Approximate Model-Predictive Control (AMPC)

• Approximate Model-Predictive Control (AMPC) is a family of receding-horizon methods that replace expensive online optimization components with tractable approximations.
• AMPC techniques include value-function approximations, implicit policy surrogates, and robust safety filtering to balance efficiency and constraint satisfaction across applications.
• Recent research emphasizes integrating learning-based surrogates with robust design to preserve MPC guarantees while reducing computational complexity.

Searching arXiv for recent and foundational papers on Approximate Model-Predictive Control to ground the article. Search results indicate multiple distinct usages of “AMPC” across control: adaptive MPC, approximate MPC via value-function approximation, imitation learning of MPC policies, Gaussian-process / neural surrogates, and stochastic / safety-augmented variants. I’ll synthesize only claims explicitly supported by the provided paper data. Approximate Model-Predictive Control (AMPC) denotes a family of receding-horizon control methods that preserve the structural logic of Model Predictive Control while replacing one or more expensive ingredients of the exact online controller by tractable approximations. Depending on the paper, the approximated object may be a value function, a dual-information propagation step, a terminal safe set, the implicit MPC feedback map, or even the nonlinear program itself. Recent arXiv literature places under this label polynomial Hamilton–Jacobi–Bellman approximations solved by sum-of-squares programming, scenario-tree rollout approximations for dual stochastic MPC, Gaussian-process and neural surrogates of implicit MPC laws, and safety-augmented learned controllers with feasibility checks and fallback policies (Jones et al., 2024, Arcari et al., 2019, Rose et al., 13 May 2026, Clement et al., 25 Mar 2026). The acronym is also used non-uniformly: in the V-formation literature, “AMPC” explicitly refers to “Adaptive MPC,” where the receding horizon, particle swarm size, and neighborhood size are varied online (Grosu et al., 2020).

1. Terminology, scope, and problem setting

The common problem addressed by AMPC is the online complexity of MPC. Standard MPC repeatedly solves a finite-horizon optimal control problem under dynamics and constraints, often nonlinear, hybrid, stochastic, or mixed-integer. AMPC methods reduce this burden by moving computation offline, by replacing difficult subproblems with local approximations, or by embedding the controller inside a filtered or adaptive architecture that is cheaper to execute online. This broad interpretation is visible across nonlinear optimal control, stochastic dual control, autonomous driving, robotics, energy systems, and explicit MPC approximation (Hose et al., 2024, Liu et al., 22 Oct 2025, Trinh et al., 2016, Pillitteri et al., 4 Jun 2026).

A central distinction is between methods that still solve a reduced online optimization and methods that output control directly. The former include one-step or reduced-order QPs, Taylor-polynomialized HJB approximations, and rollout-style stochastic programs. The latter include neural or Gaussian-process surrogates of the implicit MPC law, piecewise nonlinear explicit controllers, and imitation-learned microgrid dispatch policies. A second distinction is between approximation of the control law itself and approximation of auxiliary MPC ingredients such as value functions, terminal costs, or control-invariant sets. This distinction matters because guarantees typically attach to the auxiliary object or to a safety wrapper, not to the approximator in isolation (Joa et al., 2024, Lunardi et al., 2023, Hertneck et al., 2018).

A recurrent misconception is that AMPC is a single method. The literature instead uses the term for multiple methodological families united by receding-horizon deployment and by approximation of computationally dominant MPC components. A plausible implication is that comparisons between AMPC papers must be made at the level of approximation target, constraint architecture, and guarantee mechanism rather than by acronym alone.

2. Main approximation targets and mathematical constructions

A useful organizing principle is the object being approximated. Some methods approximate a value function and synthesize control from it; others approximate the optimizer of the MPC problem; others approximate feasibility certificates such as safe sets.

Approximation target	Representative construction	Papers
Value function / cost-to-go	Taylor polynomialization, SOS HJB subsolutions, LP-based terminal cost updates, switched quadratic approximations	(Jones et al., 2024, Joa et al., 2024, Chacko et al., 2023)
Dual / stochastic lookahead	Scenario-tree rollout with Bayesian posterior updates and truncated dual horizon	(Arcari et al., 2019)
Implicit MPC law	Piecewise nonlinear explicit maps, GP regressors, MLP/RNN imitators, residual-corrected QP parameters	(Trinh et al., 2016, Rose et al., 13 May 2026, Clement et al., 25 Mar 2026, Pillitteri et al., 4 Jun 2026)
Safety or terminal objects	Learned approximate control-invariant sets, robust safe candidate sequences, terminal fallback controllers	(Lunardi et al., 2023, Clement et al., 25 Mar 2026)

