Real-Time Model Predictive Control (MPC)

• Real-Time MPC is a control strategy that solves finite-horizon optimization problems in real time to handle constraints, nonlinearity, and uncertainty.
• It employs specialized solvers and algorithmic structures, such as prediction-correction and real-time iteration, to meet strict computational deadlines.
• Applications span robotics, automotive, and manufacturing, achieving sub-millisecond to millisecond solve times on resource-constrained hardware.

Real-Time Model Predictive Control (MPC) is the deployment of finite-horizon optimization-based feedback control strategies with sufficiently low latency to enable their use as the main loop for fast, nonlinear, and/or constrained dynamic systems. Real-time MPC leverages online, receding-horizon solution of optimization problems to regulate the system while explicitly considering dynamic constraints, state-input saturations, and—when computationally feasible—uncertainty and learning-driven model updates. The defining characteristic is that all requisite computations for each control interval (solution of the dynamic optimization, model/estimate updates, and warm start or code generation steps) must fit within a strict deadline compatible with the system’s sampling period or cycle time, often on resource-constrained hardware and for systems exhibiting significant nonlinearities or disturbances. Recent literature demonstrates the state of the art ranges from high-frequency motion control in robotics, process and chemical system regulation, coordinated vehicle and energy system operation, to advanced manufacturing and digital-twin environments (Chen et al., 10 Jan 2025, Paternain et al., 2019, Han et al., 2024, Lee et al., 2022, Kumar et al., 5 Jun 2025, Srikanthan et al., 2024, Schwenkel et al., 2018).

1. Mathematical and Algorithmic Foundations

At its core, real-time MPC solves—at each sampling instant $k$—a finite-horizon optimal control problem of the form

$$\min_{u_{k:k+N-1}} \sum_{i=1}^{N} \ell(x_{k+i}, u_{k+i-1}) + \ell_N(x_{k+N})$$

subject to

$$x_{k+i+1} = f(x_{k+i}, u_{k+i}), \quad x_{k} \text{ given}, \quad u_{min} \le u_{k+i} \le u_{max}, \quad x_{min} \le x_{k+i} \le x_{max}$$

and possibly additional pointwise and/or stagewise constraints over the horizon. Here, $f$ may represent first-principles, data-driven, or hybridized models and the cost structure encodes regulation, tracking, energy, or risk/robustness criteria (Paternain et al., 2019, Lee et al., 2022).

In real-time contexts, the optimization problem is typically transcribed to a mathematical program amenable to fast solution:

• Linear/quadratic programs (LP/QP) for linear or linearly-parameter-varying (LPV) models (Kumar et al., 5 Jun 2025, Zou et al., 2018, Srikanthan et al., 2024).
• Nonlinear programs (NLP) for explicit nonlinear or learned models, optionally with surrogate approximations or multiconvex reformulations (Chen et al., 10 Jan 2025, Salzmann et al., 2022, Masnavi et al., 2021).
• Multiparametric or explicit solutions where feasible for embedded, short-horizon applications (Tian et al., 2020, Jiang et al., 2019).

The receding-horizon, closed-loop nature of MPC is realized by applying the first input $u_{k}^*$ and then shifting the horizon forward—carrying state, warm-started variables, and optimizer information to the next interval.

2. Real-Time Algorithmic Structures

To ensure stringent computational deadlines, real-time MPC deploys several specialized algorithmic structures:

