Computational Control Overview

• Computational control is a multidisciplinary framework that integrates controller design, numerical tractability, and formal computation within a unified control paradigm.
• It addresses real-time optimization challenges, scalable numerical reconstruction in PDEs, and uncertainty in stochastic as well as biological control processes.
• The approach extends to robotics, systems biology, and digital infrastructures, emphasizing trade-offs between computational precision and performance.

Computational control denotes a family of control-theoretic practices in which the central problem is inseparable from computation itself: the online solution of optimization problems, the numerical reconstruction of controls, the certified elimination of algebraic conditions, the simulation of controlled multiphysics systems, or the explicit treatment of finite-precision and inference costs. Recent work uses the term for cloud-based nonlinear model predictive control using a parallel multi-block ADMM scheme (Ma et al., 2022), for distribution-level control of stochastic gene regulatory networks via a Partial Integro-Differential Equation approximation of the Chemical Master Equation (Faquir et al., 2024), for PINN-based reconstruction of controls in nonlinear PDEs (Kurbanov et al., 24 Apr 2026), and for formal analyses of when a controller counts as a computing system under Abstraction/Representation Theory (Horsman et al., 17 Apr 2026). This suggests that “computational control” is not a single method but an umbrella under which controller design, numerical tractability, implementation fidelity, and formal notions of computation are jointly studied.

1. Real-time optimization as a control bottleneck

A recurrent theme is that control performance is often limited less by modeling than by online computation. In nonlinear model predictive control, heavy computational load for solving nonconvex problems for large-scale systems or systems with real-time demands at each sample step has been identified as one of the factors preventing wider application, motivating a cloud-based computational MPC scheme that uses a parallel multi-block ADMM algorithm tailored to a nonconvex problem with nonlinear constraints (Ma et al., 2022).

The same issue appears in robot control formulated as a sequence of quadratic programs. In whole-body humanoid control, the QP variable can be written as

$$x = \begin{bmatrix} \ddot{q} & f_c \end{bmatrix}^\top,$$

with objectives for left hand motion, center-of-mass position, body posture, and contact-force regularization, subject to equations of motion, feet contact constraints, friction cone constraints, and joint acceleration and torque limits (Chefchaouni et al., 2024). The distinctive contribution of that study is not a new controller architecture but an explicit analysis of how much solver accuracy is actually needed. The paper reports no visible effect on motion accuracy for solution perturbations as large as $10^{-2}$, and desired motion accuracy still being satisfied even for values as large as $10^0$ (Chefchaouni et al., 2024). It also reports that updating costly QP matrices only every $r=20$ control steps, corresponding to matrix updates at $10\,\text{Hz}$, causes no visible degradation in hand tracking while reducing computational effort by almost $15\times$, and that reducing the control frequency to $50\,\text{Hz}$ lowers total computational effort from $62.5\,\text{ms}$ per second of motion to $2.8\,\text{ms}$ per second of motion (Chefchaouni et al., 2024).

Safety-critical quadrotor control exhibits the same trade-off in a more explicitly MPC-oriented form. A 12-state 6-DoF nonlinear rigid-body model is rewritten in State-Dependent Coefficient form,

$$\dot{\mathbf{x}}=\mathbf{A}(\mathbf{x})\mathbf{x}+\mathbf{B}(\mathbf{x})\mathbf{u},$$

so that the online nonlinear program is replaced by a convex QP with frozen coefficients, while retaining terminal ingredients from SDRE/SDLE constructions and explicit obstacle-avoidance constraints (Omidi, 29 Sep 2025). In the reported simulations, average CPU time per control step decreases from $10^{-2}$0 for NMPC to $10^{-2}$1 for nominal SDC-MPC and $10^{-2}$2 for robust SDC-MPC, with comparable tracking error and total cost (Omidi, 29 Sep 2025). A plausible implication is that, in many real-time settings, the decisive design choice is not whether optimization is used, but how much nonlinear structure can be retained while keeping the online problem in a numerically predictable class.

