Managing Constraint Infeasibility in Safety Critical Control

123 concepts111 papers3 proposals

In autonomous systems—such as self-driving cars, quadrotors, and robotic manipulators—controllers must simultaneously achieve performance goals (like reaching a destination) and guarantee safety (like avoiding obstacles). Modern control frameworks achieve this by solving optimization problems, such as Quadratic Programs (QPs) or Model Predictive Control (MPC), in real time. These frameworks translate safety requirements into mathematical constraints that restrict the allowable control actions at any given moment.

However, a fundamental vulnerability haunts these systems: optimization infeasibility. In complex, highly constrained, or uncertain environments, the mathematical constraints can become mutually exclusive. When an optimizer encounters a state where no valid control command satisfies all constraints simultaneously, it fails to return a solution. This mathematical deadlock nullifies safety guarantees, causing the system to freeze, behave erratically, or experience catastrophic failure.

The Geometry of Constraint Conflict

To understand why optimization-based control fails, it helps to look at the geometry of the control space. At each time step, the controller searches for a control input $u$ (such as motor torque or steering angle). This search is bounded by two primary forces:

  1. Actuator Limits: Physical hardware constraints (e.g., maximum motor torque or braking capacity) define a fixed, bounded region in the control space.
  2. Safety Boundaries: Frameworks like Control Barrier Functions (CBFs), pioneered by Aaron D. Ames et al. in Control Barrier Functions: Theory and Applications, define a "safe set" in the system's state space. To ensure the system never leaves this safe set, the CBF generates a half-space constraint on the control input $u$, dictating how aggressively the system must steer away from danger.
Infeasible Scenario:
+-----------------------+
|  Actuator Limit Box   |
|                       |   <--- No Overlap!
+-----------------------+
       \
        \  CBF Safety Constraint (Must stay below this line)
         \__________________________________________________

Point-wise infeasibility occurs when the safety half-space and the actuator bounding box do not intersect. For instance, if a vehicle approaches an obstacle too quickly, the safety constraint might demand a deceleration force that exceeds the physical limits of the brakes. As highlighted by Hengbo Ma et al. in Learning Differentiable Safety-Critical Control using Control Barrier Functions for Generalization to Novel Environments, if the parameters governing these boundaries are not perfectly tuned to the environment's geometry, the controller will fail to find a valid command precisely when a critical safety maneuver is required.

This issue is exacerbated by three main factors:

  • High Relative Degree and Multiple Inputs: When different inputs affect safety at different rates, or when systems are highly complex, coordinating inputs becomes difficult. As Wei Xiao et al. demonstrate in Control Barrier Functions for Systems with Multiple Control Inputs, restricting safety enforcement to a single control component without cross-input coordination frequently triggers mathematical infeasibility.
  • Uncertainty and Worst-Case Margins: When a robot operates in an imperfectly modeled world, controllers must incorporate uncertainty bounds. Fernando Castañeda et al. show in Pointwise Feasibility of Gaussian Process-based Safety-Critical Control under Model Uncertainty that incorporating probabilistic or worst-case uncertainty margins can expand safety requirements so conservatively that they completely swallow the available control authority, leaving the system with zero feasible actions.
  • Environmental Crowding: In dense multi-agent environments, localized safety constraints from surrounding obstacles or agents can directly contradict one another, as explored in Safety Barrier Certificates for Collisions-Free Multirobot Systems. Avoiding obstacle A might force the robot into a trajectory that violates the safety margin for obstacle B.

The Consequences: Deadlocks and Opposing Gradients

When an optimization problem becomes infeasible, the immediate consequence is a lack of control signal. As Kaiyuan Xu et al. observe in Feasibility Guaranteed Traffic Merging Control Using Control Barrier Functions, this failure nullifies the mathematical guarantees of safety, often resulting in collisions because the system cannot execute reactive adjustments.

Even when the optimization remains technically feasible, a related failure mode emerges: spurious local equilibria. This happens when the performance gradient (driving the robot toward its goal) directly opposes the safety gradient (pushing it away from an obstacle). Pol Mestres et al. analyze this in Optimization-Based Safe Stabilizing Feedback with Guaranteed Region of Attraction, demonstrating how robots become permanently trapped in deadlocks away from their targets because the geometry of the safety constraints perfectly balances out the stability goals.

