Domain Randomization (DR)

The core mechanism of DR is the deliberate, systematic variation of specific simulation parameters to create a diverse training distribution. This forces the model to learn task-relevant features that are invariant to these changes.

• Visual Parameters: Textures, colors, lighting (position, intensity), camera noise, and background scenes.
• Dynamics Parameters: Mass, friction, damping, actuator latency, and motor strength.
• Environmental Parameters: Object positions, gravity, and wind forces.

By sampling these parameters from a defined parameter distribution (e.g., uniform, Gaussian), each training episode presents a unique environment.

DR is often categorized by the type of parameters being randomized, each addressing different aspects of the reality gap.

• Visual Domain Randomization: Focuses on randomizing perceptual inputs. This trains vision models to be robust to changes in appearance, lighting, and camera artifacts, which is critical for tasks like object detection in varying conditions.
• Dynamics Randomization: Focuses on randomizing the physics of the simulation. This trains control policies (e.g., for robotics) to be robust to variations in mass, friction, and actuator dynamics encountered in the real world.

A comprehensive DR strategy often employs both to cover perceptual and physical discrepancies.

The primary objective of DR is to enable zero-shot sim-to-real transfer. Instead of training in a single, high-fidelity simulation that attempts to perfectly match reality—a difficult and often impossible task—DR embraces low-fidelity simulation.

By exposing the model to an extremely broad distribution of simulated worlds, the real world becomes just another unlikely sample from that distribution. The model learns a robust policy that generalizes across this distribution, thereby performing effectively when deployed on physical hardware, closing the reality gap.

Through exposure to randomized environments, models are compelled to perform invariant feature learning. They learn to discard irrelevant, randomized features (like a specific texture color) and focus on essential, task-relevant features (like object shape or dynamics).

• This is the conceptual opposite of overfitting to a specific simulation configuration.
• In computer vision, a model might learn to recognize an object by its geometric edges rather than its specific color or material.
• In robotics, a policy might learn to grasp an object based on its center of mass and contact points, not the exact friction coefficient of a particular tabletop.

Early DR used manually defined, static randomization ranges. Advanced methods now automate and optimize this process.

• Systematic Domain Randomization: Parameters are varied in a controlled, often factorized manner to ensure comprehensive coverage of the parameter space without over-randomization.
• Automatic Domain Randomization (ADR): An algorithm (e.g., using reinforcement learning) actively searches for the most challenging parameter ranges that still allow learning, optimizing the randomization schedule for maximum robustness.
• Curriculum Randomization: The range or difficulty of randomization is progressively increased, allowing the model to learn basic skills in easier environments before tackling greater variability.

Sim-to-Real Transfer is the ultimate objective of Domain Randomization: deploying a model trained entirely in simulation to perform effectively on a physical system. Success is measured by Sim2Real Performance, a key metric quantifying the policy's effectiveness in the real world. The goal is often Zero-Shot Sim-to-Real, where the model works without any fine-tuning on real data. This process directly addresses the Reality Gap—the performance drop caused by discrepancies between the simulated training environment and physical reality.

The Domain Gap is the statistical discrepancy between data distributions in a source domain (e.g., simulation) and a target domain (e.g., reality). When this gap causes model failure, it manifests as the Reality Gap. Domain Randomization aims to minimize this by:

• Training the model across a vast distribution of simulated domains.
• Encouraging Invariant Feature Learning, where the model extracts task-relevant features that are consistent despite visual or physical variations.
• The technique's success is evaluated by its ability to achieve Cross-Domain Generalization.

These are the two primary axes of parameter variation in DR.

Visual Domain Randomization randomizes perceptual properties to build robustness against appearance changes:

• Textures, colors, and object materials
• Lighting conditions, shadows, and camera parameters (noise, blur)
• Background scenes and distractors

Dynamics Randomization varies physical parameters to build robustness against real-world physics:

• Object mass, inertia, and dimensions
• Actuator dynamics (motor strength, latency)
• Surface friction, restitution, and damping coefficients

Advanced strategies to optimize the randomization process.

Automatic Domain Randomization (ADR) algorithmically searches for the most effective randomization parameters, continuously adapting the Parameter Distribution during training to maximize robustness, reducing the need for manual tuning.

Curriculum Randomization employs a Randomization Schedule that progressively increases the range or difficulty of parameters. Training starts with a narrow, easy distribution and gradually expands to more challenging variations, helping the model learn coherent policies before confronting extreme scenarios and avoiding Over-Randomization.

This is the core mechanistic layer of DR. Parameter Perturbation involves the deliberate variation of specific simulation parameters. The Parameter Distribution defines the statistical range from which these values are sampled (e.g., uniform, Gaussian). A Randomization Pipeline automates this sampling and configures simulation instances.

Systematic Domain Randomization applies perturbation in a controlled, often factorized manner to ensure broad and efficient coverage of the parameter space. The Physics Randomization Engine is the software component within a simulator that dynamically alters these physical parameters.

Frameworks and methods to test and deploy randomized models.

A Domain Randomization Benchmark is a standardized test suite (simulator + real-world tasks) for evaluating different DR methods.

Hardware-in-the-Loop (HIL) Randomization integrates DR into a testing setup where a physical robot controller interacts with a randomized simulation in real-time, providing a critical bridge between virtual training and physical hardware validation.

Randomized-to-Canonical is an alternative approach where a model learns to map randomized observations back to a canonical, non-randomized representation, explicitly learning invariant features.