Tube-Based Robust Nonlinear MPC

• Tube-based robust nonlinear MPC is a predictive control strategy that constructs a control-invariant tube around a nominal trajectory to manage disturbances and model uncertainty.
• It ensures recursive feasibility and robust constraint satisfaction by tightening state and input constraints and employing an ancillary feedback controller.
• Recent advances integrate learning-based models, adaptive strategies, and distributional robustness to reduce conservatism and enhance closed-loop performance.

A tube-based robust nonlinear model predictive control (MPC) scheme encodes model uncertainty and external disturbances through the concept of a control-invariant “tube” surrounding a nominal system trajectory. Such frameworks guarantee recursive feasibility and robust constraint satisfaction for a class of nonlinear plants under bounded uncertainty. This methodology has emerged as the de facto standard for computationally tractable robust NMPC design, enabling explicit handling of hard state and input constraints as well as principled trade-offs between conservatism and closed-loop performance. Recent advances further enable adaptation, scalable optimization, learning-based modeling, control contraction metrics, and distributional robustness within the tube-based paradigm.

1. Mathematical Formulation and Core Methodology

Consider a general nonlinear system under additive disturbance,

$$x_{k+1} = f(x_k, u_k) + w_k, \qquad w_k \in \mathcal{W} \subset \mathbb{R}^n$$

where $x_k\in \mathbb{R}^n$, $u_k\in \mathbb{R}^m$, $\mathcal{W}$ is a known compact convex disturbance set. State and input constraints,

$x_k \in X, \quad u_k \in U$

must be enforced for all $k$ and all admissible $w_k$.

Tube-based MPC parametrizes the true state as

$x_{i|k} = z_{i|k} + e_{i|k}$

with $z_{i|k}$ the nominal prediction and $e_{i|k}$ the tracking error. The tube cross-section is

$x_{i|k} \in z_{i|k} \oplus S$

where $S$ is an RPI (robust positively invariant) set for the ancillary error dynamics, which are typically linearized or over-approximated as

$$e_{k+1} = A(z_k, v_k)\, e_k + B(z_k, v_k)\, K e_k + w_k$$

with a static or dynamic ancillary feedback law $u_k = v_k + K e_k$. The set $S$ is computed so that $e_k \in S \implies e_{k+1} \in S$ for all $w_k \in \mathcal{W}$ (Carecci et al., 3 Jan 2026, Köhler et al., 2019, Doff-Sotta et al., 1 Feb 2026).

The robust MPC is then formulated as an optimization over only the nominal trajectory: $$\begin{aligned} \min_{v_{0:N-1}} \ & \sum_{i=0}^{N-1} \ell(z_{i|k}, v_{i|k}) + V_f(z_{N|k}) \ \text{subject to:} \ & z_{0|k} = x_k\ & z_{i+1|k} = f(z_{i|k}, v_{i|k})\ & z_{i|k} \in X \ominus S\ & v_{i|k} \in U \ominus K S\ & z_{N|k}\in X_f \end{aligned}$$ Here, $\ominus$ denotes Pontryagin set difference, tightening constraints to account for tube deviations (Carecci et al., 3 Jan 2026, Bokor et al., 20 Nov 2025, Doff-Sotta et al., 1 Feb 2026). At each timestep, the plant is controlled by

$u_k = v_{0|k}^* + K(x_k - z_{0|k}^*)$

where $z_{0|k}^*, v_{0|k}^*$ are the optimizer's initial values.

2. Tube Dynamics, Uncertainty Modeling, and Feedback Design

Tube cross-section and invariance: The RPI set $S$ may be polyhedral, ellipsoidal, or parameterized by incremental Lyapunov functions or contraction metrics (Carecci et al., 3 Jan 2026, Köhler et al., 2019, Sasfi et al., 2022).

Ancillary controller: A linear or nonlinear local feedback map $K$ (often LQR or based on contraction/ISS rates) confines disturbances within the tube. Recent methods use dynamic tube geometry with sliding-mode/boundary-layer control, solving ODEs for tube thickness according to local uncertainty bounds (Lopez et al., 2019).

Uncertainty sets: The tube can adapt its size locally based on state-dependent disturbance bounds $\Delta(x)$, as in the DTMPC scheme, which relates tube geometry to local modeling error and disturbance (Lopez et al., 2019). Tube propagation using min-max DIs in continuous time supports arbitrary convex tubes under input-affine uncertainties (Villanueva et al., 2016).

Constraint tightening: Both state and input constraints are "tightened" (shrunk) by the maximal error induced by the tube cross-section. For general nonlinear constraints $g(x_k, u_k) \leq 0$, a Lipschitz or incremental Lyapunov characterization provides an explicit tightening parameter (Köhler et al., 2019, Sasfi et al., 2022).

3. Extensions: Learning, Distributional Robustness, and Adaptation

Learning-augmented tube MPC: Data-driven surrogate models (e.g., Random Fourier Features, Koopman operators, DC programming, neural net residuals) yield more accurate nominal predictions and tighter tube computations (Bokor et al., 20 Nov 2025, Zhang et al., 2021, Doff-Sotta et al., 1 Feb 2026, Xie et al., 2022). For example, residual learning with RFF can decrease tube volume by up to 50% relative to linear baselines in nonlinear vehicle path tracking (Bokor et al., 20 Nov 2025).

