Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Conformal Safety Shielding for Imperfect-Perception Agents

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper proposes a run-time safety shield for imperfect-perception agents: conformalize the perception model into a set of likely states, then allow an action only if a perfect-perception shield would allow it for every state in that set.

desk verdict Theorem 1 is a solid formal result worth publishing, but the abstract overstates the conformal shield's safety guarantee, which Section 2.3 explicitly leaves unproven. read the letter →

arxiv 2506.17275 v2 pith:ICYBLDFW submitted 2025-06-12 eess.SY cs.AIcs.LGcs.SY

William Scarbro , Calum Imrie , Sinem Getir Yaman , Kavan Fatehi , Corina S. Pasareanu , Radu Calinescu , Ravi Mangal This is my paper · ORCID
classification eess.SYcs.AIcs.LGcs.SY
keywords conformalpredictionsafetyshieldingimperfectperceptionMarkovdecisionprocessprobabilisticruntimeverificationautonomoustaxiingstuckstates
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Autonomous agents that perceive the world through neural networks can misjudge their state, and safety shields that assume perfect perception may then approve dangerous actions. This paper proposes a shield that conformalizes the perception model so that, for each observation, it outputs a set of states guaranteed to contain the true state with probability at least $1-\alpha$, and then permits an action only if the perfect-perception shield would allow it for every state in that set. Because the construction never needs to model how observations are generated or how the perception network maps them to estimates, it promises local safety for agents whose perception is flawed. The paper's formal proof, however, covers perfect perception: Theorem 1 bounds the probability of reaching unsafe states within $n'$ steps by $1-(1-\lambda)^{n'}$, assuming the initial state is safe and the agent never encounters a stuck state. The local-safety guarantee for the conformal shield itself is stated as the design goal and explicitly left for future work.

What carries the argument

The load-bearing mechanism is the occupancy-vector analysis behind Theorem 1. Each state of the system is represented by a vector $d\in\mathbb{R}^S$ of probabilities; unsafe states are made absorbing in a transformed model $\Gamma^\psi_{\mathrm{Perf}}$, so the probability of satisfying the safety property equals the probability of ending in the unsafe set. The shield is defined by a threshold on $\sigma_\psi(s,a)$, the worst-case probability of reaching unsafe states after taking $a$ from $s$, and the proof tracks $g(d)=\mathbf{C}_{\mathbf{S}^{\lambda,n}_{\Delta}}^{\top} d$, the probability mass on states where the shield is nonempty, i.e. on states that are not stuck. Under the no-stuck and initially-safe assumptions, the shield condition yields $g(d_{i+1})\ge(1-\lambda)g(d_i)$, and induction gives the bound. For the conformal shield, the additional machinery is a conformal prediction set that replaces the point estimate, together with the intersection operation over that set.

What would settle it

Collect the observations made by the shielded agent during deployment and check how often the true state is inside the conformal prediction set; if the empirical coverage drops below $1-\alpha$, the shield's premise fails. A sharper test is to search the reachable state space for a state whose conformal set excludes the true state while the intersection of the perfect-perception shields still allows an action that is unsafe for the true state—if such a state is reachable with positive probability, the shield can permit a dangerous action.

Watch

Extended reading notes

Core claim

The paper's central claim is that safety under imperfect perception can be enforced by composing a perfect-perception shield with a conformalized perception function. The conformalized function outputs a set of possible states rather than a single estimate, and the conformal shield is the intersection $\hat{\Delta}_{\psi,\lambda}(\mathbf{s})=\bigcap_{s\in\mathbf{s}}\Delta_{\psi,\lambda}(s)$: an action is allowed only if the perfect-perception shield allows it for every state in the predicted set. Since the conformal guarantee says the true state lies in the predicted set with probability at least $1-\alpha$, the action is then safe in the same sense as the underlying shield, provided that guarantee holds on the deployed system. The paper proves a global finite-horizon bound for the perfect-perception case: under initial safety and no-stuck assumptions, the worst-case probability of reaching unsafe states within $n'$ steps is at most $1-(1-\lambda)^{n'}$. The local-safety guarantee for the conformal shield itself is stated as the design objective and explicitly left for future work; the case study shows that the conformal shield reduces the probability of reaching fail states compared with a non-conformalized baseline, while introducing stuck states.

