Curriculum Randomization

The core mechanism involves starting with a narrow, easy distribution of randomized parameters and gradually expanding the range or complexity over training epochs. This prevents the model from being overwhelmed initially, a common failure mode in standard Domain Randomization known as over-randomization.

• Example: Training a robot arm to grasp objects might start with randomized lighting only, then progressively add variations in object texture, mass, and finally, background clutter and camera noise.

Instead of sampling parameters from a fixed statistical distribution, Curriculum Randomization dynamically adjusts the sampling distribution. Parameters like friction coefficients or lighting intensity are sampled from distributions whose mean shifts or variance increases according to a predefined schedule or an automated performance metric.

• Key Benefit: This ensures the model is continuously challenged at an appropriate level, promoting steady learning progress and preventing plateaus.

Advanced implementations tie the curriculum progression directly to the model's validation performance. The randomization range expands only when the model achieves a target success rate or accuracy on a held-out validation set within the current difficulty level.

• Mechanism: This creates a closed-loop system where the training difficulty auto-tunes to the learner's capability, similar to adaptive testing. It is a step towards Automatic Domain Randomization (ADR).

By exposing the model to a controlled, expanding set of variations, it learns to discard spurious correlations and extract task-relevant, invariant features. The curriculum guides the model to first solve the core task, then gradually ignore irrelevant randomized distractors.

• Result: The final model exhibits stronger cross-domain generalization, as its internal representations are robust to the wide range of perturbations seen at the end of training.

Curriculum Randomization directly addresses the sim-to-real transfer problem. By ending training with a broad parameter distribution that encompasses potential real-world conditions, the model is less likely to be surprised by domain shift upon deployment.

• Contrast with Static DR: Standard Domain Randomization uses a fixed, broad distribution from the start, which can be inefficient. Curriculum Randomization is a more sample-efficient path to the same goal of robustness.

Curriculum Randomization is often combined with Systematic Domain Randomization. Parameters are varied in a structured, factorized way (e.g., varying lighting independent of textures), and the curriculum controls the number of factors randomized simultaneously or their individual ranges.

• Engineering Practice: This allows for precise control over the complexity space, ensuring comprehensive coverage of the parameter space by the end of training without chaotic early-stage learning.

Training a robot arm to grasp diverse objects by progressively randomizing simulation parameters.

• Start: Narrow randomization of object color and table texture.
• Progress: Expand to randomize object shape, size, and mass.
• Final Stage: Introduce lighting direction, camera noise, and distractor objects.

This curriculum forces the policy to learn fundamental geometry and physics before adapting to perceptual noise, resulting in a grasp success rate transferable to a physical robot with minimal fine-tuning.

Training vision models for self-driving cars using synthetic data with a curriculum of increasing visual complexity.

• Phase 1: Randomize time of day and basic weather (clear, overcast).
• Phase 2: Introduce rain intensity, fog density, and windshield artifacts.
• Phase 3: Add sensor degradation (lens flare, motion blur) and unusual lighting (tunnels, glare).

This approach builds a perception system robust to the long tail of rare but critical visual conditions encountered in real-world driving.

Teaching a quadcopter to fly through obstacle courses by gradually increasing physical and visual randomness.

• Easy: Randomize obstacle placement in a simple corridor with consistent lighting.
• Medium: Vary obstacle shapes/sizes and introduce mild wind gusts.
• Hard: Apply full Dynamics Randomization (mass, thrust, drag) alongside Visual Domain Randomization (textures, shadows) on obstacles.

The curriculum ensures the drone first masters basic flight dynamics before learning to compensate for unpredictable physical and perceptual disturbances.

Preparing a robot for precise part insertion by progressively varying tolerances and environmental factors.

• Foundation: Train with perfect part alignment and rigid fixtures.
• Variation: Randomize part position/orientation within a small tolerance and fixture stiffness.
• Robustness: Introduce cable slack, sensor miscalibration, and part deformation models.

This method yields a controller that can perform delicate assembly tasks despite mechanical wear and minor part imperfections on a real factory floor.

Training a bipedal robot to walk by carefully escalating the difficulty of its simulated terrain and dynamics.

• Initial: Flat ground with nominal actuator strength and joint damping.
• Intermediate: Add gentle slopes, small obstacles, and randomize ground friction.
• Advanced: Implement uneven terrain, external pushes, and broad Dynamics Randomization (link masses, motor delays).

The curriculum prevents early training failure and leads to a stable, adaptive walking policy capable of handling real-world disturbances.

Improving the robustness of AI models for diagnosing conditions from X-rays or MRIs using synthetic data curricula.

• Base Training: Use high-fidelity synthetic scans with standard anatomical variations.
• Progressive Challenge: Systematically introduce and increase imaging artifacts (motion blur, noise), equipment variations (contrast, resolution), and patient positioning differences.

This process helps the model learn to ignore irrelevant scanner-specific features and focus on pathological biomarkers, improving generalization across different hospital imaging systems.

Domain Randomization is the foundational simulation-based training technique where a wide range of environmental parameters (e.g., lighting, textures, physics) are varied during training. This forces the model to learn features and policies that are invariant to these changes, improving robustness and enabling sim-to-real transfer. It is the parent strategy from which Curriculum Randomization evolves.

Automatic Domain Randomization is an advanced, algorithmic extension of DR. Instead of manually defining parameter ranges, ADR uses a search process (often driven by the model's performance) to automatically discover and apply the most effective randomization parameters. It optimizes the randomization schedule to maximize policy robustness, reducing the need for extensive manual tuning.

A Randomization Schedule is the strategic plan governing how simulation parameters are varied over the course of training. Key types include:

• Static Schedule: Parameters are randomized from a fixed distribution throughout training.
• Dynamic Schedule: The distribution changes based on training progress or other metrics.
• Curriculum Schedule: The core of Curriculum Randomization, where the range or difficulty of randomization is progressively increased from easy to hard.

In Domain Randomization, the Parameter Distribution defines the statistical range from which simulation parameters are sampled. Common distributions include:

• Uniform Distribution: Parameters are sampled with equal probability across a defined min/max range.
• Gaussian Distribution: Parameters are sampled from a normal distribution defined by a mean and standard deviation. The choice of distribution and its bounds is critical for effective randomization and avoiding over-randomization.

Sim-to-Real Transfer is the overarching goal of deploying a model trained in simulation to perform effectively in the real world. The discrepancy between simulation and reality is called the reality gap or domain gap. Techniques like Domain Randomization and Curriculum Randomization are explicitly designed to bridge this gap by training models on a diverse set of simulated conditions, promoting cross-domain generalization for zero-shot sim-to-real deployment.

Over-Randomization is a critical failure mode in Domain Randomization where the applied parameter variations are too extreme. This can make the core task impossible to learn or cause the model to fail to converge on any coherent policy, ultimately degrading sim2real performance. Curriculum Randomization is specifically designed to mitigate this risk by starting with manageable, easy variations before introducing more challenging ones.