⚡ electrical engineering

Integral Control Barrier Functions with Input Delay: Prediction, Feasibility, and Robustness

This paper proposes a robust safety-critical control framework for nonlinear systems with input delay that integrates predictor feedback with Integral Control Barrier Functions (ICBFs) to simultaneously compensate for delays, guarantee feasibility under joint state and input constraints, and ensure forward invariance.

Original authors: Adam K. Kiss, Ersin Das, Tamas G. Molnar, Aaron D. Ames

Published 2026-03-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Adam K. Kiss, Ersin Das, Tamas G. Molnar, Aaron D. Ames

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are driving a car that has a very strange, broken steering wheel. Every time you turn the wheel, the car doesn't react immediately. Instead, it waits exactly 1.2 seconds before it actually starts to turn.

This is the problem of Input Delay. In the real world, this happens with self-driving cars, robots, and drones. There is always a tiny lag between when the computer says "turn left" and when the motors actually do it.

If you are driving fast and a wall appears, and you wait 1.2 seconds to turn, you might crash. If you try to turn too hard to make up for the delay, you might spin out of control.

This paper is about building a "smart safety system" for these laggy machines. Here is how they solved it, using simple analogies:

1. The Problem: The "Blind" Driver

Standard safety systems act like a driver who only looks at the road right now.

The Scenario: You see a wall 10 meters away.
The Mistake: The computer says, "Brake!" But because of the delay, the brakes don't actually clamp down until you are already 5 meters from the wall.
The Result: You crash.

2. The Solution: The "Crystal Ball" (Predictor Feedback)

The authors' first big idea is to stop looking at the now and start looking at the future.

They give the computer a Crystal Ball (mathematically called a "Predictor").

Instead of asking, "Where am I right now?", the computer asks, "Where will I be in 1.2 seconds if I keep doing what I'm doing?"
The safety system then checks if that future position is safe.
The Analogy: Imagine you are throwing a ball to a friend who is running away. You don't aim at where they are standing now; you aim at where they will be when the ball arrives. This paper teaches the robot to "aim" its safety commands at the future, not the present.

3. The Double Constraint: The "Tightrope Walker"

Real machines have two rules they must follow at the same time:

Don't hit the wall (State Constraint).
Don't break the engine (Input Constraint). For example, the brakes can only push so hard before they overheat or the wheels lock up.

Usually, safety systems check these rules separately. But what if the only way to avoid the wall is to slam the brakes so hard that you break the engine? The system gets stuck. It's like a tightrope walker who is told, "Don't fall off the rope" AND "Don't step on the rope." If the rope is the only thing keeping them up, they are in a paradox.

The Paper's Fix:
They created a special mathematical "compatibility check." Before the robot tries to move, it checks: "Is there a way to avoid the wall without breaking the engine?"

If yes, it finds that perfect path.
If no, it knows the situation is impossible and stops trying to force a solution that doesn't exist. This prevents the computer from freezing or making a crazy, dangerous move.

4. The "Crystal Ball" is Sometimes Wrong (Robustness)

Here is the tricky part: The Crystal Ball isn't perfect.

Maybe the delay isn't exactly 1.2 seconds; maybe it's 1.3 seconds.
Maybe the road is slippery, and the car doesn't slide exactly as the math predicted.

If the Crystal Ball is wrong, the robot might think it's safe when it's actually about to crash.

The Paper's Fix:
They added a "Safety Buffer" (Robustness).

Instead of checking if the future position is exactly safe, they check if it's safe even if the prediction is a little off.
The Analogy: Imagine walking through a dark tunnel. A normal person walks right down the middle. A "Robust" person walks slightly to the side, keeping a hand on the wall, just in case they stumble.
The paper creates a mathematical "cushion" around the safety zone. If the prediction is slightly wrong, the robot is still safe because it was already keeping a little distance.

Summary: The Three Magic Tricks

Look Ahead: Don't react to the present; react to the future (Predictor Feedback).
Check the Math: Make sure the rules for "don't crash" and "don't break" don't fight each other (Feasibility).
Keep a Buffer: Assume your crystal ball might be slightly wrong, and stay extra safe to compensate (Robustness).

The Real-World Test

They tested this on a Self-Driving Car (Adaptive Cruise Control).

Without their system: The car saw a slow car ahead, waited too long to brake, and crashed into it.
With their system: The car predicted where it would be in 1.2 seconds, calculated the perfect braking force that wouldn't lock the wheels, and smoothly slowed down, keeping a safe distance even though the brakes were "slow."

In short, this paper teaches robots how to drive safely even when they have a slow reaction time and imperfect information, by always looking a little bit into the future and keeping a little extra space just in case.

1. Problem Statement

The paper addresses the challenge of safety-critical control for nonlinear systems subject to input delays and joint state and input constraints.