Load-bearing premise

The local-safety argument for the conformal shield assumes that the conformal prediction set contains the true state with probability at least $1-\alpha$ on the data the shield actually encounters once deployed, even though the calibration data come from an unshielded agent choosing actions uniformly at random; once the shield changes the agent's behavior, exchangeability between calibration and deployment is no longer guaranteed.

Editorial extensions

If this is right

  • If the conformal shield works as claimed, imperfect-perception agents can be given run-time safety filtering without a model of how observations are generated or how the perception network maps them to state estimates.
  • For perfect-perception agents, any shield of the form in Eq. (8) carries a finite-horizon worst-case bound of $1-(1-\lambda)^{n'}$ on unsafe-state reachability, independent of the lookahead $n$ used to synthesize it.
  • The hyperparameters $\alpha$ and $\lambda$ trade failure probability against stuck probability: stronger conformal guarantees and stricter shields reduce unsafe actions but make empty-shield states more likely.
  • Stuck states are not an artifact of imperfect perception; the paper constructs a perfect-perception system where a shielded controller reaches a stuck state with probability arbitrarily close to one.
  • The conformal shield can be synthesized offline or online; online construction is attractive because the shield is only needed for the sets of predicted states that occur in operation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Our inference: if the iterative recalibration loop sketched in Section 2.3 is run to convergence, the calibration data would come from the shielded deployment distribution, and the local-safety property of Eq. (14) would follow from the conformal coverage guarantee; testing whether that loop converges is a natural next experiment.
  • Our inference: the intersection operation is a conservative conjunction over the whole predicted set; a risk-weighted variant that allows an action when a weighted average of $\sigma_\psi$ over the set stays below the threshold could reduce stuck states while keeping a probabilistic bound.
  • Our inference: since Theorem 1's bound does not involve the shield lookahead $n$, designers should not expect longer lookahead to improve worst-case safety guarantees; any benefit must come from the specific system dynamics rather than from the theorem.
  • Our inference: the same scheme applies to regression-based estimators: conformalize the regressor to obtain an interval or box of possible states and intersect the allowed actions over that set, which would extend the approach beyond discrete classification.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a conformal safety shield for discrete autonomous agents with imperfect perception. It models the system as an MDP whose states are triples (actual state, observation, estimate), and combines a perfect-perception shield with a conformalized perception function that outputs a set of plausible states. The shield then intersects the allowed actions over all states in the conformal set. The theoretical part proves a global probabilistic safety bound for perfect-perception shields (Theorem 1) under 'initially safe' and 'no stuck' assumptions, and states a local safety property (Eq. 14) for the conformal shield. The practical part evaluates the approach on an autonomous taxiing case study using a probabilistic abstraction of perception. The paper is clearly written and provides reproducible code, but the advertised 'local safety' guarantee is explicitly left open in Section 2.3.

Significance. The conformal shield idea is a natural and potentially valuable bridge between distribution-free uncertainty quantification and runtime safety enforcement, and it avoids modeling the observation process. Theorem 1 is a clean, self-contained result for perfect-perception shielding, with a proof in the appendix that appears correct and a useful discussion of the unavoidable trade-off leading to stuck states. The case study is informative and the code is publicly available. However, the central claimed safety guarantee for the conformal shield (Eq. 14) is not proven; the paper itself states this is future work. Since the abstract and introduction assert this guarantee, the present manuscript overstates its contributions and requires correction.

