Sim-to-Real Deployment and Continual Correction › Limitations and Impact

The experiment we built has specific limitations that it is important to name, because a reader who takes the wrong lesson from a toy is worse off than a reader who takes no lesson at all.

This is a simulator inside a simulator. DEPLOY and SHIFT are synthetic plants. The predictors use different forms, including analytic clipped dynamics and polynomial feature maps, but the plant generator remains the same oscillator family. Sensor delay changes observation timing; reduced authority is disclosed to both planner and plant. The example does not test an unmodeled actuator limit or omitted contact, thermal coupling, backlash or sensor drift. It establishes neither physical transfer nor a general limit on residual correction.

Zero process noise is a favorable assumption. With noiseless transitions and exactly representable dynamics, identification and residual fitting are essentially exact. This makes open-loop prediction easier, but does not produce a closed-loop tracking advantage for the fitted models here. We did not rerun with process noise, so no change in ranking or noise robustness is established. Noisy adaptation would require separate checks of estimation error, regularization, and monitoring calibration.

The randomization family is simple. Twenty-four draws from independent uniform parameter intervals do not test correlated, heavy-tailed or multimodal variation. Such distributions can occur in deployment, but none necessarily makes a parameter-blind predictor worse: conditional input information, shared dynamics and evaluation weighting matter. Compare them in a separate controlled experiment rather than inferring a ranking from the shape of a marginal distribution.

Shared scenarios are not identical trajectories. Matching initial conditions, references and planner seeds permits paired episode-level comparisons even when selected actions and states diverge. A per-step state difference compares policy outcomes, not predictions from identical states. Repeated independent paired scenarios are one way to estimate empirical uncertainty in a cost difference. Dependence-aware estimates or calculations under a justified known probability law are alternatives. This example evaluates one fixed scenario and reports descriptive costs, not an estimated sampling variance.

The safety evidence is limited. The watchdog has no real-time guarantee; envelope measurements concern a synthetic plant. The gate applies a fixed rule to one held-out episode, without an uncertainty estimate. The reference quantile is a cutoff, while the measured same-condition fraction counts sample exceedances, not alarm events. A sequential guarantee requires a stated procedure and justified null assumptions, including dependence. Those requirements are not established here.

Coverage and observation timing are separate constraints. This chapter does not vary adaptation-data quantity, so the delay sweep cannot establish that more data would never help. It shows the behavior of fixed predictors under changed observation timing. Whether a shortfall comes from missing coverage, unavailable history, a model-class limit, or the control objective requires a targeted experiment, rather than a diagnosis from tracking cost alone.

The executable example establishes a limited set of facts: noiseless identification and residual fitting can recover this deployment plant; accurate open-loop prediction does not automatically minimize this sampled controller's tracking MSE; and a held-out gate can reject a candidate on a criterion distinct from average rollout fit. The zero-delay curve is a reference, not a bound. These CPU results do not establish that any method improves physical deployment, and they do not replace timing, fault-injection, or supervised hardware trials.

The final chapter of this part examines what remains open after the work on failure modes, robustness and safe control, training and inference systems, governance, and deployment. It asks which assumptions the methods depend on and proposes experiments that can test them. We will close the handbook there.