• Efficient Receding-Horizon Solvers: Advanced first- and second-order methods (e.g., projected gradient, SQP, ADMM, L-BFGS, Differential Dynamic Programming) are structured for incremental, warm-started, or inexact solutions that guarantee feasibility, contraction, and often closed-loop stability with a fixed or bounded number of iterations per control interval (Paternain et al., 2019, Chen et al., 10 Jan 2025, Lee et al., 2022, Srikanthan et al., 2024, Fedorová et al., 2023).
• Prediction-Correction and Real-Time Iteration (RTI): Schemes that track the optimizer trajectory as plant parameters drift, using prediction (Newton or parameter update) and correction (few fast steps) to remain within the quadratic convergence region and guarantee stability and bounded suboptimality (Paternain et al., 2019, Fedorová et al., 2023).
• Parallel and Explicit MPC: Offloading computational burden through offline parametric solution maps for submodules or entire problems; parallel evaluation of piecewise-affine policies or state partitioning engines accelerates evaluation for high-dimensional or networked systems (Jiang et al., 2019).
• Split-Multi-convex and Custom Decompositions: Reformulations of complex or nonconvex constraints allow for alternating minimization over simple blocks (multi-convexity) and fast cycling via methods like the split-Bregman approach (Masnavi et al., 2021).
• Learning-augmented and Surrogate-Driven Strategies: Online integration of learned models—deep neural surrogates, Gaussian processes, or memory-based value function approximators—either as primary dynamics, to augment constraints, or to furnish initializer and warm start sequences, while preserving hard real-time properties (Chen et al., 10 Jan 2025, Salzmann et al., 2022, Kobayashi et al., 2022, Schwenkel et al., 2018).

3. Surrogate Modeling and Data Integration

Recent progress in embedding neural surrogates and learning modules into real-time MPC enables constraint-aware control of nonlinear and data-rich systems with fast or complex dynamics:

• One-shot Multi-step Surrogates: Rather than rolling out single-step predictions, models such as TiDE encode entire state trajectories within a prediction horizon as a vector, allowing the inner MPC optimization to operate on the control trajectory block-wise through a smooth, differentiable mapping (Chen et al., 10 Jan 2025).
• Online Model Updating and Digital Twins: Real-time MPC benefits from embedding digital twins that combine process measurements, model updating, and anomaly/quality detection with predictive control (Chen et al., 10 Jan 2025).
• Memory-based and Adaptive Warm-starts: Storage and reuse of optimizer trajectories acquired during operation are leveraged to construct warm-starts that asymptotically converge to the true solution on the limit set of the closed-loop, diminishing suboptimality with operations (Schwenkel et al., 2018).
• Sampling-based Control with Statistical Acceleration: Reverse-KL and mirror descent strategies facilitate mode-seeking, robust sampling for control distribution updates, achieving black-box MPC performance at real-time rates even with highly complex or high-dimensional models (Kobayashi et al., 2022).

4. System Applications and Implementation Domains

Real-time MPC is deployed in diverse domains characterized by algorithmic, system-level, and latency constraints:

• Advanced Manufacturing: TiDE-powered real-time MPC delivers melt pool thermal and geometric regulation in additive manufacturing, tracking temperature while proactively avoiding porosity defects with sub-second solve times, outperforming classical PID schemes (Chen et al., 10 Jan 2025).
• Robotic Manipulation and Motion: Adaptive real-time MPC with fast hierarchical or DDP-based optimizers enables <1 ms closed-loop updates, multi-contact handling, hybrid force/motion tasks, and singularity-avoidance in high-DoF robotic arms and humanoid locomotion (Han et al., 2024, Lee et al., 2022, Kim et al., 26 May 2025).
• Automotive and Transportation: Real-time LPV-MPC schemes deliver <10 ms cycle time for autonomous vehicle tracking across aggressive maneuvers, leveraging horizon-aligned weight adaptation and parallel state estimation/control stacks (Kumar et al., 5 Jun 2025). Traffic networks employ distributed computation via online active set strategies to homotopically track QP solutions, balancing compute load across intervals to guarantee feasibility (Zou et al., 2018).
• Energy and Power Systems: Real-time MPC frameworks integrated with synthetic-data-augmented state estimation and online model identification, e.g., via diffusion models and PEM techniques, achieve robust operational cost savings and predictive accuracy for grid-in-the-loop applications (Xu et al., 13 May 2025).

