Safety Governor: Control Framework Overview

• Safety Governor is a supervisory control layer that augments a nominal controller by dynamically modifying commands to maintain safety constraints.
• It encompasses diverse designs such as reference, action, time, and ethical governors, each with unique signal manipulation and safety certification methods.
• It employs techniques like online optimization, invariant set methods, and data-driven models to guarantee recursive feasibility and extend safe system operation.

Searching arXiv for recent and relevant papers on “safety governor” and closely related governor-based safety control frameworks. A Safety Governor is a supervisory control mechanism that augments a nominal controller, planner, or learning policy by modifying references, actions, or timing variables so that safety constraints remain satisfied online. In the recent control literature, the term spans reference governors, action governors, robust action governors, safety command governors, time governors, path feasibility governors, and neuro-symbolic ethical governors. Across these variants, the common structure is an add-on layer that sits above or between planning and actuation, monitors candidate commands against a certified safety model, and passes through, reshapes, or overrides those commands only as needed to preserve constraint satisfaction, recursive feasibility, or a prescribed risk bound (Li et al., 2022, Li et al., 2022, Aueawatthanaphisut et al., 15 Mar 2026).

1. Scope and taxonomy

The literature uses “Safety Governor” as an umbrella for several supervisory patterns that differ mainly in which signal is manipulated and how safety is certified. A reference governor filters desired references before they reach a stabilizing controller, an action governor filters control actions after a controller or RL policy proposes them, and time- or path-based governors regulate progression along a planned path or trajectory rather than directly modifying the state feedback law (Kim et al., 2023, Li et al., 2022, Arslan, 2022).

Governor family	Manipulated signal	Representative mechanism
Reference Governor / Explicit Reference Governor	Reference or auxiliary reference	Admissible-set filtering or explicit update law
Action Governor / Robust Action Governor	Nominal control action	Minimal action modification subject to safe-set membership
Time Governor / Path Feasibility Governor	Path parameter or path progress	Advance only when predicted motion or terminal condition is safe
Safety Command Governor	Actuator command	QP-based correction of nominal input
Ethical Governor	Planned action description	Risk scoring with threshold-based override

This diversity is not incidental. In multirotor perception control, the governor operates at the reference level to maintain visibility constraints (Kim et al., 2023). In safe RL and uncertain piecewise-affine systems, the governor acts on the control input itself (Li et al., 2022, Li et al., 2021). In high-order path following and planner–MPC integration, the governed variable is a scalar time or path parameter (Arslan, 2022, Zhang et al., 12 Jul 2025). In autonomous robotic manipulation, the governor can even operate on language-described actions by classifying ethical risk before execution (Aueawatthanaphisut et al., 15 Mar 2026).

A common misconception is that Safety Governor is synonymous with Reference Governor. The broader literature does not support that equivalence. Reference governors are a central subclass, but the same supervisory idea also appears as action filters, time-parametrization layers, path-progress governors, and explicit safety command filters (Li et al., 2022, Chen et al., 2024, Arslan, 2022).

2. Control architecture and canonical formulations

Most formulations share a two-layer architecture. A nominal controller, planner, or learned policy produces a candidate command, and the governor computes the actually applied command from the current state and the candidate. In the generalized Action Governor formalism, this is written as

$u(t)=\kappa\big(x(t),u_1(t)\big),$

where $u_1(t)$ is the action before adjustment (Li et al., 2022). In Robust Action Governor designs, the filtered action is typically the solution of a one-step optimization that minimizes deviation from the nominal command while ensuring that the next state lies in a robust safe set under all admissible disturbances (Li et al., 2022).

Reference-governor formulations instead evolve an auxiliary reference. In the Explicit Reference Governor family, a standard structure is

$\dot v = \Delta(x,v)\,\rho(v,r),$

where $\rho$ is a Navigation Field pointing toward the desired reference $r$ and $\Delta$ is a Dynamic Safety Margin that scales motion toward $r$ according to the available safety margin (Gautam et al., 12 Apr 2025). A related optimization-free CBF-based ERG formulation uses an augmented system $\dot g=\rho$ and a closed-form projected update for $\rho^\star$ derived from a barrier inequality, with the key property that $\rho=0$ is always feasible on the safe set (Nakano et al., 5 Apr 2026).

