Safety-Critical Control via Recurrent Tracking Functions

arXiv:2510.01147 · eess.SY, cs.SY · Submitted 2025-10-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Safety-Critical Control via Recurrent Tracking Functions".

Dev: Here's the gist of the paper. Safety-Critical Control via Recurrent Tracking Functions addresses the challenge of synthesizing safety-critical controllers for high-order nonlinear systems by introducing Recurrent Tracking Functions (RTFs) to relax the monotonic decay requirement in traditional Lyapunov tracking functions, thereby allowing transient deviations while ensuring safety The zero-superlevel set SV is control τ-recurrent, and safety is guaranteed for all initial states in such a set when RTFs are satisfied Layered Control Architecture The paper begins by establishing the setting of a layered control system consisting of a full-order model (FoM) and a reduced-order model (RoM) that interacts through projections and feedback laws The objective is to guarantee the safety of the FoM by designing a RoM feedback law and compact set S such that x ∈ S ⇒ xcl(t) ∈ SFoM,∀t ≥ 0 Safety for the FoM is defined based on the projection of the RoM's safe set, specifically SFoM:= The FoM’s safe set with boundary ∂SFoM = Π−1(h−1(0)) Recurrent Tracking Function Framework Recurrent Tracking Functions (RTFs) are introduced as a relaxation of standard Lyapunov decrease conditions, requiring only that the decrease condition recurrently holds on time instances separated by at most a uniform bound τ Any norm of the tracking error is considered a valid RTF for Lipschitz continuous systems with exponentially stable tracking error The key condition for an RTF is the β-exponential τ-recurrence, which states that min t∈T eβtV(zcl(t), e˙) ≤ V(z0,e˙) Recurrent Control Barrier Function Construction Theorem 3 shows that augmenting a RoM CBF with a valid RTF yields an RCBF hV (z, e˙) = −V (z, e˙) + αeh(z), where SV:= The zero-superlevel set SV is control τ-recurrent This construction guarantees safety for all initial conditions in SV when the condition β > α is satisfied Robustness to Disturbances When modeling model mismatch via a bounded disturbance d, the RTF condition is modified to account for the disturbance-induced offset, leading to an Input-to-State Safety (ISSf) property by defining an enlarged recurrent set Sd = The ISSf for all t ≥ 0, where γ(∥d∥∞) = (2β − α)ι(∥d∥∞)/α This robust safety guarantee ensures that the trajectory remains in the safe state set SFoM despite disturbances Numerical Validation and Application The approach is validated through numerical experiments on a 2D double-integrator example, demonstrating that when the CBF constraint is inactive, this choice of parameters leads to an exponentially convergent tracking error with β = 2.45 and M = 3.24 The results show that only when Theorem 3 and Theorem 4 are satisfied with a valid CBF in the RoM, can we ascertain that the trajectory is safe all the time Future work will focus on developing learning-based approaches for constructing RTFs directly from trajectory data The framework successfully develops a recurrence-based method for layered safety-critical control, using RTFs to relax the classical Lyapunov requirement and guarantee safety through recurrent CBFs The condition β > α ensures that the tracking error converges faster than the system approaches the safety boundary, thereby maintaining h(zcl(t)) ≥ 0 for all time despite the recurrent nature of hV This guarantees that xcl(t) ∈ SFoM,∀t ≥ 0 Acknowledgments The authors would like to thank Aaron Ames for suggesting this problem as an application of recurrence-based analysis The paper was supported by the NSF Global Centers program under Grant No.

Rosa: 2330450 and by the DOE Office of Science (ASCR) under Award No.

Dev: 826565 References

1: W.

Rosa: D.

Title and authors: Rosa: So we're talking about "Safety-Critical Control via Recurrent Tracking Functions" by Jixian Liu and Enrique Mallada. It sounds really technical, but the core idea is making safety-critical control simpler to design for complex systems than before.

Dev: The implication here is that you don't need these massive, complicated full models to guarantee safety if you can handle the tracking error with this new recurrence condition.

Taro: For me, what's interesting is that this allows them to build safety certificates even when they can't construct traditional Control Barrier Functions for the whole system because it gets computationally impossible.

Rosa: Right, so instead of stopping when the math gets too hard to check, you use this recurrence condition to keep the safety guarantee alive even if things drift a bit away from perfect tracking.

Dev: They are essentially replacing that strict monotonic decrease requirement with a weaker finite-time recurrence condition. This is what lets them move past the limitations of traditional Lyapunov functions for these high-order nonlinear systems.

The paper's summary: Rosa: The paper explains that they start by setting up a layered control system, using a full model and a reduced model that talk to each other through some projections and feedback laws.

