Every field has its jargon, and Physical AI borrowed from three at once: robotics, neuroscience and chip design. After 25 years of moving between the three, I end up translating a lot. So here are the terms I get asked about most, in plain words. No equations, I promise.
Physical AI · Physical intelligence · Embodied AI · Edge AI · Latency · Sim2Real · Digital twin · World model · Robot foundation model · Vision-language-action model · Teleoperation · Dexterous manipulation · Sensor fusion · Humanoid robot · Continual learning · Catastrophic forgetting · Backpropagation · Adaptive Resonance Theory · Mismatch negativity · Neuromorphic computing · Spiking neural network · Spike-timing-dependent plasticity · Memristor · In-memory computing · Biological intelligence · Leaky integrator
Physical AI
Artificial intelligence that perceives, learns and acts in the physical world, inside a machine with a body, sensors, motors and a limited power budget. A chatbot can take a few seconds to answer and nothing bad happens. A robot arm, a drone or a rover cannot. Read more: my Physical AI page.
Physical intelligence
What a machine has to understand to get things done in the real world: where things are, how they move, what happens if it pushes, grabs or lets go. Animals have it long before they have language. Most AI today has it the other way around.
Embodied AI (embodied intelligence)
AI that learns and acts through a body. The body is not a detail: it gives the system constant feedback and punishes bad predictions right away. The term comes from robotics and cognitive science, and it overlaps almost completely with Physical AI. Read more: What happens when AI gets a body?.
Edge AI
AI that runs on the device itself (a camera, a phone, a robot) instead of a data center. I like to say that Edge AI is not a place, it is a constraint: what counts is how much latency, power, memory and connectivity you can afford. Read more: Edge AI is not a place, it’s a constraint.
Latency
The delay between sensing something and acting on it. For a chatbot, a second is fine. For a robot catching a falling glass, a tenth of a second is already too late. In Physical AI, latency is part of the spec.
Sim2Real (simulation to reality)
Training a robot in a simulated world, then running it in the real one. Simulation is cheap and safe, and you can run a thousand experiments overnight. The catch is the gap: leave friction or sensor noise out of the simulator and the robot becomes a champion of a world that does not exist.
Digital twin
A virtual copy of a real machine, factory or environment, kept up to date with data from the real thing. In robotics it is used to test and train before touching the hardware.
World model
A model, learned from data, that predicts what will happen next in the world, including what happens if the machine takes a certain action. Brains do this all the time. It lets a robot try things in its head before trying them with its arm.
Robot foundation model
A large model trained on data from many robots and many tasks, meant to be a general starting point for new robots and new jobs, the way large language models are for text.
Vision-language-action (VLA) model
A model that takes camera images and an instruction in plain language (“put the cup on the shelf”) and outputs the robot’s movements directly. It is the current favorite recipe for general-purpose robots.
Teleoperation
A person driving a robot remotely, often to collect examples of how a task should be done. A lot of today’s robot training data comes from people wearing headsets and moving robot arms by hand.
Dexterous manipulation
Handling objects with fingers the way people do: turning a key, folding a shirt, picking a raspberry without crushing it. It is one of the hardest open problems in robotics, because it mixes vision, touch and very fast control.
Sensor fusion
Combining signals from several sensors (cameras, touch, sound, motion, position) into one picture of what is going on. No single sensor is reliable all the time. Together they can be.
Humanoid robot
A robot with a human-like shape: two legs, two arms, a head. The shape fits a world built for people. But a human shape does not mean a human mind… building the brain is the hard part, not the body. Read more: Brains, Not Bodies.
Continual learning (lifelong learning)
Learning new things over time, one at a time, without retraining from scratch and without losing what was learned before. Animals do it from birth. For most AI it is still an open problem, and for machines in the field it is a daily need. Read more: What happens when AI never stops learning?.
Catastrophic forgetting
What happens when a neural network learns something new and, in the process, wipes out what it knew before. You learned a new face this morning and you still remember your mother’s. Many networks cannot say the same. Read more: Overcoming the Catastrophe of Forgetful AI.
Backpropagation
The standard way to train deep networks: measure the error at the output and send corrections backward through every layer. It works very well in a data center. It is heavy, slow and hungry, which makes it a poor fit for learning on a small device.
Adaptive Resonance Theory (ART)
A family of brain models, started by Stephen Grossberg in the 1970s and developed with Gail Carpenter, in which learning happens only when what comes in from the senses matches what the brain expects. That match-first rule is what lets the system learn new things without erasing old ones. Read more: Learning That Never Stops.
Mismatch negativity
A dip in the brain waves (EEG) that shows up roughly a tenth to a quarter of a second after something unexpected, like an odd tone in a series of identical ones. It is a visible trace of the brain’s prediction being wrong. Read more: Learning That Never Stops.
Neuromorphic computing
Chips and systems designed to work like nervous systems: many simple units, memory sitting next to computation, activity only when something happens. The name goes back to Carver Mead in the late 1980s. The promise is a lot of intelligence on very little power. Read more: The next AI revolution will be hardware + software.
Spiking neural network
A neural network whose units talk through short pulses (spikes), like real neurons, instead of continuous numbers. If nothing changes, nothing fires and almost no energy is spent.
Spike-timing-dependent plasticity (STDP)
A learning rule found in real synapses: if one neuron fires just before another, the connection between them gets stronger; if it fires just after, weaker. It uses only local information, which makes it a natural fit for hardware.
Memristor
An electronic component whose resistance depends on the current that has flowed through it, so it remembers its past. Proposed by Leon Chua in 1971 and built at HP Labs in 2008. It behaves a little like a synapse, which is why we used it for brain-inspired hardware.
In-memory computing
Doing the computation where the data is stored, instead of shuttling data back and forth between memory and processor. That shuttling is where much of a computer’s energy goes. Brains never had the problem: synapses store and compute in the same place. Read more: More data centers is the answer. But what was the question?.
Biological intelligence
The kind of intelligence that evolved in animals. It runs on very little energy, learns all the time from a handful of examples, and exists to keep a body alive in a world that keeps changing. It is the benchmark I measure AI against: a squirrel on a few nuts a day still beats most robots.
Leaky integrator
The simplest useful model of a neuron: it adds up its inputs over time, and the total slowly leaks away when nothing new arrives. I promised no equations, so this is the only one: dx/dt = −Ax + I, where x is the activity, I the input and A how fast it leaks. Most models of spiking neurons start here.
I keep everything on Physical AI in one place, from papers and patents to 15 years of press: Physical AI.