Tube-based Robust Model Predictive Control

• Tube-based RMPC is a control method that separates system trajectories into a nominal part and a bounded error 'tube' to ensure robust feasibility.
• It employs various tube designs—rigid, homothetic, and variable-shape—to manage both additive and multiplicative uncertainties through tightened constraints.
• Advanced implementations leverage convex reformulations and real-time adaptations to achieve computational efficiency and robust performance in complex applications.

Tube-based Robust Model Predictive Control (RMPC) is a class of control methodologies in which the evolution of system states is robustified against bounded disturbances and uncertainties by enforcing that the true trajectory remains within a “tube” surrounding a nominal, disturbance-free trajectory. Tube-based schemes are especially powerful in guaranteeing recursive feasibility and robust satisfaction of hard state and input constraints, while preserving computational tractability—often via convex reformulation—especially in systems with additive and, more generally, multiplicative uncertainties.

1. Core Concepts and Theoretical Underpinnings

At the heart of tube-based RMPC is the decomposition of system trajectories into a nominal (disturbance-free) component and a bounded error component (the "tube"), maintained by ancillary robust feedback. For a discrete-time linear system with additive disturbance: $$x_{k+1} = A x_k + B u_k + w_k, \quad w_k \in \mathcal{W}$$ the nominal trajectory $\bar{x}_k$ evolves under: $\bar{x}_{k+1} = A \bar{x}_k + B \bar{u}_k$ and the tracking error $e_k = x_k - \bar{x}_k$ is stabilized via a tube feedback policy, usually $u_k = \bar{u}_k + K(x_k - \bar{x}_k)$ for some stabilizing $K$. The error dynamics become: $e_{k+1} = (A + BK) e_k + w_k$ The “tube” is a robust control invariant (RCI) set $S$ for the error dynamics, i.e., $e_k \in S$ for all $k$ if $e_0 \in S$. This set is constructed so that, for all admissible disturbances,

$(A + BK) S \oplus \mathcal{W} \subseteq S$

With this separation, the original robust MPC problem—otherwise a min-max optimization—is converted into a tractable nominal MPC with constraints “tightened” by $S$ to ensure all real system trajectories remain robustly feasible (Sieber et al., 2021).

For nonlinear, parametrically uncertain, or output feedback systems, analogous decompositions are constructed using local linearization, incremental Lyapunov methods, or set-theoretic estimation, with extensions to piecewise tube cross-sections, state-dependent tubes, or data-driven tube construction (Köhler et al., 2019, Mulagaleti et al., 2021).

2. Tube Construction: Invariant Sets, Tube Types, and Constraint Tightening

The selection of tube geometry is central to both conservatism and online complexity. Key tube designs include:

• Rigid tubes: The tube cross-section is fixed throughout prediction (e.g., a precomputed polytope or ellipsoid). Simpler but often conservative, suited for static additive uncertainties (Lorenzetti et al., 2019).
• Homothetic tubes: The cross-section is a scaled (and sometimes translated) fixed polytope or ellipsoid. Widths are optimized online, yielding improved performance over rigid tubes and tractable constraint tightening (Massera et al., 2020, Han et al., 2024, Saccani et al., 2023).
• Variable-shape tubes: Tube cross-sections, described by parameterized polytopes (joint facet/vertex parameterization) or ellipsoids, are optimized at each time step to minimize conservatism and maximize constraint admissibility (Villanueva et al., 2022, Parsi et al., 2022).
• Dynamic tubes with state-dependent geometry: In advanced settings, the error tube is endowed with its own dynamics, often as additional optimization states, capturing local uncertainty levels (e.g., via sliding-mode error dynamics or data-driven disturbance bounds) (Lopez et al., 2019, Han et al., 2024, Bokor et al., 20 Nov 2025, Kiani et al., 2023).

Tightening of state and input constraints for the nominal problem is performed via Minkowski/Pontryagin difference with the projected tube set, leading to: $$\bar{x}_k \in \mathcal{X} \ominus S, \quad \bar{u}_k \in \mathcal{U} \ominus K S$$ for state/input constraint sets $\mathcal{X}, \mathcal{U}$, ensuring robustness by construction.

