Simulation Fidelity Trade-off

This is the core quantitative trade-off. High-fidelity simulations use complex physics engines (e.g., MuJoCo, NVIDIA Isaac Sim, PyBullet) with fine-grained timesteps, accurate contact models, and detailed sensor models. This accuracy comes at a steep computational cost, limiting the scale of parallel training. Lower-fidelity simulations simplify physics (e.g., using kinematic models or coarser collision meshes) and sensor rendering, enabling orders-of-magnitude more training episodes per dollar. The key is determining the minimum fidelity required for the task; a policy for object pushing may not need photorealistic rendering, while a policy for delicate assembly might.

A high-fidelity simulation that closely matches a single, static real-world setup risks training a policy that overfits to that specific digital twin. The policy learns the exact friction coefficients, lighting conditions, and object masses of the simulation, failing on any physical deviation. Domain randomization applied to a lower-fidelity sim explicitly prevents this by treating the simulation not as a replica, but as a vast, variable training arena. By randomizing parameters (mass, friction, textures, lighting), the policy is forced to learn the underlying task invariants, resulting in a robust policy that generalizes to the messy, variable real world.

High-fidelity simulations require precise system identification—the process of measuring and calibrating real-world parameters (e.g., motor torque constants, link inertia, camera distortion) to populate the simulation model. This is a labor-intensive, expert-driven process and any miscalibration widens the reality gap. Lower-fidelity simulations with broad randomization reduce this burden. Instead of painstakingly matching reality, engineers define wide, plausible ranges for parameters. The policy learns to be agnostic to the exact value, effectively making the system identification problem easier by requiring only bounds, not precise values.

Domain randomization is the primary technique for managing the fidelity trade-off. It strategically uses lower-fidelity simulation to maximize robustness.

• Physics Randomization: Varies dynamics parameters (mass, friction, damping) to teach policies force-invariant strategies.
• Visual Randomization: Alters textures, lighting, and camera properties to create perception systems robust to visual domain shift.
• Sensor Noise Randomization: Injects realistic noise into simulated IMU, LiDAR, or joint encoder readings. The randomization distribution (uniform, Gaussian) and its bounds are critical hyperparameters, balancing diversity with physical plausibility.

The optimal point on the fidelity spectrum is dictated by the task.

• High-Fidelity Preferred: Tasks involving delicate contact, complex deformable objects, or precise high-frequency dynamics (e.g., surgical robotics, drone acrobatics).
• Lower-Fidelity + Randomization Sufficient: Tasks defined by higher-level strategy or where dynamics can be abstracted (e.g., mobile robot navigation, bin picking, grasping).
• Hybrid Approaches: Often, a multi-fidelity strategy is best. Training might start in a cheap, low-fidelity sim for coarse skill acquisition, then transfer to a higher-fidelity sim for final refinement, a process related to curriculum randomization.

Evaluating the trade-off requires rigorous metrics. The ultimate test is real-world validation and the Sim2Real success rate—the percentage of successful task executions on physical hardware. This is the only true measure of a policy's out-of-distribution (OOD) robustness. Internally, engineers use simulation-only benchmarks that test policy performance across a held-out randomized simulation ensemble representing a worst-case domain. The goal is not to achieve perfect simulation accuracy, but to maximize this real-world success rate within computational budgets.

The reality gap, also known as the simulation-to-reality gap, is the performance discrepancy between a model trained in simulation and its performance when deployed on a physical system. This gap is caused by modeling inaccuracies in the simulator, such as imperfect physics, simplified sensor models, or missing environmental noise.

• Primary Cause: Inevitable simplifications in the simulation engine.
• Direct Relationship: The goal of techniques like domain randomization is to explicitly bridge this gap by training for robustness across a wide range of simulated conditions.

System identification is the process of building or calibrating a mathematical model of a dynamic system (e.g., a robot's dynamics) from observed real-world input-output data. It is a complementary approach to domain randomization for addressing the fidelity trade-off.

• High-Fidelity Path: Uses real data to calibrate simulator parameters (mass, friction, motor constants) to closely match the physical robot, reducing the reality gap.
• Contrast with Randomization: While system ID aims for accuracy, domain randomization embraces uncertainty by training across many possible parameter sets.

Domain shift refers to the degradation in model performance caused by differences in data distributions between the training environment (source domain, e.g., simulation) and the deployment environment (target domain, e.g., the real world). The simulation fidelity trade-off is fundamentally a strategy to mitigate domain shift.

• Core Problem: A high-fidelity sim may reduce shift but is computationally expensive and can overfit to its specific inaccuracies.
• Robustness Solution: Lower-fidelity sims with broad randomization explicitly train the model to be invariant to distributional shifts, improving out-of-distribution generalization.

Out-of-distribution (OOD) robustness is a model's ability to maintain performance when presented with inputs that differ significantly from its training data distribution. This is the explicit target when opting for lower-fidelity, highly randomized simulations.

• Trade-off Manifestation: Sacrificing precise fidelity for a wider, more varied training distribution.
• Key Technique: Domain randomization synthetically creates a vast set of 'distributions' during training, forcing the policy to learn a core solution that works across many of them, thereby enhancing OOD robustness.

Zero-shot transfer is the deployment of a simulation-trained policy directly onto a physical robot without any fine-tuning on real-world data. It is the ultimate goal that motivates the simulation fidelity trade-off analysis.

• High-Fidelity Risk: A policy trained in a single, precise simulation may fail upon encountering unmodeled real-world dynamics.
• Randomization Benefit: Policies trained with domain randomization in lower-fidelity sims are explicitly designed for this scenario, as they have never seen the 'true' parameters but have learned to handle variation.

Sim2Real success rate is the key empirical metric for evaluating the simulation fidelity trade-off. It measures the proportion of successful task executions when a simulation-trained policy is deployed on a physical robot.

• Primary Benchmark: The definitive measure of whether the chosen balance between fidelity and randomization was effective.
• Informs the Trade-off: A low success rate may indicate insufficient randomization (overfitting to sim) or excessively unrealistic randomization (training on non-physical scenarios). This metric directly guides the adjustment of fidelity parameters and randomization bounds.