Physical AI from A to Z.

A component that turns energy into motion -- for example, an electric motor in a robot joint.

A collaborative robot designed to work alongside people without separating safety fences.

A robot that uses AI models to perceive, learn, and adapt to changing situations instead of only replaying hard-coded routines.

The time needed for one complete work sequence -- a key benchmark between people, classic systems, and robots.

The ability to decide where your data is stored and processed, who uses it, and for what.

The number of independent ways a robot can move. A typical industrial arm has six; a biped often more than twenty.

A recorded run of a task with sensor data and movements -- the unit of training in imitation learning.

A camera that also measures distance for every pixel, for example via stereo vision or infrared patterns.

A virtual replica of a facility or environment in which processes can be simulated, tested, and used to generate training data.

Processing data directly on the robot or in your own plant instead of in a remote cloud.

Related term for Physical AI that emphasizes intelligence being tied to a body. The origin of the name EmAI.

The tool at the end of the arm: gripper, suction cup, robotic hand, or special tool.

Adapting a pre-trained model to a specific task with comparatively little data of your own.

Measures forces and torques, usually at the wrist -- important for assembly and safe collaboration.

A large, broadly pre-trained AI model that serves as the starting point for many specific tasks.

Protective functions that reliably prevent hazards to people -- such as emergency stop and speed and force limits.

An operating model in which a person monitors, approves, or takes over the robot’s decisions when needed.

Learning by imitation: a model learns a task from human demonstrations.

Running a trained model in operation -- on a robot, often many times per second.

A person stepping in to correct a robot or fix an error. A key indicator of a system’s maturity.

A sensor that scans the surroundings with lasers and produces a 3D point cloud.

A robot arm on a wheeled platform that can carry out tasks in changing locations.

An AI model whose trained weights are freely available. It can be inspected, adapted, and run yourself.

IT security for production systems (operational technology): protecting machines and controllers against attacks.

AI systems that perceive and act in the physical world through a body -- such as cognitive robots.

A robot’s learned control model: it maps current perception to the next action.

Learning by trial and reward: the robot tries actions and reinforces what leads to the goal.

A systematic analysis of a machine’s hazards and the measures against them. Mandatory before commissioning.

Robot Operating System: a widely used open-source framework with tools and libraries for robotics software.

Transferring behavior learned in simulation to the real robot.

Simultaneous localization and mapping: the robot builds a map of its surroundings while determining its position in it.

The share of runs in which a robot completes a task without human help.

Sensors on fingers or grippers that detect touch and pressure, enabling delicate grasping.

Remote control of a robot by a person, for example via VR headset or a leader arm. The main source of training data.

Total cost over the service life: purchase, integration, energy, maintenance, training, and data preparation.

Dependency on a supplier whose technology can only be replaced at great effort.

Vision-language-action model: processes camera images and language instructions and outputs robot actions directly.

A model that predicts how the environment changes as a result of an action -- the basis for planning ahead.