acceptodds
Under review as a conference paper at ICLR 2027

Event-Triggered Structured Active Exploration with Gaussian Process Safety Filters

Abstract

For robotic systems operating under uncertain environmental conditions, accurate system models are often unavailable. To ensure their safety despite this uncertainty, the integration of Gaussian process (GP) models into safety filters based on control barrier function (CBF) has been shown to be promising. Since these GP-CBF safety filters become infeasible when the model uncertainty is too large, they are typically updated online, but existing strategies for data generation insufficiently cover the control space, introduce excessive conservatism, and temporarily compromise safety. We address this problem by proposing an event-triggered structured exploration approach for safe and sample-efficient learning of CBF-based safety filters using GPs. Before infeasibility issues can occur, our approach activates a safe exploratory control sequence, which is constructed via orthogonal perturbations of a safe control direction. Under standard regularity assumptions and sufficient CBF robustness, we prove high-probability recursive feasibility and safety, as well as finite-trigger and thereby sample complexity guarantees. Experiments on two safe robot control benchmarks show improved safety rates with lower conservatism and lower update frequency than state-of-the-art baselines. Manipulator simulations demonstrate that the GP model underlying the safety filter can be trained solely using distance-derived labels for certain system classes.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Gaussian processes, control barrier functions, safety filters, safe learning, active exploration, robot control

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_AcbJOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper describes a new strategy for event-triggered active exploration of a control-affine nonlinear system with Gaussian Processes. While there is no Knowledge About a global Control Barrier Function, active Exploration is required in order to learn the Dynamics while maintaining safety at all times. The active Exploration triggering is required in order to obtain feasibility of the safety filter given by a Control Barrier Function. The authors extend Prior work by introducing a more structured Exploration scheme in the space of Control inputs with finite trigger times and thereby providing sample complexity guarantees and recursive feasibility.

Critical strengths
  • The authors provide a thorough review of each decision made. They discuss assumptions and embed their work thoroughly into the current research state.
  • The proposed method leads in the given examples to no safety violations and therefore safe autonomous Operation
  • The method has sample complexity bounds
Critical weaknesses
  • I do not understand Figure 1. Please elaborate more on the main differences and the fundamental different uncertainties in the four cases. Why do the arrows start in the unsafe area? You could delete the outline for the paper in lines 105-107 if you need more space.
  • The improvement seems incremental, comparing to Lederer2026c and Castaneda2025.
  • The implementation of highest uncertainty directions, as meantioned in future work, is from my point of view a good additional ablation for the current paper.
  • There seems to be no code given
Questions for authors
  • The introduction of a full orthonormal basis is a scaling issue, when considering high dimensional systems. Where do you see the boundary for your method ?
  • The same holds for the restriction of the ∥⋅∥2\Vert \cdot \Vert_2 unit ball. Have you seen issues with neglecting possible control Inputs which are in the ∥⋅∥∞\Vert \cdot \Vert_\infty unit ball, but not in the ∥⋅∥2y\Vert \cdot \Vert_2y unit ball ?
3 comments
AuthorsOpenReview ↗

Response Part 1

Clarification of Figure 1

We thank the reviewer for this suggestion. Fig. 1 illustrates the exploration strategies and corresponding GP predictive uncertainty distributions in the control-input space at a fixed system state. The circular boundary represents the control-input magnitude constraint, the background color indicates the GP predictive uncertainty of the CBF derivative (blue: low; red: high), and the shaded region denotes unsafe control inputs. The arrows represent control-input vectors drawn from the origin, with their endpoints corresponding to the applied inputs. Thus, an arrow passing through the shaded region does not imply that the intermediate unsafe inputs are executed.

The four strategies differ in their exploration patterns: Time-triggered learning does not actively select exploration directions; Single Point primarily reduces uncertainty near a single safe direction; Bandit explores multiple directions but may select unsafe inputs; and our Structured Exploration uses orthogonal perturbations around a safe anchor direction to cover multiple control directions while maintaining safety.

We have clarified the graphical elements of Fig. 1 in the Introduction, immediately before the discussion of Single Point and Bandit exploration.

Novelty and Significance:

We thank the reviewer for the comment on the novelty of our work. We believe that, compared with Lederer et al. and Castañeda et al., our approach represents a substantial advance in exploration design, theoretical guarantees, and empirical performance, rather than an incremental improvement.

1. Our experimental results demonstrate significant performance improvements. As shown in Table 1, compared with the Single Point method, our approach reduces the average trigger frequency from 13.81 Hz to 1.22 Hz on the Quadrotor and from 15.40 Hz to 0.043 Hz on the AUV, while reducing CBF violations from 3 and 32 to zero, respectively. These results demonstrate that our method substantially reduces exploration frequency while improving safety.

2. Our work introduces a fundamentally different exploration mechanism with new theoretical guarantees. Unlike existing approaches that primarily explore along a single safe control direction to maintain safety-filter feasibility, our method employs structured multidirectional exploration to systematically reduce GP uncertainty across the entire control-input space. Through a novel GP variance decomposition, we establish a uniform posterior-variance bound over the bounded control-input domain and further derive a finite bound on the number of exploration triggers, together with a corresponding sample-complexity guarantee. These theoretical results are not provided by either of the two aforementioned works.

3. Our approach improves practical applicability and broadens experimental validation. Compared with prior approaches relying on idealized continuous-time or instantaneous exploration assumptions and exploration designs without predefined control-amplitude bounds, our method explicitly accounts for safety over finite sampling intervals and constructs safe exploration sequences within bounded control-input constraints. We also extend the experimental evaluation from relatively simple cruise-control and kinematic-vehicle models to more complex robotic systems, including a Quadrotor, an AUV, and a seven-degree-of-freedom Franka manipulator.

