The Simulation Gap Never Fully Closes

Training a policy in simulation is cheap, parallelizable, and safe. Millions of episodes run overnight on hardware that costs less than a single robot arm. Then you deploy to the physical machine and the policy falls apart in ways that make no sense.

The culprit is almost always unmodeled physics: friction that varies with temperature, backlash in a gearbox that was not in the URDF, a servo that runs three percent fast, a camera with rolling shutter the simulator rendered as global. Each individual gap is small. Compounded across a trajectory, they are fatal.

Domain randomization helps by forcing policies to be robust across a distribution of physics rather than one idealized instance. It does not eliminate the gap. Every deployed system still needs a calibration pass against the real hardware, and anyone selling you a pipeline without one has not shipped.