Approximation everywhere, compounding together
Every simulator is, by construction, an approximation of the physical world it represents. Contact solvers use simplified friction models, sensor renderers approximate real optics and noise characteristics, and actuators are often modeled as producing exactly the torque commanded with no delay, when real motors, gearboxes, and control loops introduce latency and nonlinear response. None of these individual approximations is necessarily large, but a controller trained in simulation is implicitly relying on all of them being close enough to reality to not matter, and that assumption gets tested the moment the controller runs on real hardware.
The problem is that these small discrepancies don't stay small, they compound. A slightly wrong friction coefficient changes exactly when a foot starts to slip, which changes the robot's balance recovery, which changes the trajectory it takes, which puts it into a slightly different, unmodeled regime where the next small error also applies. Cable drag on a real arm that was never modeled in simulation adds an unaccounted force. Actuator delay means a control signal that simulation assumed took effect instantly actually arrives a few milliseconds late, which is enough to destabilize a policy tuned to react at simulation speed. Individually these are minor omissions, together they can produce meaningfully different dynamics, and that mismatch between simulated and real-world behavior is what's called the reality gap.