AuthorsOpenReview ↗

Response Part 2

Ablation on Maximum-Uncertainty Directions

We thank the reviewer for this suggestion. We have conducted an additional ablation study on the 7-DoF Franka Panda manipulator, where the structured exploration inputs are prioritized according to their GP predictive uncertainty. After each GP update, the original triggering condition is re-evaluated to enable early stopping.

Compared with full structured exploration, this heuristic reduces the number of online training samples by 40.40% and 26.70% for the Closed-Form CBF and Point-Wise CBF Evaluation settings, respectively, with zero CBF violations and SOCP infeasibilities in all 100 trials per setting.

These results demonstrate the potential of uncertainty-prioritized exploration with early stopping to reduce online sampling requirements. We have added the experimental procedure and quantitative results in Appendix G.5 (Uncertainty-Prioritized Exploration with Early Stopping). Theoretical guarantees for this heuristic extension remain an interesting direction for future work.

The quantitative results of this ablation are summarized in the following table.

Table: Ablation of uncertainty-prioritized exploration with early stopping on the Franka Panda manipulator.

MetricClosed-Form CBF (Full)Closed-Form CBF (Early)Point-Wise CBF Eval. (Full)Point-Wise CBF Eval. (Early)
Online Training Samples6.51±3.626.51 \pm 3.623.88±3.013.88 \pm 3.0155.44±24.5355.44 \pm 24.5340.64±17.9240.64 \pm 17.92
CBF Violations0000
SOCP Infeasibilities0000
Path Overhead Ratio1.127±0.0281.127 \pm 0.0281.106±0.0191.106 \pm 0.0191.172±0.0551.172 \pm 0.0551.148±0.0431.148 \pm 0.043
Minimum Clearance (cm)2.59±0.492.59 \pm 0.492.20±0.642.20 \pm 0.643.65±1.013.65 \pm 1.013.59±0.753.59 \pm 0.75
Trigger Frequency (Hz)0.186±0.1030.186 \pm 0.1030.352±0.2480.352 \pm 0.2480.396±0.1750.396 \pm 0.1751.000±0.4071.000 \pm 0.407
Min. Trigger Interval (steps)8.60±2.178.60 \pm 2.173.88±2.853.88 \pm 2.8520.33±8.1920.33 \pm 8.192.72±2.242.72 \pm 2.24

Scalability to High-Dimensional Control Spaces

We thank the reviewer for this question. We acknowledge that constructing a full orthonormal basis introduces additional exploration costs as the control dimension increases. For an mm-dimensional control input, each triggered exploration requires mm control inputs over a total duration of mTsmT_s. Higher dimensions may therefore require a smaller sampling period TsT_s to limit state variations during exploration, while GP computational complexity also needs to be considered.

Our approach primarily targets robotic systems with moderate control dimensions, which represent an important class of practical applications. The experiments on the 7-DoF Franka Panda manipulator further demonstrate its applicability in such settings.

Moreover, the additional ablation in Appendix G.5 shows that uncertainty-prioritized exploration with early stopping reduces online training samples by 26.70%–40.40%, indicating a promising way to mitigate exploration costs and potentially extend the approach to higher-dimensional control spaces. We therefore consider moderate-dimensional robotic systems the primary application domain of the current method, with further scalability depending on exploration length and computational costs.

Unit Euclidean Ball Constraint

We thank the reviewer for this question. The ℓ2\ell_2 unit ball is primarily adopted to exploit its geometric structure for constructing safe orthogonal exploration directions. This exploration scheme can also be combined with a standard GP-CBF safety filter operating over the full ℓ∞\ell_\infty box.

Specifically, structured exploration establishes a uniform GP posterior variance bound over the (ℓ2\ell_2) unit ball using observations along orthogonal directions. Combined with the robust feasibility margin in Assumption 3, this bound ensures the existence of a feasible input within the ball after exploration, under the stated sufficient conditions. Since U2⊆U∞\mathcal U_2\subseteq\mathcal U_\infty, this input also remains feasible for the box-constrained safety filter, allowing normal control to exploit additional inputs outside the ball. By extending the GP confidence and sampling-time drift bounds to the box domain, the resulting control inputs can satisfy the same GP-CBF safety condition.

We have clarified this extension in Section 3.1, following Eq. (4), and Appendix D.5.

Code Availability

We thank the reviewer for highlighting the importance of code availability. As stated in our Reproducibility Statement, we plan to release the source code under an open-source license upon acceptance.

Thank you for the rebuttal. I have read it and also the remaining reviews and author rebuttals. The clarifications on Figure 1, the added ablation and the explicit feasibility statement address my questions.

I tend towards acceptence, while acknowleding the paper as borderline. I will stick to my score.

Reviewer_qPAkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose an event-triggered exploration strategy for learning Gaussian-process-based control barrier function safety filters, using structured control perturbations to reduce uncertainty while maintaining safety. The effectiveness of the approach is illustrated on several benchmarks, showing improved empirical performance.

Critical strengths
  • The problem considered is an important one
  • The proposed solution is original, and it shows good empirical performance
  • The approach is generally sound
Critical weaknesses
  • The theoretical bounds depend on several constants whose practical computation is not clearly explained. The authors should clarify how valid bounds on these quantities can be obtained, particularly when the underlying dynamics are unknown.

  • Relatedly, the experiments do not clearly report how these constants are determined or whether the selected values satisfy the assumptions and sufficient conditions of the theoretical results. When heuristic estimates are used, it is important to assess how using certified bounds would affect performance. At a minimum, the authors should distinguish the empirical findings from the formal guarantees and explain what conclusions remain justified when the required conditions have not been verified.

