Glossary
Domain randomization
Randomize simulation parameters during training so the policy works across many possible realities — including the real one.
Domain randomization is a sim-to-real technique: at training time, randomize the parameters of the simulation (friction, mass, motor strength, lighting, observation noise, sometimes geometry) across wide ranges. The policy that survives has learned to be robust to any setting in those ranges, including, hopefully, the real-world value.
Domain randomization is the workhorse behind nearly every modern legged locomotion result. Quadrupeds (ANYmal, Unitree Go), bipeds (Cassie, humanoids), and increasingly manipulation arms all use it.
The catch: you can only randomize what you can simulate. Phenomena your sim doesn't model at all (cable hysteresis, sensor calibration drift, certain contact dynamics) can't be randomized over, and the policy may fail in the real world for reasons sim couldn't expose.