Sim-to-Real Deployment and Continual Correction › Summary
Michael BrenndoerferAugust 10, 20263 min read
World Models HandbookAll chapters Part 12 of 12Reliable World Models
Sim-to-Real Deployment and Continual Correction › SummaryOne section · Sections in this chapter12 / 13 ↓
Sections in this chapter13
Reading tools Definitions & inline links
Choose your expertise level to adjust how many terms are explained. Beginners see more tooltips, experts see fewer to maintain reading flow. Hover over underlined terms for instant definitions.
Article links
Make inline references clickable
- Separate dynamics, observations and interfaces. Parameter mismatch may be recoverable with identifiable data. Structural mismatch is relative to a specified class, not proof of a positive error floor. Sensor delay can leave the plant dynamics valid while changing the information available to a controller.
- Randomization supplies variation, not an automatic guarantee. Training a policy across plants, conditioning a predictor on parameters and pooling transitions without context are distinct operations. Family membership does not establish performance, and wider coverage need not reduce precision. Conditioned prediction requires usable context, from measurement, configuration or identification.
- Rollout error has a conditional upper bound. Under common actions, matched initial states, a fixed norm, uniform defect bound , and a true-map Lipschitz bound valid along both trajectories, . The bound, not necessarily actual error, grows geometrically for and , and approaches for . One-step error, recursive prediction error and closed-loop cost can rank models differently.
- Delay changes the observation channel. This controller receives aged states but does not use intervening action history to reconstruct current state. Its MSE decreases with delay on this one synthetic task, which is not a universal benefit. A time-aware estimator is a separate proposed remedy, not a method tested here.
- Residual correction needs alignment, an appropriate class and informative coverage. None alone is sufficient. Generalization beyond visited instances depends on structural assumptions and needs validation. State reconstruction from delayed observations depends on history, dynamics and disturbances.
- Detection is not diagnosis. Different faults can yield similar residuals. Distinguishing them requires discriminating information, which may come from passive measurements, additional sensors or approved active probing.
- A reference quantile is an empirical cutoff. Measure held-out sample exceedances separately from sustained alarm events. A sequential false-alarm guarantee requires a justified null law or robust null class; a percentile label alone is not one.
- Gates must be able to reject. Freeze the acceptance rule before looking at the numbers, evaluate on held-out rollout error rather than training one-step error, keep the prior model for rollback, and do not inspect the final test set during development.
- Test stages provide scoped evidence. SIL exercises code, PIL the target processor, and HIL target hardware against a simulated plant. Representative tests alone do not prove all-workload timing, simulator fidelity or physical safety. Power-HIL setups can contain physical hazards too.
- Safe data collection is a constrained procedure, not blind exploration: a declared envelope, independent safeguards, supervised runs, bounded purpose-built excitation, and coverage of the conditions you intend the model to handle. Update policies must be versioned, gated, auditable, and reversible, and must guard against forgetting and quiet drift.