Glossary

Short definitions of the terms used across this hub, in alphabetical order. Each links to the page where the idea is explained.

Action head

The last part of a vision-language-action model, which turns the model's internal state into motor commands. See world models and VLAs.

CFD (computational fluid dynamics)

Simulation of gases and liquids, and usually heat with them, by dividing the volume into cells. OpenFOAM, STAR-CCM+ and Ansys Fluent are CFD tools. See traditional physics-based simulation.

Contact solver

The routine in a physics engine that finds the forces between touching objects so they do not pass through each other. See what is a solver?.

Digital twin

A simulation of one specific real machine, kept in sync with it through live sensor data. See digital twin vs. simulation.

Domain randomization

Training a policy while randomly varying the simulator's parameters (mass, friction, motor strength, textures) so that the real values fall within what the policy has seen. See the sim-to-real gap.

FEA (finite element analysis)

Simulation of solid parts (bending, vibration, heating) by dividing the part into small elements. Ansys Mechanical and CalculiX are FEA tools.

GPU

Graphics processing unit. A chip designed to do thousands of small calculations at once, originally for drawing images, now used to run thousands of simulated robots in parallel.

Isaac Sim, Isaac Lab

NVIDIA's robot simulator and its companion training framework. See which simulators exist today?.

MJCF

MuJoCo's XML file format for describing a robot: its parts, joints, masses and shapes.

MPC (model predictive control)

A control method that, at every moment, simulates a short stretch of the future with a physics model and picks the action that works best over it. An alternative and a complement to learned policies.

MuJoCo

Multi-Joint dynamics with Contact. A physics engine and simulator maintained by Google DeepMind, known for accurate contact modelling. See which simulators exist today?.

Newton

An open-source GPU physics engine developed by NVIDIA, Google DeepMind and Disney Research, hosted by the Linux Foundation, built on NVIDIA Warp. See what is a physics engine?.

Neural operator

A neural network that learns to map one whole field (a shape, a boundary condition) to another (a pressure or temperature map), used to replace slow physics solvers. See AI-accelerated physics.

Physical AI

NVIDIA's term for AI systems that perceive and act in the physical world: robots, autonomous vehicles and the models that drive them.

Physics engine

The part of a simulator that computes movement, collisions and friction, one small time step at a time. See what is a physics engine?.

PINN (physics-informed neural network)

A neural network trained so that it is penalised whenever its output violates the physics equations, which lets it learn from less data.

Policy

The rule, usually a neural network, that maps what a robot senses to what it does. The thing that reinforcement learning trains. See what is reinforcement learning?.

PPO (Proximal Policy Optimization)

The reinforcement-learning algorithm most robots are trained with, published by OpenAI in 2017 and still the default in 2026.

Reinforcement learning (RL)

Learning by trial and error: the robot acts, receives a score (reward) and adjusts to score higher. See what is reinforcement learning?.

Reward

The number a reinforcement-learning system receives after each action, saying how good that action was.

ROS 2

Robot Operating System. The open-source software framework most robots are built on; Gazebo is its traditional simulator.

Sim-to-real gap

The difference between how a robot behaves in simulation and in reality. See the sim-to-real gap.

Solver

The numerical procedure that turns physics equations into numbers, step by step. See what is a solver?.

Surrogate model

A neural network trained on the results of a physics solver so that it can predict similar results in a fraction of the time.

Synthetic data

Training data produced by a simulator instead of collected in the real world: rendered images with automatic labels, or simulated robot trajectories.

Teleoperation

A human controlling a robot remotely, often to record demonstrations that a policy is later trained on.

Time step

The small slice of time a physics engine advances in one calculation, typically a thousandth of a second.

URDF

Unified Robot Description Format. The XML file format from ROS that describes a robot's parts and joints.

USD (OpenUSD)

Universal Scene Description. A file format created by Pixar for describing 3D scenes, used by NVIDIA Omniverse, Isaac Sim and Newton.

VLA (vision-language-action model)

A neural network that takes camera images and a text instruction and outputs robot movements directly. See world models and VLAs.

Warp

NVIDIA Warp. A Python framework that compiles simulation code into GPU programs, with support for automatic differentiation. The foundation of Newton and MuJoCo Warp, and of Invisible Force's modules.

World model

A neural network that predicts what happens next given the current state and an action, learned from video or simulation. Can act as a learned simulator. See world models and VLAs.