Domain randomization

The engineering practice of deliberately varying simulation parameters (friction coefficients, object masses, sensor noise levels, lighting conditions) during training or testing so that the resulting learned policy or tuned control law is robust across a range of real-world conditions. First formalized in robot learning research (Tobin et al., OpenAI, 2017) and implemented as a first-class feature in NVIDIA Isaac Lab.

A simulation technique in which physical parameters (friction, mass, lighting, and similar factors) are randomized across simulated training episodes, so a learned policy is exposed to a wide range of conditions and becomes more robust to the mismatch between simulation and reality. Automatic Domain Randomization, used in OpenAI's Dactyl project, progressively increases this randomization as the policy improves.

A training strategy in which simulation parameters such as friction, mass, sensor noise, and environment geometry are deliberately varied across a wide range during training, so that the resulting policy is robust to the range of real-world parameter values it will encounter. The primary research-validated approach to reducing the sim-to-real gap for learned policies.

A sim-to-real bridging strategy that randomizes physics parameters, sensor noise properties, and visual conditions during training so that the resulting policy is robust to the real-world values of those parameters, which fall somewhere within the randomized range.

A simulation technique that generates many variations of a world (random obstacle positions, random friction values, random lighting) to train robot policies that transfer to real-world variation. Extends the fixed test world concept from this topic toward large-scale training environments.