Distributionally robust and ambiguity tube MPC: Distributional ambiguity in the disturbance model is managed by propagating sets of probability measures ("ambiguity tubes") in the Wasserstein distance, with risk-adjusted constraints and costs. The corresponding MPC problem optimizes over tubes of distributions and ancillary control laws, providing stochastic stability guarantees via supermartingale Lyapunov arguments (Wu et al., 2022).

Adaptive and set-membership approaches: Robust adaptive MPC combines online parameter set estimation (set-membership) with tube construction based on the current (shrinking) uncertainty set. Contraction-based and incremental Lyapunov tube formulations enable recursive tightening of constraints and reduction in tube conservatism as model knowledge improves online (Sasfi et al., 2022, Köhler et al., 2019).

Offset-free tracking: Tube-based MPC architectures can be augmented with dynamic integral action and learning-based modeling (e.g., NNARX) to provide robust, offset-free regulation even in the presence of plant-model mismatch (Xie et al., 2022).

4. Computational Tractability and Practical Scheme Design

Complexity control: Tube-based nonlinear MPC introduces only a minor computational overhead compared to nominal MPC, especially when using scalar tube radii (incremental Lyapunov or contraction metric parameterizations) or simplex parameterizations of cross-sections (Doff-Sotta et al., 1 Feb 2026, Köhler et al., 2019, Carecci et al., 3 Jan 2026). Dynamic tube MPC and other state-dependent schemes only require a handful of extra scalar states and constraints per sample (Lopez et al., 2019).

Offline/online workload split: Ancillary feedback gains, tube cross-sections, RPI sets, and model residual error bounds are computed offline. Online, the robust constrained problem is a nominal MPC with tightened constraints and, optionally, tube-dynamics variables (Bokor et al., 20 Nov 2025, Doff-Sotta et al., 1 Feb 2026).

Sequential convex programming and DC representations: For general nonlinear (possibly nonconvex) plant models, difference-of-convex programming—linearizing only the concave part—enables single convex program solves per timestep, with recursive feasibility and stability guarantees (Doff-Sotta et al., 1 Feb 2026).

5. Performance, Trade-offs, and Theoretical Guarantees

Recursive feasibility: Tightened constraints and RPI tube construction guarantee that, given feasibility at time zero, the MPC problem remains feasible for all future $k$ under admissible uncertainties (Carecci et al., 3 Jan 2026, Köhler et al., 2019).

Robust constraint satisfaction: By construction, real trajectories are confined to the nominal trajectory plus the RPI tube, ensuring that original state and input constraints are preserved for all admissible disturbances and model errors (Carecci et al., 3 Jan 2026, Bokor et al., 20 Nov 2025, Subramanian et al., 2022).

Stability: The Lyapunov function or contraction-metric-based cost is strictly decreasing, yielding robust asymptotic or practical asymptotic stability. When combined with integral action or online parameter adaptation, this framework ensures offset-free regulation and convergence of the parameter set (Xie et al., 2022, Köhler et al., 2019, Sasfi et al., 2022).

Conservatism vs. computational cost: Classic tube MPC is often conservative due to worst-case constant tubes. Dynamic and adaptive tube constructions (e.g., homothetic tubes, set-membership updates, or DTMPC) significantly reduce conservatism by accounting for local uncertainty, with only marginal increases in computational burden (Lopez et al., 2019, Sasfi et al., 2022, Köhler et al., 2019).

6. Applications and Empirical Insights

Tube-based robust NMPC has been demonstrated across domains:

• Process industries: Multi-stage polymerization processes using tube-enhanced multi-scenario NMPC achieve zero constraint violations and fast economic batch times, with substantial cost reductions over full scenario-tree NMPC (Subramanian et al., 2022).
• Bioprocess control: Anaerobic co-digestion plants leveraging tube-based NMPC achieved up to 35% reduction in toxic concentration peaks (VFA) at a modest (≈12%) degradation in methane-tracking accuracy compared to classical NMPC (Carecci et al., 3 Jan 2026).
• Autonomous vehicles/robotics: RFF-augmented tube MPC for nonlinear bicycle models achieved up to 70% smaller tracking errors and half the tube volume relative to classical linear tubes, with all hard constraints satisfied over 50 Monte Carlo trials (Bokor et al., 20 Nov 2025, Sun et al., 2017).
• Adaptive control and learning: Tube-based RAMPC frameworks have been validated in nonlinear uncertain systems, showing robust performance with shrinking conservatism over time as parameter sets are learned (Köhler et al., 2019).

7. Future Directions and Comparative Perspectives

Ongoing research in tube-based robust nonlinear MPC investigates:

• Online model adaptation/integration of high-dimensional learning surrogates (e.g., RFF, kernel models, Koopman operators, DCNNs) (Bokor et al., 20 Nov 2025, Zhang et al., 2021, Doff-Sotta et al., 1 Feb 2026),
• Distributionally robust ambiguity tubes for risk-sensitive control under uncertain disturbance distributions (Wu et al., 2022),
• Homothetic tube propagation based on contraction metrics and fully nonlinear model dynamics (Sasfi et al., 2022),
• Balancing scenario-tree complexity vs. tube conservatism in industrial and economic NMPC (Subramanian et al., 2022),
• Efficient computation of RPI sets and constraint tightening for large-scale systems (Doff-Sotta et al., 1 Feb 2026, Villanueva et al., 2016),
• Embedded real-time implementations and extensions to hybrid/discrete-time switching systems.

These advances strengthen the tube-based robust nonlinear MPC paradigm as the backbone of modern safety-critical and uncertain process control, supported by theoretical guarantees and scalable, data-driven optimization.