Learning-Augmented Control

• Learning-Augmented Control is a hybrid framework that supplements traditional control methods with data-driven models to enhance performance without compromising safety.
• It involves techniques such as prediction blending, residual model augmentation, and parameter tuning to improve transient and steady-state performance in various domains.
• LAC retains foundational control structures while integrating learning elements certified by methods like Lyapunov functions, competitive ratios, and safe action sets.

Searching arXiv for papers on Learning-Augmented Control and closely related formulations. Learning-Augmented Control (LAC) denotes a family of control architectures in which a conventional control backbone—such as Model Predictive Control (MPC), primal-dual dynamics, robust $H_\infty$ synthesis, impedance control, iterative learning control, or online networked control—is augmented by learned models, learned predictions, learned confidence variables, or learned controller parametrizations, while explicit control-theoretic structure remains central to synthesis and certification. Across formulations, the learned component may estimate unknown dynamics, blend untrusted predictions with nominal models, tune cost and constraint parameters, or reparameterize the control input; the recurring objective is to improve transient or average performance when data are informative without relinquishing safety, feasibility, stability, steady-state optimality, or competitive guarantees (Yu et al., 10 Mar 2026, Li, 19 Jul 2025).

1. Definition and conceptual scope

The term "Learning-Augmented Control" is used explicitly in multiple, partially overlapping senses. In constrained nonlinear MPC, it denotes the integration of untrusted machine learning predictions together with an online confidence parameter that balances ML and nominal predictions and yields "best-of-both-worlds" guarantees (Li, 19 Jul 2025). In power-system frequency regulation, it denotes embedding learning into a primal-dual controller through a change of variables $u=f(s)$, so that transient metrics can be improved in a data-driven manner while preserving asymptotic stability and steady-state optimality (Yu et al., 10 Mar 2026). In water infrastructures, it denotes wrapping a safe control prior with safe action sets so that ML advice can reduce energy and environmental cost while preserving per-round safety constraints (Yang et al., 24 Jan 2025). In mission-aligned autonomy, it denotes a two-level optimization scheme that combines planning, control, and learning, with a lower MPC layer and an upper classical-planning layer integrated with RL-style adaptation (Kungurtsev et al., 6 Jul 2025).

Earlier formulations do not always use the label LAC, but they instantiate closely related ideas. "Dual Control with Active Learning using Gaussian Process Regression" (Alpcan, 2011) explicitly trades off control performance and information acquisition by scoring actions with predicted tracking error and GP uncertainty. "Anticipating the Long-Term Effect of Online Learning in Control" (Capone et al., 2020) designs control parameters by minimizing an anticipated expected cost in which the GP posterior evolves with the trajectory induced by the policy. This suggests that LAC is better understood as a design pattern—learning coupled to control synthesis under explicit structural commitments—rather than as a single algorithmic template.

2. Architectural patterns

Across the literature, LAC systems differ mainly in where the learned object is inserted into the loop. Some methods learn predictions of exogenous parameters and then regulate how much those predictions are trusted; some learn residual dynamics and inject them into a nominal model; some learn controller parameters inside a fixed control architecture; and some learn a structured change of variables that preserves the optimizer of an underlying control problem.

A recurring pattern is prediction blending inside MPC. In (Li, 19 Jul 2025), the controller introduces a confidence variable $\lambda_t\in[0,1]$ and uses blended predictions

$$\phi_{\tau|t}^{(\lambda_t)}=\lambda_t\,\phi_{\tau|t}+(1-\lambda_t)\,\kappa_{\tau|t},$$

so that the first control move is computed from an MPC problem driven by a convex combination of ML and nominal forecasts. In (Yang et al., 24 Jan 2025), the learned action $\tilde u_h$ is not trusted directly; instead it is accepted only if it lies in a safe action set

$$\mathcal{U}_{\lambda,h}=\Big\{u_h\in\mathcal U \,\Big|\, R_h+\phi_h(u_h)\le (1+\lambda)R_h^{\pi^\dagger}\Big\},$$

and otherwise it is projected or linearly mapped back toward the prior action.

A second pattern is residual or uncertainty-model augmentation of a nominal plant. In quadrotor control, the rigid-body model is augmented by a recursively learned GP drag model, so that

$$\dot{\mathbf x}=f_{\mathrm{phys}}(\mathbf x,\mathbf u)+f_{\mathrm{RGP}}(\mathbf x;\theta),$$

with the learned term inserted only in the translational acceleration channel (Smid et al., 2023). In learning-enhanced robust control, GP confidence bands over an uncertain nonlinearity are converted into IQC-compatible sector bounds, after which controller synthesis proceeds through an LFR/IQC robust-control pipeline rather than through end-to-end policy learning (Fiedler et al., 2021).