major comments (3)
  1. [2.3] The local safety guarantee claimed in the Abstract and Introduction is not proven. Section 2.3 states: 'We leave the investigation of such a procedure and its ability to help establish the desired local safety guarantee as future work.' Equation (14) is a desired property, not a theorem. This is load-bearing because the paper's headline contribution is the conformal shield's safety.
  2. [2.3] The step from conformal coverage (Eq. 13) to local safety (Eq. 14) is invalid as presented. Equation (13) holds for a fresh pair exchangeable with the calibration distribution, which Section 2.3 explicitly draws from the unshielded, non-conformalized MDP Γ with uniformly random actions. Once the shield is deployed, the action distribution changes, so the joint distribution over observations and states along trajectories changes and exchangeability is lost. No argument is given that calibration coverage transfers to the deployment distribution; the iterative recalibration idea is left to future work. Thus Eq. (14) currently lacks support.
  3. [1 and 2.2] The abstract and introduction state that Theorem 1 proves a global safety property of 'existing shield constructions for perfect-perception agents.' However, Theorem 1 applies to the new absolute shield defined in Eq. (8), which blocks all actions with risk above λ, not to previous relative shields such as [12] that filter actions based on the safest action from a state. The authors themselves note this distinction later in Section 2.2. The phrasing should be corrected to avoid misrepresenting the scope of Theorem 1.
minor comments (5)
  1. [Abstract] The phrase 'resulting in local safety' in the Abstract should be hedged (e.g., 'with the aim of local safety') because Section 2.3 explicitly leaves this guarantee open.
  2. [Eq. (1)] In Eq. (1), the target state uses 'm′' but the quantification uses 'o′'; the arrow label uses 'a′' for the action, which is inconsistent with the P(ŝ,a,ŝ′) notation. Please unify the notation.
  3. [Section 2.3] The MDP state in Eq. (11) is written (s,m,s), while Section 2.1 uses (s,o,s′) and Section 2.4 uses (s,s). Using three notations for the same object is confusing.
  4. [Theorem 1] The wording 'maximum probability ... is given by' followed by an inequality should be 'is bounded by'.
  5. [References] Reference [24] is the authors' own arXiv preprint of the same paper; if this is a submission, it should be removed or cited as a related self-reference.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the conformal shield construction and Theorem 1 are derived from stated definitions and external conformal guarantees; the unproven local-safety guarantee is explicitly flagged as future work.

full rationale

The paper's derivation chain is not circular. The perfect-perception shield is defined in Eq. 8 as the set of actions with one-step unsafety σψ(s,a) ≤ λ, and Theorem 1 is proved from this definition by an occupancy-vector induction in Appendix A under the explicitly stated 'initially safe' and 'no stuck' assumptions; the bound is not presupposed. The conformal shield is defined in Eq. 12 as the intersection of perfect-perception shields over the predicted state set; if the true state were covered by the conformal set with probability at least 1−α under the deployment distribution, Eq. 14 would follow directly from Eqs. 12 and 13 by monotonicity. However, the paper does not claim to have proved Eq. 14 in general: Section 2.3 states that 'the calibration dataset is initially drawn from the distribution over (oi,si) pairs induced by the path distribution for the imperfect-perception MDP Γ that is neither shielded nor conformalized and where the actions in each step are drawn uniformly at random,' and it explicitly leaves 'the investigation of such a procedure and its ability to help establish the desired local safety guarantee as future work.' The abstract's phrase 'resulting in local safety' therefore overstates what is formally established, but that is a correctness/limitation gap rather than a circular reduction. The empirical evaluation uses confusion matrices measured on a held-out test split and PRISM model checking; no safety outcome is fitted and then reported as a prediction. Self-citations to [3,20] are used to import an abstraction tool and a case-study setup, not to supply the central safety theorems, so no load-bearing self-citation is present.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The theoretical results depend on standard conformal prediction assumptions and two explicit conditions (initially safe, no stuck) stated in Theorem 1. The empirical evaluation additionally assumes the confusion-matrix abstraction is faithful and that the hand-set taxiing dynamics are accurate. The conformal shield's headline safety guarantee is not derived; it is left as future work.

free parameters (5)
  • shield threshold λ = 0.7, 0.8, 0.9 in experiments
    User-specified threshold on allowed unsafety probability; the safety bound in Theorem 1 depends on it.
  • conformal error α = 1-α ∈ {0.95, 0.99, 0.995}
    User-specified coverage level of the conformal prediction set.
  • shield lookahead n = 5
    Horizon used to compute σψ in Eq. 9; fixed in the case study.
  • mission horizon n' = 1 to 30
    Horizon for the safety property and model checking.
  • dynamics success probability = 0.9 (dyn_suc), 0.05 (dyn_fail)
    Hand-set probabilities for the taxiing dynamics model in the PRISM model (Listing 1.1).
assumptions (5)
  • domain assumption Calibration and test observation-state pairs are exchangeable
    Standard conformal prediction assumption stated in Section 2.3; the paper notes the deployed policy changes the distribution, so this is violated in practice.
  • ad hoc to paper The system never reaches a stuck state (no stuck)
    Assumption in Theorem 1 and Appendix A.4 (eq. 40); the paper shows stuck states are unavoidable in general.
  • domain assumption The initial state is safe (has a shield action)
    Assumption in Theorem 1 and Appendix A.4 (p6.5).
  • domain assumption The probabilistic abstraction ν built from the confusion matrix accurately represents the observation and perception processes
    Used in Section 2.4 to construct Γ^∆_Conf for model checking; inherited from prior work [3,20].
  • domain assumption The safety-relevant fragment of the MDP (dynamics) is accurately modeled
    Noted in Section 3.4; the dynamics are hand-set in the PRISM model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conformal Safety Shielding for Imperfect-Perception Agents." pith.science (2026). https://pith.science/paper/ICYBLDFW

