Reality Gap

Physics engines approximate real-world dynamics using simplified equations and numerical solvers. Key approximations include:

• Contact dynamics: Simulated collisions and friction models (e.g., penalty-based, impulse-based) are computationally efficient but often fail to capture the complex, multi-point contact and stiction of real objects.
• Actuator dynamics: Motors, hydraulics, and gears have non-linear properties like backlash, saturation, and thermal effects that are rarely modeled with high fidelity.
• Material properties: Simulated rigidity, elasticity, and deformation are often homogeneous and fail to capture the heterogeneity of real materials.

Real-world sensors (cameras, LiDAR, IMUs, force-torque sensors) introduce noise, drift, and latency absent in simulation.

• Perceptual noise: Cameras have motion blur, rolling shutter effects, and varying exposure. LiDAR suffers from beam divergence and multi-path reflections.
• Proprioceptive noise: Joint encoders have quantization error. IMUs experience bias and drift. Force sensors have electrical noise.
• System latency: The end-to-end loop from sensor reading to actuator command involves processing and communication delays, creating a temporal mismatch with instantaneous simulation feedback.

The visual domain gap is the difference between rendered synthetic imagery and real-world camera feeds.

• Texture and lighting: Simulated textures are often procedurally generated and lack the microscopic detail and wear of real surfaces. Global illumination and shadows are approximated, not physically measured.
• Sensor simulation: Perfect pinhole camera models ignore lens distortion, chromatic aberration, and sensor noise patterns.
• Object diversity: Simulated object meshes and placements lack the vast morphological variation and clutter found in natural environments.

Every physical robot has dynamics that are impractical to model perfectly in simulation.

• Cable management: The drag and inertia of cables and hoses are rarely simulated.
• Thermal effects: Motor performance degrades with heat, and material dimensions change with temperature.
• Wear and tear: Components like gears and belts degrade over time, changing friction and backlash properties.
• Communication buses: Delays and packet drops on internal networks (e.g., CAN, Ethernet) create jitter in control signals.

Real-world environments are non-stationary and stochastic in ways that are difficult to simulate exhaustively.

• Unpredictable interactions: Human presence, moving objects, wind gusts, and changing lighting conditions create open-world disturbances.
• Surface variability: The coefficient of friction for a 'wooden floor' varies with polish, dust, and humidity.
• Object property variance: Two visually identical real-world objects can have different masses, centers of gravity, or compliance.

Simulations operate on discretized time and space, introducing numerical artifacts.

• Time-stepping: Fixed or variable integration steps approximate continuous dynamics. Large steps can cause tunneling (objects passing through each other) or energy drift.
• Collision mesh resolution: Simplified collision geometries (convex hulls, primitive shapes) fail to capture fine geometric details, leading to inaccurate contact points.
• Floating-point precision: Cumulative rounding errors in long simulations can cause divergent behavior from theoretical continuous models.

The overarching process of deploying a model or policy trained in a simulated environment onto a physical system. Its success is directly measured by how effectively it overcomes the Reality Gap. Core approaches include:

• Zero-Shot Transfer: Direct deployment with no real-world fine-tuning.
• Fine-Tuning Transfer: Limited adaptation using real-world data.
• The field's goal is to make this transfer robust, efficient, and safe, minimizing the Performance Drop.

A primary technique for bridging the visual and dynamic Reality Gap. Instead of perfect simulation fidelity, it trains a policy across a vast distribution of randomized simulation parameters to encourage robustness. Key randomized elements include:

• Visual Properties: Textures, lighting, colors.
• Physics Parameters: Mass, friction, actuator delays.
• Sensor Noise: Camera distortion, LiDAR dropout patterns.
• The policy learns invariant features, improving its chance of generalizing to unseen real-world conditions.

The process of building or refining a mathematical model of a physical system's dynamics by observing its input-output behavior. It directly attacks the dynamic component of the Reality Gap. The workflow is:

1. Execute control commands on the real robot and record its state.
2. Fit simulation parameters (e.g., inertia, motor gains) to this real-world data.
3. Use the calibrated simulation for more accurate training. This reduces the mismatch between simulated and real physics, leading to more transferable policies.

A machine learning subfield focused on transferring knowledge from a source domain (simulation) to a different but related target domain (reality). Techniques aim to learn domain-invariant representations. Key methods include:

• Domain-Adversarial Training: A discriminator network tries to identify the domain of features, forcing the main model to produce features that confuse it.
• CycleGAN: An unsupervised image-to-image translation model used to make simulated visuals look photorealistic, or to map real images back to a simulated style for paired training.

The degree to which a simulation replicates the visual, physical, and behavioral characteristics of the target real-world system. It exists on a spectrum:

• Low-Fidelity: Fast, abstracted physics and simple graphics. Useful for rapid prototyping and Domain Randomization.
• High-Fidelity: Computationally expensive, with accurate Physics Engine dynamics, photorealistic rendering, and detailed sensor models. Aims to minimize the Reality Gap by construction. The choice involves a trade-off between computational cost, development time, and the required accuracy for transfer.

The quantitative metric that defines the Reality Gap. It is the degradation in task performance (e.g., success rate, reward, precision) observed when a policy trained in simulation is executed on the physical hardware. A 30% drop means a policy with 100% success in simulation achieves only 70% in reality. This metric drives the entire Sim-to-Real research agenda, as the goal is to engineer techniques that bring this drop as close to 0% as possible.