Domain Randomization: Training Robust Policies Through Simulation Diversity

Domain randomization (DR) is a sim-to-real transfer technique that trains robot policies on simulations where visual and physical parameters are randomly varied at each episode or even each timestep. The core idea is simple but powerful: if the policy can succeed in simulations with randomly sampled textures, lighting, object sizes, friction values, and camera configurations, then the real world is just one more point in this already-broad training distribution, and the policy should handle it without explicit adaptation.

DR operates on two categories of parameters. Visual domain randomization varies the appearance of the simulated scene: object textures (sampled from random noise, internet images, or procedural generators), lighting conditions (direction, intensity, color temperature, number of light sources), camera parameters (position jitter, field of view, image noise), and background elements (random objects, distractors, floor textures). This is computationally cheap because rendering with randomized textures costs little more than rendering with fixed textures.

Physics domain randomization varies the physical properties of the simulation: object masses, friction coefficients, joint damping values, actuator delays, motor strengths, and gravity vectors. This is more delicate because unrealistic physics parameters produce training experiences that do not exist in the real world. If the simulation trains the robot on objects with negative friction (objects that slide uphill), the resulting policy may learn behaviors that are physically impossible. Physics randomization must be centered on measured or estimated real-world values with conservative variance.

The effectiveness of domain randomization depends on the gap being primarily parametric — the real world differs from the simulation in ways that can be expressed as variations of known parameters. For visual appearance, this assumption holds well: lighting, textures, and camera positions are parametric. For physics, the assumption is weaker: the real world contains phenomena (deformable contacts, granular materials, aerodynamic effects) that are not just different parameter values but fundamentally different physics than what the simulator models.

Implementing domain randomization requires choosing which parameters to randomize, what ranges to use, and how to structure the training pipeline.

For visual DR, the minimum viable implementation randomizes: lighting (2-8 point lights with random positions, intensities, and colors), textures (random RGB values or sampled from a texture dataset for tabletop, objects, and walls), camera position (uniform noise of +/-3cm translation, +/-5 degrees rotation around the nominal position), and image augmentation (color jitter, Gaussian noise, random crops). Libraries like NVIDIA Isaac Sim and MuJoCo with dm_control provide built-in visual randomization. More sophisticated approaches use neural style transfer or diffusion-based texture generation for more realistic appearance variation.

For physics DR, randomize around measured values: object mass (+/-20-30%), surface friction (0.3-1.0 for most household objects), joint damping (+/-20%), and gripper friction (+/-25%). Avoid extreme ranges that produce physically implausible scenarios. Measuring real-world values (weighing objects, measuring friction with simple tests) and centering the randomization around those values significantly improves transfer.

Training with DR typically requires 3-5x more simulation episodes than training without DR because the policy must learn to handle the full range of conditions rather than optimizing for a single configuration. On GPU-accelerated simulators like Isaac Gym, this is manageable — training 10,000 environments in parallel for 1,000 episodes per environment takes hours rather than days.

The practical architecture for production sim-to-real transfer in 2025-2026 is: (1) Train in simulation with aggressive visual DR and conservative physics DR for 100,000-1,000,000 episodes. (2) Collect 500-2,000 real-world demonstrations on the target hardware. (3) Fine-tune the DR-pretrained policy on real data with a lower learning rate. This pipeline consistently outperforms either pure simulation training or pure real-world training.

Claru provides the real-world data layer that completes the domain randomization pipeline. Teams that pretrain policies with DR in simulation need high-quality real-world demonstrations for the fine-tuning stage that closes the remaining gap. Claru delivers these demonstrations with the quality guarantees that determine whether fine-tuning succeeds or fails.

Our data captures the real-world conditions that simulators cannot randomize toward: genuine object textures and materials, natural lighting variation, real sensor noise characteristics, and true contact physics. For teams using NVIDIA Isaac Sim or MuJoCo for DR pretraining, Claru's real-world datasets provide the complementary data source needed for production deployment.