In the value-function family, a prominent construction polynomializes non-polynomial dynamics and costs by Taylor expansion around the current state–time point, solves a relaxed HJB inequality with SOS programming, and obtains a polynomial sub-value function $P_d(x,t)\le V(x,t)$. For input-affine costs and dynamics over $U=[-1,1]^m$, this yields a bang-bang law of the form

$$u_i^{\mathrm{AMPC}}(x,t)= -\operatorname{sign}\!\big(c_i(x,t)+\nabla_x \hat V(x,t)^\top f_i(x)\big),$$

where $\hat V=P_d$ is the approximate value function computed from the local polynomialized problem (Jones et al., 2024). In a different direction, single-step MPC for stochastic infinite-horizon control uses a sampled approximation of $\mathbb{E}[Q(x^+)]$ and a convex LP-defined value function

$$Q^j(x)=\min_{\lambda^j}\ \mathbf{J}^{j\top}\lambda^j \quad\text{s.t.}\quad \mathbf{X}^j\lambda^j=x,\ \lambda^j\ge 0,\ \mathbf{1}^\top\lambda^j=1,$$

thereby turning terminal cost construction into a data-driven safe-set interpolation problem (Joa et al., 2024). Another approximate-dynamic-programming variant represents the value function as the pointwise minimum of quadratic forms $\bar x^\top \bar P_i \bar x$ obtained from switched affine approximations and Riccati recursions, then uses that approximation in a one-step or two-stage online search over quantized inputs (Chacko et al., 2023).

Scenario-based dual stochastic AMPC occupies a different niche. There the approximation is not a static surrogate of the control law but a rollout truncation of exact stochastic dynamic programming. A short “dual” phase is represented by a scenario tree with Bayesian posterior updates of unknown parameters, whereas the remaining horizon is treated as an exploitation phase with open-loop optimization per scenario leaf. The resulting controller is still receding-horizon, but the approximation lies in the replacement of the exact nested $\min$-$\mathbb{E}$ dynamic program by one nonlinear optimization defined over the sampled tree (Arcari et al., 2019).

This taxonomy suggests that “approximate” in AMPC can mean approximation in dynamics, in cost-to-go, in information state propagation, in explicit policy representation, or in invariant-set geometry. The practical consequences differ sharply across these cases.

3. Learned surrogates of the implicit MPC law

A large branch of AMPC replaces the online optimizer by a learned map from the current information state to the first control move, or to the full horizon sequence. The most direct formulation learns $\pi_{\mathrm{MPC}}$ by supervised imitation. In the statistical-guarantee framework for nonlinear systems, the learned controller $U=[-1,1]^m$0 is required to satisfy

$U=[-1,1]^m$1

along validated closed-loop trajectories, so that the approximation error can be interpreted as an admissible input disturbance for a robust MPC design (Hertneck et al., 2018). In neural parameter-adaptive AMPC, two networks are trained: one for the nominal MPC law and one for its parametric sensitivity. The deployed controller is then

$U=[-1,1]^m$2

which enables online tuning across plant or cost parameters without regenerating the training set (Hose et al., 2024).

More recent work pushes this surrogate viewpoint in several directions. Sequential-AMPC uses a recurrent horizon policy rather than a horizon-wide MLP. One RNN cell with shared parameters generates the sequence $U=[-1,1]^m$3, so the parameter count is independent of horizon length $U=[-1,1]^m$4, unlike a feedforward architecture whose last layer scales with $U=[-1,1]^m$5 (Clement et al., 25 Mar 2026). Gaussian-process AMPC for autonomous vehicles exploits a curvilinear-coordinate reformulation and a nominal feedforward term

$U=[-1,1]^m$6

then learns only the residual control as a function of $U=[-1,1]^m$7. The online law is

$U=[-1,1]^m$8

with $U=[-1,1]^m$9 given by GP posterior means (Rose et al., 13 May 2026). In imitation learning for mixed-integer microgrid EMPC, the MLP outputs only the continuous control vector $$u_i^{\mathrm{AMPC}}(x,t)= -\operatorname{sign}\!\big(c_i(x,t)+\nabla_x \hat V(x,t)^\top f_i(x)\big),$$0, after which a projection onto a hyperrectangular admissible set enforces the input bounds (Liu et al., 22 Oct 2025).

