← Glossary · Concept
Physical AI — when an agent reaches into the world
Physical AI connects models, robots, simulation and actions in the real environment. It is not about a cute robot demo, but about who carries the risk when a model starts moving things.
What it is
Physical AI is the umbrella term for systems where a model does not only handle text, images or code, but affects the physical world. A robotic arm, humanoid, warehouse cart, home device, drone or autonomous lab. The model reads state, plans an action and turns it into movement through software.
Sometimes this is a real robot. Sometimes it is simulation, a world model or a video agent learning how the physical world should respond. The boundary is blurry, because product demos love selling simulation as almost-reality. Product-friendly, but safety-hostile.
How it differs from computer-use agents
A computer-use agent clicks through a digital environment. If it fails, it breaks a ticket, spreadsheet or account. Physical AI adds space, time, material and human safety. The failure is no longer just bad JSON.
So “the model can plan” is not enough. It needs sensors, control layers, force limits, geofencing, fallback modes, audit logs and often local decisions without the cloud. Reality has latency and sharp edges.
Why it is coming back now
Better multimodal models can read images, video, instructions and environment state in one context. Simulation and synthetic data make training cheaper. MCP and tool use make it easier to attach a body to external capabilities: maps, weather, databases, calendars and internal APIs.
The important shift is that the robot does not have to carry every capability inside one application. The body can stay relatively dumb while capabilities arrive as services. Practical. Also explosive from a safety point of view.
Where the risk is
- A demo environment is not production. A tidy table is not a warehouse, kitchen or street.
- Simulation is not the world. The model may learn rules that do not hold in reality.
- Tools extend reach. A robot with API access is not just a robot, but an action node in a system.
- Responsibility splits across the model maker, hardware maker, integrator and operator.
- Without local limits, a cloud brain becomes one very pretty failure point.
What to watch
Ask whether the system can fail safely. Whether it separates planning from low-level control. Whether it has action logs, realistic tests and clear permissions. Most of all: whether the metric measures the real task or just a shiny video.
Physical AI will matter. But every “the model finally acts in the world” has a second sentence: the world is not a sandbox.