Enforcing Safety Guarantees Under Actuator Saturation

35 concepts24 papers1 proposal

Imagine driving a car toward a cliff. A mathematical safety filter tells you, "You are safe as long as you do not cross the edge." If you approach the cliff at 100 mph, the filter remains silent until the very last millisecond, at which point it commands infinite braking force. In theory, you stop instantly. In reality, your brakes lock up, your tires slip, and you sail over the edge.

This failure mode highlights a critical gap in traditional safety-critical control: the assumption of infinite control authority. In real-world robotics, actuators have physical limits—motors saturate, thrusters hit maximum thrust, and fuel valves can only open so wide. When a safety filter ignores these limits, it generates commands that are mathematically sound but physically impossible, causing the system's actual trajectory to diverge from its safe guarantees.

To bridge this gap, a growing body of research focuses on input-constrained safety synthesis. The goal is simple yet profound: design safety filters that inherently understand the physical limits of the hardware they control, ensuring that a safe corrective action always exists.

The Geometry of Input-Constrained Safety

To guarantee safety, control systems often rely on the concept of control invariance. A set of states (such as a collision-free region) is considered control-invariant if, for every state within that set, there exists at least one valid control action that keeps the system inside it.

Standard Safe Set (Spatial Boundary Only)
[───────────────────────────────| WALL ]
       ▲
       │ (Filter assumes infinite braking power;
          commands impossible deceleration too late)

Input-Constrained Safe Set (Reshaped by Velocity & Saturation)
[───────────────|────────────────| WALL ]
                ▲
                │ (Boundary pushes back: must brake early 
                   because max torque is limited)

When actuators can saturate, the geometry of the safe set changes dramatically. As pointed out by Cao et al. and Brunke, if a robot's momentum carries it toward an unsafe region and its actuators lack the power to counteract that motion, adjusting safety thresholds online cannot prevent a violation. The system enters an unrecoverable state long before it physically crosses the safety boundary.

Similarly, Goarin et al. demonstrate that ignoring high-order dynamics and motor thrust limits results in commanded maneuvers that hardware cannot execute, leading to collisions during aggressive maneuvers. There is a fundamental mismatch between the aggressiveness of theoretical safety constraints and the actual tracking capabilities of physical actuators (the authors). Therefore, safety frameworks must incorporate hard state and control constraints simultaneously to be viable for practical deployment (Chow et al.).

Reshaping Boundaries via Input-Aware Barrier Functions

The most prominent tool for enforcing safety is the Control Barrier Function (CBF), which acts as a mathematical fence around the safe set. To account for actuator saturation, researchers modify these functions so that the "fence" expands or contracts based on the available control authority.

Several distinct mechanisms have emerged to achieve this:

  • Barrier Pairs: He et al. introduce a two-layer framework that pairs intent inference with a robust "Barrier Pair" control layer. By explicitly accounting for actuator limits during synthesis, it ensures safety guarantees remain valid even when a robot operates at the edge of its performance envelope under external forces.
  • Fixed-Time Recovery: Garg et al. integrate actuator power limits directly into Fixed-Time Barrier Functions (FxTBF). This defines a realistic domain of attraction, ensuring the controller never attempts a maneuver exceeding the hardware's physical capabilities while guaranteeing the system returns to a target state within a fixed time.
  • High-Order Penalty Chains: For complex systems where inputs do not immediately affect safety (high relative degree), Xiao et al. introduce penalty coefficients into a chain of barrier functions. These coefficients serve as tuning parameters to reconcile hard safety boundaries with real-world hardware saturation.

Mathematical Feasibility and Optimization

When physical control limits are too restrictive, standard safety optimization problems can become mathematically infeasible, leaving the controller with no valid action (Wang et al.). To prevent this, the safety function itself must be synthesized alongside the control constraints.

