Sim-to-real is the process of developing robot behavior in a simulated world and transferring it to a real robot. The appeal is obvious: simulated robots can fall, fail, and repeat tasks millions of times without breaking hardware.
The problem is that simulation is always wrong somewhere. Real floors have friction. Real motors have heat, backlash, latency, and manufacturing variation. Real cameras see glare, shadows, dust, and occlusion. The gap between the clean simulated world and the messy physical world is the sim-to-real gap.
Good sim-to-real work tries to make simulation messy on purpose. Researchers randomize lighting, textures, object positions, masses, friction, and sensor noise so the trained behavior is less brittle when it reaches hardware.
For humanoid companies, sim-to-real claims are useful but incomplete. The question is not whether something worked in simulation. The question is whether the same behavior has been shown repeatedly on physical robots, under conditions the company did not perfectly control.