Domain Randomization

Instead of training on a single, meticulously calibrated simulation, domain randomization defines probability distributions over environment parameters and samples randomly at each training step. Key parameters include:

• Noise floor: Uniformly sampled from -120 dBm to -80 dBm
• Delay spread: Log-uniform from 10 ns to 10 µs
• Doppler shift: Uniform from -500 Hz to +500 Hz
• Carrier frequency offset: Uniform from -10 ppm to +10 ppm
• IQ imbalance: Gain mismatch up to 2 dB, phase mismatch up to 5 degrees

The model never sees the same environment twice, preventing overfitting to any specific channel condition.

The core mechanism: by maximizing the variance of nuisance parameters during training, the neural network is forced to discard them as predictive features. The model learns that:

• Signal modulation type is independent of noise floor
• Emitter identity is independent of Doppler shift
• Symbol rate is independent of delay spread

This drives the feature extractor to latch onto the truly discriminative signal characteristics—such as cyclostationary signatures and transient shapes—that persist across all environmental conditions.

When randomization ranges are sufficiently broad, the real-world deployment environment falls within the convex hull of the training distribution. This enables zero-shot transfer: the model generalizes to live over-the-air signals without any fine-tuning on real data.

Critical requirements for zero-shot success:

• Randomization ranges must exceed expected real-world variation
• The simulator must capture the functional form of physical effects, even if parameters are randomized
• Avoid unrealistic parameter combinations that could create physically impossible signals and mislead the learner

Naive uniform randomization can waste training time on trivial or impossibly difficult environments. Advanced strategies include:

• Curriculum learning: Gradually expand randomization ranges as model performance improves, starting with mild conditions and progressing to extreme noise and distortion
• Adaptive domain randomization: Dynamically adjust sampling distributions based on the model's current performance gaps, focusing training on the most challenging parameter regions
• Adversarial domain randomization: Use a second network to actively search for environment parameters that maximize the primary model's loss, creating an automatic curriculum of hardest cases

Domain randomization and domain adaptation are complementary strategies for closing the sim-to-real gap:

• Domain randomization is a data-side solution: modify the training distribution to encompass the target domain
• Domain adaptation is a model-side solution: explicitly align feature distributions between source and target domains using techniques like gradient reversal layers (GRLs) or maximum mean discrepancy (MMD) loss

In practice, combining both—randomizing simulation parameters while also applying unsupervised domain adaptation on unlabeled real captures—often yields the most robust RF models.

Domain randomization is not a universal solution. Key failure modes include:

• Simulation bias: If the simulator's physics engine cannot model certain real-world effects (e.g., specific non-linear amplifier behaviors), no amount of randomization will bridge the gap
• Over-randomization: Excessively wide parameter ranges can create unlearnable training distributions where signal structure is destroyed, preventing convergence
• Computational cost: Each training batch requires on-the-fly simulation with unique parameters, significantly increasing per-epoch compute compared to static datasets
• Reality gap in dynamics: Randomizing static parameters does not address discrepancies in temporal dynamics between simulation and reality

The performance discrepancy observed when a model trained on synthetic RF data from a channel emulator is deployed in a live over-the-air environment. This gap arises from unmodeled physical imperfections—such as amplifier non-linearity, phase noise, and antenna coupling effects—that are absent or idealized in simulation. Domain randomization explicitly widens the simulation distribution to encompass these unknown real-world factors, forcing the model to treat them as irrelevant nuisance variables rather than spurious features.

The algorithmic modeling of physical propagation effects to augment clean RF signals with realistic environmental distortions. Key impairments include:

• Multipath fading: Rayleigh and Rician statistical models replicating signal reflections
• Doppler shift: Frequency offsets simulating relative transmitter-receiver motion
• Thermal noise: Additive white Gaussian noise (AWGN) at varying power levels
• Delay spread: Time-dispersive effects captured via power delay profiles Domain randomization randomizes these parameters across extreme ranges during training.

A transfer learning technique that mitigates the distribution shift between a labeled source domain (e.g., simulation) and an unlabeled target domain (e.g., real-world RF channel). Unlike domain randomization—which ignores the target distribution entirely—domain adaptation explicitly aligns feature representations using techniques like gradient reversal layers (GRLs) and maximum mean discrepancy (MMD) minimization. The two approaches are complementary: randomization provides broad invariance, while adaptation fine-tunes to a specific deployment environment.

A regularization technique that injects maliciously perturbed examples into the training set to harden a model against adversarial RF attacks. In the context of domain randomization, adversarial training serves a dual purpose: it simultaneously improves robustness to both intentional jamming waveforms and naturally occurring channel distortions. The randomized parameter space can include adversarial perturbation budgets, ensuring the model learns decision boundaries that are stable under worst-case environmental variations.

A high-fidelity, software-based virtual replica of a physical RF environment used to generate massive volumes of realistic synthetic data. Unlike simple channel models, a digital twin incorporates:

• Geospatial ray tracing for site-specific propagation
• Dynamic emitter modeling with realistic mobility patterns
• Hardware-in-the-loop integration with actual front-ends Domain randomization operates within the digital twin by systematically varying its environmental parameters—building materials, emitter positions, and atmospheric conditions—to produce a diverse training distribution.

The capacity of a trained neural network to maintain high classification or regression accuracy on previously unseen RF data distributions and channel conditions. Domain randomization directly targets generalization by preventing the model from overfitting to any single simulated environment. The key metric is out-of-distribution (OOD) robustness: a successfully randomized model exhibits minimal accuracy degradation when encountering novel SNR regimes, modulation formats, or interference patterns not present in the nominal training set.