@misc{pith2026250617275,
  author       = {Pith},
  title        = {Pith review of: Conformal Safety Shielding for Imperfect-Perception Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ICYBLDFW}},
  note         = {Machine review of arXiv:2506.17275}
}
read the original abstract

We consider the problem of safe control in discrete autonomous agents that use learned components for imperfect perception (or more generally, state estimation) from high-dimensional observations. We propose a shield construction that provides run-time safety guarantees under perception errors by restricting the actions available to an agent, modeled as a Markov decision process, as a function of the state estimates. Our construction uses conformal prediction for the perception component, which guarantees that for each observation, the predicted set of estimates includes the actual state with a user-specified probability. The shield allows an action only if it is allowed for all the estimates in the predicted set, resulting in local safety. We also articulate and prove a global safety property of existing shield constructions for perfect-perception agents bounding the probability of reaching unsafe states if the agent always chooses actions prescribed by the shield. We illustrate our approach with a case-study of an experimental autonomous system that guides airplanes on taxiways using high-dimensional perception DNNs.

Figures

Figures reproduced from arXiv: 2506.17275 by the authors.

Figure 1
Figure 1. Overall pipeline for our approach. ΓP erf refers to the MDP formulated for the perfect-perception agent (Section 2.2) while Γ ∆ Conf refers to the MDP for the imperfect-perception agent with the conformalized shield and a probabilistic abstraction of the observation and perception processes (Section 2.4). evaluation, we also measure the overall safety of the shielded imperfect-perception agent via model checking. To… view at source ↗
Figure 2
Figure 2. The simulation setup of the Turtlebot4 in Gazebo (left), and a sample image [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Results for properties Pfail, Pstuck, and Psuccess as n ′ increases. Black lines represent the baseline results. increase α ′ will inevitably result in fewer unique sets being produced as a result of the sets including more and more of the states. Therefore, this will result in a smaller model for analysis purposes. This is observed as α ′ increases from 0.95 to 0.99 with the accompanying increase in time, and then … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Time to perform model checking on each property as [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: A simple example demonstrating worst case safety. [PITH_FULL_IMAGE:figures/full_fig_p028_5.png]
Figure 6
Figure 6. Figure 6: An example demonstrating the maximal stuck likelihood. [PITH_FULL_IMAGE:figures/full_fig_p029_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 22 canonical work pages

  1. [12]

    Jansen, N., Könighofer, B., Junges, S., Serban, A.C., Bloem, R.: Safe reinforcement learning via probabilistic shields (2019), https://arxiv.org/abs/1807.06096

    arXiv 2019

  2. [1]

    arXiv preprint arXiv:2107.07511 (2021)

    Angelopoulos, A.N., Bates, S.: A gentle introduction to conformal prediction and distribution-free uncertainty quantification. arXiv preprint arXiv:2107.07511 (2021)

    arXiv 2021

  3. [2]

    Athena Scientific, 4th edn

    Bertsekas, D.P.: Dynamic Programming and Optimal Control. Athena Scientific, 4th edn. (2017)

    work page 2017

  4. [3]

    IEEE Transactions on Software Engineering (2024)

    Calinescu, R., Imrie, C., Mangal, R., Rodrigues, G.N., Păsăreanu, C., Santana, M.A., Vázquez, G.: Controller synthesis for autonomous systems with deep-learning perception components. IEEE Transactions on Software Engineering (2024)

    work page 2024

  5. [4]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Carr, S., Jansen, N., Junges, S., Topcu, U.: Safe reinforcement learning via shielding under partial observability. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 14748–14756 (2023)

    work page 2023

  6. [5]

    IEEE Robotics and Automation Letters 7(2), 1904–1911 (2022)

    Dawson, C., Lowenkamp, B., Goff, D., Fan, C.: Learning safe, generalizable perception-based hybrid control with certificates. IEEE Robotics and Automation Letters 7(2), 1904–1911 (2022)

    work page 2022

  7. [6]