3. Algorithms and Implementation Strategies

A canonical tube-based RMPC solves, at each sampling instant, a tractable finite-horizon OCP for the nominal trajectory $(\bar{x}_k, \bar{u}_k)_{k=0}^N$, subject to tightened constraints and often with a quadratic or economic cost. The control input applied is

$u_0 = \bar{u}_0^* + K(x_0 - \bar{x}_0^*)$

where starred variables denote OCP minimizers.

Advanced computational approaches have been developed to further enhance tractability and parallelism:

• Parallel explicit tube MPC: Decomposes the OCP into “decoupled” subproblems (parametric explicit QPs for each stage) and a structured coupling QP, achieving small latency even with long horizons (Wang et al., 2019).
• System-level tube MPC: Optimizes the tube feedback gain or system-level responses directly online, minimizing conservatism relative to offline-fixed feedbacks but with only mild complexity increase (Sieber et al., 2021).
• Configuration-constrained and concentric-container tube MPC: Directly optimize variable-shape polytopic tubes or generalized containers, reducing conservatism and constraint count relative to classical approaches (Villanueva et al., 2022, Han et al., 2024).

For nonlinear and uncertain systems, linearization-based tube MPC, dynamic tube MPC with tube shape as an optimization state (e.g., via sliding-mode or learning-based error bounds (Lopez et al., 2019, Bokor et al., 20 Nov 2025, Kiani et al., 2023)), and adaptive/learning RMPC (with online disturbance model identification or set-membership parameter estimation (Köhler et al., 2019, Mulagaleti et al., 2021, Kiani et al., 2023)) are state-of-the-art strategies to maintain robust performance with nonlinearity and uncertainty. Pre-solving OCPs “ahead of time” and using state predictions to avoid input delay has also been introduced for robotic systems (Luo et al., 2024, Luo et al., 2021).

4. Extensions: Multiplicative Uncertainty, Learning, and Output Feedback

Modern tube-based RMPC is not limited to additive disturbances:

• Multiplicative/model parametric uncertainties: Several algorithms, including homothetic tube MPC (Massera et al., 2020), concentric container and varying-tube approaches (Han et al., 2024), and ellipsoidal tube RMPC (Parsi et al., 2022), address structured time-varying or parametric uncertainties.
• Data-driven and learning-based tube synthesis: Online estimation of disturbance bounds or system parameters via Gaussian Processes, Random Fourier Features, or set-membership estimation directly shrinks tube size, reducing conservatism and enabling less restrictive operation (Bokor et al., 20 Nov 2025, Kiani et al., 2023, Mulagaleti et al., 2021).
• Robust output feedback MPC: For systems with incomplete state measurements and measurement noise, tube-based RMPC incorporates observer error within the tube and tightens constraints accordingly, achieving robustness while maintaining small to moderate computational load (Lorenzetti et al., 2019).

5. Theoretical Guarantees: Recursive Feasibility, Constraint Satisfaction, and Performance

Tube-based RMPC generally inherits strong theoretical properties:

• Recursive feasibility: The "shifted" solution to the nominal OCP (with appropriate error tube update) guarantees that feasibility at time $k$ implies feasibility at time $k+1$, for all admissible disturbance sequences (Sieber et al., 2021, Lorenzetti et al., 2019).
• Robust constraint satisfaction: By construction, the true trajectory remains within the feasible set for all disturbances, provided the tube and tightened constraints are properly computed (Lorenzetti et al., 2019, Han et al., 2024, Parsi et al., 2022).
• Robust stability: Most tube-based RMPCs guarantee robust asymptotic or input-to-state practical stability (ISpS); the system converges to a disturbance-invariant set around the equilibrium or reference (Sieber et al., 2021, Massera et al., 2020).

Recent developments enable non-conservative feasible regions, sometimes matching those of optimal (but intractable) min–max formulations, via tube shape optimization, system-level or configuration-constrained parameterizations, or real-time tube adaptation (Sieber et al., 2021, Bokor et al., 20 Nov 2025, Villanueva et al., 2022).