5. Computational Constraints, Complexity, and Performance

Guaranteeing real-time feasibility requires explicit attention to:

• Bounded Online Complexity: Algorithm design ensures total solve time per interval fits the available time window, by (a) limiting iteration count, (b) parallelizing or exploiting problem structure, and/or (c) precomputing and evaluating explicit solution maps (Fedorová et al., 2023, Srikanthan et al., 2024, Zou et al., 2018, Jiang et al., 2019).
• Guaranteed Stability and Performance Bounds: Theoretical results establish that finite-iteration, suboptimal real-time MPC will yield contractive or Lyapunov-stable closed-loop behavior, with explicit error or suboptimality bounds computable offline as functions of convergence rates and margin parameters (Fedorová et al., 2023, Srikanthan et al., 2024, Jiang et al., 2019).
• Empirical Timing Results: Application demonstrators show that, for properly designed schemes, solve times are:
  • 0.26 s mean (max 0.54 s) per tick for DED melt-pool MPC with TiDE on a Threadripper 32-core platform, with room for explicit/NN-solver-based further speedup (Chen et al., 10 Jan 2025).
  • 0.7–0.9 ms per QP for manipulator MPC (1 kHz) with horizon size 10–15 (Lee et al., 2022).
  • <10 ms per cycle for LPV-MPC in automotive tracking (20 Hz at 50 km/h) (Kumar et al., 5 Jun 2025).
  • 0.3 ms (MLP warm start), 9 ms (DDP whole-body MPC) per cycle in humanoid walking tasks with over 4000 variables/constraints per horizon (Kim et al., 26 May 2025).
  • Sub-millisecond worst-case for explicit MPC, as in Li-ion charging on embedded targets (Tian et al., 2020).
• Hardware Compatibility: Real-time MPC is realized on embedded CPUs (Raspberry Pi, custom ARM, Jetson NX), GPUs (NVIDIA RTX/Orin), FPGAs, and resource-constrained microcontrollers, modularly exploiting warm start, parallel, and structure-exploiting code generation (Lee et al., 2022, Salzmann et al., 2022, Chen et al., 10 Jan 2025, Jiang et al., 2019).

6. Robustness, Learning, and Advanced Capabilities

• Stability/Robustness with Uncertainty: Incorporating Bayesian learning and risk-averse optimization, real-time MPC can provide rigorous closed-loop guarantees even under online parameter learning and model uncertainty, by constructing credible-interval-based ambiguity sets and risk-averse Lyapunov arguments (Li et al., 26 Nov 2025).
• Self-reflective and Learning-based Designs: By integrating experiment-design-based terms into the cost, self-reflective MPC schemes purposefully excite the plant to improve state and parameter estimates, improving long-term performance by reducing estimation uncertainty at small computational overhead (Feng et al., 2016).
• Hierarchical and Task-Priority Control: Hierarchical linearizations and null-space techniques in real-time MPC frameworks allow simultaneous enforcement of strict task-priority structures, singularity avoidance, and constraint satisfaction with fast receding-horizon corrections (Lee et al., 2022, Han et al., 2024).
• Explicit Safety and Constraint Satisfaction: Real-time variants with structure-preserving constraint handling, projection-based steps, and explicit region partitioning guarantee recursive feasibility and input/state constraints at all times (Jiang et al., 2019, Srikanthan et al., 2024, Tian et al., 2020).

---

In summary, real-time MPC is realized through a complex interplay of problem formulation, surrogate and learning model integration, algorithmic specialization (prediction-correction, explicit/parallelization, suboptimal iteration bounds), and system-level engineering. The current state of the art reflects rigorous theoretical analysis, empirical timing validation, and application demonstrations across multiple industrial and scientific domains (Chen et al., 10 Jan 2025, Paternain et al., 2019, Lee et al., 2022, Kumar et al., 5 Jun 2025, Han et al., 2024, Salzmann et al., 2022, Srikanthan et al., 2024, Schwenkel et al., 2018).