Time- and path-based variants govern trajectory progression rather than the control signal directly. In Time Governors for safe path following, the governed variable is the path parameter $u_1(t)$0, and the update law is

$u_1(t)$1

so progress along the path occurs only when the predicted motion set remains inside free space (Arslan, 2022). In PathFG, the governor advances the path parameter through a terminal-set feasibility test on the predicted terminal MPC state, thereby integrating path planning and nonlinear MPC while preserving recursive feasibility (Zhang et al., 12 Jul 2025).

Threshold logic appears in risk-based governors. The neuro-symbolic ethical governor computes

$u_1(t)$2

with $u_1(t)$3, $u_1(t)$4, and $u_1(t)$5, and then applies the decision rule

$u_1(t)$6

Actions above threshold are replaced by a safe pose or conservative primitive (Aueawatthanaphisut et al., 15 Mar 2026).

These formulations differ in interface, but not in intent. Each converts an unconstrained or weakly constrained nominal decision process into a constrained closed loop by embedding an explicit supervisory map between intention and actuation.

3. Safety certificates, invariance, and recursive feasibility

The formal backbone of Safety Governor design is set invariance. Different papers instantiate this with different objects, but the recurring themes are robust control invariance, maximal output admissible sets, safe/returnable sets, invariant Lyapunov sublevel sets, and terminal-set feasibility conditions.

In Action Governor theory, the governor relies on a compact safe/returnable set $u_1(t)$7 associated with a nominal policy $u_1(t)$8. The central guarantees are all-time safety, eventual feasibility, and, in the linear positively invariant case, recursive feasibility (Li et al., 2022). In Robust Action Governor formulations for uncertain PWA systems, the corresponding safe object is a robust viable or invariant set $u_1(t)$9, computed through robust predecessor operations over the PWA partition and non-convex state constraints. If the initial state lies in $\dot v = \Delta(x,v)\,\rho(v,r),$0 and the projection condition

$\dot v = \Delta(x,v)\,\rho(v,r),$1

holds, then recursive feasibility and all-time constraint satisfaction follow (Li et al., 2022).

Reference-governor designs for linear systems often use maximal output admissible sets. For polynomial or uncertain polynomial constraints, the MOAS is computed in lifted coordinates, yielding finitely determined inner approximations that remain positively invariant under the governed closed loop (Kim et al., 2023, Schieni et al., 2022). The visibility-constrained multirotor governor extends this logic to time-varying references through a lifted representation whose inner MOAS remains finitely determined despite marginally stable lifted dynamics, using steady-state tightening (Kim et al., 2023).

Several nonlinear Safety Governors replace polyhedral invariant sets with Lyapunov sublevel sets. In safe navigation with local obstacle sensing, invariant ellipsoids bound the output of a feedback-linearized nonlinear system under bounded disturbances. Safety is enforced by requiring the invariant output ellipsoid radius, measured in a quadratic norm, to remain below the sensed obstacle distance around the governor state (Li et al., 2020). In ERG-CBF synthesis, the governor constructs a smooth barrier from softmin-aggregated Dynamic Safety Margins and steady-state admissibility constraints, and then enforces

$\dot v = \Delta(x,v)\,\rho(v,r),$2

for the augmented state $\dot v = \Delta(x,v)\,\rho(v,r),$3, with feasibility guaranteed because the trivial update $\dot v = \Delta(x,v)\,\rho(v,r),$4 always satisfies the barrier inequality on the safe set (Nakano et al., 5 Apr 2026).

