Autonomous mobile robots move goods in warehouses and hospitals, and pick-and-place arms handle mixed items.
- Facilities. Cleaning, security patrol and delivery robots in malls, offices and hotels.
- Healthcare. Assistive and logistics robots, where safety and data rules are strict.
Most Singapore businesses won't build robots. They will buy them, or buy camera systems with AI inside. The questions that matter are the same either way: where the AI runs, what happens when the network drops, who updates the models, and how the system fails safely.
What should a business check before a physical AI project?
- Latency budget. How fast must the system react? Anything under a second points to on-device compute.
- Power and heat. What can the machine or enclosure supply and dissipate?
- Sensor count. Every extra camera adds bandwidth and compute. Plan the interfaces before choosing the computer.
- Training and simulation. Who trains and updates the models, and on what hardware?
- Safety. The AI should never be the only safety layer. Hardware limits and emergency stops must work without it.
- Lifecycle. Edge computers sit in machines for years. Check the module's availability and software support period.
Buying in Singapore
For learning and prototypes, the NVIDIA Jetson Orin Nano Super 8GB Developer Kit (S$696) is the entry point. For multi-camera robots, there's the Jetson AGX Orin 32GB (S$3,285) and Jetson AGX Orin 64GB (S$5,871.60) developer kits. For generative AI on the robot itself, the NVIDIA Jetson AGX Thor 128GB Developer Kit (S$9,075.60). For model development, the NVIDIA DGX Spark 4TB (S$9,450) and ASUS Ascent GX10 128GB/1TB (S$10,350) share an Arm-based software stack with Jetson. For simulation and training, the NVIDIA RTX PRO 6000 Blackwell Workstation Edition (S$29,100) (listed 11 Nov 2026; prices change; ask for a current quotation). DGX Station GB300 systems such as the ASUS ExpertCenter Pro ET900N G3 are price on request. We don't list robots, lidar or industrial cameras, so ask us if you need them sourced. To start, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
What is physical AI in simple terms? Physical AI is AI that senses the real world through cameras and other sensors, decides what to do, and acts through a machine such as a robot, vehicle or inspection system.
Is embodied AI the same as physical AI? They overlap. Embodied AI is the research term for AI that has a body and learns by interacting with its environment. Physical AI is the broader industry term for AI that perceives and acts in the physical world.
Why can't a robot just use a cloud AI service? Robots must react in milliseconds and keep working when the network drops. A round trip to a cloud server is too slow and unreliable for perception and control, so the AI runs on a computer inside the machine.
What is NVIDIA Jetson Thor? Jetson Thor is NVIDIA's robot computer module. It offers up to 2,070 FP4 TFLOPS and 128 GB of memory at 40 to 130 W, and NVIDIA says it delivers 7.5 times the AI performance of Jetson AGX Orin.
What hardware does a physical AI project need? Usually three tiers: workstations or servers for simulation and training, a developer desktop such as DGX Spark for testing models, and an efficient edge computer such as Jetson on the machine itself.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
