Sim-to-real means letting a robot practise in software, where it can try a task thousands of times safely and cheaply, then running the learned skill on real hardware. The hard part is the “reality gap”: simulated physics, lighting and sensors never match the real world exactly.
A common fix is domain randomisation, which deliberately varies colours, textures, lighting and physics in simulation so the real world looks like just one more variation. You meet the term whenever a robotics lab reports training mainly in simulation.