Mechanisms for Guaranteeing Feasibility

To prevent these failures, researchers have developed several core strategies designed to ensure that the optimization problem remains mathematically solvable at every step.

1. Dynamic and Optimal Decay Rates

Traditional CBFs use a fixed linear class-$\mathcal{K}$ function (a tracking parameter denoted as $\alpha$) to dictate how fast a system is allowed to approach its safety boundary. If $\alpha$ is too aggressive, it demands control inputs that the actuators cannot deliver.

To solve this, Bike Zhang et al. introduce an "optimal-decay" framework in Safety-Critical Control using Optimal-decay Control Barrier Function with Guaranteed Point-wise Feasibility. Instead of treating the decay rate as a rigid constant, they turn $\alpha$ into a dynamically optimized variable. If the system approaches an obstacle too quickly or faces tight actuator saturation, the optimizer automatically relaxes the decay rate. This dynamically reshapes the safety half-space, ensuring it always intersects with the physical actuator limits and guaranteeing point-wise feasibility. A similar philosophy of treating decay rates as dynamic variables is leveraged by Jinhao Liu et al. in Flexible Active Safety Motion Control for Robotic Obstacle Avoidance: A CBF-Guided MPC Approach to maintain agility without violating hard boundaries.

2. Hierarchical Relaxation and Resilient Control

When a system encounters extreme external disturbances that exceed its design limits, satisfying every single operational constraint may be physically impossible. In these scenarios, the controller must be allowed to compromise gracefully.

Luiz F. O. Chamon formalizes this concept in Resilient Control: Compromising to Adapt, shifting the perspective from rigid constraint satisfaction to a trade-off framework where constraint violations are penalized rather than treated as hard failures.

In practice, this is often implemented via a prioritized "rulebook." For example, Wei Xiao et al. propose a recursive optimization framework in Rule-based Optimal Control for Autonomous Driving. If a vehicle faces an unavoidable conflict, the controller systematically relaxes lower-priority rules (such as maintaining a speed limit or staying perfectly centered in a lane) to ensure that the highest-priority constraint—absolute collision avoidance—remains mathematically feasible and fully satisfied.

3. Predictive Safety Filtering and Look-Ahead Constraints

Point-wise feasibility guarantees that a valid control input exists right now, but it does not prevent the system from steering into a "blind alley"—a state that is safe today but guarantees infeasibility tomorrow.

To address this long-horizon challenge, Kim P. Wabersich et al. develop "predictive control barrier functions" in Data-Driven Safety Filters: Hamilton-Jacobi Reachability, Control Barrier Functions, and Predictive Methods for Uncertain Systems. By combining soft constraints with iterative tightening over a predictive horizon and anchoring the trajectory with a terminal CBF, their safety filter ensures that the system remains within a region of the state space where future optimization problems are guaranteed to stay feasible.

Predictive Safety Filter:
Current State ---> [ Look-Ahead Horizon Evaluation ] ---> Terminal CBF (Guaranteed Safe Zone)
                        |
                        +---> Iterative Tightening prevents entering "blind alleys"

Similarly, Kaiyuan Xu et al. inject explicit control-dependent feasibility constraints directly into the optimization process in Feasibility Guaranteed Traffic Merging Control Using Control Barrier Functions, mathematically blocking the system from entering states where future actuator saturation would conflict with safety boundaries.

By shifting from rigid, isolated constraints to adaptive, hierarchical, and predictive optimization structures, these approaches ensure that safety-critical controllers remain mathematically viable and operationally safe under real-world conditions.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 100 papers
  2. Version 1 · 2026-10-04 · 172 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_optimization_infeasibility_f29f233b,
  title = {optimization infeasibility},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optimization-infeasibility/rdr_f29f233b7f6b40c4a3a8f0f90a3c48d8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_optimization_infeasibility_f29f233b_v2,
  title = {optimization infeasibility},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optimization-infeasibility/rdr_f29f233b7f6b40c4a3a8f0f90a3c48d8/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake