Reality Gap
The reality gap is the gap between how a robot behaves in a computer simulation and how it behaves in the real world. Engineers often practice and train a robot inside a simulator first — a kind of detailed video game of the robot and its surroundings — because it is fast, cheap, and safe: you can run a thousand attempts overnight without breaking a single motor. But when you take the skill that worked perfectly in the simulation and drop it into a physical robot on a real floor, it often stumbles, slips, or fails outright. That drop-off in performance is the reality gap.
The gap opens up because no simulation captures every messy detail of the real world. The simulator might assume the floor has a certain friction, the motors respond instantly, the camera never blurs, and there is no dust, glare, wobble, or worn-out gear — but reality is full of exactly those imperfections. A robot tuned to the clean, predictable physics of the simulation can be caught off guard by them. Roboticists try to shrink the gap in two main ways: making simulations more realistic, and deliberately randomizing the simulation (varying friction, lighting, weights, and delays across many runs) so the robot learns to cope with a wide range of conditions and is less thrown by whatever the real world actually serves up.
A four-legged robot learns to run flawlessly in simulation, but on a real lawn the grass is springier than expected and one foot keeps slipping, so it tips over within a few steps until engineers retrain it across many randomized surfaces.
Perfect in the simulator, wobbly on the lawn — that drop in performance is the reality gap in action.
The phrase "sim-to-real" describes the act of transferring a behavior from simulation to a real robot; the reality gap is the obstacle that act has to overcome.