domain adaptation (RL)

Domain adaptation handles the case where you trained in one setting — a simulator, or one particular robot — and must now perform in a related but different one, such as the real world or a new robot. Rather than retrain from scratch, you adapt the skill you already learned to the new domain.

Common approaches align the two domains so the policy cannot tell them apart: learn features that are invariant to the gap, translate observations so rendered images look real (or the reverse), or fine-tune the policy on a small amount of target-domain data. The aim is to reuse the skill while swapping out only the parts that depend on domain-specific surface details.

Adaptation assumes the task is fundamentally the same and only the domain differs. If the new domain changes the underlying dynamics or the reward itself, adaptation alone will not save you — you are really facing a new problem that needs more than a surface alignment.