Actuator-Aware Feedback Control

• Actuator-aware feedback logic is a design principle that tailors control laws to the actuator’s native command space, ensuring that control decisions respect the physical and digital bounds of the actuator.
• It encompasses co-design of sensors and actuators, where actuator selection, sensor placement, and execution characteristics are jointly optimized to enhance system robustness and efficiency.
• The approach addresses practical challenges like hybrid actuation, saturation, quantization, and hysteresis by modeling execution uncertainties and using specialized feedback loops to maintain stability and safety.

Searching arXiv for the cited topic papers to ground the article in current arXiv metadata. arxiv_search(query="Actuator-aware feedback logic mobile robot brakes static output feedback sensor actuator selection piezoelectric actuator hysteresis proprioceptive feedback", max_results=10) Actuator-aware feedback logic denotes a class of control formulations in which the feedback law is designed around the actual affordances, limits, availability, and execution fidelity of actuators rather than around an idealized input channel. Across the literature, this includes hybrid continuous/discrete actuation, actuator and sensor co-design, saturation-aware hybrid control, digital controllers whose state is the actuator code itself, feedback built from actuator-execution mismatch, and even material systems in which the actuator’s own motion modulates its stimulus. Taken together, these works suggest that actuator-aware feedback logic is less a single algorithm than a design principle: the controller should be expressed in the actuator’s native command space and should remain valid under the physical mechanisms by which actuation is actually produced (Nikshi et al., 2019, Nugroho et al., 2018, Jankovic, 27 Mar 2026).

1. Conceptual scope and defining features

In the cited work, actuator awareness appears in several distinct but related senses. In one sense, it means that the control output is synthesized directly in actuator-native coordinates. The Mixed conventional/braking Actuation Mobile Robot uses a continuous longitudinal drive force $F_d$ together with discrete brake states $F^{(1)},F^{(2)}\in\{0,1\}$, so the controller must decide among left-brake, right-brake, or neither, rather than outputting an artificial steering angle (Nikshi et al., 2019). In a second sense, actuator awareness means that actuator existence or placement is itself a design variable. Static output feedback with simultaneous sensor and actuator selection embeds binary activation matrices $\Pi$ and $\Gamma$ into the closed-loop conditions, so stabilizability depends jointly on the chosen actuator subset, the chosen sensor subset, and the resulting feedback gain (Nugroho et al., 2018). In a third sense, actuator awareness concerns the discrepancy between commanded and realized action. Machine proprioceptive feedback computes a fast residual from $u^{act}-u^*$ and feeds only that unexpected portion back into a safety controller so that other channels can compensate (Jankovic, 27 Mar 2026).

The same theme extends beyond conventional rigid-body systems. In digital laser frequency stabilization, the actuator code itself becomes the Markov state, and the lock is modeled as random transitions among discrete actuator settings rather than as a continuous servo signal (Banik et al., 8 Apr 2026). In opto-mechanical soft matter, the actuator’s deformation changes the amount of light it receives through a mechanically coupled baffle, so feedback is embedded physically in the actuator-stimulus path (Yang et al., 2024). A plausible implication is that actuator-aware feedback logic should be understood as a cross-cutting systems concept spanning robotics, networked control, soft matter, photonics, and safety-critical mobility rather than as a subfield tied to one control architecture.

2. Native actuator command spaces and hybrid actuation

A canonical example is the Mixed conventional/braking Actuation Mobile Robot. Mechanically, it has one rear omni-directional drive wheel and two front freely rolling conventional wheels, each equipped with a brake. Steering is not produced by a continuously actuated steering linkage. Instead, the rear omni wheel generates force only along the robot longitudinal axis $x_r$, while the front wheels are not driven and can only be locked or unlocked. With positive drive force, activating the left brake causes one turning direction and activating the right brake causes the opposite; for negative drive force the mapping reverses. The resulting actuation structure is hybrid and asymmetric: one continuous channel $F_d$, and two binary channels $F^{(1)},F^{(2)}\in\{0,1\}$ (Nikshi et al., 2019).

