System-Level Tube MPC (SLTMPC)

• SLTMPC is a robust control methodology that simultaneously synthesizes nominal trajectories and dynamic tubes using system-level parameterizations.
• It reduces conservativeness and enlarges the feasible region in constrained linear and uncertain systems through online optimization of tube feedback laws.
• SLTMPC leverages affine system responses, disturbance reachable sets, and asynchronous computation to enable scalable, real-time implementation in diverse applications.

System-Level Tube Model Predictive Control (SLTMPC) is a robust control methodology that synthesizes and optimizes both nominal trajectories and dynamic tubes in model predictive control (MPC) using system-level parameterizations. Distinct from conventional tube-MPC, SLTMPC allows concurrent, often online, optimization of the nominal plan and the tube laws themselves, significantly reducing conservativeness and enlarging the feasible region for constrained linear and uncertain systems. The approach leverages affine or block-lower-triangular system responses, disturbance reachable sets, and recent computational enhancements—such as asynchronous decomposition—to achieve scalability and practical real-time implementation in a broad class of applications (Sieber et al., 2021, Sieber et al., 2021, Sieber et al., 2024, Sieber et al., 2022, Chen et al., 2019).

1. System-Level Parameterization and Core Formulation

SLTMPC employs the system level parameterization (SLP), representing the closed-loop map from stacked initial state and sequence of disturbances $\tilde{\boldsymbol\delta}$ to the predicted state-input trajectory. This parameterization replaces classical static feedback gains $K$ or disturbance feedback polices with block-lower-triangular matrices $(\bm{\Phi}_x, \bm{\Phi}_u)$. The defining affine constraint is: $$\left[I - \mathcal{Z}\mathcal{A} \;|\; -\mathcal{Z}\mathcal{B}\right] \begin{bmatrix} \bm{\Phi}_x \ \bm{\Phi}_u \end{bmatrix} = I.$$ Here, $\mathcal{Z}$ is the block-down-shift (or shift) operator over horizon $N$; $\mathcal{A}, \mathcal{B}$ are block-diagonal lifted system matrices (Sieber et al., 2021, Sieber et al., 2021). The block partitions

$$\bm{\Phi}_x = [1\ |\ \bm{\phi}_z \ \bm{\Phi}_e],\;\; \bm{\Phi}_u = [\bm{\phi}_v\ \bm{\Phi}_k]$$

separate the nominal trajectory terms and the error-to-disturbance maps.

The SLTMPC program treats $(\bm{\Phi}_x, \bm{\Phi}_u)$ as decision variables, optimizing both nominal trajectory and feedback response simultaneously in a convex program subject to system dynamics, state/input tightening, finite impulse response (FIR) or Toeplitz structure, and terminal invariance constraints (Sieber et al., 2021, Sieber et al., 2022).

2. Disturbance Reachable Sets and Tube Construction

A central construct is the system-level disturbance reachable set (SL-DRS), defined for each horizon step $i$ as

$K$0

where $K$1 is the admissible disturbance set. Imposing a FIR constraint ensures the entire future error trajectory is contained in a finite, convex family of tube cross-sections, which are then used for robust constraint tightening via Pontryagin difference: $K$2 for state and $K$3 for input constraints. These sets can be synthesized online or precomputed offline for use in standard tube-MPC (Sieber et al., 2021).

When model uncertainty is present (e.g., polytopic parametric uncertainty), SLTMPC extends tube construction using homothetic or affine disturbance over-approximations in the auxiliary coordinate. The affine representation, combined with the system-level synthesis theorem, ensures inclusion of all uncertainty realizations via linear constraints on the tube and tightening variables (Sieber et al., 2024).

3. SLTMPC Optimization Problem Structure

At each timestep, SLTMPC solves a finite-horizon convex program. The key decision variables are (i) the nominal trajectory sequence $K$4, (ii) the system response (tube feedback) matrices $K$5, and, where applicable, tube inflations or over-approximation variables for parametric uncertainty.

A typical SLTMPC program (Sieber et al., 2021, Sieber et al., 2022, Sieber et al., 2024) is: $K$6 This guarantees that, for all admissible disturbances and model uncertainties, trajectories remain within tightened constraints and the closed-loop is input-to-state stable (Sieber et al., 2021, Sieber et al., 2022, Sieber et al., 2024).

4. Asynchronous Computation and Practical Implementations

Real-time deployment of SLTMPC is enabled by asynchronous computation schemes that decouple tube synthesis and nominal trajectory planning. The computation is split into two parallel processes:

• The secondary (low-frequency) process optimizes (or updates) tube feedback matrices—potentially expensive, e.g., 19 ms per solve for horizon $K$7, or 700 ms for uncertain systems.
• The primary (high-frequency) process uses stored tube data, introducing convex combination weights over multiple tube candidates, and solves a much smaller QP for the nominal plan each sample (e.g., 1.5–16 ms per step) (Sieber et al., 2022, Sieber et al., 2024).