2. Numerical reconstruction and scalable computation

Another major meaning of computational control concerns control problems that are analytically well posed but numerically difficult. A recent example is the reconstruction of unknown controls for nonlinear PDEs from partial observations and endpoint or boundary information. For semilinear heat and wave equations with either internal control $10^{-2}$3 or bilinear control $10^{-2}$4, a PINN-based framework called WeightedPINN introduces trainable weight networks that multiply PDE residuals and boundary or terminal residuals inside a min–max loss (Kurbanov et al., 24 Apr 2026). The method is analyzed under globally Lipschitz nonlinearities and activation regularity assumptions, and the paper proves both approximation-type bounds of the form

$10^{-2}$5

for the weighted loss and stability results stating that small loss implies small reconstruction error in $10^{-2}$6 and $10^{-2}$7 (Kurbanov et al., 24 Apr 2026). Numerical experiments in dimension $10^{-2}$8, with ADAM, 10,000 training iterations, and 1,000 interior or boundary training points per iteration, show that WeightedPINN consistently achieves lower final loss than standard PINN, especially in equation error (Kurbanov et al., 24 Apr 2026).

High-dimensional stochastic control under uncertainty raises a different numerical difficulty: the state dimension is effectively multiplied by the ensemble size. A scalable algorithm based on direct single- or multi-shooting addresses this by vectorized propagation of all samples, exact gradients through computational graphs, and adaptable memory management via common sub-expression elimination (Lambrianides et al., 2019). The discrete dynamics are enforced with sample-wise time stepping, while multi-shooting continuity is imposed through a trajectory-label-agnostic ensemble constraint,

$10^{-2}$9

rather than a separate equality for every sample (Lambrianides et al., 2019). The framework is demonstrated on uncertain UGV and UAV path planning and on a nonlinear advection-reaction-diffusion PDE, with the control remaining deterministic while uncertainty enters through random initial conditions and parameters (Lambrianides et al., 2019).

Singular stochastic control for reflected Brownian motion provides another example of a computationally reformulated problem. There the original infinite-horizon singular control problem on $10^0$0 is approximated by a bounded drift-control problem and then solved through a neural-network approximation of the value function and its gradient using a stochastic identity along simulated trajectories of a reference process (Ata et al., 2023). The paper states that the resulting method is feasible for dimensions up to $10^0$1 or more, reports objective error below $10^0$2 in a one-dimensional benchmark with $10^0$3, and presents a 30-dimensional example computed in about 24 hours on a 20-CPU machine (Ata et al., 2023). In this line of work, computational control refers less to the controller’s architecture than to a tractable surrogate of a mathematically harder control class.

3. Stochastic, biological, and embodied regulation

In systems biology and neuroscience, computational control frequently means controlling distributions, attractors, or embodied sensorimotor loops rather than deterministic trajectories alone. For stochastic gene regulatory networks, one framework replaces the Chemical Master Equation by a PIDE for the probability density $10^0$4 of protein copy numbers,

$10^0$5

and then formulates optimal control and MPC directly at the level of the PDF (Faquir et al., 2024). The objective is terminal distribution matching,

$10^0$6

with gradients computed by an adjoint equation and optimization carried out by nonlinear conjugate gradient with Hager–Zhang updates and projected step constraints (Faquir et al., 2024). The paper emphasizes the ability to shape bimodal populations and to track moving target distributions under intrinsic molecular noise, with reported reachability-test times of about 3.3 seconds in the main text examples and MPC runs of approximately 100 seconds for a $10^0$7 min tracking task (Faquir et al., 2024).

A complementary biological use of computational control appears in Boolean molecular network models. There, node and edge interventions are encoded algebraically by introducing binary control variables into polynomial dynamical systems over $10^0$8, and Gröbner bases are used to solve for controllers that force desired steady states, block transitions, or eliminate undesirable attractors (Murrugarra et al., 2015). The method identifies both single and combinatorial interventions, including in a 16-node p53-mdm2 DNA damage response network where 945 control solutions are reported for enforcing a cell-death fixed point, and in a reduced 16-node T-LGL leukemia survival signaling network where combinations such as Fas ON + FLIP OFF and sFas OFF + FLIP OFF are returned as controls that guarantee disease states are not attractors (Murrugarra et al., 2015). In this setting, computational control is an algebraic search over intervention structure rather than a time-discretized feedback law.