  • Section 3 states that the proposed approach guarantees recursive feasibility. This is an important property and should be explicitly included in a theorem statement in the main text, together with its precise conditions. In the current version of the main text of the paper, I could not find a formal statement and proof of the recursive feasibility of the resulting algorithm.

Questions for authors
  • Can you clarify how the various constants in various theorems can be formally computed? Also, how do you compute them in the experiments? If heuristics are employed, please explicitly discuss how these affect the conclusions and results of the paper.

  • Can you add a formal theorem and proof about recursive feasibility?

3 comments
AuthorsOpenReview ↗

Response Part 1

Computation of constants for theoretical guarantees

We thank the reviewer for pointing out the missing clarity on how the constants in our theoretical guarantees can be obtained. Given full knowledge of the system dynamics and the constraints, all parameters can be directly determined or rigorously estimated. For the RKHS norm bound BB, we refer to Appendix A.2, which discusses this in detail. The constants ϵ,θ\epsilon, \theta (Assumption 3) can be found via an optimization problem. Finally, parameters such as the dynamics variation bound Δ\Delta and CBF constraint Lipschitz constants Lϕ,LψL_{\phi},\boldsymbol L_\psi can be directly determined using knowledge of the dynamics and the CBF. Based on these parameters, all conditions of our theoretical results can be straightforwardly checked. Note that extensions to sufficiently large data sets instead of exact knowledge of the dynamics are possible.

However, our approach is designed for scenarios in which we neither have large prior data sets nor accurate knowledge of the uncertain dynamics components. In these scenarios, it is practically impossible to obtain verifiable constants. We want to emphasize that the required prior knowledge that we impose seems unavoidable by any safe learning algorithm in such scenarios. Without a regularity condition such as the RKHS norm bound BB (Assumption 2), a data point provides no information anywhere except at the exact location that it was collected. Similarly, any model error can cause a safety violation if the CBF condition has no robustness margin (Assumption 3). This issue of constants that are difficult or impossible to verify is typically dealt with by considering them as design parameters in the literature on safety using kernel-based methods, see e.g., [1-4]. We also handle it this way.

While it may seem that such a parameter tuning approach leads to a heuristic algorithm, this is not true. By specifying the parameters β,ϵ,η1,η2\beta, \epsilon, \eta_1, \eta_2 for Algorithm 1 directly, we implicitly describe for which class of dynamics and which robustness margin range our safety guarantees hold. Importantly, this approach guarantees that the safety guarantees hold for some class of nonlinear dynamical systems under weak requirements on the parameters. If the unknown ground-truth parameters that we cannot verify in practice fall into this class, safety is guaranteed. However, a violation does not immediately imply safety constraint violations since our conditions are only sufficient for safety. Given the absence of knowledge of the ground-truth parameters, this approach additionally allows us to empirically tune the parameters used in Algorithm 1 to obtain strong empirical performance as it is similarly done in related work concerning safety [1-4]. Therefore, using such design parameter is an intuitive and practical approach to handling such uncertainty in our opinion.

To illustrate the computation of all relevant constants, we conducted additional experiments with a simplified dynamical system and constraints in an idealized scenario, in which knowledge of the ground-truth dynamics is available. The results (Appendix H) demonstrate that Algorithm 1 is implementable under the assumptions posed in our theoretical results and works as expected. Thereby, these results complement our prior experimental results that show that Algorithm 1 is also effective at maintaining safety in practice when these assumptions cannot be verified and potentially do not hold.

AuthorsOpenReview ↗

Response Part 2

Constants used in experiments

We are grateful for the reviewer pointing out the missing details about the employed constants in our numerical experiments. As explained before, we consider the parameters of Algorithm 1 as design choices that we manually tuned. The detailed description of the process we followed for tuning the parameters is provided in Appendix B. As we pointed out before, the theoretical guarantees stay intact with such manually tuned parameters for some class of dynamical systems and CBF constraints, but we did not rigorously check or verify that the considered robots and safety constraints lie in this class since we also would not be able to do that in practice. Thus, no guarantees hold for the specific robots in our experiments in a strict sense. However, we want to stress that this is also the case in existing works on kernel-based methods for safety critical learning, e.g., [1-4], and we are currently not aware of any work that verifies required constants for more complex sequential decision making problems with similar limitations on prior knowledge or in a practical scenario. If the reviewer is aware of such papers, we would appreciate hearing about them.

We want to highlight that our motivation for the numerical experiments was to evaluate how Algorithm 1 performs in practical settings. Note that only such scenarios make a comparison of safety constraint violations meaningful as otherwise our theoretical guarantees (and those of comparative approaches) would exclude them in the first place. We point this out more clearly to distinguish explicitly between empirical findings and theoretical guarantees in the revised version. In the discussion of numerical results, we revised certain statements to keep a clear separation between empirical findings and theoretical guarantees by merely suggesting that our theory can provide valuable insight on certain observed effects. Additionally, we demonstrate the usage of verified ground-truth parameters in a newly added problem with simplified setting mentioned earlier.


Additional References

[1] As, Yarden, Lenart Treven, Carmelo Sferrazza, Stelian Coros, and Andreas Krause. "Actsafe: Active exploration with safety constraints for reinforcement learning." In Proceedings of International Conference on Learning Representations, pp. 29752-29776. 2025.

[2] Wendl, Manuel, Yarden As, Manish Prajapat, Anton Pollak, Stelian Coros, and Andreas Krause. "Safe Exploration via Policy Priors." In Proceedings of International Conference on Learning Representations, 2026.

