Sources of the Reality Gap

Simulators attempt to replicate real-world dynamics with models representing various aspects of physics, such as rigid-body dynamics, batteries, chaotic nature, and stochasticity. Each of these components can contribute to the reality gap.

A fundamental source of the dynamics gap arises from the incorrect parameterization of the physical model. Assigning accurate values to parameters like friction, mass, and inertia can be challenging.

The choice of numerical integration methods (e.g., Euler, Runge-Kutta) and discretization can introduce discrepancies between the simulated and the real system, with a trade-off between fidelity and computation time.

Modeling complex, context-dependent, and often irrational human behaviors in simulation presents unique challenges that create substantial reality gaps.

Physical phenomena like wear and tear, material fatigue, and thermal effects are often overlooked or simplified in simulation, causing significant sim-to-real differences.

Simplified digital assets (meshes, SDFs) used for computational efficiency often lack the complexity and properties of real-world environments, leading to inaccurate geometric assumptions.

Simulation models are designed to mimic the physical characteristics of real-world sensors, but often omit real-world artifacts such as lens flares, chromatic aberration, and sensor-specific interference.

Real-world sensor noise is often complex, non-Gaussian, and state-dependent, while many simulators use simple Gaussian noise models, leading to a significant gap.

Using low-resolution assets, simplified scene graphs, or generic materials can fail to capture fine-grained perceptual cues such as surface textures, reflectance, and subtle geometry.

Simulated robot models often simplify or omit important physical details like manufacturing tolerances, material wear, and mechanical backlash, which can cause self-collisions or unstable motions.

Simulators use simplified geometric approximations for collision detection, which limits the accuracy of contact modeling, especially in tasks involving fine-grained manipulation.

Real actuators behave as higher-order systems with non-linearities like dead-zones and backlash, while simulators often treat them as simple first-order systems.

Real robots have low-level control layers with hidden filters and safety logic that are often not modeled in simulation, creating a significant actuation gap.

Motor drivers and power electronics introduce latency, command quantization, and current/voltage caps that are typically absent in simulation.

Real-world communication can have delays and packet loss, with fallback mechanisms that are almost never modeled in simulation.

Real-world safety mechanisms like virtual walls are often not present in simulation, which can change the robot's behavior and widen the reality gap.

Simulated POMDPs may use privileged information for rewards and termination, or infeasible resets, creating state-action distributions unlike those in the real world.

Discrepancies in the granularity of discretization, or different frequencies for control loops between simulation and the real world can significantly impact the reality gap.