Start with a real failure
A warehouse detector repeatedly misses dark packaging against a dark conveyor. Generating a million additional bright, unobstructed boxes would enlarge the dataset without addressing that failure. The useful question is which variations the model needs to encounter and whether training on them changes its behavior on held-out real observations.
Synthetic perception data consists of rendered or generated sensor inputs with labels derived from the scene. It can make difficult labels and unusual conditions easier to obtain. It can also reproduce a designer's mistaken assumptions at enormous scale. We would begin with a defined task, a real evaluation set and an explicit list of failure conditions before buying more rendering capacity.
What the research actually demonstrates
Tobin and colleagues' 2017 paper demonstrates simulation-to-real transfer for an object-localization task using domain randomization. It is evidence that deliberately varying simulated appearances can help in a particular setting. It is not evidence that synthetic training eliminates real-world evaluation across every perception task.
NVIDIA's Replicator documentation describes randomizing scene properties such as lighting, object poses and textures. That establishes an available workflow. A documentation example does not establish customer savings, safety acceptance or independent comparative accuracy.
Coverage is the hard part
A visually convincing frame can have the wrong distribution of objects, motion or sensor artifacts. Conversely, a deliberately unrealistic texture may be useful when it prevents dependence on an irrelevant visual cue. Fidelity should be evaluated against the task rather than treated as a universal ranking of simulators.
The test set must remain independent of the choices made while constructing training scenes. Otherwise, repeated tuning can make a pipeline appear to generalize when it has effectively learned the evaluation procedure. Results should be broken out by meaningful conditions, including rare failures, rather than compressed into one average score.
Count useful examples and engineering time
Consider an illustrative pipeline that costs $12,000 to prepare and render a dataset. If validation accepts 60,000 examples, its cost is $0.20 per accepted example. If only 15,000 are useful after inspection, the cost becomes $0.80. These are invented inputs, not a supplier quotation or measured customer result.
The comparison should include asset creation, scenario design, labeling checks, compute and real-world validation. A low rendering cost can coexist with an expensive overall pipeline. The more relevant economic outcome is an improvement in the deployed task at an acceptable total cost, including the errors that remain.
Keep perception data separate from control transfer
This guide addresses training inputs for perception. Sim-to-real transfer addresses the broader question of whether a learned behavior survives contact, timing and physical-model errors on hardware. Better images do not by themselves validate a robot controller.
The former synthetic-data economics guide is consolidated here because its purchasing question belongs with the same validation workflow. We retain the distinction between technical feasibility and commercial attribution: use of a simulation tool does not establish a separately disclosed revenue stream, customer contract or durable supplier margin.
The evidence that would change our view
We would favor reproducible comparisons showing where synthetic augmentation improves a fixed real-world test, how much engineering effort it requires and which failure modes remain. A useful case explains the alternative, including collecting additional real observations or changing the model and sensor configuration.
The case weakens if improvements vanish outside the designed scenes, if real validation dominates the cost, or if simpler augmentation delivers the same result. Our judgment is conditional: synthetic data can be valuable coverage infrastructure, but the investment case rests on demonstrated downstream utility.