[3] Wu, Zhixuan, Rui Yang, Lei Zheng, and Hui Cheng. "Safe learning-based feedback linearization tracking control for nonlinear system with event-triggered model update." IEEE Robotics and Automation Letters, 7(2), 3286-3293, 2022.

[4] Capone, Alexandre, Ryan Cosner, Aaron Ames, and Sandra Hirche. "Learning safe control via on-the-fly bandit exploration." Proceedings of the International Conference on Machine Learning, 2025.


Clarification of Recursive Feasibility

We thank the reviewer for this suggestion. The feasibility of the GP-CBF-SOCP is an essential part of the probabilistic safety analysis in Appendix D.4. To make this guarantee explicit, we have rephrased Theorem 1 (Recursive Feasibility and Probabilistic Safety) in Section 4.1 to explicitly state that the GP-CBF-SOCP remains feasible whenever it is invoked by Algorithm 1. We have also added the corresponding feasibility argument more explicitly to the proof in Appendix D.4, alongside the existing safety analysis of the structured exploration sequence.

The revised statement of Theorem 1 is provided below.

Theorem 1 (Recursive Feasibility and Probabilistic Safety). Consider the continuous-time system (1) and a continuously differentiable CBF h(x)h(\mathbf{x}). Suppose Assumptions 1, 2 and 3 are satisfied so that Γ=θ−2β(m+1)3σϕ,02+2(m+1)3ϵ2σ‾2>0.\Gamma = \theta - 2\beta \sqrt{(m+1)^3\sigma_{\phi,0}^2 + \tfrac{2(m+1)^3}{\epsilon^2}\underline{\sigma}^2}>0. Furthermore, assume that the GP mean and variance defined in Eq. (3) are based on differentiable stationary kernels kϕ(⋅,⋅),kψ1(⋅,⋅),…,kψm(⋅,⋅)k_\phi(\cdot,\cdot), k_{\psi_1}(\cdot,\cdot),\ldots,k_{\psi_m}(\cdot,\cdot). If at most Nˉ∈N\bar{N}\in\mathbb{N} training samples are generated and Eq. (6b) does not hold at t=0t=0, there exist thresholds Ξ1(Nˉ,δ),Ξ2(Γ,β,ϵ),Ξ3(Ts),Ξ4(Ts,β)∈R≥0\Xi_1(\bar{N},\delta),\Xi_2(\Gamma,\beta,\epsilon),\Xi_3(T_s),\Xi_4(T_s,\beta)\in\mathbb{R}_{\geq0}, such that the control signal (7) and the event trigger (6) resulting from Algorithm 1 with parameters β≥Ξ1(Nˉ,δ),Ts≤Ξ2(Γ,β,ϵ),η1≥Ξ3(Ts),η2≥Ξ4(Ts,β)\beta\geq \Xi_1(\bar{N},\delta),\qquad \sqrt{T_s}\leq \Xi_2(\Gamma,\beta,\epsilon),\qquad \eta_1\geq \Xi_3(T_s),\qquad \eta_2\geq \Xi_4(T_s,\beta) ensures that, at every sampling instant tkt_k at which Algorithm 1 invokes the GP-CBF-SOCP (4) in line 16, a feasible control input exists. Moreover, the safety of the closed-loop system is guaranteed, i.e., x(t)∈C\mathbf{x}(t)\in\mathcal{C} for all t≥0t\geq0, with probability 1−δ1-\delta.

We sincerely thank the reviewers for their thoughtful feedback. We have provided point-by-point responses and revised the manuscript to address the concerns raised. We would greatly appreciate any further feedback, particularly if any concerns remain unresolved, and would be happy to provide additional clarification.

Reviewer_8tySOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an event-triggered structured exploration method for GP-based CBF safety filters with control-dependent model uncertainty. High-probability safety and recursive feasibility are established throughout both normal control and exploration. The sample complexity of event-triggered learning is also analyzed, and an upper bound on the total number of exploration events is derived. Extensive simulation experiments demonstrate the effectiveness of the proposed method. Overall, this paper is well-written.

Critical strengths
  1. The paper addresses the important problem of maintaining safety during online learning under control-dependent model uncertainty.

  2. The structured exploration design is intuitive and exploits the control-affine structure to collect informative data across multiple control directions while preserving safety.

  3. The paper provides theoretical justification for the safety and recursive feasibility of the proposed method, together with a sample-complexity analysis of event-triggered learning.

  4. Extensive simulation experiments are conducted to evaluate the proposed method against representative baselines.

Critical weaknesses
  1. The connection between the theoretical safety guarantees and the practical parameter choices is not sufficiently clear.

  2. There are several issues in the theoretical analysis that require clarification or correction. Specific concerns are detailed in the Questions section.

Questions for authors
  1. How is the maximum GP standard deviation in the triggering condition (6b) computed in practice?

  2. Given the switching between normal control and exploratory inputs in Eq. (7), could the authors explain why the trajectories in Fig. 2 appear smooth without noticeable disturbances caused by exploration?

  3. Could the authors provide more details on how the choice of β\beta in Eq. (12) guarantees the uniform high-probability error bound in Eq. (14)?

  4. The transition from Eq. (99) to Eq. (100) is unclear. How does the upper bound on the accumulated information gain lead to the bound on the number of exploration events?

  5. Could the authors clarify the derivation of Eq. (102), particularly how the definition of Ξ~4\widetilde{\Xi}_4 in Eq. (97) yields the stated lower bound?