Input Delay: In real-world systems (e.g., autonomous vehicles, robotics), actuation, computation, and communication introduce time delays ( $\tau$ ). A controller evaluating safety conditions based on the current state $x(t)$ fails because the control input $u(t)$ only affects the system at $t+\tau$ . This mismatch often leads to safety violations (state constraint breaches) or controller infeasibility (input constraint breaches).
Joint Constraints: Standard Control Barrier Functions (CBFs) typically handle state constraints. However, systems often have simultaneous limits on both states (e.g., safe distance) and inputs (e.g., actuator saturation). When these are combined in a Quadratic Program (QP), the problem can become infeasible even if individual constraints are satisfiable, leading to a loss of safety guarantees.
Uncertainty: Practical delays are often uncertain. Using a predicted state based on an incorrect delay estimate introduces a prediction error, acting as a disturbance that can destabilize the system or violate safety.

2. Methodology

The authors propose a framework integrating Predictor Feedback with Integral Control Barrier Functions (ICBFs) to handle delays and joint constraints simultaneously.

A. Predictor Feedback Integration

To compensate for the delay $\tau$ , the framework employs predictor feedback. Instead of evaluating safety at the current state $x(t)$ , the controller evaluates conditions on a predicted state $x_p(t) = x(t+\tau)$ , obtained by forward-integrating the system dynamics over the delay interval using the known input history.

The control law is dynamically defined (integral controller): $\dot{u} = \phi(x_p, u) + v$ , where $v$ is the safety correction computed via a QP.
This shifts the safety evaluation to the time when the input will actually take effect, effectively creating a delay-free loop for the safety filter.

B. Integral Control Barrier Functions (ICBFs)

The paper utilizes ICBFs, which extend standard CBFs to handle input constraints and actuator dynamics.

State Constraint ( $h_x$ ): Encoded via an extended barrier function $h_e(x, u)$ that depends on both state and input.
Input Constraint ( $h_u$ ): Directly encoded as a barrier on the input $u$ .
Safety Filter: A QP is formulated to find the minimum-norm correction $v$ that satisfies the derivative conditions for both $h_e$ and $h_u$ simultaneously.

C. Feasibility Analysis

A major theoretical contribution is the derivation of a closed-form feasibility condition for the two-constraint QP.

Using a variation of Farkas' Lemma, the authors derive conditions under which the QP is guaranteed to have a solution.
They identify that standard ICBF conditions are insufficient for joint constraints; a specific compatibility condition (Condition 34c in the paper) must be met when the gradients of the state and input barriers are anti-parallel. This condition guides the design of the nominal controller $\phi$ to ensure the QP remains feasible.

D. Robustness to Prediction Errors

To handle uncertainty in the delay estimate ( $\hat{\tau} \neq \tau$ ), which causes a prediction error $d(t)$ , the authors introduce Tunable Robust ICBFs.

They modify the QP constraints by adding robustness terms ( $r_i$ ) that depend on the barrier value and the magnitude of the prediction error bound.
These terms are tunable (via parameters $\mu_i, \sigma_i$ ) to concentrate robustness near the safety boundary while minimizing conservatism in the interior of the safe set.
Theoretical guarantees show that if the prediction error is small, the system remains strictly safe. If the error is large, the system remains within a bounded "inflated" safe set (Input-to-State Safety).

3. Key Contributions

Predictor-Based ICBF Framework: The first integration of predictor feedback into the ICBF framework, enabling safety-critical control for nonlinear systems with input delays and joint constraints.
Closed-Form Feasibility Condition: Derivation of a necessary and sufficient condition (Proposition 1) for the feasibility of the two-constraint QP. This ensures that a compatible pair of ICBFs exists, guaranteeing that the safety filter does not fail due to conflicting constraints.
Robust ICBF Formulation: Development of a robust safety filter that treats prediction errors as disturbances, providing formal safety guarantees even when the delay estimate is inaccurate.
Validation: Demonstration of the framework on an Adaptive Cruise Control (ACC) system with actuation delay.

4. Results

The framework was validated using an Adaptive Cruise Control (ACC) simulation with an actuation delay of $\tau = 1.2$ seconds.

Delay Compensation: A naive controller (ignoring delay) violated both state (safe distance) and input (acceleration) constraints. The proposed predictor-based controller successfully compensated for the delay, maintaining safety and replicating the behavior of a delay-free system.
Feasibility: The simulation confirmed that the specific design of the barrier function (including a braking-distance term) satisfied the derived feasibility condition, preventing QP infeasibility during aggressive braking.
Robustness: In a scenario with a 50% underestimation of the delay ( $\hat{\tau} = 0.6$ s), the non-robust controller failed (violated the safe distance). The proposed robust ICBF controller, utilizing the tunable robustness terms, proactively adjusted the control input to maintain both constraints despite the significant prediction error.

5. Significance

This work bridges a critical gap between theoretical safety control and practical implementation in systems with delays.

Safety Assurance: It provides a rigorous mathematical guarantee that safety constraints can be satisfied even when control inputs are delayed and actuators are saturated.
Practical Applicability: By addressing prediction errors, the method is robust enough for real-world deployment where exact delay models are rarely available.
Theoretical Advancement: The derivation of the closed-form feasibility condition for joint constraints solves a long-standing issue in CBF literature regarding the compatibility of multiple constraints in optimization-based control.

In summary, the paper presents a comprehensive solution for safe control of delayed nonlinear systems, combining prediction, dynamic control laws, and robust optimization to ensure forward invariance of safe sets under realistic operating conditions.