That actuation structure determines both the model and the controller. The robot is modeled on $SE(2)=\mathbb{R}^2\times\mathbb{S}^1$, and the front-wheel geometry yields the nonholonomic constraint

$\dot{y}_r + x_r^{(i)}\dot{\theta}=0.$

Brake friction is explicitly multiplied by the discrete brake-state variables,

$F^{(1)},F^{(2)}\in\{0,1\}$0

and steering moment arises only from brake friction at lateral offsets. Because braking both generates yaw torque and dissipates longitudinal motion, the paper treats brake selection itself as the central feedback decision. The parking controller is therefore sequential: first a fuzzy controller drives $F^{(1)},F^{(2)}\in\{0,1\}$1 and $F^{(1)},F^{(2)}\in\{0,1\}$2 to target values while $F^{(1)},F^{(2)}\in\{0,1\}$3 is held constant, then a proportional controller with saturation regulates $F^{(1)},F^{(2)}\in\{0,1\}$4. The fuzzy input is

$F^{(1)},F^{(2)}\in\{0,1\}$5

with

$F^{(1)},F^{(2)}\in\{0,1\}$6

and the post-defuzzification logic maps the fuzzy output directly into brake commands: $F^{(1)},F^{(2)}\in\{0,1\}$7 This is actuator-aware in a literal sense: the fuzzy system does not output a virtual steering command, but a decision already matched to the discrete brake hardware (Nikshi et al., 2019).

A related but more extreme formulation appears in digitally stabilized lasers. There, the actuator state is

$F^{(1)},F^{(2)}\in\{0,1\}$8

with discrete index $F^{(1)},F^{(2)}\in\{0,1\}$9, and bang-bang sign logic updates the next actuator code according to the sign of the sampled, quantized discriminator. Transition probabilities

$\Pi$0

define a Markov chain over actuator codes. Here the actuator is not merely constrained; it is the modeled state of the closed loop (Banik et al., 8 Apr 2026). This suggests a broader interpretation: once actuation is quantized, discrete, or hybrid, actuator-aware feedback logic often replaces smooth control synthesis by logic defined directly on attainable actuator transitions.

3. Architecture-conditioned feedback and co-design

A second major line of work treats actuator awareness as an architectural issue. In static output feedback with simultaneous sensor and actuator selection, the plant is

$\Pi$1

with block-diagonal $\Pi$2 and $\Pi$3, and binary variables

$\Pi$4

activate or deactivate local actuators and sensors. These appear through

$\Pi$5

and the closed loop becomes $\Pi$6. The point is not merely sparse hardware selection; it is that the existence of a stabilizing static output feedback gain depends on the chosen architecture itself. The paper formulates this as a mixed-integer nonlinear matrix inequality problem and gives both an equivalent MI-SDP reformulation using integer/disjunctive programming and an exact binary-search-like algorithm using repeated static-output-feedback feasibility checks. It also uses a sufficient LMI condition attributed to Crusius and Trofino (Nugroho et al., 2018).

For a 10-subsystem mass-spring system with $\Pi$7 and a logistic constraint requiring at least 2 activated sensors and 2 activated actuators, all three reported methods found solutions with 2 sensors and 2 actuators, i.e.

$\Pi$8

The reported outcomes are summarized below (Nugroho et al., 2018).

Method	Reported result	Runtime
MI-SDP	max closed-loop real part $\Pi$9	$\Gamma$0 s
BSA-SDP	max closed-loop real part $\Gamma$1, $\Gamma$2 iterations	$\Gamma$3 s
BSA-PBH	max closed-loop real part $\Gamma$4, $\Gamma$5 iterations	$\Gamma$6 s

A more dynamic version of the same idea appears in self-tuning network control architectures with joint sensor and actuator selection. There, the active actuator set $\Gamma$7 and sensor set $\Gamma$8 are themselves feedback decisions, and the architecture policy maps available information to the actuator subset, sensor subset, and control input. In the full-state linear-quadratic case, the Bellman recursion minimizes jointly over architecture and control, and the paper proves that the optimal value function is piecewise quadratic and the policy piecewise linear over a finite partition of the state space. In a 50-node unstable network with $\Gamma$9, the fixed architecture with actuators at nodes 1 and 2 was reported as 80x worse than the self-tuning architecture. In a second 50-node output-feedback experiment with exactly 5 active sensors and 5 active actuators, the fixed architecture failed to stabilize the system over the horizon, while the self-tuning architecture stabilized it (Ganapathy et al., 2024).

Actuator placement in distributed systems shows the same nonseparability. For vortex-shedding suppression in the linearized 2D cylinder wake, the optimal actuator for full-state information control is not generally the optimal actuator for one-sensor output feedback. The paper compares OE, FIC, CIOC, IOC$u^{act}-u^*$0, and IOC$u^{act}-u^*$1, and shows that choosing the OE-optimal sensor together with the FIC-optimal actuator can be far worse than a jointly optimized or collocated design because convective time lag dominates the sensor-actuator loop. At $u^{act}-u^*$2, IOC$u^{act}-u^*$3 is reported as about 593% worse than CIOC (Jin et al., 2021). The general lesson is that actuator-aware feedback logic often requires co-design of architecture, not merely gain tuning after the architecture is fixed.

