The questions I get most about Physical AI, with short answers. Each group links to a page where I take more time. If yours is not here, send it to me.
Last updated: 12 November 2026.
Physical AI
Is Physical AI the same as robotics?
Robotics is a major part of it. Physical AI also includes intelligence integrated into instruments, industrial equipment, vehicles, and other systems that sense and affect the physical world. The important question is how perception, decisions, and physical consequences are connected.
Does Physical AI require a large language model?
No. A language model can help interpret an instruction or plan a sequence, but a useful physical system may rely on smaller perception models, state estimation, conventional control, or combinations of these. Its suitability depends on measurable behavior in the intended environment.
Why does simulation not settle the problem?
A simulator is a model of the world. Errors in friction, contact, lighting, timing, or sensing can be enough to break a policy that looked reliable on screen. Sim2Real requires tests on hardware, measurements of the remaining gap, and operating limits.
Does neuromorphic computing mean copying a whole brain?
No. Biological principles can inform a circuit, learning rule, architecture, or sensing method without reproducing an entire nervous system. The engineering case rests on what those choices deliver for the task and its resource budget.
How does my NASA and DARPA work relate to today’s Physical AI?
It connects through a persistent problem: getting learning systems to perceive, adapt, and act under real constraints. The NASA accounts, university reporting, collaborative papers, and patent records above document parts of that history. They do not imply that every subsequent system uses the same architecture or that research intended for planetary navigation flew on a Mars rover.
More: Physical AI.
Physical AI and generative AI
Is Physical AI the same as generative AI?
No. Generative AI produces content: text, images, code. Physical AI produces behavior: it senses, decides and moves a machine in the real world, on time and on a power budget.
Is a humanoid robot generative AI or Physical AI?
Physical AI. It may use a generative model to understand an instruction, but its job is to act in the world, and that is where the hard limits are.
Can generative AI run a robot?
Partly. Vision-language-action models connect the language ability of generative AI to motors. The robot still needs fast, low-power perception and control on board, because a late answer is a wrong answer.
Which one is harder?
Different kinds of hard. Writing a sonnet turned out to be easier for machines than folding a shirt. Almost nobody saw that coming, except maybe the roboticists.
More: Physical AI and generative AI.
Physical AI, embodied AI, robotics and edge AI
Is Physical AI the same as robotics?
No. Robotics is the engineering of machines that move and act, and a robot can work with no learning at all. Physical AI is about the intelligence inside the machine: perceiving, learning and deciding in the physical world. Many Physical AI systems are robots, but not all robots use AI.
Is Physical AI the same as embodied AI?
Almost. Embodied AI is the older, academic name for AI that learns and acts through a body, and much of that research is done in simulation. Physical AI is the name industry uses today, and it puts more weight on real machines, real sensors and real constraints such as power and latency.
Is Physical AI the same as edge AI?
No. Edge AI says where the computing happens: on the device and not in a data center. A device can run AI at the edge without acting on the world at all. Most Physical AI needs edge AI, because a machine that moves cannot wait for an answer from far away.
More: Physical AI, embodied AI, robotics and edge AI.
Continual learning and forgetting
What is continual learning?
Continual learning is the ability of a system to learn new things over time, one after the other, without retraining from scratch and without losing what it learned before. It is also called lifelong learning.
What is catastrophic forgetting?
Catastrophic forgetting is what happens when a neural network is trained on something new and, in the process, overwrites what it already knew. The term goes back to a 1989 paper by McCloskey and Cohen, who called it catastrophic interference.
Why does continual learning matter for robots?
A robot works in a world that keeps changing: new objects, new lighting, new tasks. If every change means sending data away and retraining the whole network, the robot is always a step behind. A robot that learns on the spot, and keeps what it knew, is useful for much longer.
How do you reduce catastrophic forgetting?
There are three main families of methods. Replay: mix old examples, or generated ones, with the new. Protection: make the connections that mattered for old tasks harder to change. Structure: give new knowledge its own room in the network, and learn only when the input matches what the system expects, as in Adaptive Resonance Theory.
More: Continual learning and forgetting.
Working in Physical AI
What does a Physical AI engineer do?
A Physical AI engineer builds the intelligence of machines that sense and act in the real world: robots, drones, vehicles, industrial equipment. The work sits between machine learning, robotics and embedded systems.
What skills does a Physical AI engineer need?
Machine learning for perception and control, real-time and control engineering, embedded computing under tight power and memory budgets, sensors and signal processing, simulation, and disciplined testing on real hardware.
What should I study to work in Physical AI?
There is no single degree. People come from robotics, computer science, electrical engineering, physics and computational neuroscience. What they share is that they have built something that had to work outside a computer.
More: Working in Physical AI.
The leaky integrator
What is a leaky integrator?
A leaky integrator is a system that adds up its input over time while steadily losing part of what it has accumulated. It is described by the equation dx/dt = −Ax + I, where x is the activity, I is the input and A is the leak rate.
What is the time constant of a leaky integrator?
The time constant is 1/A. It is the time the system takes to cover about 63% of the distance to its new resting level after the input changes.
What is the difference between a leaky integrator and a leaky integrate-and-fire neuron?
A leaky integrate-and-fire neuron is a leaky integrator with two extra rules: when the activity crosses a threshold the neuron emits a spike, and the activity is reset.
Is a leaky integrator a low-pass filter?
Yes. It follows slow changes in its input and smooths out fast ones. In discrete time it is the same as exponential smoothing, also called an exponential moving average.
More: The leaky integrator.