6 comments
AuthorsOpenReview ↗

Response Part 1

Connection between theoretical safety guarantees and practical parameter choice

We thank the reviewer for highlighting the insufficiently explained connection between our theoretical guarantees and practical parameter choices. It is possible to estimate all of these parameters in principle when enough information about the ground-truth dynamical system is available, e.g., via a global optimization problem for θ,ϵ\theta, \epsilon (Assumption 3) or as discussed in Appendix A.2 for the RKHS norm BB (Assumption 2). However, there is admittedly a gap between the parameters considered for the theoretical results and the practical parameter choice, which mainly comes from the fact that we cannot verify the value of these parameters in practice in our setting because we consider no/very little prior data and unknown dynamic components. Note that such regularity knowledge (BB) is essential to extrapolate from data, while the robustness margin is crucial to admit small learning errors. To see this, observe that the model error can change arbitrarily without regularity assumptions, such that it is impossible to guarantee any constraint on Eq. (1) when away from training samples. Similarly, without any robustness margin, an arbitrarily small learning error can make it impossible to satisfy the CBF constraint.

We address this issue of unknown parameters that are required for Algorithm 1 by treating them as design choices. We want to stress that this is the common approach in the literature on kernel-based methods for safe learning, see e.g., [1-4]. Additionally, we treat the constants η1,η2\eta_1, \eta_2 for compensating the effects of digital implementation as tuning parameters since Lipschitz-based time discretization bounds [5] tend to be overly conservative [6]. We describe in detail in Appendix B how we tune these parameters in practice. Moreover, we provide a numerical analysis how their choice affects the performance of our algorithm in Subsection 5.2 and Appendix F illustrating that our method is relatively insensitive to these parameters.

While it may seem that such a parameter tuning approach leads to a heuristic algorithm, this is not true. By specifying the parameters β,ϵ,η1,η2\beta, \epsilon, \eta_1, \eta_2 for Algorithm 1 directly, we implicitly describe for which class of dynamics and which robustness margin range our safety guarantees hold. Importantly, this approach guarantees that the safety guarantees hold for some class of nonlinear dynamical systems under weak requirements on the parameters. If the unknown ground-truth parameters that we cannot verify in practice fall into this class, safety is guaranteed. However, a violation does not immediately imply safety constraint violations since our conditions are only sufficient for safety. Given the absence of knowledge of the ground-truth parameters, this approach additionally allows us to empirically tune the parameters used in Algorithm 1 to obtain strong empirical performance as it is similarly done in related work concerning safety [1-4]. Therefore, using such design parameters is an intuitive and practical approach to handling such uncertainty in our opinion.

Since we cannot verify the ground-truth system parameters in practice and the considered problems exhibit dynamics and constraints with considerable complexity, we did not check the satisfaction of the theoretical conditions in our numerical experiments. To illustrate that our results are not vacuous, we conducted additional experiments with a simplified dynamical system and constraints that allow us to obtain all parameters analytically from the ground-truth dynamics (Appendix H). These results demonstrate that Algorithm 1 is implementable under the assumptions posed in our theoretical results and works as expected. Thereby, these results complement our prior experimental results that show that Algorithm 1 is also effective at maintaining safety in practice when these assumptions cannot be verified and potentially do not hold.

In the revised paper, we have added a remark at the end of Section 4 to briefly discuss the transfer of the theoretical results into practice along the lines presented here. Moreover, we explicitly motivate the numerical evaluation in Section 5 as an investigation of Algorithm 1 in practical scenarios that do not allow verification of the conditions of Theorems 1 and 2 in advance. We stress that a comparison of constraint violations is only meaningful in such a setting as otherwise our theoretical guarantees (and those of comparative approaches) would exclude them in the first place. In the discussion of numerical results, we revised certain statements to keep a clear separation between empirical findings and theoretical guarantees by merely suggesting that our theory can provide valuable insight on certain observed effects. Finally, we add the evaluation in a simplified scenario with verifiably satisfied assumptions in Appendix H and refer to it in the main paper.

AuthorsOpenReview ↗

Response Part 2

Additional References

[1] As, Yarden, Lenart Treven, Carmelo Sferrazza, Stelian Coros, and Andreas Krause. "Actsafe: Active exploration with safety constraints for reinforcement learning." In Proceedings of International Conference on Learning Representations, pp. 29752-29776. 2025.

[2] Wendl, Manuel, Yarden As, Manish Prajapat, Anton Pollak, Stelian Coros, and Andreas Krause. "Safe Exploration via Policy Priors." In Proceedings of International Conference on Learning Representations, 2026.

[3] Wu, Zhixuan, Rui Yang, Lei Zheng, and Hui Cheng. "Safe learning-based feedback linearization tracking control for nonlinear system with event-triggered model update." IEEE Robotics and Automation Letters, 7(2), 3286-3293, 2022.

[4] Capone, Alexandre, Ryan Cosner, Aaron Ames, and Sandra Hirche. "Learning safe control via on-the-fly bandit exploration." Proceedings of the International Conference on Machine Learning, 2025.

[5] Breeden, Joseph, Kunal Garg, and Dimitra Panagou. “Control Barrier Functions in Sampled-Data Systems.” IEEE Control Systems Letters, 6:367–372, 2022. doi: 10.1109/LCSYS.2021.3076127.

[6] Taylor, Andrew J., Victor D. Dorobantu, Ryan K. Cosner, Yisong Yue, and Aaron D. Ames. “Safety of Sampled-Data Systems with Control Barrier Functions via Approximate Discrete Time Models.” In Proceedings of the 61st IEEE Conference on Decision and Control (CDC), pp. 7127–7134, 2022. doi: 10.1109/CDC51059.2022.9993226.

AuthorsOpenReview ↗

Response Part 3

Practical Computation of the Maximum GP Standard Deviation

The GP posterior variance in Eq. (3) is a quadratic function of the control input. We compute its maximum over the unit ball by eigendecomposing the control-dependent covariance block and solving a one-dimensional scalar equation. This yields the global maximum without requiring general non-convex optimization. We have clarified this computation in the revised manuscript and provide the detailed derivation in Appendix C.1 (Practical Computation of the Maximum GP Standard Deviation).