A second misconception is that governors only guarantee instantaneous correction. In the cited works, the stronger objective is usually recursive feasibility: the governor must choose a command that keeps the next state inside a set from which the same problem remains feasible at the next step. This is explicit in AG, RAG, MOAS-based RG, PathFG, and PET-ERG analyses (Li et al., 2022, Li et al., 2022, Zhang et al., 12 Jul 2025, Nakano et al., 5 Apr 2026).

4. Representative instantiations across domains

The concept has been specialized to markedly different safety semantics. In robotic manipulation, the neuro-symbolic ethical governor inserts a language-grounded supervisory layer between planner and controller. A fine-tuned DistilBERT classifier trained on the ETHICS commonsense dataset estimates unsafe-action probability from natural-language task descriptions, which is then combined with uncertainty and class-probability variance into a continuous ethical risk metric. In simulation, training loss decreased from approximately $\dot v = \Delta(x,v)\,\rho(v,r),$5 to below $\dot v = \Delta(x,v)\,\rho(v,r),$6, validation loss increased to approximately $\dot v = \Delta(x,v)\,\rho(v,r),$7, overall accuracy was approximately $\dot v = \Delta(x,v)\,\rho(v,r),$8–$\dot v = \Delta(x,v)\,\rho(v,r),$9, and ethical risk scores formed bimodal peaks near $\rho$0–$\rho$1 and $\rho$2–$\rho$3, indicating discrimination between routine low-risk and human-proximal or high-risk contexts (Aueawatthanaphisut et al., 15 Mar 2026).

In vehicle control, the Safety Command Governor based on a deep Koopman model and CBFs filters driving torque to preserve lateral stability while keeping steering unchanged. The learned lifted linear dynamics transform the safety filter into a convex QP, and the reported average solve time is approximately $\rho$4 ms at $\rho$5 Hz (Chen et al., 2024). In adaptive cruise control, the Time Shift Governor augments an MPC-CBF controller by making the lead-vehicle reference itself a decision variable through a scalar time shift $\rho$6. In the reported simulations, baseline MPC-CBF achieved $\rho$7 success with $\rho$8 collisions, whereas the TSG-guided version achieved $\rho$9 success and $r$0 collisions (Kee et al., 30 Jun 2025).

Aviation and attitude-control applications show that governor ideas are not limited to Euclidean state spaces. The visibility-constrained multirotor reference governor enforces camera field-of-view inequalities, positive depth, and state limits with mean computation time $r$1 ms and maximum $r$2 ms under a $r$3 ms sample time (Kim et al., 2023). For constrained rigid-body attitude control directly on $r$4, PET-ERG updates an auxiliary attitude reference only at sampled instants when a robust safety condition

$r$5

is satisfied. The resulting scheme enforces torque saturation and geometric pointing constraints while establishing asymptotic stability and exponential convergence for almost all initial configurations (Nakano et al., 5 Apr 2026). A related velocity-free spacecraft attitude governor uses MRPs, an immersion-and-invariance observer, and an explicit reference governor that tightens the invariant threshold with a dynamic factor linked to angular-velocity estimation error (Dang et al., 2024).

Mobile robot navigation has produced two distinct governor traditions. One uses invariant ellipsoids and a virtual governor state to regulate progression along a geometric path in unknown cluttered environments with local sensing (Li et al., 2020). The other uses time governors and prediction sets such as Lyapunov ellipsoids and Vandermonde simplexes to regulate the path parameter directly (Arslan, 2022). PathFG generalizes this line by integrating geometric path planning with nonlinear MPC and expanding the region of attraction from the MPC $r$6-step reachable set to all states connected to the goal by a feasible path, with reported PathFG overhead of approximately $r$7 s and PathFG+MPC total computation of approximately $r$8 s per step (Zhang et al., 12 Jul 2025).

5. Learning, data-driven models, and safe adaptation

Safety Governors have become a standard mechanism for reconciling learning with hard constraints. In safe RL with RAG, the RL agent proposes an action, the governor solves a minimally invasive MIQP or QP to keep the next state in a robustly safe set, and training proceeds on the filtered closed-loop data. In the adaptive cruise control case study, safe RL with RAG achieved zero violations throughout training, whereas conventional RL exhibited safety violations during exploration (Li et al., 2021). The broader generalized AG framework extends this principle to both safe Q-learning and safe data-driven Koopman control, with the governor remaining agnostic to how the nominal action was generated (Li et al., 2022).

