Tube-based Model Predictive Control

• Tube-based MPC is a robust control strategy that decomposes control into a nominal predictive plan and an ancillary feedback law to manage disturbances and model mismatches.
• It employs explicit tube parameterizations—such as polyhedral, zonotopic, and ellipsoidal—to tighten constraints and maintain recursive feasibility and closed-loop stability.
• Recent advances include learning-based refinement and system-level optimization, which expand feasible regions and reduce conservatism in high-dimensional and nonlinear systems.

Tube-based Model Predictive Control (MPC) is a foundational paradigm in constrained robust and adaptive control for handling multivariable systems with plant-model mismatch, exogenous disturbances, and dynamic uncertainty. In Tube-based MPC, the feedback law is decomposed into a nominal predictive controller and an ancillary feedback law that robustly confines the true system trajectories within time-varying sets, or "tubes," around the nominal trajectory, ensuring satisfaction of all hard constraints and providing rigorous stability and feasibility guarantees across a wide spectrum of uncertainty models and system classes.

1. Fundamental Principles of Tube-based MPC

At the core of tube-based MPC is the system decomposition into a nominal trajectory and a deviation (error) system. For a general discrete-time (possibly nonlinear) system with additive disturbance:

$x_{k+1} = f(x_k, u_k) + w_k,$

$$x_k \in \mathcal{X}, \quad u_k \in \mathcal{U}, \quad w_k \in \mathcal{W},$$

the control law is implemented as

$u_k = v_k + K(x_k - z_k),$

where $z_k$ is the nominal state, $v_k$ the nominal input from the MPC optimizer, and $K$ the ancillary feedback. The deviation $e_k = x_k - z_k$ is confined within a robust positively invariant (RPI) set or a parameterized family of sets called the tube, which is propagated using the error dynamics. Constraint satisfaction is enforced for all $x_k$ and $u_k$ in the tube via tightened sets: $z_k \in \mathcal{X} \ominus \mathcal{T}_k$, $$x_k \in \mathcal{X}, \quad u_k \in \mathcal{U}, \quad w_k \in \mathcal{W},$$0.

Tube-based MPC is distinct from standard MPC in two critical aspects: (i) explicit constraint tightening accounts for worst-case deviations, and (ii) robust recursive feasibility and closed-loop stability can be rigorously ensured via tailored tube parametrizations (Sieber et al., 2021, Sieber et al., 2021, Diaconescu et al., 24 Sep 2025).

2. Tube Construction, Parameterizations, and Complexity

A spectrum of tube parameterizations has been introduced to balance conservatism with computational tractability and region-of-attraction size:

• Polyhedral / Homothetic Tubes: The tube cross-section at stage $$x_k \in \mathcal{X}, \quad u_k \in \mathcal{U}, \quad w_k \in \mathcal{W},$$1 is modeled as a homothetic or polytopic set (scaled or shifted proto-set, e.g., $$x_k \in \mathcal{X}, \quad u_k \in \mathcal{U}, \quad w_k \in \mathcal{W},$$2). Homothetic tubes require only scaling factors, greatly reducing the number of decision variables compared to full polyhedral tubes (Gao et al., 6 May 2025, Hanema et al., 2019).
• Zonotopic Tubes: Elastic or scaled zonotope parameterizations allow the tube shape to adapt at runtime, facilitating efficient inclusion testing and supporting high-dimensional systems. For $$x_k \in \mathcal{X}, \quad u_k \in \mathcal{U}, \quad w_k \in \mathcal{W},$$3, the tube shape is defined by generator scaling—drastically improving real-time tractability relative to polyhedral constraints (Diaconescu et al., 24 Sep 2025, Ghiasi et al., 24 Dec 2025).
• Ellipsoidal Tubes: Tubes parameterized by ellipsoidal sets provide advantages for large-scale systems, as their complexity scales linearly in the system order and the key inclusions reduce to semidefinite constraints (Parsi et al., 2022).

The choice of tube structure governs the trade-off between computational complexity, feasible region volume, and conservatism. Polyhedral tubes yield maximal regions but scale poorly. Zonotopic and ellipsoidal approaches, especially with precomputed inclusion certificates, achieve real-time feasibility in higher-dimensional regimes (Diaconescu et al., 24 Sep 2025, Parsi et al., 2022, Ghiasi et al., 24 Dec 2025).