A separate line keeps an online optimization layer but amortizes the nonlinear MPC into a learned state-dependent QP. For input-affine systems, a state-dependent analytic baseline yields initial QP cost parameters, and a neural network predicts residual corrections to the Cholesky factor and linear term of the QP cost. The first control move is then obtained by a differentiable interior-point QP layer, so hard linear constraints remain explicit in the online computation (Pillitteri et al., 4 Jun 2026). This architecture differs from direct policy regression because the network does not output $$u_i^{\mathrm{AMPC}}(x,t)= -\operatorname{sign}\!\big(c_i(x,t)+\nabla_x \hat V(x,t)^\top f_i(x)\big),$$1 itself; it outputs cost corrections for an optimization problem that is still solved online.

These learned variants share an offline/online split, but they differ substantially in how much MPC structure is preserved: direct imitators preserve only the input-output map, residual-QP and GP-feedforward methods preserve analytic structure, and sensitivity-based methods preserve local parametric information about the original nonlinear program.

4. Robustness, feasibility, and certification mechanisms

AMPC does not automatically inherit the guarantees of the original MPC. The guarantee mechanisms reported in the literature are therefore architectural rather than incidental. One approach is robust-design plus statistical validation. A robust MPC is first designed to tolerate bounded input disturbances $$u_i^{\mathrm{AMPC}}(x,t)= -\operatorname{sign}\!\big(c_i(x,t)+\nabla_x \hat V(x,t)^\top f_i(x)\big),$$2 with $$u_i^{\mathrm{AMPC}}(x,t)= -\operatorname{sign}\!\big(c_i(x,t)+\nabla_x \hat V(x,t)^\top f_i(x)\big),$$3; the learned approximation error is then treated as such a disturbance. Hoeffding’s inequality is used to validate, from sampled trajectories, that the approximation remains within the admissible error bound with high confidence, yielding a closed-loop statistical guarantee on stability and constraint satisfaction for a fraction of initial conditions (Hertneck et al., 2018).

A second approach is local parametric robustness. Parameter-adaptive AMPC assumes an input-robust MPC and uses local solution sensitivities of the underlying nonlinear program. Under a bounded nominal approximation error and bounded sensitivity approximation error, a theorem identifies a neighborhood $$u_i^{\mathrm{AMPC}}(x,t)= -\operatorname{sign}\!\big(c_i(x,t)+\nabla_x \hat V(x,t)^\top f_i(x)\big),$$4 in parameter space within which the adaptive AMPC remains stable (Hose et al., 2024). The scope of the guarantee is explicitly local and relies on no active-set change in the parametric NLP. This directly addresses a common misconception: sensitivity-based AMPC is not a global replacement of re-training, but a first-order local predictor around $$u_i^{\mathrm{AMPC}}(x,t)= -\operatorname{sign}\!\big(c_i(x,t)+\nabla_x \hat V(x,t)^\top f_i(x)\big),$$5.

A third approach is explicit safety filtering. Safe Sequential-AMPC wraps a learned horizon policy inside a feasibility and cost-improvement gate. At each time, the neural proposal $$u_i^{\mathrm{AMPC}}(x,t)= -\operatorname{sign}\!\big(c_i(x,t)+\nabla_x \hat V(x,t)^\top f_i(x)\big),$$6 is accepted only if it belongs to the feasible sequence set $$u_i^{\mathrm{AMPC}}(x,t)= -\operatorname{sign}\!\big(c_i(x,t)+\nabla_x \hat V(x,t)^\top f_i(x)\big),$$7 and if its MPC cost does not exceed that of a stored safe candidate $$u_i^{\mathrm{AMPC}}(x,t)= -\operatorname{sign}\!\big(c_i(x,t)+\nabla_x \hat V(x,t)^\top f_i(x)\big),$$8; otherwise the controller reuses the shifted safe candidate, appended with a terminal controller (Clement et al., 25 Mar 2026). The guarantee comes from the robust NMPC and the wrapper, not from the RNN. In the same spirit, receding-constraint MPC with a learned approximate control-invariant set $$u_i^{\mathrm{AMPC}}(x,t)= -\operatorname{sign}\!\big(c_i(x,t)+\nabla_x \hat V(x,t)^\top f_i(x)\big),$$9 guarantees recursive feasibility under an $\hat V=P_d$0-step control-invariance assumption that is weaker than exact invariance, and guarantees safety under even weaker assumptions by triggering a safe task-abortion strategy when risk is detected (Lunardi et al., 2023).

The single-step stochastic infinite-horizon formulation gives another guarantee pattern. There, the approximate value function $\hat V=P_d$1 and safe set $\hat V=P_d$2 are updated episode by episode, and the resulting MPC with horizon $\hat V=P_d$3 is shown to satisfy recursive feasibility, robust constraint satisfaction, and convergence in probability to the target set, while the estimated value function converges locally to the optimal value function as the disturbance discretization is refined (Joa et al., 2024).

