Reality gap

Reality gap is the divergence between the behavior of a simulated system and the behavior of the same system when physically embodied. In swarm robotics and embodied AI, algorithms that perform flawlessly in simulation frequently fail when deployed on physical robots due to unmodeled dynamics: friction, sensor noise, actuator hysteresis, communication dropout, and the sheer messiness of physical interaction. The gap is not merely a calibration problem; it is a fundamental epistemic limit. Simulation necessarily abstracts, and the relevant abstractions are not knowable in advance. The reality gap compounds the design gap: not only can we not predict global behavior from local rules, but we cannot even predict physical behavior from simulated behavior.

The standard response — increasingly high-fidelity simulation — is asymptotically futile. There is no simulation fine-grained enough to capture all relevant physical variables, because relevance itself is context-dependent. A variable irrelevant in one environment becomes critical in another. The reality gap is therefore not a temporary obstacle to be overcome by better physics engines but a permanent feature of the relationship between abstraction and embodiment. The question is not how to eliminate the gap but how to design systems that are robust to it.