    In: Learning for Dynamics and Control

    Dean, S., Matni, N., Recht, B., Ye, V.: Robust guarantees for perception-based control. In: Learning for Dynamics and Control. pp. 350–360. PMLR (2020)

    work page 2020

  8. [7]

    Dean, S., Taylor, A., Cosner, R., Recht, B., Ames, A.: Guaranteeing safety of learned perceptionmodulesviameasurement-robustcontrolbarrierfunctions.In:Conference on Robot Learning. pp. 654–670. PMLR (2021)

    work page 2021

Show all 32 references
  1. [8]

    Fremont,D.J.,Chiu,J.,Margineantu,D.D.,Osipychev,D.,Seshia,S.A.:Formalanal- ysisandredesignofaneuralnetwork-basedaircrafttaxiingsystemwithverifai.In:In- ternationalConferenceonComputerAidedVerification.pp.122–134.Springer(2020)

    2020

  2. [9]

    Proceedings of the AAAI Conference on Artificial Intelligence32(1) (Apr 2018)

    Fulton, N., Platzer, A.: Safe reinforcement learning via formal methods: Toward safe control through proof and learning. Proceedings of the AAAI Conference on Artificial Intelligence32(1) (Apr 2018). https://doi.org/10.1609/aaai.v32i1.12107, https://ojs.aaai.org/index.php/AAAI...

    2018 doi

  3. [10]

    In: Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems

    Goodall, A.W., Belardinelli, F.: Leveraging approximate model-based shielding for probabilistic safety guarantees in continuous environments. In: Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems. p. 2291–2293. AAMAS ’24, Internationa...

    2024

  4. [11]

    https://doi.org/10.1109/TCAD.2022.3197508 Conformal Safety Shielding for Imperfect-Perception Agents 19

    Hsieh, C., Li, Y., Sun, D., Joshi, K., Misailovic, S., Mitra, S.: Verifying con- trollers with vision-based perception using safe approximate abstractions (2021). https://doi.org/10.1109/TCAD.2022.3197508 Conformal Safety Shielding for Imperfect-Perception Agents 19

    2021

  5. [13]

    In: Bloem, R., Dimitrova, R., Fan, C., Sharygina, N

    Kadron, I.B., Gopinath, D., Pasareanu, C.S., Yu, H.: Case study: Analysis of autonomous center line tracking neural networks. In: Bloem, R., Dimitrova, R., Fan, C., Sharygina, N. (eds.) Software Verification - 13th International Conference, VSTTE 2021, New Haven, CT, USA, Octo...

    2021

  6. [14]

    In: Margaria, T., Steffen, B

    Könighofer, B., Lorber, F., Jansen, N., Bloem, R.: Shield synthesis for reinforcement learning. In: Margaria, T., Steffen, B. (eds.) Leveraging Applications of Formal Methods, Verification and Validation: Verification Principles. pp. 290–306. Springer International Publishing,...

    2020

  7. [15]

    Innovations in Systems and Software Engineering19(4), 379–394 (2023)

    Könighofer, B., Rudolf, J., Palmisano, A., Tappler, M., Bloem, R.: Online shielding for reinforcement learning. Innovations in Systems and Software Engineering19(4), 379–394 (2023)

    2023

  8. [16]

    In: Gopalakrishnan, G., Qadeer, S

    Kwiatkowska, M., Norman, G., Parker, D.: PRISM 4.0: Verification of probabilistic real-time systems. In: Gopalakrishnan, G., Qadeer, S. (eds.) Proc. 23rd International Conference on Computer Aided Verification (CAV’11). LNCS, vol. 6806, pp. 585–591. Springer (2011)

    2011

  9. [17]

    https://doi.org/10.1109/OJCSYS.2024.3385348

    Lindemann, L., Robey, A., Jiang, L., Tu, S., Matni, N.: Learning robust output control barrier functions from safe expert demonstrations (2021). https://doi.org/10.1109/OJCSYS.2024.3385348

    2021

  10. [18]

    arXiv preprint arXiv:2409.00536 (2024)

    Lindemann, L., Zhao, Y., Yu, X., Pappas, G.J., Deshmukh, J.V.: Formal verification and control with conformal prediction. arXiv preprint arXiv:2409.00536 (2024)

    2024 arXiv

  11. [19]

    In: Principles of Verification: Cycling the Probabilistic Landscape: Essays Dedicated to Joost-Pieter Katoen on the Occasion of His 60th Birthday, Part III, pp

    Mitra, S., Păsăreanu, C., Prabhakar, P., Seshia, S.A., Mangal, R., Li, Y., Watson, C., Gopinath, D., Yu, H.: Formal verification techniques for vision-based autonomous systems–a survey. In: Principles of Verification: Cycling the Probabilistic Landscape: Essays Dedicated to Jo...