4. Execution-aware control: saturation, quantization, and actuator mismatch

Another major strand of actuator awareness concerns the fact that realized actuation may differ from commanded actuation. In control design under actuator saturation and multi-rate sampling, the plant is

$u^{act}-u^*$4

and the controller maintains an internal state $u^{act}-u^*$5 that evolves as a copy of the plant driven by the same saturated input,

$u^{act}-u^*$6

When asynchronous measurements arrive on different channels, only the corresponding subvectors of $u^{act}-u^*$7 are reset. The deadzone

$u^{act}-u^*$8

is retained explicitly in the hybrid model, and the paper derives matrix-inequality conditions guaranteeing regional exponential stability of

$u^{act}-u^*$9

together with an ellipsoidal estimate of the basin of attraction. This is actuator-aware because the predictor state and the stability proof are both consistent with what the actuator actually delivers, not with the unsaturated command (Ferrante et al., 2022).

Machine proprioceptive feedback pushes that idea further by closing a fast loop around actuator execution error. The safe-set dynamics are written as

$x_r$0

where $x_r$1 is not an external disturbance but the unexpected portion of actuator behavior. The key residual filter is

$x_r$2

with $x_r$3. In full-MPF, the complete residual vector is fed back into the barrier-function-based safety filter. In split-MPF, each actuator-specific controller sees the deficiencies of the other channels but not its own, with $x_r$4. The paper states that full-MPF offers the stronger safety guarantee when the fast loop is asymptotically stable and the actuator uncertainty class is input-strict-passive, while split-MPF is much more tolerant of delays but weaker under simultaneous multi-channel attenuation. In 16-vehicle lane-swap simulations, no-MPF produced 32% runs with CBF violations in a propulsion-loss case, whereas both MPF variants maintained a reported minimum $x_r$5 m; under filtered steering and acceleration, full-MPF showed major safety failures while split-MPF performed much better (Jankovic, 27 Mar 2026).

Quantization and digital update logic introduce another form of execution awareness. In digitally stabilized laser frequency control, the state is the discrete actuator code and the steady-state distribution is obtained from the unit-eigenvalue solution of the Markov transition matrix. For white frequency noise under decorrelated sampling and update schemes, the Markov formulation is exact; correlated discriminator sampling leaves the actuator mean essentially unchanged but inflates its standard deviation by factors approximately $x_r$6, $x_r$7, and $x_r$8 for progressively more correlated implementations (Banik et al., 8 Apr 2026). This makes a specific point: actuator-aware feedback logic may need to model not only the plant but the discrete stochastic evolution of actuator settings induced by sampling, quantization, and decision logic.

5. Materially embedded and model-based actuator awareness

Some work moves actuator awareness into the actuator model itself, or even into the actuator’s material architecture. For magnetic shape memory actuators with large, non-smooth input hysteresis, the proposed 2DOF architecture combines an inversion-free hysteresis-compensating feedforward law,

$x_r$9

with a robustly loop-shaped PI controller

$F_d$0

where $F_d$1 and $F_d$2. The identified linear transfer function is

$F_d$3

and the low-pass filter is

$F_d$4

with cutoff frequency $F_d$5. The paper’s design logic is explicitly actuator-aware: feedforward handles branch-dependent hysteresis and near-zero or effectively infinite local gain, while feedback is kept simple because displacement sensing is noisy and filtered (Ruderman et al., 25 Feb 2025).

A related but more sensing-constrained case is electromechanical switching devices such as solenoids and relays. There, differential flatness is used to generate a nonlinear feedforward voltage trajectory from the switching trajectory, while adaptation occurs only from one operation to the next. The electrical and mechanical model includes

$F_d$6

with

$F_d$7

Because real-time position sensing is not technically or economically feasible, the adaptation uses a cycle-level cost rather than within-cycle state feedback. In simulation, after about 100 switching operations the cost is about half that of the uncontrolled scenario, and after about 200 operations it is about one tenth. In hardware, after 250 operations all tested relays performed better than the uncontrolled case (Moya-Lasheras et al., 2023). This is a borderline case for the encyclopedia topic, but it shows that actuator-aware logic can replace unavailable state feedback by strong within-cycle actuator modeling plus repetitive operation-level correction.

