Domain Randomization

This category randomizes parameters that affect how objects and scenes look, decoupling the model from specific textures, colors, and lighting conditions.

• Textures & Materials: Applying random, often unrealistic, colors, patterns, and surface properties (e.g., wood, metal, plastic) to all objects.
• Lighting: Varying the number, type, position, color, and intensity of light sources in the scene.
• Camera Properties: Altering parameters like field of view, focal length, exposure, white balance, and sensor noise to mimic different hardware.
• Backgrounds: Replacing scene backgrounds with random images or synthetic patterns.

Example: Training a robotic grasping model with objects that appear as neon green checkered cubes, matte purple spheres, and glossy polka-dotted cylinders under randomly colored lighting.

This involves randomizing the physical shape, arrangement, and quantity of elements in the simulation to prevent overfitting to a specific configuration.

• Object Poses: Randomizing the position, orientation (6D pose), and scale of target objects and distractors.
• Object Shapes & Sizes: Using a diverse set of 3D models or randomly perturbing the dimensions of base models.
• Scene Layout: Varying the placement of furniture, walls, and other environmental structures.
• Object Count: Changing the number of instances of objects present in a scene.

Example: For an autonomous vehicle perception model, randomizing the number of cars on a road, their makes/models, their distances from each other, and the curvature of the road itself.

This randomizes the laws of motion and interaction within the simulator, forcing the model to adapt to different physical realities.

• Mass & Inertia: Varying the mass and inertial properties of objects.
• Friction Coefficients: Randomizing static and dynamic friction for object-object and object-environment interactions.
• Motor Dynamics: Applying noise or delay to actuator commands and varying force/torque limits.
• Gravity & Drag: Altering the strength of gravity or adding random wind forces.

Example: Training a drone flight controller in a simulator where gravity randomly varies between 0.5g and 1.5g, and rotor thrust efficiency changes between 70% and 110%.

This injects realistic imperfections into the model's observations and actions, mimicking the noise and latency of real-world hardware.

• Sensor Noise: Adding Gaussian noise, dropout, or bias to camera pixels, LiDAR point clouds, joint encoders, and IMU readings.
• Latency & Delay: Simulating variable communication delays between sensor perception and actuator commands.
• Calibration Errors: Introducing systematic offsets to sensor measurements (e.g., a camera always tilted 2 degrees).
• Actuator Saturation: Modeling the non-linear response and limits of real motors.

Example: Providing a robot arm with joint angle readings that are jittery (±1 degree) and slightly biased, while the motor commands experience random 10-50ms delays.

This targets high-level, semantic variations that represent different 'domains' or operational conditions the model may encounter.

• Weather & Atmospheric Conditions: Simulating rain, fog, snow, or dust on camera lenses and sensors.
• Time of Day: Cycling through lighting conditions representing dawn, day, dusk, and night.
• Object Degradation: Modeling wear and tear, such as scratches on objects or faded text.
• Adversarial Conditions: Adding occlusions (e.g., a hand in front of a camera) or visual distractors.

Example: Training a warehouse robot's vision system with randomized levels of simulated dust in the air, flickering fluorescent lights, and boxes that are sometimes torn or have obscured barcodes.

This defines how parameters are randomized, which is as critical as which parameters are chosen.

• Uniform vs. Structured Distributions: Choosing between sampling parameters from a wide uniform range or a more structured, curriculum-based distribution.
• Per-Episode vs. Per-Step: Deciding if parameters are randomized once at the start of a training episode (static) or change dynamically at every timestep (dynamic).
• Curriculum Randomization: Gradually widening the randomization distribution as training progresses, starting with easier, narrower domains.
• Asymmetric Randomization: Applying different levels of randomization to the training environment versus the evaluation environment within the simulator.

Core Principle: The strategy should create a distribution of simulations so broad that the real world appears as just another sample from it.

The overarching goal of training a model in a simulated environment and deploying it in the physical world. Domain Randomization is a primary strategy within this paradigm. The core challenge is the reality gap—the discrepancy between simulation and real-world physics, visuals, and dynamics. Success is measured by zero-shot transfer, where a model works in reality without any real-world fine-tuning.

• Key Techniques: Include system identification (calibrating sim parameters to real data), domain adaptation, and progressive networks.
• Applications: Robotics, autonomous vehicles, and any task where real-world trial-and-error is costly or dangerous.

The process of calibrating a simulation's parameters (e.g., friction coefficients, motor dynamics, material properties) to closely match real-world system behavior. It is often contrasted with Domain Randomization.

• Domain Randomization deliberately varies parameters widely to force invariance.
• System Identification seeks the precise parameter values for high fidelity.
• Hybrid Approaches: Use a narrow, identified parameter distribution as a starting point, then apply randomization around it for robust generalization.

The fundamental performance drop experienced when a model trained in simulation is deployed in the real world. This gap is caused by mismatches in:

• Visual Domain: Textures, lighting, and rendering artifacts.
• Dynamics: Inaccurate physics modeling (friction, collisions).
• Actuation & Sensing: Noise and latency not present in sim.

Domain Randomization directly attacks this gap by exposing the model to such a vast diversity of simulated conditions that the real world appears as just another variation.

A set of techniques where a model is adapted from a source domain (e.g., simulation) to a target domain (e.g., reality) using a small amount of labeled or unlabeled target data. This differs from Domain Randomization's zero-shot approach.

• Supervised Domain Adaptation: Uses a small set of labeled real-world data.
• Unsupervised Domain Adaptation: Uses unlabeled real data to align feature distributions.
• Relation to DR: Domain Randomization can be seen as a form of multi-source domain adaptation, where the source is a massively diverse set of simulated domains.

The general practice of artificially expanding training datasets by applying coordinated transformations across multiple data types (modalities). Domain Randomization is a specialized form of MMDA applied to the simulation parameters themselves.

• Synchronized Augmentation: Applying the same semantic transform (e.g., a crop) to all modalities in a sample.
• Modality Dropout: Randomly omitting an input modality to force robust cross-modal learning.
• Cross-Modal Mixup: Blending features or data from different multimodal samples.

A technique that generates challenging, model-specific synthetic data to improve robustness. While Domain Randomization uses random parameter variations, adversarial augmentation actively searches for perturbations within a simulation that fool the current model.

• Mechanism: Uses gradients or reinforcement learning to find simulation parameters that maximize model error.
• Goal: To expose and correct specific model weaknesses, creating a curriculum of increasing difficulty.
• Outcome: Often leads to more sample-efficient training than pure randomization, but is computationally more intensive.