Domain Randomization

The core objective is to close the reality gap—the discrepancy between simulated training environments and the noisy, unpredictable real world. Instead of building a perfect simulator, domain randomization deliberately varies simulator parameters (lighting, textures, physics, noise) to such an extreme degree that the real world appears to the model as just another variation. This forces the network to learn features that are invariant to irrelevant visual or physical distractors, focusing on the fundamental task structure that persists across all domains.

The technique requires defining a parameterization of the simulator and sampling from it randomly during training. Key categories include:

• Visual parameters: Lighting position, color temperature, camera angle, object textures, background scenes, and post-processing effects like motion blur or defocus.
• Physical parameters: Object mass, friction coefficients, joint damping, actuator latency, and sensor noise models.
• Environmental parameters: Gravity vector, wind forces, and obstacle placement. The distributions are typically uniform or log-uniform, with bounds set wide enough to encompass real-world variance.

A specialized subset focused on physical system parameters rather than visual appearance. In robotics, this involves randomizing mass, center of mass, motor gains, joint friction, and contact dynamics during each training episode. The resulting policy learns to adapt to a wide range of underlying dynamics, making it robust to unmodeled physical effects and hardware variations. This was famously demonstrated by OpenAI's Dactyl system, which trained a robotic hand entirely in simulation with randomized dynamics and transferred successfully to a physical hand without any real-world fine-tuning.

An advanced variant where the randomization curriculum is learned rather than manually specified. The system progressively increases the difficulty of the domain distribution based on the agent's current performance, ensuring the model is never overwhelmed but is constantly challenged. This addresses the primary weakness of naive randomization: if the distribution is too wide, the task becomes impossible to learn; if too narrow, the policy fails to generalize. Bayesian optimization or reinforcement learning is often used to search the space of randomization parameters for the boundary of learnability.

In wireless communications, domain randomization translates to training RF fingerprinting models on signals convolved with a massive variety of synthetic channel impulse responses. By randomizing multipath profiles, Doppler spreads, delay spreads, and noise floors during training, the model learns to extract hardware-specific impairments that are invariant to propagation effects. This is critical because a fingerprinting model trained in one static channel environment will catastrophically fail when deployed in a different physical location with new reflective surfaces and interference patterns.

A key distinction exists between randomizing what the agent sees and what the agent feels.

• Observational randomization: Modifies sensor inputs—camera noise, lighting, textures. The underlying physics remain constant.
• Dynamics randomization: Modifies the physical laws of the simulation—friction, mass, latency. The agent must learn a control policy robust to a family of dynamical systems. Combining both is often necessary for contact-rich manipulation tasks where visual appearance and tactile response both vary significantly between simulation and reality.

A neural network training methodology that forces the feature extractor to produce representations that are indistinguishable across domains. An auxiliary domain classifier attempts to predict the source domain, while a gradient reversal layer inverts its loss signal, pushing the main network to discard channel-specific information. This creates a minimax game where the feature extractor learns to fool the domain classifier, resulting in channel-invariant embeddings that generalize to unseen environments.

A self-supervised paradigm that learns robust representations without labeled data by maximizing agreement between differently augmented views of the same signal. Key components include:

• Positive pairs: Two augmented versions of the same emitter's signal
• Negative pairs: Signals from different devices
• InfoNCE loss: Pulls positives together, pushes negatives apart The resulting embedding space naturally clusters by device identity rather than channel conditions, making it ideal for few-shot enrollment scenarios.

A complementary technique to domain randomization that applies label-preserving transformations to real captured signals. Augmentations include:

• Additive white Gaussian noise injection
• Multipath fading simulation via tapped delay lines
• Frequency offset and Doppler shift application
• IQ imbalance and phase noise modeling By training on signals degraded with diverse synthetic impairments, the model learns to treat channel effects as nuisance variables and focuses on the underlying hardware fingerprint.

A kernel-based statistical measure that quantifies the distance between two probability distributions in a reproducing kernel Hilbert space. In channel-robust learning, MMD is used as a regularization term added to the primary task loss to explicitly align feature distributions across different channel conditions. Unlike adversarial methods, MMD provides a stable, non-adversarial optimization objective. Common kernel choices include the Gaussian RBF kernel and the multiple kernel MMD variant for capturing distribution differences at multiple scales.

The process of decomposing a learned representation into independent, semantically meaningful factors of variation. In RF fingerprinting, the goal is to separate:

• Device-specific factors: Hardware impairments, oscillator characteristics, amplifier non-linearities
• Channel-specific factors: Multipath profile, path loss, Doppler spread Techniques include variational autoencoders with structured latent spaces, adversarial disentanglement with separate encoders, and mutual information minimization between factor groups. Disentangled representations enable robust authentication even when channel conditions change dramatically.

Correlation Alignment loss minimizes the difference between second-order statistics (covariance matrices) of source and target domain features. Unlike MMD which aligns distributions, CORAL specifically matches the feature correlations across domains. The loss is computed as the squared Frobenius norm of the difference between covariance matrices. CORAL is parameter-free, requiring no additional network branches or hyperparameter tuning, making it a lightweight alternative to adversarial domain adaptation for channel-robust fingerprinting.