Reality Gap
The reality gap is the performance degradation that occurs when a control policy or behavior learned in simulation is deployed on a physical robot. The gap arises because simulation cannot fully replicate the complexity, noise, and dynamics of the physical world: friction varies with temperature and wear, sensor readings contain unmodeled noise, actuators have backlash and hysteresis, and the environment includes unmodeled contacts, deformations, and disturbances. What works in simulation fails in reality — not because the learning algorithm is wrong, but because the simulator is wrong.
The reality gap is the central obstacle in robotics and embodied AI. It is the practical expression of Moravec's Paradox: the sensorimotor intelligence that evolution spent hundreds of millions of years building cannot be replicated in simulation because simulation strips away the very complexity that makes physical interaction intelligent. A robot that learns to walk in a physics engine has learned to walk in that engine's approximation of physics, not in actual physics.
Sources of the Gap
Systematic errors: Simulators use approximate physics models — rigid-body dynamics with simplified contact models, constant friction coefficients, idealized actuators. These approximations introduce systematic biases: the robot in simulation experiences different dynamics than the robot in reality.
Unmodeled dynamics: Real systems exhibit phenomena that simulators do not model: cable flex, gear backlash, motor heating, battery voltage sag, sensor drift. These dynamics are often small in magnitude but large in effect for precise control.
Stochastic variation: Real-world variability — floor texture, lighting conditions, payload mass, ambient temperature — creates a distribution of environments that a single simulation cannot capture. A policy that overfits to the simulated environment fails when the real environment deviates.
Perception mismatch: Simulated vision uses rendered images that lack the noise, motion blur, lens distortion, and lighting complexity of real cameras. A vision policy trained on synthetic data may fail on real images because the feature distributions differ.
Bridging Strategies
Domain randomization: Training policies across a distribution of simulated environments with randomized parameters — friction, mass, actuator gain, sensor noise — can produce policies robust to the reality gap. The intuition is that if the policy sees enough variation in simulation, the real world is just another sample from the distribution. This works surprisingly well for some tasks but has limits: the randomization distribution must include the real world's parameters, and for complex dynamics, the required distribution may be too broad to learn from.
System identification: Identifying the real robot's dynamics from data and updating the simulator to match. This reduces the gap but cannot eliminate it: some dynamics are unobservable, some are time-varying, and some are too complex to model.
Sim-to-real adaptation: Using real-world data to fine-tune a simulation-trained policy. This requires real-world interaction, which is expensive and risky for physical robots. The adaptation must be sample-efficient because real-world samples are scarce.
Hardware-in-the-loop: Combining simulation with real hardware for specific components — using real sensors with simulated actuators, or vice versa — to isolate sources of mismatch.
The Deeper Problem
The reality gap is not merely an engineering obstacle. It is an epistemological one. It reveals that intelligence is not a disembodied algorithm that can be trained in simulation and transferred to reality. It is a property of the coupled brain-body-environment system, and it depends on the specific physics of that coupling. A simulation is always an abstraction, and abstraction necessarily discards information. When the discarded information is irrelevant to the task, transfer succeeds. When it is relevant — as it is for virtually all sensorimotor tasks — transfer fails.
This suggests that the path to general embodied intelligence is not better simulators but better hardware: robots that can learn in the real world, safely and efficiently, without requiring simulation as a crutch. The reality gap will shrink as robots become cheaper, more robust, and more capable of real-world learning. But it will never close entirely, because the gap is not a bug in our simulators. It is a feature of reality: the world is always more complex than any model of it.
The reality gap is the reminder that intelligence is not software running on generic hardware. It is the specific way a specific body couples to a specific world. Until we build robots that can learn from that coupling directly, we are training ghosts and wondering why they cannot touch the ground.