Across these papers, the technical lesson is consistent: guarantees survive approximation only when the approximation is explicitly bounded, filtered, embedded in a robust outer design, or attached to an auxiliary object with its own monotonicity or invariance properties.

5. Applications and reported empirical performance

AMPC has been applied to qualitatively different control regimes. In coordinated vehicle control for distributed-drive electric vehicles, an adaptive path-tracking controller updates the prediction horizon $\hat V=P_d$4, the lateral-error weight $\hat V=P_d$5, and the steering-rate weight $\hat V=P_d$6 as functions of longitudinal speed. When coordinated with direct yaw moment control, the reported improvements were $\hat V=P_d$7 in yaw stability and $\hat V=P_d$8 in path-tracking accuracy under variable speed and low adhesion, and $\hat V=P_d$9 and $\mathbb{E}[Q(x^+)]$0, respectively, under high speed and low adhesion relative to the stated baselines (Wu et al., 2023).

For embedded autonomous vehicle trajectory tracking, GP-based AMPC in curvilinear coordinates was deployed on a Raspberry Pi and compared with an ACADOS real-time-iteration MPC. The GP approximation computed control inputs about five times faster while achieving similar closed-loop tracking performance; the average closed-loop cost was $\mathbb{E}[Q(x^+)]$1 for the GP-based controller versus $\mathbb{E}[Q(x^+)]$2 for MPC, and the path deviation remained below $\mathbb{E}[Q(x^+)]$3 for both on the circular track (Rose et al., 13 May 2026). In nonlinear robotic-arm tracking, amortized NMPC based on a learned state-dependent QP achieved orders-of-magnitude speedup over IPOPT while maintaining comparable tracking performance, with reported mean per-step times of $\mathbb{E}[Q(x^+)]$4 ms for the amortized controller, $\mathbb{E}[Q(x^+)]$5 ms for RTI, and $\mathbb{E}[Q(x^+)]$6 ms for IPOPT (Pillitteri et al., 4 Jun 2026).

In safety-critical learning-based NMPC, Sequential-AMPC improved open-loop feasibility and closed-loop safety relative to a feedforward AMPC baseline. On the quadcopter benchmark, open-loop feasibility increased from $\mathbb{E}[Q(x^+)]$7 to $\mathbb{E}[Q(x^+)]$8, and closed-loop safety under the wrapper increased from $\mathbb{E}[Q(x^+)]$9 to $$Q^j(x)=\min_{\lambda^j}\ \mathbf{J}^{j\top}\lambda^j \quad\text{s.t.}\quad \mathbf{X}^j\lambda^j=x,\ \lambda^j\ge 0,\ \mathbf{1}^\top\lambda^j=1,$$0; the same study reported $$Q^j(x)=\min_{\lambda^j}\ \mathbf{J}^{j\top}\lambda^j \quad\text{s.t.}\quad \mathbf{X}^j\lambda^j=x,\ \lambda^j\ge 0,\ \mathbf{1}^\top\lambda^j=1,$$1 fewer training epochs on that benchmark for the sequential architecture (Clement et al., 25 Mar 2026). In hardware tuning without retraining, Bayesian optimization over the parameterized AMPC family improved performance of neural approximate MPC on an inverted cartpole and an under-actuated balancing unicycle within about 20 experiments, using the same pre-trained networks and tuning only the parameter vector $$Q^j(x)=\min_{\lambda^j}\ \mathbf{J}^{j\top}\lambda^j \quad\text{s.t.}\quad \mathbf{X}^j\lambda^j=x,\ \lambda^j\ge 0,\ \mathbf{1}^\top\lambda^j=1,$$2 (Hose et al., 16 Dec 2025).

