Vary conditions with a reason
Domain randomization changes aspects of a simulated environment so a model does not train on only one setup. The original domain-randomization research by Tobin and colleagues studied visual randomization for transfer. That result does not establish that arbitrary variation will improve every robotics task.
Start with the conditions expected to differ in deployment. Write down the uncertainty you want to address and a way to evaluate whether the variation helps.
Separate variation families
| Family | Example | Check before using it |
|---|---|---|
| Appearance | Surface color or lighting | The target remains observable in the intended sensor view |
| Placement | Object pose or clutter | The task still has a valid start and reachable goal |
| Geometry | Object dimensions | Important clearances and mechanism relationships remain valid |
| Dynamics | Mass or contact parameters | The range has a defensible source and produces meaningful behavior |
These are experiment design options, not a promise that every generator or library package exposes all parameters.
Build controlled experiments
Start with a baseline scene. Change one family of conditions, keep a record of the range and seed, and compare against the baseline. Expand only after inspecting representative outputs and task results.
Check for invalid starts, intersecting objects, disconnected components and unreachable goals. Simply producing more scenes can multiply such problems. Keep the rules that reject invalid variations visible and versioned with the experiment.
For repeatable synthetic-data workflows, Replicator's examples provide a reference for configuring scene changes and outputs. Use the tooling appropriate to your runtime and version.
Evaluate outside the training conditions
Hold back layouts, objects or parameter combinations for evaluation. Decide in advance which task measures count as improvement. If possible, include real-world checks with the relevant hardware rather than using visual diversity as the success metric.
simgenerator provides assets and environments and an environment generation workflow. Inspect each resulting scene against your own task requirements. For a controlled family of scenes based on a particular facility, discuss the authoring scope, including which variations and acceptance checks are required.