Smoothness of Exploration Trajectories

We thank the reviewer for this question. Fig. 2 shows the robots' spatial trajectories. Although the control inputs in Eq. (7) switch at discrete sampling instants, both benchmarks exhibit second-order position dynamics, where the control inputs affect the position trajectories through two integrations. Consequently, the switching may cause discontinuities in acceleration, but position and velocity remain continuous, resulting in smooth spatial trajectories. Moreover, each exploratory input is applied for a short sampling period (0.01 s for the quadrotor and 0.02 s for the AUV), making the resulting local changes less noticeable at the spatial scale of Fig. 2. We have clarified this point in Section 5.1, where the trajectories of the two robotic benchmarks are discussed.

Uniform High-Probability Error Bound

We thank the reviewer for this question. We have clarified the derivation from Eq. (14) to Eq. (16) in the revised manuscript.

Under the RKHS norm bounds in Assumption 2, applying the uniform GP error bound of Chowdhury and Gopalan (2017, Theorem 2) yields

∣μi(x,u)−ΔH(x,u)∣≤(B+R2γi+1+log⁡(1/δ))σi(x,u),|\mu_i(x,u)-\Delta H(x,u)| \leq \left(B+R\sqrt{2\gamma_i+1+\log(1/\delta)}\right)\sigma_i(x,u),

which holds with probability at least 1−δ1-\delta, simultaneously for all inputs and data updates.

By the GP posterior mean expression in Eq. (3) and the definitions of the CBF coefficients, the estimation error of the full CBF derivative can be expanded as

&\hat{\phi}_i(x)+\hat{\psi}_i^\top(x)u-\phi(x)-\psi^\top(x)u \\\\ &=\Delta\tilde{\phi}(x)+\Delta\tilde{\psi}^\top(x)u-\Delta\phi(x)-\Delta\psi^\top(x)u \\\\ &=\mu_i(x,u)-\Delta H(x,u). \end{aligned}$$ The corresponding GP posterior standard deviation is $\sigma_i(x,u)$. Since the maximum information gain satisfies $\gamma_i\leq\gamma_{\bar N}$ for all $i\leq\bar N$, the fixed $\beta$ in Eq. (14) uniformly bounds the confidence coefficient in the above error bound, directly yielding Eq. (16). We have added these intermediate derivation steps to the proof of Lemma 2 in the appendix and checked the corresponding theorem citation.
Reviewer_PSt7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper builds on an existing Gaussian-process-based control-barrier-function (GP-CBF) safety filter framework and proposes a new algorithm to extend its feasibility guarantees through active event-triggered exploration. The rationale is to have the system detect when the GP uncertainty is growing so large that feasibility of the runtime program may be lost and prevent this (with high probability) by actively choosing control actions to reduce this uncertainty.

Critical strengths

The paper presents a detailed mathematical analysis of the interplay between online GP inference and runtime robust safety filtering. The formulation is quite general (although still grounded in the control-affine structure of traditional model-based CBF work) and allows control-dependent modeling error through bilinear interaction between control and uncertainty.

In addition, the use of active perception/dual control effect to prevent the growth in uncertainty from voiding the safety filter's computed safety guarantee is conceptually attractive. I would encourage the authors to also engage with the implications of homoscedastic GP uncertainty reduction, which only depends on where we sample and not on what we discover when we do (there exists some limited work on heteroscedastic GPs to tackle this issue in, e.g. Bayesian optimization, but my sense is this is still quite under-explored in safe learning).

Critical weaknesses

Clarity and rigor. The technical exposition and overall logical flow suffer from repeated non sequiturs as well as some incorrect or seemingly self-contradicting statements. For example:

  1. The use of Gaussian processes in safety filters is motivated by claiming "intuitive representation of uncertainty" as the basis for their use in "safety-critical applications" (not quite the required standard of rigor).
  2. "This approach can be shown to eventually recover feasibility after it is lost." This cannot possibly be the case: in a CBF safety filter, feasibility a the critical invariant to ensure all-time safety. Loss of feasibility means that the stem may subsequently flow into a catastrophic failure despite all future efforts, which is in contradiction with any provable eventual recovery of feasibility.
  3. While it is widely recognized that attempting to use a "nominal" CBF filter based on an inaccurate dynamics model can "pose significant risks for safety", this does not imply the need to "collect data for learning to compensate for mismatches". In fact, data-driven error reduction is neither strictly necessary nor generally sufficient to mitigate the above risk. Note that even if we somehow manage to learn a perfectly accurate correction to the nominal CBF derivative from online data, this does not always allow us to prevent catastrophic outcomes: unless we are very lucky, we might learn—to our horror—that our "nominal CBF" is not in fact a valid CBF under the true system dynamics, and we no longer have a known control strategy to avoid certain failure.
  4. The proposed method is presented as though a key feature is that it triggers exploration preemptively, "before excessive model uncertainty makes the safety filter infeasible"; however, Assumption 3 already makes such infeasibility impossible regardless of GP uncertainty (by "assuming away" the above-mentioned not-a-valid-CBF problem). As detailed below, this in fact brings the method's safety benefits into question.