A third pattern is controller-parameter augmentation under preserved structure. In secondary frequency regulation, the learned element is a strictly monotone map $u=f(s)$ inserted into a primal-dual controller, which acts as a nonlinear preconditioner in the $s$-channel while keeping the equilibrium aligned with the KKT conditions of the original steady-state optimization (Yu et al., 10 Mar 2026). In neural-network-augmented iterative learning control, the learned element is a lateral neural network that predicts the nonlinear component of the converged ILC effort from reference position and velocity, while ILC remains the robustness backbone (Mashhadireza et al., 14 Nov 2025). In dexterous manipulation, SCAPE augments operational-space impedance control by learning state-dependent stiffness and stiffness limits, rather than learning torques directly (Kim et al., 2021). In L-Learning, the learned object is the Lagrangian itself, from which $\hat D(q)$, $u=f(s)$0, and $u=f(s)$1 are extracted and then used in a Lyapunov-certified tracking law (Quan et al., 26 May 2026).

Mechanism	Learned object	Backbone retained
Confidence blending	$u=f(s)$2, forecast trust	MPC
Safe projection of advice	$u=f(s)$3 with safe action mapping	safe prior policy
Residual model augmentation	drag or uncertainty model	physics model / robust synthesis
Change-of-variables preconditioning	strictly monotone $u=f(s)$4	primal-dual control
Feedforward augmentation	nonlinear ILC effort	ILC + feedback
Structured gain modulation	stiffness $u=f(s)$5, $u=f(s)$6	impedance control
Energy-model learning	Lagrangian $u=f(s)$7	Lyapunov-based tracking

3. Guarantees, certificates, and performance criteria

The distinctive feature of LAC, relative to purely predictive or purely end-to-end learned control, is that the learned element is typically analyzed through a certificate already native to control theory. In the primal-dual frequency-regulation setting, the controller

$u=f(s)$8

is accompanied by a Lyapunov function

$u=f(s)$9

whose derivative satisfies

$\lambda_t\in[0,1]$0

Under strict monotonicity, Lipschitz continuity, and $\lambda_t\in[0,1]$1, the paper states uniqueness of the optimizer, equivalence to the original steady-state program, KKT-optimal equilibria, and global asymptotic convergence (Yu et al., 10 Mar 2026).

In online MPC with untrusted predictions, guarantees are phrased in competitive-analysis terms. The delayed confidence-learning controller in (Li, 19 Jul 2025) establishes a competitive ratio bound for nonlinear systems,

$\lambda_t\in[0,1]$2

and a tight bound for the LQ case. The stated interpretation is that accurate ML predictions drive the controller toward clairvoyant performance, whereas adversarial predictions drive $\lambda_t\in[0,1]$3 toward nominal prediction use, so that performance degrades gracefully while recursive feasibility and safety are preserved.

In safe online control for water infrastructures, the principal guarantee is per-round relative safety: $\lambda_t\in[0,1]$4 The safe action set construction ensures that the prior action $\lambda_t\in[0,1]$5 is always feasible, so the controller never loses a guaranteed fallback (Yang et al., 24 Jan 2025).

Other formulations use distinct performance notions. UCB-NCS proves high-probability regret of order $\lambda_t\in[0,1]$6 relative to the known-parameter optimal networked controller (Singh et al., 2020). AntLer proves that its sample-average approximation converges almost surely to the optimal anticipated-learning design, both in value and in argmin, under GP and regularity assumptions (Capone et al., 2020). Learning-enhanced robust control maps GP confidence intervals to sector-IQCs and then proves, with probability at least $\lambda_t\in[0,1]$7, that the synthesized controller robustly stabilizes the true system and achieves the certified robust performance level (Fiedler et al., 2021).

Transient-performance metrics also play a central role. In frequency regulation, the learning objective explicitly targets the exponentially weighted metric

$\lambda_t\in[0,1]$8

the frequency nadir $\lambda_t\in[0,1]$9, and the time-averaged control cost

$$\phi_{\tau|t}^{(\lambda_t)}=\lambda_t\,\phi_{\tau|t}+(1-\lambda_t)\,\kappa_{\tau|t},$$0

The paper further states that if $$\phi_{\tau|t}^{(\lambda_t)}=\lambda_t\,\phi_{\tau|t}+(1-\lambda_t)\,\kappa_{\tau|t},$$1, then $$\phi_{\tau|t}^{(\lambda_t)}=\lambda_t\,\phi_{\tau|t}+(1-\lambda_t)\,\kappa_{\tau|t},$$2 (Yu et al., 10 Mar 2026).

