Domain Randomization

Randomizing appearance and physics in simulation during training so the real world looks like just another variation.

Domain randomization was proposed by Josh Tobin and colleagues at OpenAI in 2017 (IROS 2017). Instead of trying to make simulation look exactly like reality, it heavily randomizes rendering parameters — textures, lighting, camera settings — and dynamics parameters such as mass, friction, and damping (the latter often called dynamics randomization) during training, so the model sees enough variation that the real world just looks like one more instance of it. The original paper trained an object-localization network purely on randomized simulated RGB images, and after transferring to a real robot, achieved about 1.5 cm of localization accuracy. It's one of the most common techniques for sim-to-real transfer, and frameworks such as Isaac Lab include it out of the box. The difficulty is that the randomization range is usually tuned by hand — too wide, and the policy becomes overly conservative; automatic domain randomization lets the range expand on its own as training progresses.

ExampleWhen training a quadruped locomotion policy, randomize ground friction, body mass, and motor delay independently for every parallel environment, so the resulting policy is less likely to fall over on a real robot.

Also called

DR

Related

Sim-to-Real Gap (Reality Gap) · Sim-to-Real Transfer · Dynamics Randomization · Visual Randomization · Automatic Domain Randomization · System Identification

Sources

Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World (arXiv 1703.06907)
Lilian Weng: Domain Randomization for Sim2Real Transfer