Charlot Lab · PINN × MPC · verify-first control

A fast surrogate you can trust.

A learned surrogate of the plant's dynamics (a DD-PINN-style flow map — predict the next state in one pass, initial condition exact by construction) is the model inside model-predictive control. It plans by rolling the surrogate forward. The verify-first twist: every plan carries a certified trust band — a rigorous bound on how far the surrogate can drift from the true dynamics — so you know exactly how far ahead to believe it, and the certificate voids the moment the state leaves the region the surrogate was trained on.

true trajectory (RK4 plant) MPC plan (surrogate rollout) + certified trust band trust region (in-distribution) certificate void (out-of-distribution)
next-step certified error ≤ · horizon(H) band surrogate: TRUSTED reached upright: 0 surrogate 1-step RMSE 0.0026 vs RK4-20
CEM-MPC (K=48, H=12) plans on the surrogate; the TRUE plant is RK4 with 20 substeps/step

The plan is the gold dashed line; the gold tube around it is the certified trust band, which starts a hair wide (the next executed step is certified to ≤ ~0.02) and fans out with the horizon — so you trust the immediate action tightly and re-plan every step rather than believe a distant prediction. Press kick to shove the state past the dashed trust region: the surrogate is now extrapolating, the certificate voids (red), and the honest move is to distrust the plan until the state returns in-distribution. This is verify-first applied to the MODEL: a learned surrogate is only as good as its certificate, and the certificate names its own domain of validity. Honest scope: on this 2-D toy the surrogate is not faster than the true integrator — the large surrogate speedups (≈466× soft-robot MPC, ≈44,000× Cosserat-rod) are real only for stiff, soft, or high-DOF plants; here we demonstrate the certified-surrogate-in-the-loop mechanism, not the speedup. Nano, single-seed, simulation.