reality gap

In robot learning, the reality gap is the drop in performance you see when a skill that was trained in simulation is moved onto a real robot. The robot learned its behavior — its policy, the rule that turns what it senses into what it does — inside a simulator, where the physics are tidy and predictable. Then that exact same policy is loaded into a physical machine, and it suddenly does worse: it slips, hesitates, overshoots, or fails outright at the very task it had mastered in the simulator. That falloff between simulated competence and real-world competence is the reality gap, and it is the central reason a policy that looks finished still can't be trusted on hardware.

The gap exists because no simulation copies reality perfectly. The simulator may assume a fixed floor friction, instant motor response, a never-blurry camera, and no dust, flex, or worn gears — while the real robot lives among all those imperfections at once. A learned policy quietly tunes itself to whatever the simulator happened to assume, so even small mismatches between the two worlds throw it off, and the mistakes can compound as the robot drifts into states it never practiced. The wider the gap, the more a perfect simulation score overstates how the robot will really behave.

The reality gap is what every transfer technique is fighting. You shrink it from two directions: make the simulation more faithful by measuring the real robot and correcting the model, and make the policy more robust by randomizing the simulation so it learns to cope with a whole spread of conditions rather than one idealized world. A policy trained to survive a hundred different simulated frictions barely notices the single real friction it eventually meets.