Introduction

Popular technique for improving domain transfer, often used in zero-shot setting when the target domain is unknown or cannot easily be used for
training.

An ideal agent learns to generalise across domains.

DR uniformly randomises environment parameters (i.e. friction, torque) in predefined ranges after every training episode.

It is hoped that by randomising every current present variable that is shared with target environment, agent will view target domain as just another variation.

Sample complexity grows exponentially with number of randomisation parameters.

Methodology

Simulators - Domain randomization uses simulation environments where elements can be modified easily.

Randomized Parameters - Key parameters within the simulation that affect the model's learning are randomized. These could include:

• Visual: Lighting, textures, object colors, camera angles, background clutter
• Physics: Friction, gravity, object mass, motor strengths
• Task: Goal positions, object configurations

Model Training - The machine learning model (like a robotic control system) is trained repeatedly over many iterations of the simulation, each iteration presenting a different randomized combination of parameters.

Importance

The "Reality Gap" - Machine learning models trained solely on perfect simulation data often fail when placed in the messy, unpredictable real world.

Closing the Gap - Domain randomization forces the model to learn the core features and principles of a task independent of the specifics of a single environment.

Safer, Faster Learning - Training directly in the real world for things like robotics can be expensive, slow, and potentially dangerous.

Example - Robotics

Imagine training a robotic arm to grasp objects. Domain randomization might involve:

• Changing the textures and colors of the objects
• Randomly placing objects within its reach
• Altering lighting conditions in the simulation
• Slightly modifying the simulated friction of the arm joints or strength of the gripper