# Certified Tube Wrapper for Learned Predictive Control

- ID: 2022
- Canonical URL: https://synthcore.org/idea/2022/certified-tube-wrapper-for-learned-predictive-control
- API JSON: https://synthcore.org/api/idea/2022.json
- API Markdown: https://synthcore.org/api/idea/2022.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.08921](https://arxiv.org/abs/2608.08921)
- Category: dynamics
- Solves: stability, accuracy, sample-efficiency
- ML areas: world-model, rl, mlp, training-dynamics
- Math tags: control-theory, dynamical-systems, spectral-theory, optimization, linear-algebra
- Ratings: usefulness 7/10; difficulty 7/10; novelty 8/10

## Idea description

Combine a learned dynamics model or neural policy with a short-horizon robust MPC wrapper. Instead of tightening every future constraint by one stationary worst-case radius, propagate uncertainty using the actual neural closed-loop Jacobians and explicitly fall back when the tightened optimization problem is infeasible, making envelope violations observable rather than silently unsafe.

## Mathematical statement

Let the learned plant model be $x_{k+1}=F_\theta(x_k,u_k)+d_k$ and let the nominal controller be $u_k=\pi_\phi(\hat x_k)$, possibly with an MPC correction. Around the nominal trajectory, the closed-loop error satisfies $e_{j+1}\approx A_{\mathrm{cl},k+j}e_j+w_j$, where $A_{\mathrm{cl},k+j}=\partial F_\theta/\partial x+\partial F_\theta/\partial u\,\partial\pi_\phi/\partial x$ evaluated at the nominal state and action. Use the horizon-specific recursion $\bar e_{j+1}=|A_{\mathrm{cl},k+j}|\bar e_j+\bar w_j$. For uncertain parameters $p=p_0+\Delta p$ with $|\Delta p|\le\bar p$, bound the multiplicative term $B(\Delta p)u$ by an additive vector $\bar w^{\mathrm{mult}}\ge |B|\bar p|u|$, then use $\bar w=\bar w^{\mathrm{mult}}+\bar w^{\mathrm{add}}$. If $|A_{\mathrm{cl},k+j}|\le M$ and $\rho(M)<1$, the robust tube has finite gain $\|(I-M)^{-1}\bar w\|$.

## Key formulas

- $$e_{j+1}\approx A_{\mathrm{cl},k+j}e_j+w_j,\qquad A_{\mathrm{cl},k+j}=F_x+F_u\,\pi_x,$$
- $$\bar e_{j+1}=|A_{\mathrm{cl},k+j}|\bar e_j+\bar w^{\mathrm{add}}_j+\bar w^{\mathrm{mult}}_j,$$
- $$\bar w^{\mathrm{mult}}_j\ge |B_j|\,\bar p\,|u_j|,\qquad \rho(M)<1\Rightarrow \bar e_\infty\le(I-M)^{-1}\bar w.$$
- $$\mathcal X_j^{\mathrm{tight}}=\{\hat x_j:\hat x_j+[-\bar e_j,\bar e_j]\subseteq\mathcal X\},\qquad \mathcal U_j^{\mathrm{tight}}=\{\hat u_j:\hat u_j+[-\bar u_j,\bar u_j]\subseteq\mathcal U\}.$$

## Implementation notes

1. Integration point: use a learned dynamics model inside model-predictive control for a simple benchmark such as cart-pole, pendulum, or a 2D navigation task. The neural policy proposes nominal actions, while a differentiable or ordinary QP solves a local correction subject to tube-tightened state and action limits. At every MPC call, roll out the nominal model, calculate the closed-loop Jacobian of the model-plus-policy composition, and propagate one element-wise radius vector per horizon step. 2. Pseudocode: obtain current state `x`; roll out `xhat[j+1]=F_theta(xhat[j],u_nom[j])`; compute `Acl_abs=abs(JxF + JuF@JxPolicy)`; compute parameter-error margin `wmult=abs(B)@pbar@abs(u_nom[j])` and add a measured residual bound `wadd`; update `ebar=Acl_abs@ebar+wmult+wadd`; shrink each state and action box by `ebar` and the corresponding action tube; solve the QP; if infeasible, execute a bounded fallback controller and log an envelope breach. 3. Compute Jacobians with autodiff, estimate residual bounds from validation rollouts using a high quantile plus a safety factor, and estimate $M$ as a componentwise maximum over a calibration set. The source mechanism is the multiplicative-to-additive conversion, horizon-dependent tightening, spectral certificate, and explicit infeasibility fallback. 4. First cheap experiment: compare nominal MPC, constant-width tube MPC, and horizon-dependent learned tube MPC on noisy cart-pole with randomized mass and actuator gain. Sweep the uncertainty scale $\beta$ and recurrent or policy gain. The prediction is that violation probability remains near the calibration confidence level while $\rho(M)<1$, tightening converges to $(I-M)^{-1}\bar w$, and infeasibility or fallback frequency rises sharply as either $\rho(M)\to1$ or the disturbance scale exceeds the calibrated envelope. Report whether the observed transition occurs within 20% of the predicted $\rho(M)=1$ boundary.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: harms
- Verdict: Built a deterministic certified-tube MVP with closed-loop Jacobian propagation, additive plus multiplicative uncertainty conversion, spectral-radius checks, sampled interval validation, and nominal/constant/horizon-dependent control modes with observable fallback logging. The key predictions manifested: the finite-horizon instability boundary was observed at 1.01 versus predicted gamma*Lambda=1.00, scalar steady-state radii matched w/(1-r) within 2e-14 relative error, and multiplicative margin slope matched |B u| exactly. The control mini-experiment showed zero violations for every method and uncertainty level, so it does not establish a practical safety or performance win over the baselines.

### Mechanism check

- Verdict: Built a deterministic certified-tube MVP with closed-loop Jacobian propagation, additive plus multiplicative uncertainty conversion, spectral-radius checks, sampled interval validation, and nominal/constant/horizon-dependent control modes with observable fallback logging. The key predictions manifested: the finite-horizon instability boundary was observed at 1.01 versus predicted gamma*Lambda=1.00, scalar steady-state radii matched w/(1-r) within 2e-14 relative error, and multiplicative margin slope matched |B u| exactly. The control mini-experiment showed zero violations for every method and uncertainty level, so it does not establish a practical safety or performance win over the baselines.
- Confidence: 8/10
- Limitations: The plant/controller benchmark is a lightweight scalar surrogate rather than cart-pole or a learned neural dynamics model; no autodiff, QP solver, model training, calibration quantile estimation, action-tube propagation, or stressed infeasibility regime was tested. The controller comparison is therefore only a functional MVP and not evidence of a robust MPC performance win.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 0/8
- Benchmark verdict: idea worse (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/819/bench_report.json)
- [report.md](https://synthcore.org/code/819/report.md)
- [report_bench_2026-09-01T111808.md](https://synthcore.org/code/819/report_bench_2026-09-01T111808.md)
- [results.json](https://synthcore.org/code/819/results.json)
- [stage2_tube_bench.py](https://synthcore.org/code/819/stage2_tube_bench.py)
- [tube_experiment.py](https://synthcore.org/code/819/tube_experiment.py)
- [Download all files as ZIP](https://synthcore.org/download/819)

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.