4. Representative realizations and reported outcomes

Reported implementations span secondary frequency regulation, water pumping systems, quadrotor trajectory tracking, networked control over lossy links, Lorentz-force motion systems, dexterous manipulation, and robot tracking with learned Lagrangians (Yu et al., 10 Mar 2026, Yang et al., 24 Jan 2025, Smid et al., 2023, Quan et al., 26 May 2026, Mashhadireza et al., 14 Nov 2025, Kim et al., 2021). The breadth of domains reinforces the view that LAC is defined more by architecture and certification strategy than by any single plant class.

Domain	LAC mechanism	Reported outcome
Secondary frequency regulation	monotone $$\phi_{\tau|t}^{(\lambda_t)}=\lambda_t\,\phi_{\tau|t}+(1-\lambda_t)\,\kappa_{\tau|t},$$3 inside primal-dual control	convergence “rate (s)” $$\phi_{\tau|t}^{(\lambda_t)}=\lambda_t\,\phi_{\tau|t}+(1-\lambda_t)\,\kappa_{\tau|t},$$4 vs. $$\phi_{\tau|t}^{(\lambda_t)}=\lambda_t\,\phi_{\tau|t}+(1-\lambda_t)\,\kappa_{\tau|t},$$5, frequency nadir $$\phi_{\tau|t}^{(\lambda_t)}=\lambda_t\,\phi_{\tau|t}+(1-\lambda_t)\,\kappa_{\tau|t},$$6 vs. $$\phi_{\tau|t}^{(\lambda_t)}=\lambda_t\,\phi_{\tau|t}+(1-\lambda_t)\,\kappa_{\tau|t},$$7, accumulated cost $$\phi_{\tau|t}^{(\lambda_t)}=\lambda_t\,\phi_{\tau|t}+(1-\lambda_t)\,\kappa_{\tau|t},$$8 vs. $$\phi_{\tau|t}^{(\lambda_t)}=\lambda_t\,\phi_{\tau|t}+(1-\lambda_t)\,\kappa_{\tau|t},$$9
Water infrastructures	ML advice filtered through safe action sets	LAOC $\tilde u_h$0: energy $\tilde u_h$1, carbon $\tilde u_h$2 kg, max risk ratio $\tilde u_h$3; zero violations
Quadrotor MPC	recursive GP drag augmentation	random trajectory, $\tilde u_h$4: nominal $\tilde u_h$5 mm, pretrained GP $\tilde u_h$6 mm, RGP $\tilde u_h$7 mm
2-DOF robot arm	learned Lagrangian with Lyapunov controller	sine tracking: RMSE $\tilde u_h$8, ITAE $\tilde u_h$9; PID RMSE $$\mathcal{U}_{\lambda,h}=\Big\{u_h\in\mathcal U \,\Big|\, R_h+\phi_h(u_h)\le (1+\lambda)R_h^{\pi^\dagger}\Big\},$$0; SAC (10k) RMSE $$\mathcal{U}_{\lambda,h}=\Big\{u_h\in\mathcal U \,\Big|\, R_h+\phi_h(u_h)\le (1+\lambda)R_h^{\pi^\dagger}\Big\},$$1; TD3 (10k) RMSE $$\mathcal{U}_{\lambda,h}=\Big\{u_h\in\mathcal U \,\Big|\, R_h+\phi_h(u_h)\le (1+\lambda)R_h^{\pi^\dagger}\Big\},$$2

Other case studies are more qualitative but still structurally informative. In SCAPE, position control fails to learn grasping in the Block task and fails in NuFingers, whereas stiffness modulation allows safe manipulation and sim-to-real transfer; the method combines augmented demonstrations, Q-filtering, and imitation regulation rather than relying on stiffness demonstrations (Kim et al., 2021). In neural-network-augmented ILC, the initial MSE at reference changes is reported as “significantly lower,” and convergence is faster because the NN supplies an approximate nonlinear feedforward term immediately after a new reference is commanded (Mashhadireza et al., 14 Nov 2025). In networked control, UCB-NCS is theoretical rather than empirical, but it provides a concrete optimistic controller that learns plant and channel parameters jointly (Singh et al., 2020).

5. Relation to adjacent paradigms and recurrent misconceptions

LAC is not synonymous with end-to-end reinforcement learning. Several representative methods retain a fixed control architecture and learn only selected ingredients inside it: confidence weights in MPC, residual aerodynamic terms in a rigid-body model, stiffness values in impedance control, or a Lagrangian from which a Lyapunov-certified controller is derived (Li, 19 Jul 2025, Smid et al., 2023, Kim et al., 2021, Quan et al., 26 May 2026). This is structurally different from direct policy search over raw control inputs.

LAC is also not identical to classical adaptive control, though there is overlap. AntLer explicitly positions itself as complementary to adaptive control: the unknown dynamics are modeled nonparametrically with GPs, and the policy parameters are optimized by anticipating how future learning changes future cost (Capone et al., 2020). Likewise, the robust IQC framework does not merely update a nominal model; it converts statistical uncertainty into an admissible uncertainty block for robust synthesis (Fiedler et al., 2021).

