Randomize one factor
Step
- Does
- turns on domain randomization one factor at a time and compares the backends under each.
- Used in
- Port from PhysX
- Passes when
- each factor's samples and deltas stay within the tolerance set at nominal parity, or the factor's difference has a named cause.
Inputs. Nominal parity passed (check-parity); the task’s randomization factors (masses, friction, restitution, gains, delays, pushes, initial states). Tools. The application’s randomization config and backend switch.
- With nominal settings matching, enable one factor; keep every other factor at its nominal value and the seeds fixed.
- Run both backends and compare per-sample values and deltas, not only mean return: the sampled values each solver actually reads, the trajectories, terminations and contact behavior.
- When a factor differs, ablate inside it (for materials: friction alone, restitution alone) until one property explains it. Check that the property reaches the solver after randomization and that a solver option for it is not silently dropped.
- Only after every factor passes alone, run the full randomization as its own labeled comparison.
Rules. A choice that fixes one task (for example restitution off) can break another; make task-specific choices explicit in that task’s launcher, not a shared default.
Outputs. One row per factor: samples compared, deltas, cause if any, fix or acceptance.
Source: agents/steps/randomize-one-factor/README.md, the copy agents read.
Steel docs and tutorials are licensed under Apache-2.0. Third-party notices. "Skild", "Skild AI" and the Skild logo are trademarks of Skild AI and are not licensed under Apache-2.0.