A policy learned in simulation rarely survives contact with the real thing, because the two worlds do not behave identically. Friction differs, delays differ, the mass is not quite what the model assumed. This is the dynamics mismatch, and it is why so much reinforcement learning stays in the simulator.

The setting here is online dynamics adaptation, which is the realistic version of the problem: unlimited data in the source domain where experimenting is cheap, and only a small budget of interactions in the target domain where it is not. You get to practise as long as you like, then a handful of real attempts.

Earlier approaches lean on domain classifiers or value-guided data selection, essentially working out which source experience resembles the target closely enough to trust. DADiff uses a diffusion model to drive that adaptation instead, which is a different bet: rather than filtering the experience you already have, generate what the target domain would plausibly have produced.