A second misconception is that LAC always means optimism or active exploration. Some formulations do rely on optimism or exploration bonuses, such as GP dual control and UCB-NCS (Alpcan, 2011, Singh et al., 2020). Others are fundamentally conservative wrappers around learned advice: LAOC projects the learned action into a safe set, and delayed-confidence MPC reduces reliance on ML when realized errors are large (Yang et al., 24 Jan 2025, Li, 19 Jul 2025). Still others improve transients without changing the optimizer at all, as in monotone change-of-variables preconditioning for primal-dual control (Yu et al., 10 Mar 2026).

A third misconception concerns terminology. One strand uses LAC in a narrow algorithmic sense associated with untrusted predictions, online confidence learning, competitive ratio, and "best-of-both-worlds" behavior (Li, 19 Jul 2025). Another uses it more broadly for any controller in which learning augments a formal control structure while preserving physical or optimization-based semantics, including robust control with IQCs, impedance control with learned stiffness, and mission-aligned hierarchical planning-plus-control (Fiedler et al., 2021, Kim et al., 2021, Kungurtsev et al., 6 Jul 2025). This suggests that the term currently has both narrow and broad usages.

6. Limitations, assumptions, and open directions

The strongest results in LAC are assumption-heavy. Frequency-regulation guarantees depend critically on strict monotonicity, Lipschitz continuity, and $$\mathcal{U}_{\lambda,h}=\Big\{u_h\in\mathcal U \,\Big|\, R_h+\phi_h(u_h)\le (1+\lambda)R_h^{\pi^\dagger}\Big\},$$3 for the learned change of variables (Yu et al., 10 Mar 2026). Competitive-ratio guarantees for delayed-confidence MPC assume stabilizability, control-invariant sets, SSOSC, LICQ, and Exponentially Decaying Perturbation Bounds (Li, 19 Jul 2025). Safe-action-set results for water systems assume Lipschitz dynamics and non-negative, $$\mathcal{U}_{\lambda,h}=\Big\{u_h\in\mathcal U \,\Big|\, R_h+\phi_h(u_h)\le (1+\lambda)R_h^{\pi^\dagger}\Big\},$$4-strongly convex, $$\mathcal{U}_{\lambda,h}=\Big\{u_h\in\mathcal U \,\Big|\, R_h+\phi_h(u_h)\le (1+\lambda)R_h^{\pi^\dagger}\Big\},$$5-smooth risk functions (Yang et al., 24 Jan 2025). GP-to-IQC guarantees require RKHS norm bounds, subgaussian noise, structural correctness of the uncertainty block, and a domain restriction containing all relevant signals (Fiedler et al., 2021).

Many practical formulations remain only partially certified. AntLer explicitly does not claim closed-loop stability or safety guarantees, focusing instead on optimal parameter selection via anticipated cost minimization (Capone et al., 2020). The data-augmented quadrotor MPC uses the GP mean only and does not incorporate GP variance into cautious, tube, or chance-constrained MPC; the paper states that it does not provide formal guarantees on stability, recursive feasibility, or constraint satisfaction (Smid et al., 2023). In frequency regulation, explicit actuator limits, ramp-rate constraints, line limits, and robustness to significant model mismatch are not part of the main results, although extensions are described as structurally compatible (Yu et al., 10 Mar 2026).

Several open directions recur across the literature. One is constraint-rich augmentation: incorporating hard operational constraints into learned primal-dual or Lyapunov-based designs (Yu et al., 10 Mar 2026, Quan et al., 26 May 2026). Another is distribution shift and partial observability, especially for hierarchical mission-aligned settings in which planning, state abstraction, and control interact under uncertain perception (Kungurtsev et al., 6 Jul 2025). A third is scalability: sparse GP methods, dynamic IQC multipliers, higher-dimensional robust synthesis, multi-predictor aggregation, and online re-synthesis are all identified as important extensions (Fiedler et al., 2021, Li, 19 Jul 2025). A fourth is safe exploration: several papers note barrier-function filters, Lyapunov-based regularization, or feasible-exploration mechanisms as natural complements, but not yet as fully integrated solutions (Yang et al., 24 Jan 2025, Kungurtsev et al., 6 Jul 2025).

Taken together, these works depict LAC as a technically heterogeneous but conceptually coherent research program: retain a structured control or planning backbone, place learning in a carefully delimited role, and use control-theoretic certificates—Lyapunov functions, KKT structure, robust invariant sets, IQCs, regret bounds, or competitive ratios—to quantify what the learned augmentation may improve and what it must not compromise.