Training
Domain Randomization
2017ActiveUpdated: 28 August 2026Published
Key
innovation
Randomizing simulation parameters (textures, lighting, physics) so a model trained in simulation generalizes to the real world.
Category
Training
Abstraction level
Pattern
Operation level
Architecture blockModel
Use cases
Sim-to-real transfer in roboticsTraining manipulation and locomotion policiesPerception robust to lighting/texture changesReinforcement learning in simulators (e.g. MuJoCo, Isaac)
How it works
In each training episode the simulator parameters are sampled from chosen distributions; the real world becomes 'just another sample' from that distribution for the model, improving robustness and generalization without needing a faithful model of reality.
Problem solved
Models trained only in simulation usually perform poorly in reality due to the reality gap in the appearance and dynamics of the environment.