** Soundness and relevance.** The proposed method's safety relies heavily on the paper's Assumption 3, which, as the text acknowledges, is already made in prior work with similar purpose. The paper credits the proposed event-triggered data gathering strategy (Alg. 1) with enforcing probabilistic safety of the closed-loop system—however, Assumption 3 appears to bear the critical role, rendering the method itself redundant:

  • With Assumption 3 in place, the controller already knows crucial parameters that globally bound the true rate of descent of the nominal CBF under tightened control authority—any system identification/dynamics model learning process capable of supplying such a global bound would, in the process, determine which control values enforce it at each state. That is, the byproduct of any realistic instantiation of Assumption 3 would render the proposed GP-based method redundant (in fact, it would, by assumption, enjoy categorical guarantees, rather than merely probabilistic ones).
  • Without Assumption 3, the proof of Theorem 1 doesn't work, and therefore the proposed GP-based method does not in fact guarantee safety. Given this, it is unclear that the method actually offers safety benefits with respect to existing robust CBF approaches. This is exacerbated by the fact that experimental results only compare against other exploration approaches, but not against other robust safety filtering approaches.
Questions for authors

The main insight from the authors that could influence my assessment of the paper would be the following:

  1. Can you articulate what may be a realistic class of uncertain robot dynamics (e.g., some full-order robot model with well-defined uncertainty bounds) in which (a) your proposed approach can provably maintain safety (in particular, Assumption 3 provably holds) and yet (b) it is not possible to also prove (without a GP model) what control satisfies the robust feasibility condition at a given state?

  2. As I understand it, only a system in which both (a) and (b) happen at once would benefit from the proposed method's probabilistic safety guarantee—are there other classes of systems where you believe the method would provide unique added value in terms of safety?

4 comments
AuthorsOpenReview ↗

Response Part 1

Role of Assumption 3 and a Concrete Robotic Example

We thank the reviewer for raising this important question. Assumption 3 guarantees the existence of a control input satisfying the robust CBF constraint of the true system at every safe state. However, when the control-dependent dynamics are unknown, knowing the robustness margin θ\theta does not reveal which control direction satisfies this constraint. We illustrate this distinction through a concrete robotic example in Appendix H.

Assumption 3 vs. GP-CBF-SOCP Feasibility: Assumption 3 concerns the true CBF constraint, guaranteeing that for every safe state, there exists a control input satisfying

\quad \|\mathbf{u}\|\leq 1-\epsilon.$$ Here, $\phi(\mathbf{x})$ and $\boldsymbol{\psi}(\mathbf{x})$ are determined by the true system dynamics. In contrast, the GP-CBF-SOCP relies on **learned estimates** $\hat{\phi}(\mathbf{x})$ and $\hat{\boldsymbol{\psi}}(\mathbf{x})$ and requires $$\hat{\phi}(\mathbf{x})+\hat{\boldsymbol{\psi}}^\top(\mathbf{x})\mathbf{u} -\beta\sigma(\mathbf{x},\mathbf{u})-\eta_1\geq 0, \quad \|\mathbf{u}\|\leq 1.$$ Thus, even when the true CBF constraint admits a positive robustness margin, large GP prediction uncertainty can render the estimated robust constraint infeasible for all admissible inputs. **Assumption 3 guarantees robust feasibility of the true CBF constraint, whereas Algorithm 1 reduces GP uncertainty through event-triggered structured exploration to maintain recursive feasibility of the GP-CBF-SOCP.** To illustrate this distinction, we have added a **two-dimensional robotic example in Appendix H**, featuring an unknown actuator coupling matrix $M$ and a circular safe region of radius $0.4 \mathrm m$. The available prior information is $0.9\leq\sigma_{\min}(M)\leq\sigma_{\max}(M)\leq1.1$. We analytically establish that every system satisfying these bounds fulfills Assumption 3 with the same exact global robustness margin $\theta=0.0384$. In particular, consider $M=I_2$ and $M=-I_2$. At the safety boundary $\mathbf{x}=(0.4,0)^\top$, applying the control input $\mathbf{u}=(-0.6,0)^\top$, chosen for $M=I_2$, yields $$c(\mathbf{x},\mathbf{u};I_2)=0.0432>0,\qquad c(\mathbf{x},\mathbf{u};-I_2)=-0.0528<0.$$ In fact, at this state, the two systems require $u_1\leq-0.06$ and $u_1\geq0.06$, respectively. Hence, **no common safe control input exists for both systems, despite their identical, exactly known robustness margins.** This directly demonstrates why identifying the unknown control dynamics remains necessary for determining a safe control direction. Furthermore, Appendix H provides analytical calculations and numerical verification of the relevant theoretical constants and sufficient conditions for safety and recursive feasibility. We also conducted **160 simulations** across two actuator configurations, eight target directions, and ten random seeds. All runs maintained safety and successfully reached their targets, with no SOCP infeasibilities and only one structured exploration batch per run. These results further demonstrate the effectiveness of our approach for safe exploration under unknown control dynamics. The same distinction is relevant to our 7-DoF Franka Panda manipulator benchmark: knowing that a robustly safe torque input exists does not identify which input satisfies the CBF constraint when the control-dependent dynamics are unknown. Our GP model estimates the corresponding CBF derivative residuals online, enabling the safety filter to evaluate candidate torque inputs. The numerical verification results are summarized below. **Table: Verification of the sufficient conditions in Theorems 1–2.** |Theoretical Condition|Numerical Comparison| |---|---| |$\beta\geq\Xi_1(\bar N,\delta)$|$3.5\geq3.29279043$| |$\Gamma>0$|$0.0126804381>0$| |$\sqrt{T_s}\leq\Xi_2(\Gamma,\beta,\epsilon)$|$0.0316227766\leq0.0360137465$| |$\eta_1\geq\Xi_3(T_s)$|$0.0005\geq0.0000620747301$| |$\eta_2\geq\Xi_4(T_s,\beta)$|$0.002\geq0.00102749431$| |$T_s\leq\widetilde\Xi_2$|$0.001\leq0.00138674968$| |$\sqrt{T_s}\leq\widehat\Xi_2^2$|$0.0316227766\leq0.0401851611$| |$\eta_2\leq\widetilde\Xi_4$|$0.002\leq0.0384$| The multi-target results are summarized below. **Table: Multi-target results for the planar robot.** |Metric|$M=I_2$|$M=\operatorname{diag}(1,-1)$| |---|---|---| |Safe trajectories|80/80|80/80| |Successful target reaching|80/80|80/80| |SOCP infeasibilities|0|0| |Mean arrival time (s)|$13.428\pm0.089$|$13.434\pm0.093$| |Mean final error (mm)|1.324|1.333| |Minimum $h(\mathbf{x})$ ($\mathrm{m^2}$)|0.008743|0.008743| |Minimum true CBF constraint ($\mathrm{m^2/s}$)|0.00049887|0.00049887| |Exploration batches per run|1|1| |Total training samples per run|6|6|
AuthorsOpenReview ↗