    2024

  12. [20]

    In: Enea, C., Lal, A

    Pasareanu, C.S., Mangal, R., Gopinath, D., Getir-Yaman, S., Imrie, C., Calinescu, R., Yu, H.: Closed-loop analysis of vision-based autonomous systems: A case study. In: Enea, C., Lal, A. (eds.) Computer Aided Verification - 35th International Conference, CAV 2023, Paris, Franc...

    2023 doi

  13. [21]

    Peruffo, A., Ahmed, D., Abate, A.: Automated and formal synthesis of neural barrier certificates for dynamical models (2020), https://arxiv.org/abs/2007.03251

    2020 arXiv

  14. [22]

    In: Automated Technology for Verification and Analysis: 19th International Symposium, ATVA 2021, Gold Coast, QLD, Australia, October 18–22, 2021, Proceedings 19

    Pranger, S., Könighofer, B., Posch, L., Bloem, R.: Tempest-synthesis tool for reactive systems and shields in probabilistic environments. In: Automated Technology for Verification and Analysis: 19th International Symposium, ATVA 2021, Gold Coast, QLD, Australia, October 18–22,...

    2021

  15. [23]

    John Wiley & Sons (1994)

    Puterman, M.L.: Markov Decision Processes: Discrete Stochastic Dynamic Programming. John Wiley & Sons (1994)

    1994

  16. [24]

    arXiv preprint arXiv:2506.17275 (2025)

    Scarbro, W., Imrie, C., Yaman, S.G., Fatehi, K., Pasareanu, C.S., Calinescu, R., Mangal, R.: Conformal safety shielding for imperfect-perception agents. arXiv preprint arXiv:2506.17275 (2025)

    2025 arXiv

  17. [25]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA)

    Sheng, S., Parker, D., Feng, L.: Safe pomdp online planning via shielding. In: 2024 IEEE International Conference on Robotics and Automation (ICRA). pp. 126–132. IEEE (2024) 20 W. Scarbro, C. Imrie et al

    2024

  18. [26]

    In: 2023 IEEE International Conference on Robotics and Automation (ICRA)

    Tong, M., Dawson, C., Fan, C.: Enforcing safety for vision-based controllers via control barrier functions and neural radiance fields. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 10511–10517. IEEE (2023)

    2023

  19. [27]

    Vovk, V., Gammerman, A., Shafer, G.: Algorithmic learning in a random world, vol. 29. Springer (2005)

    2005

  20. [28]

    https://doi.org/10.48550/arXiv.2502.21308

    Waite, T., Geng, Y., Turnquist, T., Ruchkin, I., Ivanov, R.: State-dependent conformal perception bounds for neuro-symbolic verification of autonomous systems (2025). https://doi.org/10.48550/arXiv.2502.21308

    2025 doi arXiv

  21. [29]

    Watson, C., Alur, R., Gopinath, D., Mangal, R., Pasareanu, C.S.: Scenario-based compositional verification of autonomous systems with neural perception (2025), https://arxiv.org/abs/2504.20942

    2025 arXiv

  22. [30]

    https://doi.org/10.48550/arXiv.2203.02401

    Xiao, W., Wang, T.H., Chahine, M., Amini, A., Hasani, R.M., Rus, D.: Differentiable control barrier functions for vision-based end-to-end autonomous driving (2022). https://doi.org/10.48550/arXiv.2203.02401

    2022 doi arXiv

  23. [31]

    https://doi.org/10.1109/TRO.2023.3249564

    Xiao, W., Wang, T.H., Hasani, R.M., Chahine, M., Amini, A., Li, X., Rus, D.: Barriernet: Differentiable control barrier functions for learning of safe robot control (2023). https://doi.org/10.1109/TRO.2023.3249564

    2023

  24. [32]

    eventually unsafe

    Yang, S., Pappas, G.J., Mangharam, R., Lindemann, L.: Safe perception-based control under stochastic sensor uncertainty using conformal prediction. In: 2023 62nd IEEE Conference on Decision and Control (CDC). pp. 6072–6078. IEEE (2023) Conformal Safety Shielding for Imperfect-...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.