AI Trend Research

← AI glossary

AI glossary · Advanced · Robotics & physical AI · Reinforcement learning

Sim-to-Real

Sim-to-Real is a technique for training a robot or AI in a virtual environment on a computer (a simulation) and then transferring those skills to the real world.

About this term
Level
Advanced · Specialist terms that need the inner workings
Added
2026-10-08
In plain words It is like a pilot practicing countless times in a flight simulator before actual flights. But just as a pilot can be thrown off in real flight if the simulator differs even slightly from reality, a robot can also make mistakes because of the gap between virtual and real.

How it works

First, the robot and its surroundings are built inside a simulator that mimics the laws of physics. There, the robot learns movements such as walking and grasping by trying them millions of times through reinforcement learning and other methods. During training, floor friction, object weight, lighting, sensor noise and so on are deliberately and continually varied; this is called "domain randomization." A robot that has become used to varied conditions holds up well even when reality differs somewhat from the simulation. The control model (policy) obtained this way is transferred to the real robot and sometimes refined again with a small amount of data collected in the real world.

Sim-to-Real training flowSimulator → Domain randomization; Domain randomization → Reinforcement learning training; Reinforcement learning training → Policy model; Policy model → Deploy to real robot(Transfer); Deploy to real robot → Calibrate with real-world data(Small amount of data); Calibrate with real-world data → Reinforcement learning trainingSimulatorVirtual robot andenvironmentDomainrandomizationVary friction, weight,lightingReinforcementlearning trainingExtensive trial anderrorPolicy modelMotion controlDeploy to realrobotCalibrate withreal-world dataTransferSmall amount of data
Sim-to-Real training flow A policy trained in a virtual environment under varying conditions is transferred to a real robot and refined again with real-world data.

Why it matters

Repeated trial and error with real robots takes a long time and risks breaking machines or injuring people. In simulation, many robots can be trained at the same time and much faster than in reality. Falling over or dropping things costs nothing. That is why it is used as a core training method in fields where data is hard to collect, such as humanoids and robot arms.

Things to keep in mind

The gap between virtual and real is commonly called the "reality gap," and it is the biggest challenge. Objects that are hard to simulate, such as cloth or liquids, and fine contact and friction are especially likely to diverge from reality. If randomization is too broad, learning becomes difficult; if it is too narrow, the robot fails in reality. So rather than relying on simulation alone, it is often combined with real data.

Example

Robotics companies widely train basic movements such as walking and balancing first in large-scale simulation and then transfer the results to real humanoids or quadruped robots. In effect, the gait learned by thousands of virtual robots falling and getting up at the same time inside a simulator carries over to the steps of a single real robot.

Related terms

Humanoid Reinforcement learning World model Synthetic data VLA