3. Robustness, Uncertainty, and Learning-based Adaptation

Tube-based MPC is effective for both additive disturbances and broader uncertainty classes:

• Additive Bounded Disturbances: Fundamental results guarantee robust constraint satisfaction and practical stability by selecting tubes that are invariant under the closed-loop error dynamics (Gao et al., 6 May 2025, Luo et al., 2024).
• Parametric Uncertainty: For LPV or polytopic uncertainty, the tube invariance is enforced across a family of vertices representing extreme parameter values. Heterogeneous parameterizations (hybrid scenario/homothetic tubes) allow interpolation between region-of-attraction and complexity (Hanema et al., 2019, Badalamenti et al., 2024).
• Distributional Uncertainty: Data-driven or online learning-based tube MPC utilizes observed disturbances to adapt the disturbance set (e.g., updating a homothetic polytope or zonotope) using scenario-based or statistical guarantees. This approach expands feasible regions and reduces conservatism, with precise probabilistic receding-horizon feasibility bounds (Gao et al., 6 May 2025, Ghiasi et al., 24 Dec 2025, Tranos et al., 2022).
• Dynamic and Nonlinear Uncertainty: IQC-based tube MPC uses dynamic multipliers to bound deviations when the plant is subject to general dynamic uncertainty, constructing tube update laws that guarantee input-to-state stability under minimal conservatism (Schwenkel et al., 2021).

Learning-based tube refinement is typically performed via linear or convex programming and statistical scenario theory, ensuring prescribed constraint violation probabilities and recursive feasibility (Gao et al., 6 May 2025, Ghiasi et al., 24 Dec 2025, Tranos et al., 2022).

4. Advanced System-Level, Nonlinear, and Data-driven Extensions

Recent developments have greatly generalized and extended tube-based MPC.

• System-Level Tube MPC (SLTMPC): By optimizing the entire tube controller (rather than fixing $$x_k \in \mathcal{X}, \quad u_k \in \mathcal{U}, \quad w_k \in \mathcal{W},$$4 offline), system-level parameterizations flexibly trade conservativeness for tractability. SLTMPC enables concurrent, online optimization of both the nominal trajectory and the sequence of tubes (e.g., system-level disturbance reachable sets), extending region-of-attraction and performance (Sieber et al., 2021, Sieber et al., 2021, Sieber et al., 2022, Sieber et al., 2024).
• Nonlinear Systems: For nonlinear, smooth, or hybrid plants (including systems lifted via Koopman operators or nonlinear residuals), local linearization, error bounding, or lifting is combined with tube construction, producing robust feasible regions and retaining tractable optimization problems (Zhang et al., 2021, Luo et al., 2024, Doff-Sotta et al., 1 Feb 2026, Bokor et al., 20 Nov 2025). Difference-of-convex (DC) decompositions combined with sequential convex programming have been demonstrated to yield recursively feasible tube MPC for general nonlinear systems with theoretical stability guarantees (Doff-Sotta et al., 1 Feb 2026).
• Economic and Constraint-less Variants: Robust economic tube MPC leverages turnpike and dissipativity properties to ensure robust convergence in the absence of terminal costs or constraints, providing sharpened bounds on average performance and convergence to steady-state tubes (Schwenkel et al., 2019).
• Anticipatory/Asynchronous Computation: In resource-constrained or delay-prone systems, anticipatory tube-MPC or asynchronous tube design pipelines enable computational delays to be absorbed or offloaded by predictive state estimation and memory, maintaining rapid online response and recursive feasibility (Luo et al., 2024, Sieber et al., 2022, Sieber et al., 2024).

5. Algorithmic Structure and Implementation Benchmarks

A typical tube-based MPC workflow is outlined by the following steps, with data-driven or system-specific modifications as necessary:

1. Offline:
   • Parameterize and compute the primitives: tube shape (polyhedral, zonotopic, ellipsoidal), ancillary feedback ($$x_k \in \mathcal{X}, \quad u_k \in \mathcal{U}, \quad w_k \in \mathcal{W},$$5), and robust invariant/contractive sets for the error dynamics (Diaconescu et al., 24 Sep 2025, Parsi et al., 2022, Ghiasi et al., 24 Dec 2025).
   • For SLTMPC, precompute or formulate the necessary system-response mapping and robust constraint-tightening ingredients (Sieber et al., 2021, Sieber et al., 2021, Sieber et al., 2024).
2. Online (per sampling step):
   • Linearize or lift the model if nonlinear or using operator-based techniques (Zhang et al., 2021, Luo et al., 2024, Doff-Sotta et al., 1 Feb 2026, Bokor et al., 20 Nov 2025).
   • Update or refine tube sets (if data-driven or adaptive).
   • Solve the (typically quadratic, conic, or SDP) program for the nominal trajectory and tube parameters, enforcing robust reachable set/tube invariance and tightened constraints.
   • Apply the feedback law:

   $$x_k \in \mathcal{X}, \quad u_k \in \mathcal{U}, \quad w_k \in \mathcal{W},$$6

   where the $$x_k \in \mathcal{X}, \quad u_k \in \mathcal{U}, \quad w_k \in \mathcal{W},$$7 denotes the optimizer solution at time $$x_k \in \mathcal{X}, \quad u_k \in \mathcal{U}, \quad w_k \in \mathcal{W},$$8. - Shift horizon and tube for next step.