Tube memory slots ensure recursive feasibility even if some tube solutions become temporarily outdated. Robust invariance and constraint satisfaction are proven via candidate policy shifting and invariant set construction, even under asynchronous tube updates (Sieber et al., 2022, Sieber et al., 2024).

5. Extensions for Uncertain and Distributed Systems

SLTMPC generalizes directly to uncertain and distributed architectures:

• Parametric uncertainty: Affine tube over-approximation with block-wise inclusion of all vertices of the uncertainty polytope. The error-tube feedback law is synthesized to guarantee all plant uncertainty realizations are contained (Sieber et al., 2024).
• Time-varying and LTV systems: System-level parameterization allows efficient robust constraint tightening for time-varying models with norm-bounded disturbance and operator uncertainty (Chen et al., 2019).
• Distributed/multi-agent systems: Nested tubes and chain-of-tubes architectures, in which inner and outer SLTMPC controllers operate at different layers (per subsystem and over the network) to reduce conservativeness relative to single-tube methods, achieve performance near centralized MPC while preserving scalability and recursive feasibility (Hernandez et al., 2016).

6. Theoretical Properties and Computational Tradeoffs

Theoretical guarantees proved across the SLTMPC literature include:

• Recursive feasibility: Ensured by tailored invariant terminal sets, explicit Minkowski sum tightenings, and by employing shift-and-perturb arguments using the FIR property or stored tube memory (Sieber et al., 2021, Sieber et al., 2022, Sieber et al., 2024).
• Constraint satisfaction: Follows from Pontryagin difference-based constraint tightening using SL-DRS or affine tube inclusions.
• Input-to-State Stability (ISS): Shown by constructing an ISS Lyapunov function from the optimal value function; for uncertain systems, the closed-loop map is ISS w.r.t. over-approximating additive disturbance (Sieber et al., 2021, Sieber et al., 2024).
• Convexity and scalability: The central optimization remains a quadratic or conic QP with variable count linear in horizon $K$8. Compared to disturbance-feedback MPC ($K$9 variables), SLTMPC achieves near-minimal $(\bm{\Phi}_x, \bm{\Phi}_u)$0 scaling yet avoids the conservativeness of fixed-gain tube-MPC (Sieber et al., 2021).

A summary comparison of computational complexity: | Method | Controller Variables | Main Constraint Structure | |-----------------------------|---------------------|---------------------------| | Disturbance-feedback MPC | $(\bm{\Phi}_x, \bm{\Phi}_u)$1 | Large QP/LP | | SLTMPC | $(\bm{\Phi}_x, \bm{\Phi}_u)$2 | Affine SLP, QP/conic QP | | Tube-MPC, fixed $(\bm{\Phi}_x, \bm{\Phi}_u)$3 | $(\bm{\Phi}_x, \bm{\Phi}_u)$4 | Precomputed gains/Tubes |

7. Numerical Results and Observed Performance

Extensive numerical studies confirm the following:

• SLTMPC reduces average closed-loop cost significantly relative to fixed-gain tube-MPC (e.g., mean cost reduction from $(\bm{\Phi}_x, \bm{\Phi}_u)$530 to $(\bm{\Phi}_x, \bm{\Phi}_u)$626 in a double-integrator scenario), with only a modest increase in computation (e.g., 33 ms per solve vs. 53 ms for full disturbance feedback) (Sieber et al., 2021).
• SLTMPC enlarges the region of attraction, closely approaches disturbance-feedback MPC in feasible set size and cost metrics, and maintains constraint satisfaction over thousands of random trials (Sieber et al., 2021, Sieber et al., 2021).
• Asynchronous variants yield closed-loop performance and feasible region between full SLTMPC and classical tube-MPC, while per-sample computational cost matches that of tube-MPC (Sieber et al., 2022, Sieber et al., 2024).
• In distributed settings (chain-of-tubes), decentralized SLTMPC approaches centralized MPC performance (within $(\bm{\Phi}_x, \bm{\Phi}_u)$71% total cost), while single-tube decentralized approaches are much more conservative (Hernandez et al., 2016).
• In cases with significant parametric uncertainty, the receding-horizon, recursive-feasible SLTMPC formulation achieves lower average cost and superior constraint handling compared to previous methods reliant on invariant sets for the full plant (Sieber et al., 2024).

• (Sieber et al., 2021) Sieber et al., "System Level Disturbance Reachable Sets and their Application to Tube-based MPC", 2021.
• (Sieber et al., 2021) Dean et al., "A System Level Approach to Tube-based Model Predictive Control", 2021.
• (Sieber et al., 2022) Pfister & Sieber, "Asynchronous Computation of Tube-based Model Predictive Control", 2022.
• (Sieber et al., 2024) Sieber & Pfister, "Computationally Efficient System Level Tube-MPC for Uncertain Systems", 2024.
• (Chen et al., 2019) Chen & Tu, "Robust Closed-loop Model Predictive Control via System Level Synthesis", 2019.
• (Hernandez et al., 2016) Hernandez & Trodden, "Distributed Model Predictive Control Using a Chain of Tubes", 2016.