The learning-based explicit MPC literature uses a different decomposition. A neural network approximates the explicit MPC law, but a safety governor projects the NN output onto an affine family

$r$9

constrained by membership in an augmented invariant set $\Delta$0. The resulting online problem is a convex QP in the parameter $\Delta$1:

$\Delta$2

The architecture is dual-mode: once the state enters $\Delta$3, the learned controller reduces exactly to the linear feedback $\Delta$4, eliminating approximation error near the origin while preserving recursive feasibility (Mao et al., 21 Jul 2025).

Data-driven governors also appear when the safety model itself is learned. The deep Koopman Safety Command Governor learns a lifted linear prediction model from hardware-in-the-loop vehicle data and then imposes CBF constraints on the predicted next state (Chen et al., 2024). The ethical governor learns unsafe-action probability from language and then applies deterministic symbolic threshold rules (Aueawatthanaphisut et al., 15 Mar 2026). These examples show a consistent division of labor: learning supplies prediction, classification, or policy approximation, while the governor preserves a hard supervisory envelope.

This suggests a general pattern in recent work: safety-critical guarantees are usually not entrusted to the learned policy itself, but to an external supervisory layer with explicit invariance, admissibility, or threshold logic.

6. Trade-offs, misconceptions, and current directions

Safety Governors are often described as minimally invasive, but that property is only meaningful relative to a safety certificate that can itself be conservative. Conservatism arises from invariant-set approximations, polyhedral over-approximations, big-$\Delta$5 formulations, short-horizon robust predecessors, lifted polynomial representations, and softmin barrier surrogates (Li et al., 2022, Schieni et al., 2022, Nakano et al., 5 Apr 2026). PathFG reduces horizon length, but still depends on terminal-set construction and path feasibility assumptions (Zhang et al., 12 Jul 2025). Ethical governors improve interpretability relative to purely data-driven filters, but the reported validation overfitting and domain shift from ETHICS text to robotic language remain limitations (Aueawatthanaphisut et al., 15 Mar 2026).

Another misconception is that governors must always solve an online optimization. Explicit and optimization-free constructions are prominent. C-ERG updates the reference by a closed-form differential equation and avoids online optimization entirely (Gautam et al., 12 Apr 2025). PET-ERG updates only at sampled events and likewise avoids online optimization (Nakano et al., 5 Apr 2026). The optimization-free ERG-CBF framework derives a closed-form projection for the virtual reference input, again without iterative online optimization (Nakano et al., 5 Apr 2026). Conversely, QP, MIQP, and scalar bisection remain attractive when they preserve convexity and provide real-time runtimes on the target platform (Chen et al., 2024, Kim et al., 2023).

A further point of clarification is that governors are not limited to maintaining indefinite invariance. The unified safety protection and extension governor explicitly treats the case where indefinite safety is infeasible. It solves a single constrained optimization with continuous slack variables so as either to preserve invariant-set feasibility or, failing that, to maximize the time before any safety violation occurs (Li et al., 2023). In that sense, the governor framework now covers both “keep safe forever” and “extend safe operation as long as possible.”

Current work points in several directions. One is broader signal modalities: language-grounded ethical risk, perception constraints, and multimodal safety cues (Aueawatthanaphisut et al., 15 Mar 2026, Kim et al., 2023). Another is better integration with adaptive and learned controllers without sacrificing hard guarantees (Li et al., 2022, Mao et al., 21 Jul 2025). A third is extending optimization-free and event-triggered constructions to more complex manifolds and cyber-physical platforms (Nakano et al., 5 Apr 2026). Across these trends, the defining characteristic remains unchanged: a Safety Governor is a supervisory layer that makes safety a property of the closed-loop interconnection rather than a by-product of nominal control performance.