Embodied sensorimotor control broadens the term further. A recent review organizes the neural control of movement into four interacting frameworks: distributed anatomical loops, neural population dynamics on low-dimensional manifolds, optimal feedback control, and embodied musculoskeletal simulation (Almani et al., 17 Sep 2025). It treats neural activity as a dynamical system,

$10^0$9

links behavior to value functions and feedback policies from optimal control, and highlights explicit body models in OpenSim and MuJoCo together with deep reinforcement learning policies $r=20$0 as a route to controllers that are both behaviorally effective and neurally interpretable (Almani et al., 17 Sep 2025). This suggests that, in neuroscience, computational control names an integrative explanatory program spanning circuit, population, optimization, and biomechanics rather than a single algorithmic template.

4. Control when computation is part of the cost

A distinct strand of work treats computation itself as a limited resource that must be modeled inside the control problem. In one formulation of LQG control, the objective includes not only state and action penalties but also an internal cost on the relative precision of posterior beliefs, written through mutual-information terms,

$r=20$1

The paper argues that agents may rationally sacrifice task performance if doing so saves enough bits during inference, and reports phase transitions between optimal, fully reactive, fully predictive, and custom-fit suboptimal inference regimes, together with a family of equivalent suboptimal policies related by orthogonal rotations (Olivos-Castillo et al., 2024). Its summary phrase is that the agent “moves more to think less” (Olivos-Castillo et al., 2024).

A more classical control-theoretic treatment of computation as a first-class issue appears in constructive control theory. That framework is motivated by computational uncertainty: the discrepancy between an ideal mathematical controller or certificate and what can actually be computed at finite precision (Osinenko, 4 Jan 2025). The paper argues that this uncertainty should not simply be absorbed into plant noise and gives constructive versions of approximate optimal control, stability results, measurable selector theorems, approximate eigenvalue problems, Carathéodory trajectories, and Danskin’s theorem (Osinenko, 4 Jan 2025). Its approximate-optimality statement replaces existence of an exact minimizer by

$r=20$2

and its treatment of sample-and-hold stabilization emphasizes that inexact optimization can distort Lyapunov-based guarantees (Osinenko, 4 Jan 2025).

These works share a common correction to a frequent misconception: robust control against process and measurement disturbances does not automatically account for the cost or inexactness of the computation that implements the controller. One line models inference cost directly inside the objective; the other reformulates analysis so that every existence claim carries finite computational content. The two approaches differ in method, but both make computation endogenous to control rather than an external engineering afterthought.

5. Formal theories of controlled computation

Some recent literature addresses a more foundational question: when, exactly, does a control system compute? Under Abstraction/Representation Theory, a control loop contains the ingredients required for physical computation: a representation relation, an abstract evolution, a physical evolution, and an $r=20$3-commuting compute cycle (Horsman et al., 17 Apr 2026). The paper’s central identification is that the plant functions as the representational entity, or more precisely as a proxy for the engineer or user, while the controller is the physical computer. The computed quantity is typically the discrepancy between desired and actual plant state,

$r=20$4

with feedback computation of a corrective signal $r=20$5 (Horsman et al., 17 Apr 2026). The analysis is carried out for a digital thermostat, an electro-mechanical thermostat, the purely mechanical centrifugal governor, and an open-loop human-controlled heating system, and concludes that all of them perform some degree of computation (Horsman et al., 17 Apr 2026).

This result is used to address a longstanding controversy in cognitive theory: the centrifugal governor has often served as a supposed counterexample to computationalism, but the ART analysis argues that it does compute, even if the whole control system is not itself a computer in the same sense as a digital machine (Horsman et al., 17 Apr 2026). The paper’s distinction between the controller subcycle and the overall plant-controller loop is therefore central.