Dev: To guarantee safety for the big full model, they design the reduced-order model feedback law and a compact set. They define this safe set as the projection of what's safe in the reduced system.

Taro: So it’s like you have a smaller, easier-to-manage model that tells you when to keep things in a certain area so that even the giant system stays safe.

Rosa: That’s right. The paper then introduces Recurrent Tracking Functions as the key idea here because standard Lyapunov functions require that decrease to be monotonic, which isn't always possible for these systems.

Dev: They show that by using RTFs, you can relax that requirement to a weaker condition called beta-exponential tau recurrence. This means the tracking error doesn't have to go down smoothly; it just needs to satisfy this recurrence over time instances separated by at most a uniform bound tau.

The paper's improvements: Rosa: The main improvement they present is constructing these Recurrent Control Barrier Functions, or RCBFs, by augmenting the reduced-order model CBF with this valid RTF.

Dev: They get this function called hV, which they show looks like negative of the Lyapunov function plus some term involving alpha and the RTF. This construction is what allows them to build a safety certificate for all initial conditions within a specific set, which they call the zero-superlevel set, or SV.

Taro: So if you satisfy that beta is greater than alpha, then this whole setup guarantees safety for every single starting point in that set SV. That's a strong claim.

Rosa: It does sound strong, and the paper shows this works even when things get messy because it allows for transient deviations of tracking errors as long as they stay within the defined recurrent set.

Conclusion: Dev: So to wrap up, these Recurrent Tracking Functions allow them to relax the classical Lyapunov requirement and guarantee safety through these recurrent CBFs. The condition beta greater than alpha ensures that the tracking error converges faster than the system approaches the safety boundary, which keeps everything in its safe set for all time.

Rosa: It really boils down to using recurrence-based methods for layered control, letting you bypass those strict monotonic decay requirements and still get a hard safety guarantee.

Taro: From my side, I just think this is crucial because it means we can actually design these safety systems for high-order nonlinear dynamics that are too complex to verify with standard math tools.

Dev: And that's the core of it. The paper "Safety-Critical Control via Recurrent Tracking Functions" gives us a way to build safety certificates without needing those overly restrictive conditions on the tracking error decay.

Jixian Liu, Enrique Mallada

Johns Hopkins University · Department of Electrical and Computer Engineering

eess.SY, cs.SY

Submitted: 2025-10-01

Updated: 2026-11-04

Comments: 12 Pages, 2 Figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 63/100

The gist: Safety-Critical Control via Recurrent Tracking Functions addresses the challenge of synthesizing safety-critical controllers for high-order nonlinear systems by introducing Recurrent Tracking…

Key concepts

Recurrent Tracking Functions (RTFs)
RTFs relax standard Lyapunov conditions by requiring a decrease condition to hold recurrently over a time interval $\tau$. This allows the system to tolerate temporary tracking errors, provided they return to a safe state within the recurrence bound. It's a way to ensure stability without demanding strict monotonic decay.
Layered Control Architecture
This setup uses two models: a full-order model (FoM) for safety and a reduced-order model (RoM) for control. The RoM's safe set is used to constrain the FoM, ensuring that even if the RoM tracking error isn't strictly decreasing, the overall system remains safe by satisfying specific projection and feedback laws.
Recurrent Control Barrier Functions (RCBFs)
RCBFs are constructed by augmenting a standard control barrier function with an RTF. This construction yields a new function $hV$ where the zero-superlevel set SV is control $\tau$-recurrent. This guarantees safety for all initial states within this recurrent set, even under transient conditions.
Input-to-State Safety (ISSf)
This property extends the safety guarantee when the system faces bounded disturbances. By modifying the RTF condition to account for these disturbances, an enlarged recurrent set $S_d$ is defined. This ensures that the trajectory remains within a safe margin proportional to the magnitude of any external disturbance.

Terminology

Summary

Safety-Critical Control via Recurrent Tracking Functions addresses the challenge of synthesizing safety-critical controllers for high-order nonlinear systems by introducing Recurrent Tracking Functions (RTFs) to relax the monotonic decay requirement in traditional Lyapunov tracking functions, thereby allowing transient deviations while ensuring safety The zero-superlevel set SV is control τ-recurrent, and safety is guaranteed for all initial states in such a set when RTFs are satisfied

Layered Control Architecture

The paper begins by establishing the setting of a layered control system consisting of a full-order model (FoM) and a reduced-order model (RoM) that interacts through projections and feedback laws The objective is to guarantee the safety of the FoM by designing a RoM feedback law and compact set S such that x ∈ S ⇒ xcl(t) ∈ SFoM,∀t ≥ 0 Safety for the FoM is defined based on the projection of the RoM's safe set, specifically SFoM:= The FoM’s safe set with boundary ∂SFoM = Π−1(h−1(0))