One approach leverages algebraic geometry. Zhao et al. use Sum-of-Squares (SOS) programming to automatically synthesize safety indices by shifting the focus to local manifold positiveness. This ensures a feasible control action always exists within the hardware's torque and velocity envelope. Chen et al. extend this to state-dependent control spaces, allowing the safety boundary to adapt to constraints that change dynamically, such as velocity-dependent acceleration limits.

Alternatively, instead of modifying traditional CBFs, the authors propose Control Barrier Certificates (CBC). By embedding actuator saturation directly as algebraic constraints within an optimization program and focusing purely on the boundary of the safe set, CBCs offer a less conservative approach that stabilizes unstable systems even under strictly limited control authority.

Scaling to High Dimensions and Multi-Agent Swarms

While optimization-based synthesis works well for low-dimensional systems, it struggles to scale. Modern applications use machine learning and distributed architectures to approximate input-constrained safe sets for complex systems.

Neural Control Barrier Functions

Traditional hand-designed safety specifications often fail when required control efforts exceed motor or battery capacities. To address this, Liu et al. train Neural Control Barrier Functions (NCBFs) using a learner-critic architecture that explicitly accounts for actuator saturation, achieving near-perfect safety rates on high-dimensional systems. To stabilize training on highly complex systems like an F-16 jet, So et al. introduce Policy Neural Control Barrier Functions (PNCBFs), which reframe the safety problem as a policy evaluation task to bypass optimization instabilities.

Distributed Multi-Agent Safety

In multi-agent swarms, agents must avoid collisions while respecting their own limited actuation. Frameworks like MA-ICBF (Zinage et al.) combine neural learning with lightweight Model Predictive Control (MPC) safety layers and momentum-based deadlock resolution to handle thousands of agents under motor limits. Similarly, GCBF+ (So et al.) leverages Graph Neural Networks to learn distributed safety certificates that inherently account for physical actuation limits, such as maximum braking force, allowing safe scaling to massive swarms.

[Agent 1] ──(GNN Local Sensing)──► [Graph CBF Layer] ──► [Feasible Safe Command]
                                         ▲
[Agent 2] ──(Actuator Limits)────────────┘

Practical Deployment and Control Authority

Accounting for actuator constraints is not only a theoretical necessity but also changes how robots are deployed in unstructured environments:

  • Aerospace and Robotics: Wu et al. combine CBFs with Gaussian Process regression to avoid kinematic singularities under strict torque limits. In autonomous navigation, Dai integrates raw 3D point cloud processing with a preview controller that respects motor saturation to prevent robots from getting stuck in local minima.
  • Process Control: In micro gas turbine engines, exceeding acceleration or fuel limits can cause catastrophic compressor surge or flameout. Makaryants addresses this by modifying the error signal during neural network training, forcing the controller to learn the engine's narrow physical operational envelope without needing a hand-tuned supervisor.

A final, crucial insight for high-performance systems is the preservation of actuation margin. As demonstrated by Foehn et al. in time-optimal quadrotor flight, planning trajectories at the absolute physical limits of a vehicle leaves no margin for real-time error correction. Effective deployment requires capping planned actuation below maximum capacity to reserve control authority for the feedback controller. Furthermore, Mastalli emphasizes that these torque and hardware boundaries must be enforced as hard inequality constraints rather than soft quadratic penalties; soft penalties routinely allow violations of critical hardware limits during high-speed, dynamic maneuvers.

Key Foundational Work

A framework for time-consistent, risk-averse model predictive control: Theory and algorithms (2014) — contributed 1 concept elements to this direction.
Distributed controller design for a class of sparse singular systems with privacy constraints (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_actuator_saturation_constraints_fb35f5f1,
  title = {actuator saturation constraints},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/actuator-saturation-constraints/rdr_fb35f5f1d9f746a6b63c95ff8e1b6532},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_actuator_saturation_constraints_fb35f5f1_v1,
  title = {actuator saturation constraints},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/actuator-saturation-constraints/rdr_fb35f5f1d9f746a6b63c95ff8e1b6532/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake