Transfer error is the measurable performance loss—such as increased task failure rates, higher control effort, or deviation from a desired trajectory—observed when a policy or controller trained in simulation is deployed on a real physical system. This error quantifies the reality gap and serves as the primary benchmark for assessing simulation fidelity and the effectiveness of transfer methods like domain randomization. It is distinct from, but often caused by, calibration error and unmodeled dynamics in the simulator.
SIMULATION FIDELITY AND SYSTEM ID
Primary Sources of Transfer Error
Transfer error arises from systematic discrepancies between the simulated training environment and the target physical system. These sources can be categorized into modeling inaccuracies, sensory mismatches, and unaccounted-for real-world phenomena.
01
Unmodeled Dynamics
Unmodeled dynamics are physical phenomena absent from the simulation's mathematical representation. These include complex effects like:
- Non-linear friction (Stribeck effect, stick-slip)
- Flexibility and vibration in links, gears, and cables
- Actuator saturation and non-ideal torque-speed curves
- Aerodynamic forces or fluid interaction Since the policy never encounters these forces in simulation, it cannot learn to compensate for them, leading to instability or poor tracking performance in reality.
02
Simulation Bias
Simulation bias is a systematic error introduced by the simulator's numerical approximations and implementation choices. Key sources are:
- Numerical integrator error from discrete time-stepping
- Simplified contact models that approximate complex collisions
- Approximate solvers for constraints and collisions
- Steady-state assumptions (e.g., ignoring motor electrical dynamics) This bias causes the simulated state evolution to consistently diverge from real-world physics, even if parameters are nominally correct.
03
Sensor and Actuator Discrepancy
This source encompasses mismatches between simulated and real perception and actuation. Critical gaps include:
- Sensor noise characteristics (often Gaussian in sim, correlated and non-Gaussian in reality)
- Latency and dropout in real sensor data streams
- Actuator delay and command filtering
- Quantization of encoder readings and command outputs
- Render-based vs. physical camera sensors (differences in lens distortion, auto-exposure, motion blur) A policy reliant on perfect, instantaneous sensing will fail when confronted with noisy, delayed real data.
04
Parameter Calibration Error
Even with a perfect model structure, inaccurate physics parameters cause transfer error. This includes:
- Inertial parameters (mass, center of mass, inertia tensor)
- Friction coefficients (viscous, Coulomb)
- Motor constants (torque constant, back-EMF)
- Geometric parameters (link lengths, joint alignments) Calibration error is the residual mismatch after identification. Error propagates through the forward dynamics, causing the simulated robot to accelerate and move differently than its real counterpart.
05
Domain Gap in Visual Perception
For vision-based policies, a major source of error is the visual domain gap between synthetic rendering and real imagery. This includes:
- Texture and material appearance (procedural vs. real)
- Lighting conditions and shadows
- Object model fidelity and geometric detail
- Lack of real-world visual clutter and distractors A convolutional neural network trained on perfect synthetic images can suffer catastrophic performance drops when processing noisy, textured, and variably lit real camera feeds.
06
Stochastic Real-World Disturbances
The real world is inherently stochastic, while simulations are often deterministic. Unpredictable disturbances not modeled in sim cause error:
- External forces (unexpected collisions, wind gusts)
- Payload variations and changes in system configuration
- Ground compliance and terrain deformation
- Thermal effects on sensor readings and actuator performance
- Electrical noise and bus voltage fluctuations Policies trained without exposure to such variability lack robustness, as they operate outside their trained distribution.