The strongest embodiment of feedback in the actuator itself appears in light-responsive soft matter. A dye-doped liquid crystal elastomer bends under photothermal absorption; an attached opaque baffle then changes the illumination seen by the actuator. If deformation increases illumination, feedback is positive; if deformation decreases illumination, feedback is negative. The sign is therefore set geometrically by

$F_d$8

Positive feedback yields a switcher with light-controlled energy barriers, measured through

$F_d$9

and supports bistable and multistable shape morphing. Negative feedback yields homeostatic self-oscillation, including a 1.2 cm actuator oscillating at about 3.8 Hz and two robots—a locomotor and a swimmer—operating under constant illumination (Yang et al., 2024). This suggests that actuator-aware feedback logic need not always be realized in software; it can also be embedded in the physical coupling between actuator state and actuator stimulus.

6. Applications, limits, and recurrent design lessons

The applications represented in these papers are unusually broad: parking control in a nonstandard mixed-actuation mobile robot (Nikshi et al., 2019), sparse architecture design for networked linear systems (Nugroho et al., 2018), self-tuning architecture selection (Ganapathy et al., 2024), flow-control placement in cylinder wakes (Jin et al., 2021), hybrid stabilization under saturation and asynchronous measurements (Ferrante et al., 2022), digitally stabilized lasers (Banik et al., 8 Apr 2026), proprioceptive safety compensation in traffic systems (Jankovic, 27 Mar 2026), hysteretic magnetic-shape-memory positioning (Ruderman et al., 25 Feb 2025), repetitive control of solenoid switching devices (Moya-Lasheras et al., 2023), and physically embedded feedback in opto-mechanical soft actuators (Yang et al., 2024). A broad inference from this range is that actuator-aware feedback logic becomes most valuable when the actuator cannot be treated as a static, linear, always-available input operator.

Several misconceptions are corrected by the literature. First, actuator-aware feedback logic is not synonymous with fault tolerance. The MAMR paper motivates brake actuation partly by controllability under actuator failure, but it does not provide a theorem proving controllability under a specific MAMR failure scenario and explicitly leaves experimental recovery under actuator failure as future work (Nikshi et al., 2019). Second, it is not simply ordinary robust control with a new name. The cited papers repeatedly modify the state, the architecture, or the decision variables themselves to account for actuator properties: actuator selection variables $F^{(1)},F^{(2)}\in\{0,1\}$0, actuator-state Markov chains, deadzone-embedded hybrid models, or residual loops around $F^{(1)},F^{(2)}\in\{0,1\}$1. Third, it is not always strict closed-loop feedback. The sparse actuator array for tactile rendering uses calibrated spatial response fields and inverse optimization to synthesize a desired vibration energy distribution, but it is a model-based open-loop method rather than a runtime feedback controller (Li et al., 2024). That paper is relevant because it formalizes how actuator-specific spatial response maps should shape command synthesis even without online sensing.

The limitations are equally recurrent. The mixed braking robot uses a sequential parking controller rather than simultaneous regulation of all states and omits formal robustness margins (Nikshi et al., 2019). Static output feedback with sensor/actuator selection relies on a sufficient, not necessary-and-sufficient, LMI characterization, and the cheaper PBH-screening variant is explicitly not guaranteed to be optimal for SOF selection (Nugroho et al., 2018). In self-tuning architectures, exhaustive architecture search remains combinatorial and the greedy methods are presented without approximation guarantees (Ganapathy et al., 2024). Full-MPF gives stronger safety guarantees but is sensitive to non-passive actuator dynamics such as delay; split-MPF is more delay-tolerant but weaker under simultaneous multi-actuator attenuation (Jankovic, 27 Mar 2026). The MSM actuator study does not provide a comprehensive stability theorem for the full nonlinear 2DOF interconnection (Ruderman et al., 25 Feb 2025), and the opto-mechanical soft-robot paper offers only limited explicit quantitative theory despite rich experimental evidence (Yang et al., 2024).

Across the papers, several design lessons recur. The first is that the control output should be represented in the actuator’s native command space whenever possible: brake-side selection for binary brakes, code transitions for a digital laser actuator, or voltage for piezoelectric stacks and shape-memory devices. The second is that actuator awareness often forces hierarchical or hybrid designs: sequential parking, predictor-reset hybrid loops, parallel feedforward-plus-feedback structures, or outer architecture selection wrapped around inner control synthesis. The third is that actuator uncertainty is frequently environmental or interaction-dependent—friction in brake steering, conductivity and magnetic field in liquid-metal control, sampling logic in digital stabilization, or delay/passivity class in safety compensation—so directional or rule-based logic is often used where exact inversion would be fragile. The fourth is that actuator-aware feedback logic often redefines the control problem itself: from “find $F^{(1)},F^{(2)}\in\{0,1\}$2” to “choose which actuators exist,” “which actuator mode should be on,” “which discrete actuator state should follow,” or “how should the actuator’s own motion regulate its input.”