Process-control and energy-system applications show the same runtime/performance trade-off. For a nonlinear multi-tank system, ADP-based AMPC variants were reported to be approximately $$Q^j(x)=\min_{\lambda^j}\ \mathbf{J}^{j\top}\lambda^j \quad\text{s.t.}\quad \mathbf{X}^j\lambda^j=x,\ \lambda^j\ge 0,\ \mathbf{1}^\top\lambda^j=1,$$3, $$Q^j(x)=\min_{\lambda^j}\ \mathbf{J}^{j\top}\lambda^j \quad\text{s.t.}\quad \mathbf{X}^j\lambda^j=x,\ \lambda^j\ge 0,\ \mathbf{1}^\top\lambda^j=1,$$4, and $$Q^j(x)=\min_{\lambda^j}\ \mathbf{J}^{j\top}\lambda^j \quad\text{s.t.}\quad \mathbf{X}^j\lambda^j=x,\ \lambda^j\ge 0,\ \mathbf{1}^\top\lambda^j=1,$$5 faster than NMPC for three controller variants, with only moderate degradation in integral squared error (Chacko et al., 2023). In a nonlinear CSTR benchmark, a neural approximate MPC with statistical guarantees produced more than $$Q^j(x)=\min_{\lambda^j}\ \mathbf{J}^{j\top}\lambda^j \quad\text{s.t.}\quad \mathbf{X}^j\lambda^j=x,\ \lambda^j\ge 0,\ \mathbf{1}^\top\lambda^j=1,$$6 speedup over online RMPC evaluation in the reported implementation (Hertneck et al., 2018). For microgrid energy management, imitation-learned approximate EMPC achieved economic performance comparable to optimization-based EMPC while requiring only $$Q^j(x)=\min_{\lambda^j}\ \mathbf{J}^{j\top}\lambda^j \quad\text{s.t.}\quad \mathbf{X}^j\lambda^j=x,\ \lambda^j\ge 0,\ \mathbf{1}^\top\lambda^j=1,$$7 of the computation time in practice (Liu et al., 22 Oct 2025).

The application spectrum also includes more specialized uses. In V-formation control, adaptive MPC was combined with particle swarm optimization, adaptive horizon selection, and distributed neighborhood resizing, and the distributed version was reported to perform almost as well as centralized AMPC while using only local information and distributed consensus (Grosu et al., 2020). This usage underscores again that the acronym AMPC spans both approximation-oriented and adaptation-oriented receding-horizon designs.

6. Limitations, ambiguities, and current research directions

The most immediate limitation is terminological ambiguity. The same acronym denotes “Approximate MPC” in many papers and “Adaptive MPC” in others (Grosu et al., 2020). More substantively, the approximation object varies: value-function approximation, policy imitation, safe-set approximation, stochastic rollout truncation, and residual QP correction solve different problems and admit different guarantees. Treating them as interchangeable obscures the role of model structure, uncertainty, and constraint handling.

A second limitation is dimensionality. SOS-based HJB approximations and SDP relaxations scale poorly with state dimension and polynomial degree (Jones et al., 2024). Scenario-tree dual MPC grows exponentially with dual horizon length $$Q^j(x)=\min_{\lambda^j}\ \mathbf{J}^{j\top}\lambda^j \quad\text{s.t.}\quad \mathbf{X}^j\lambda^j=x,\ \lambda^j\ge 0,\ \mathbf{1}^\top\lambda^j=1,$$8 through $$Q^j(x)=\min_{\lambda^j}\ \mathbf{J}^{j\top}\lambda^j \quad\text{s.t.}\quad \mathbf{X}^j\lambda^j=x,\ \lambda^j\ge 0,\ \mathbf{1}^\top\lambda^j=1,$$9 (Arcari et al., 2019). Piecewise explicit approximations trade online speed for an offline growth in region count and storage (Trinh et al., 2016). Even learned surrogates can remain data-intensive: the unicycle fine-tuning study reports 3.5M state–action–sensitivity samples and roughly $\bar x^\top \bar P_i \bar x$0k CPU-core-hours for offline dataset generation (Hose et al., 16 Dec 2025).

A third limitation is that many learned AMPC schemes do not by themselves enforce recursive feasibility or state constraints. The microgrid imitation-learning controller projects only onto admissible input boxes and leaves state-constraint enforcement to future work (Liu et al., 22 Oct 2025). Sequential-AMPC still exhibits high intervention rates on difficult tasks, indicating that good open-loop predictions are not sufficient to satisfy terminal or cost-improvement conditions reliably (Clement et al., 25 Mar 2026). GP-based AMPC for trajectory tracking does not use posterior variance for online trust management or robust tightening (Rose et al., 13 May 2026). These observations suggest that approximate policies often require auxiliary safety layers, fallback controllers, or certified terminal ingredients.

Current directions therefore combine approximation with structure preservation and explainability. Residual-corrector QP architectures retain explicit constraint handling and KKT-aware training (Pillitteri et al., 4 Jun 2026). ExAMPC augments approximate NMPC with low-order spline embeddings, SHAP analyses, and symbolic-regression monitors to expose sensitivities of optimization behavior, continuous-time constraint violations, and execution-time peaks to physical variables and scenario parameters (Allamaa et al., 1 Mar 2025). A plausible implication is that future AMPC systems will be judged less by raw imitation error than by how effectively they preserve the semantics of MPC—feasibility, terminal behavior, dual sensitivity, and diagnosability—under approximation.

In that sense, AMPC is best understood not as a single algorithmic recipe, but as a design space for redistributing the computational load of MPC while attempting to preserve the parts of receding-horizon control that matter most for the application at hand.