Response Part 2

Safety Benefits Compared with Robust CBF Approaches

We thank the reviewer for raising this question. Previous studies have highlighted that robust CBF approaches relying on worst-case model uncertainty bounds may suffer from conservatism and feasibility issues when uncertainty is large. Specifically, Castañeda et al. (2025) point out that overly large uncertainty bounds can lead to impractically conservative control behavior, while Lederer et al. (2026a) note that worst-case backup strategies relying on prior knowledge generally deteriorate control performance.

This issue is further illustrated by our new example in Appendix H.1.2. Although both possible true systems satisfy Assumption 3, no control input can simultaneously satisfy their CBF constraints at the considered safety boundary state. This demonstrates that the existence of a safe control input for the true system does not guarantee that a feasible robust safety constraint can be constructed solely from prior uncertainty bounds.

Our approach addresses this challenge by updating the estimates of the true control dynamics and their uncertainty through online GP learning. Event-triggered structured exploration further reduces predictive uncertainty across multiple control directions to maintain recursive feasibility of the GP-CBF-SOCP. Our existing experiments primarily compare different online exploration strategies in terms of safety, conservatism, and triggering frequency, while the new example further illustrates the role of online learning when prior information alone is insufficient to identify a safe control direction.

AuthorsOpenReview ↗

Response Part 3

Clarifications of Introduction Statements

We thank the reviewer for highlighting several statements in the Introduction that require clarification. We have revised the corresponding passages to clarify their theoretical foundations and the objectives of our work.

1. Theoretical Role of GP in Safety Filtering. The key motivation for using GPs is their ability to provide uniform high-probability prediction error bounds, rather than merely an intuitive representation of uncertainty. Specifically, under the corresponding assumptions, the error bound ∣μ(x,u)−ΔH(x,u)∣≤βσ(x,u)|\mu(x,u)-\Delta H(x,u)|\leq\beta\sigma(x,u) can be directly incorporated into the CBF constraint to establish probabilistic safety guarantees. This is also the theoretical foundation adopted by existing GP-CBF approaches, such as Castañeda et al. (2025). We have clarified this point in the Introduction.

2. Clarification on Feasibility Recovery. The original statement refers to Capone et al. (2025) and it is indeed what is proposed in this paper. While it is correct that classical CBF safety filters rely on all time feasibility to ensure safety, it can easily be seen that it is merely a sufficient, but not a necessary condition. Infeasibilities do not imply safety violations. This is exactly what is exploited by this paper: When feasibility is lost, a sequence for generating data is triggered, such that the learning error bound reduces. If the error bound reduction happens before the safe set is left, feasibility is recovered. We agree that this poses a significant risk that the system may run into a safety constraint violation before feasibility is recovered. We want to emphasize that we already point this out as a weakness of the existing work by Capone et al. (2025) in our paper and our approach is constructed to maintain CBF constraint feasibility at all times.

We have further clarified this distinction in the Introduction.

3. Valid CBFs and the Role of Online Learning. We agree that inaccuracies in nominal dynamics do not necessarily require data-driven learning in general. However, as discussed by Castañeda et al. (2025) and Lederer et al. (2026a), model inaccuracies can invalidate safety guarantees, and when the control-dependent dynamics are unknown, even the directions of safe control inputs may be difficult to determine. While robust CBF approaches can rely on prior uncertainty bounds, large bounds may lead to excessive conservatism or infeasible safety filters.

Similarly, we also do not claim that data-driven error reduction is generally sufficient. Our theoretical guarantees require a valid CBF for the true system (Assumption 1) and control inputs that yield a positive robustness margin of the CBF conditions for the true system (Assumption 3). In this setting, we show that running our algorithm with suitable parameters is sufficient to ensure safety. We briefly discuss how a CBF for the true system can be found for certain classes of systems purely using system structure and without knowledge of system parameters in Appendix A.1. Thus, we do not neglect the existence of the 'horror scenario' described by the reviewer, but we want to stress that we exclude it in our theoretical analysis and empirically show on a range of robot control problems that it does not appear when running our proposed algorithm in more practical settings.

We have clarified the role of online learning in the revised manuscript.

4. Impossibility of infeasibilities We refer to our response on the Role of Assumption 3 above.

We sincerely thank the reviewers for their thoughtful feedback. We have provided point-by-point responses and revised the manuscript to address the concerns raised. We would greatly appreciate any further feedback, particularly if any concerns remain unresolved, and would be happy to provide additional clarification.

Comments on the paper

1 comment

We sincerely thank the reviewers for their constructive comments and valuable suggestions. We have revised the manuscript accordingly, and all changes are highlighted in blue in the revised PDF. For further and more specific changes made, please refer to the detailed responses to the reviewers below as well as the duly revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.