Recurrent Tracking Function Framework

Recurrent Tracking Functions (RTFs) are introduced as a relaxation of standard Lyapunov decrease conditions, requiring only that the decrease condition recurrently holds on time instances separated by at most a uniform bound τ Any norm of the tracking error is considered a valid RTF for Lipschitz continuous systems with exponentially stable tracking error The key condition for an RTF is the β-exponential τ-recurrence, which states that min t∈T eβtV(zcl(t), e˙) ≤ V(z0,e˙)

Recurrent Control Barrier Function Construction

Theorem 3 shows that augmenting a RoM CBF with a valid RTF yields an RCBF hV (z, e˙) = −V (z, e˙) + αeh(z), where SV:= The zero-superlevel set SV is control τ-recurrent This construction guarantees safety for all initial conditions in SV when the condition β > α is satisfied

Robustness to Disturbances

When modeling model mismatch via a bounded disturbance d, the RTF condition is modified to account for the disturbance-induced offset, leading to an Input-to-State Safety (ISSf) property by defining an enlarged recurrent set Sd = The ISSf for all t ≥ 0, where γ(∥d∥∞) = (2β − α)ι(∥d∥∞)/α This robust safety guarantee ensures that the trajectory remains in the safe state set SFoM despite disturbances

Numerical Validation and Application

The approach is validated through numerical experiments on a 2D double-integrator example, demonstrating that when the CBF constraint is inactive, this choice of parameters leads to an exponentially convergent tracking error with β = 2.45 and M = 3.24 The results show that only when Theorem 3 and Theorem 4 are satisfied with a valid CBF in the RoM, can we ascertain that the trajectory is safe all the time Future work will focus on developing learning-based approaches for constructing RTFs directly from trajectory data

Conclusion and Safety Guarantee

The framework successfully develops a recurrence-based method for layered safety-critical control, using RTFs to relax the classical Lyapunov requirement and guarantee safety through recurrent CBFs The condition β > α ensures that the tracking error converges faster than the system approaches the safety boundary, thereby maintaining h(zcl(t)) ≥ 0 for all time despite the recurrent nature of hV This guarantees that xcl(t) ∈ SFoM,∀t ≥ 0

Acknowledgments

The authors would like to thank Aaron Ames for suggesting this problem as an application of recurrence-based analysis The paper was supported by the NSF Global Centers program under Grant No. 2330450 and by the DOE Office of Science (ASCR) under Award No. p. 0 826565

References

[1] W. D. Compton, M. p. 5 H. Cohen, and A. p. 5 D. Ames, “Learning for layered safety-critical control with predictive control barrier functions,” arXiv preprint arXiv:2412. p. 504658, 2024

[3] C. Wang, Y. Meng, Y. Li, S. L. Smith, and J. p. 5 Liu, “Learning control barrier functions with high relative degree for safetycritical control,” in 2021 European Control Conference (ECC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 D. Compton, T. p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 G. Molnar, and A. p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411. p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 516479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 J. Tomlin, and S. p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 p. 5 L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier–value functions for safetycritical control,” in 2021 60th IEEE Conference on Decision and Control (CDC), IEEE, 2021

[9] M. H. Cohen, N. Csomay-Shanklin, W. D. Compton, T. G. Molnar, and A. D. Ames, “Safety-critical controller synthesis with reduced-order models,” arXiv preprint arXiv:2411.16479, 2024

[8] J. J. Choi, D. Lee, K. Sreenath, C. J.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Recurrent Control Barrier Functions (RCBFs) allow for safety certification of full-order nonlinear systems by relaxing the traditional requirement of monotone decay to a finite-time recurrence condition. This enables safety guarantees for high-order dynamics where constructing valid Control Barrier Functions is computationally intractable.

  2. Layered control architectures, combining a Reduced-Order Model (RoM) CBF with Recurrent Tracking Functions (RTFs), allow the full system to maintain safety even when the RoM tracking error does not decay monotonically but exhibits exponential tau recurrence. The improved AI system can guarantee that trajectories remain in the safe set SFoM for all time despite transient deviations.

  3. The framework allows for robust safety assessment under bounded disturbances by modifying the RTF condition to account for input-to-state stability (ISSf). This modification introduces an enlarged recurrent set Sd, ensuring that the closed-loop system's trajectory remains within a safety margin scaled with the disturbance magnitude.

Sources

Related papers