Empirical benchmarks demonstrate large feasible regions and real-time solution times even for moderately high-dimensional systems (e.g., solve times $$x_k \in \mathcal{X}, \quad u_k \in \mathcal{U}, \quad w_k \in \mathcal{W},$$910 ms for double-integrator $u_k = v_k + K(x_k - z_k),$0, $u_k = v_k + K(x_k - z_k),$1; feasible to $u_k = v_k + K(x_k - z_k),$2 for zonotopic tubes). Adaptive/elastic tubes and learning-based refinement significantly enlarge feasibility regions and reduce conservatism, outperforming prior fixed-tube baselines (Diaconescu et al., 24 Sep 2025, Ghiasi et al., 24 Dec 2025, Gao et al., 6 May 2025, Tranos et al., 2022).

6. Rigorous Guarantees: Feasibility, Constraint Satisfaction, and Stability

Tube-based MPC offers strong theoretical guarantees, rigorously established via the following properties:

• Recursive Feasibility: If the tube (or nominal/tube pair) is feasible at initialization, the receding-horizon policy preserves feasibility for all time under the designed uncertainty/disturbance set (Sieber et al., 2021, Sieber et al., 2021, Parsi et al., 2022, Badalamenti et al., 2024, Ghiasi et al., 24 Dec 2025).
• Robust Constraint Satisfaction: For all admissible uncertainties and disturbance realizations, all predicted and implemented $u_k = v_k + K(x_k - z_k),$3 respect the original constraints (Gao et al., 6 May 2025, Zhang et al., 2021, Bokor et al., 20 Nov 2025).
• Input-to-State/Practical Stability: Under standard assumptions (Schur/Hurwitz $u_k = v_k + K(x_k - z_k),$4, contractive/positively invariant tube), the closed-loop system is ISS or practically stable with respect to the disturbance set; state approaches a minimal tube around the origin or steady-state (Gao et al., 6 May 2025, Schwenkel et al., 2019, Ghiasi et al., 24 Dec 2025, Luo et al., 2024, Schwenkel et al., 2021). The Lyapunov decrease is typically enforced via cost-to-go or terminal quadratic constraints.
• Probabilistic Guarantees in Learning: With sufficient data, the learned disturbance sets contain the true support with high confidence, and recursive feasibility and constraint satisfaction hold with prescribed probability, supported by scenario theory (Gao et al., 6 May 2025, Ghiasi et al., 24 Dec 2025, Tranos et al., 2022).

7. Extensions, Open Problems, and Benchmark Applications

Tube-MPC continues to evolve in scope and performance along several lines:

• Distributed, Output Feedback, and Large-scale Systems: Current research extends tube-based frameworks to distributed control, output feedback, and high-dimensional plants, leveraging scalable set representations (ellipsoidal, zonotopic) and system-level synthesis (Parsi et al., 2022, Sieber et al., 2021, Sieber et al., 2024).
• Nonlinear Data-driven and Hybrid Dynamics: Ongoing work blends machine learning (e.g., RFFs, operator learning, DCNNs) with robust tube design to handle complex uncertainty and nonlinearity with minimal conservatism (Bokor et al., 20 Nov 2025, Doff-Sotta et al., 1 Feb 2026).
• Economic, Reference-tracking, and Contractive Tube Variants: Methods are available for robust economic goals, reference tracking under time-varying or persistent uncertainty, and adaptive contractive tubes with online contractivity/envelope guarantees (Schwenkel et al., 2019, Badalamenti et al., 2024, Ghiasi et al., 24 Dec 2025).
• Benchmarks: Tube-based MPC approaches are widely benchmarked on vehicle path-tracking, robot manipulators, and process control (mass-spring systems, chemical reactors, VTOL aircraft, autonomous driving) with consistent evidence of enlarged feasible regions, reduced constraint violations, and up to an order-of-magnitude speed-up versus full nonlinear or polytopic methods (Luo et al., 2021, Luo et al., 2024, Diaconescu et al., 24 Sep 2025, Bokor et al., 20 Nov 2025, Ghiasi et al., 24 Dec 2025, Parsi et al., 2022).

Comprehensive references and technical details are available in (Sieber et al., 2021, Sieber et al., 2021, Diaconescu et al., 24 Sep 2025, Gao et al., 6 May 2025, Ghiasi et al., 24 Dec 2025, Sieber et al., 2024, Parsi et al., 2022, Schwenkel et al., 2021, Bokor et al., 20 Nov 2025, Doff-Sotta et al., 1 Feb 2026).