Glossary term

Sim-to-real

Training a system in simulation and transferring it to work in the real world.

How it works

Training a robot in the real world is slow and risky, so it often learns in simulation first. The challenge is the 'reality gap': simulated physics and visuals differ from the real world.

Randomizing the simulation, for example varying friction, lighting and weights, helps the learned skill carry over.

Sim-to-real

Step 1Train in simulation
Step 2Randomize conditions
Step 3Transfer to the real robot
Step 4Fine-tune with real data

In code

Domain randomization

import random

def randomize_world():
    return {
        "friction": random.uniform(0.5, 1.5),
        "mass": random.uniform(0.8, 1.2),
        "light": random.uniform(0.3, 1.0),
    }

Why it matters

It lets robots practice millions of times safely and cheaply before touching the real world.

Sim-to-real

How it works

In code

Why it matters

Related terms

Sim-to-real

How it works

In code

Why it matters

Related terms