The hanger-codesign programme asks how good a cloth simulator must be before the hanger it designs for a laundry robot is the right one. Earlier steps scored the same arm widths on several "tiers": a fast 2D chain, a 3D spring sheet, the sheet stopped early, the sheet on another mesh, and the sheet with guessed or measured fabric stiffness. This step asks when a cheap tier's choice can be certified without running the expensive one everywhere.
The bound. Write e(d) for the expensive tier's score minus the cheap tier's at design d. The cheap choice's regret on the expensive tier is at most e at the cheap choice minus e at the true best, so at most the spread of e across designs. A cheap model that is wrong by the same amount everywhere costs nothing. If the cheap tier's own winning margin beats that spread, its choice is certified. A sharper "cone" version uses how fast e changes per centimetre of arm width.
The test. The 50 tier pairs already on disk from milestones 1b to 1d needed no new simulation, and the predictions were committed first:
- The 2D chain is off by up to 1.6 to 2.0 cm of shape error in every instance. The bound charges only its spread across widths, which is a third of that. That is still 0.6 to 1.3 cm, against the chain's own margins of 0.05 to 0.25 cm. So no 2D choice is certified in any of the 28 pairs, by either version of the bound.
- Yet the 2D choice is rarely bad. Its regret on the 3D sheet averages 0.03 cm and peaks at 0.20 cm (7.6%) on very stretchy cloth. The bound is loose because it ignores how sharply the 2D curve penalises the 3D sheet's favourite width.
- Certification works only between near-identical models. A stretch stiffness guess of 3200 N/m against the value sourced from KES data is certified on 2 or 3 of 4 fabrics. Nothing else is certified.
- Three expensive runs around the cheap choice estimate the slope of e to within a factor of two in 45 of 50 pairs. So a design loop can tell cheaply which regime it is in.
What it means. For a multi-fidelity design loop (the programme's next milestone), the cheap tier is a screen, not a judge. Under the cone, it rules out a median of 3 of the 7 widths, and the expensive tier has to score the rest. The reference tier here is still our own converged sheet rather than a validated solver, so this measures how the bound behaves, not how accurate the cloth is. Full tables: the project README, "Milestone 5".
Simulation only; produced with AI assistance.