6. Applications and Demonstrations

Tube-based RMPC has been demonstrated in a range of advanced control domains:

• Multi-agent pursuit, encirclement, and capture: Robust decentralized tube MPC enables pursuers to guarantee perpetual encirclement and eventual capture of an evader, with angle-partition based simplification of nonconvex constraints (Wang et al., 2021).
• Robotic manipulation and path tracking: Nonlinear tube MPC with hybrid learning-based prediction or anticipatory control eliminates input delays, enables fast real-time implementation, and preserves robust constraint satisfaction (Luo et al., 2024, Luo et al., 2021, Bokor et al., 20 Nov 2025).
• Energy systems, microgrid voltage regulation, aerospace, and vehicle dynamics: Data-driven tube MPC frameworks exploit learning to reduce conservatism, improve total harmonic distortion, and accommodate large uncertainties without compromising critical constraint guarantees (Kiani et al., 2023, Köhler et al., 2019).
• Economic and turnpike MPC: Tube-based methodologies enable robust economic MPC without terminal ingredients, with rigorous cost and stability guarantees using turnpike and dissipativity arguments (Schwenkel et al., 2019).

7. Comparative Performance and Computational Complexity

A central concern in tube-based RMPC is the trade-off between conservatism and online computational complexity:

• Rigid and classical homothetic tubes offer lowest online complexity but are more conservative, especially in high-dimensional or highly uncertain systems.
• Variable-shape, configuration-constrained, and ellipsoidal tubes substantially reduce conservatism at moderate computational expense, scaling linearly with horizon and system order for ellipsoidal representations (Parsi et al., 2022, Han et al., 2024, Villanueva et al., 2022).
• Dynamic tube MPCs, system-level parameterizations, and parallel explicit decompositions further enable real-time use in large-scale, uncertain, or time-varying applications (Lopez et al., 2019, Wang et al., 2019, Sieber et al., 2021).

Key computational and practical observations from the literature are summarized below.

Tube type	Conservatism	Computation	Online Variables
Rigid tube	High	Very low	O(N(n+m))
Homothetic tube	Moderate	Low	O(N(n+m)+N)
Varying-shape tube	Low	Moderate	O(N(n+m)+Np) (poly. p)
Ellipsoidal tube	Low	Moderate	O(N(n+m)); SDP per step
Data-driven/learning	Low	Low (if tube update is efficient)	O(N(n+m))
System-level tube	Very low	Moderate to high	O(N(n+m))

Contemporary research demonstrates systematic reduction in conservatism, recursive feasibility for polytopic and ellipsoidal tubes, scalability to high-dimensional and distributed systems, and robust constraint satisfaction in the presence of system identification uncertainty and data-driven disturbance models (Han et al., 2024, Villanueva et al., 2022, Sieber et al., 2021, Bokor et al., 20 Nov 2025, Kiani et al., 2023).

---

References ([arXiv ids]):

• Encirclement and cooperative pursuit: (Wang et al., 2021)
• Dynamic tube-based nonlinear MPC: (Lopez et al., 2019)
• Concentric container/varying tube RMPC: (Han et al., 2024)
• Configuration-constrained tubes: (Villanueva et al., 2022)
• Parallel explicit tube MPC: (Wang et al., 2019)
• Data-driven tube/invariant set design: (Mulagaleti et al., 2021)
• Nonholonomic robotic tracking: (Sun et al., 2017)
• Tube MPC with Random Fourier Features: (Bokor et al., 20 Nov 2025)
• Tube MPC for manipulation with anticipation: (Luo et al., 2024, Luo et al., 2021)
• Learning-based tube RMPC for microgrids: (Kiani et al., 2023)
• System-level tube MPC: (Sieber et al., 2021)
• Homothetic/ellipsoidal tube MPC: (Massera et al., 2020, Parsi et al., 2022)
• Homothetic tube with multi-step predictors: (Saccani et al., 2023)
• Simple robust output feedback tube MPC: (Lorenzetti et al., 2019)
• Nonlinear tube MPC, computationally efficient: (Köhler et al., 2019)
• Adaptive nonlinear tube RMPC: (Köhler et al., 2019)
• Robust tube-based MPC with Koopman operators: (Zhang et al., 2021)
• Robust economic tube MPC: (Schwenkel et al., 2019)