A different formalization appears in category-theoretic work on controlled circuits. Starting from a controllable prop with a distinguished involution $r=20$6, control is freely adjoined through operators $r=20$7 and $r=20$8 satisfying seven equations, including composition, identity, strength, colour change, complementarity, commutativity, and a swap coherence law (Heunen et al., 6 Oct 2025). The paper proves that this syntactic construction corresponds semantically to taking the free rig category on the base prop, and shows that freely adding control to NOT generates reversible Boolean circuits while the same algebraic backbone extends to controlled quantum circuits (Heunen et al., 6 Oct 2025). Here computational control is the algebraic structure of conditional execution itself.

Taken together, these analyses move the term beyond numerical algorithms. In one case, computational control is a representational-feedback structure instantiated in physical devices; in the other, it is the rig-theoretic structure underlying controlled gates. Both perspectives are formal rather than implementation-specific.

6. Toolchains, digital twins, and computational infrastructures

A large part of current computational control is infrastructural: frameworks that let researchers build, combine, certify, or simulate control pipelines before deployment. In robust abstraction-based synthesis, a relational-interface calculus recasts the traditional pipeline of abstraction, game solving, and refinement as a compositional architecture of interface operators such as composition, input or output hiding, coarsening, and shared refinement (Kim et al., 2019). For a Dubins vehicle, decomposing the predecessor computation reduces reported reach-game runtime from 103.09 s in the monolithic formulation to 4.42 s in the fully decomposed one (Kim et al., 2019). The significance lies less in a new control objective than in a programmable synthesis pipeline whose monotonicity properties preserve correctness under rearrangement of abstraction steps.

Certified symbolic-numeric tooling plays a comparable role in algebraic control problems. PACE.jl combines discriminant varieties, Rational Univariate Representation, multi-precision interval arithmetic, interval Newton refinement, Sturm-Habicht sequences, and partial CAD to address identification, stability analysis, and $r=20$9-norm computation (Demin et al., 14 May 2025). The paper demonstrates parameter recovery in ODE models, parameter-space decomposition for multidimensional system stability, and certified enclosures of $10\,\text{Hz}$0-norms such as $10\,\text{Hz}$1 at higher precision (Demin et al., 14 May 2025). In this usage, computational control is supported by certified elimination and root-isolation machinery.

Large-scale experimental systems motivate still richer infrastructures. The FreeGSNKE Pulse Design Tool couples an evolutive free-boundary equilibrium solver for tokamaks with a modular virtual Plasma Control System that combines feedback and feedforward control of plasma current, position, and shape while enforcing coil current and voltage limits (Pentland et al., 30 Mar 2026). The framework is validated on MAST Upgrade flat-top discharges, where reported relative mean absolute differences are below 0.5% for plasma current and below 1% for vertical position in shot 52570, and piecewise-linear modes reduce simulation time from roughly 75–81 minutes in the fully nonlinear mode to about 2–6 minutes (Pentland et al., 30 Mar 2026). This is computational control in the form of a machine-agnostic, in silico environment for scenario design, controller tuning, and uncertainty quantification.

A related but smaller-scale design-and-control infrastructure appears in untethered modular soft robotics. There the robot is represented as a skeletal graph whose edges are actuated by infrared-responsive LCE rods, and computational tools are used either to optimize a single flat-fabricated design for multiple target deformations or to co-optimize skeletal geometry and periodic gait parameters

$10\,\text{Hz}$2

for locomotion in MuJoCo (Bhargava et al., 7 Aug 2025). Fabrication-aware constraints include a $10\,\text{Hz}$3 maximum contraction and a concurrent-heating budget corresponding to a total shrink budget of 300 mm (Bhargava et al., 7 Aug 2025). The result is a design-and-actuation workflow in which morphology and control are computed together.

Across these toolchains, a common pattern emerges. Computational control increasingly denotes not only the synthesis of a controller, but also the construction of reusable computational environments in which models, algorithms, constraints, and validation procedures can be combined before the physical closed loop is ever run.