Finite-Sample Conformal Coverage Recovery via Fusion under Degraded Local Guarantees in Occupancy Map Estimation

Aneesh Raghavan; Karl Henrik Johansson; Ritvik Mahajan

REVIEW 2 major objections 4 minor 30 references

Finite-Sample Conformal Coverage Recovery via Fusion under Degraded Local Guarantees in Occupancy Map Estimation

T0 review · 2 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A distributed fusion rule over scalar e-values recovers the user-specified conformal coverage level at every robot, regardless of communication graph or sensor noise, even when each local guarantee is degraded.

desk verdict A clean fusion method with a correct conditional proof, but the headline coverage guarantee depends on an unproven local-calibration premise that needs to be either proven or transparently conditional. read the letter →

arxiv 2607.14906 v1 pith:7P2OVBNU submitted 2026-07-16 eess.SY cs.SY

Ritvik Mahajan , Aneesh Raghavan , Karl Henrik Johansson This is my paper

classification eess.SYcs.SY

keywords conformalpredictione-valuesmulti-robotoccupancymappingdistributedfusioncoverageguaranteeGaussianprocessuncertaintyquantification

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

The paper addresses a gap in multi-robot occupancy mapping: local conformal prediction sets on each robot's map have coverage guarantees that degrade because trajectories break exchangeability and each robot sees only part of the environment. The authors propose to recover the target coverage 1−α by fusing lightweight scalar e-values over a communication neighborhood, using a per-neighborhood miscoverage budget α/d and an uncertainty-attenuated threshold. They prove (Theorem 4.1) that the fused set-valued map achieves the target coverage at every agent for any graph and any noise distribution, under a local-coverage assumption. They also show a second result: a decay-based attenuation recovers coverage even when the local bound fails, under an uncertainty-graded miscoverage model. Simulations confirm the bound and show that denser topologies shrink the fraction of unclassified cells.

What carries the argument

The load-bearing object is the lifted e-value (12): each agent converts its conformal p-value at a query into a scalar e-value by dividing the violation indicator 1{p ≤ a} by a normalizer ε that upper-bounds the local miscoverage, multiplying by an observation gate (1 if within radius r of the agent's training inputs) and an exponential attenuation exp(−λσ̂) in the GP predictive standard deviation. E-values are closed under convex averaging, so the unweighted average over the neighborhood is an e-value, and Markov's inequality converts E[e]≤1 into P(Y* ∉ Γ) ≤ α. The per-neighborhood budget β=α/d ensures a single confident active agent can decide a query on its own, while the attenuation is t

What would settle it

Conduct the multi-agent mapping experiment on the same benchmark but with the held-out fold drawn from a shifted regime with larger total-variation drift δ_T (e.g., a more abrupt change in occupancy statistics between blocks), and check whether the empirical fused coverage drops below 1−α; alternatively, compute for a fixed block the empirical local miscoverage on infinitely many independent test folds and compare it to the single-fold ε_i^k to see if ε under-estimates with non-negligible probability.

Watch

Extended reading notes

Core claim

The paper's central claim is that the target miscoverage level α can be restored at every agent by a single round of e-value broadcast and weighted-average fusion, even though each local conformal guarantee may be degraded by temporal correlation and partial spatial coverage. The construction lifts each agent's conformal p-value into an e-value normalized by a per-neighborhood budget β=α/d, gated by the agent's observation region, and attenuated by the GP predictive standard deviation; because each local e-value has expectation at most one under the null, any fixed weighted average is also an e-value, and Markov's inequality yields the fused coverage guarantee. The coverage statement holds f

Load-bearing premise

The central premise is that the normalizer ε_i^k, estimated on a single held-out fold, is a genuine upper bound on the true local miscoverage at the recalibrated threshold; the paper does not prove a concentration bound for this estimate, so if the fold underestimates the true miscoverage, the e-value is invalid and the fused coverage guarantee collapses.

Editorial extensions

If this is right

Multi-robot teams can certify map predictions at a user-chosen level without exchanging raw data, only O(|Q|) scalars per block.
The coverage guarantee is independent of graph topology: ring and mesh both meet the bound; denser graphs mainly improve the classified fraction and reduce no-data abstentions.
The method applies to any likelihood-map estimator (e.g., OGM, GPOM, Hilbert maps) because it operates on the induced set-valued map, not the underlying regression.
The validity–efficiency dial λ lets a system designer choose where to sit between high coverage with large abstention and more decisive but less conservative maps.
If the local-coverage assumption fails but the uncertainty-graded model holds, attenuation can still restore the target level.

Reading between the lines

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

The reliance on a held-out fold for ε_k is the empirical pivot: if that estimate under-estimates true local miscoverage, the fused e-value is invalid and the theorem collapses; a concentration guarantee for ε would harden the result.
The observation gate restricts each agent's contribution to its observed region; one could test data-independent gating functions (e.g., GP variance) against the current radius-based gate, since Proposition 1 says any function of x alone controls the quantity Assumption 3 constrains.
The method's abstention behavior (returning {−1,+1}) could be used as an exploration heuristic: regions where fused evidence is insufficient are exactly where new measurements are most valuable.
Because the fused guarantee is marginal, per-cell conditional coverage is not claimed; a natural extension is to seek localized budgets that adapt to each query's neighborhood, or to combine with a second layer of conformal calibration on the fused scores.

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.

Desk Editor's Note

read the letter

Short version: this is a worthwhile paper with a genuinely useful algorithmic idea, but the headline guarantee is conditional on an unproven premise. If you read Theorem 4.1 as a conditional statement — “provided each local e-value is valid” — the proof is clean and the method is sound. If you read the abstract as claiming an unconditional, distribution-free coverage guarantee, it overreaches.

What’s actually new: the fusion rule is a nice design. The per-neighborhood budget alpha/d, the observation gate, and the uncertainty attenuation inside the e-value are all sensible, and the budget trick that lets a single confident agent decide a query is clever. Proposition 1 is a neat, correct result: under a shared measurement kernel, total-variation closeness of the joint law reduces to closeness of the spatial marginal. Theorem 4.1’s proof is straightforward and correct given Assumption 3, and the validity–efficiency dial with lambda is a nice practical insight. The paper is also honest about its limitations: it clearly states the marginal nature of the guarantee, the reliance on stationarity/mixing, and the communication cost.

The soft spot is exactly where the stress-test note points. Assumption 3 is the load-bearing premise, and it is asserted, not derived. Remark 2 says the held-out fold “provides” the upper bound on local miscoverage, but a single empirical miscoverage estimate from a finite fold is a random variable and is not guaranteed to dominate the true miscoverage. If epsilon_i^k comes in too low, then E[e_j^k] > 1, the fused e-value is not valid, and the Markov step in Theorem 4.1 fails. The gap is especially real when beta_i = alpha/d_i is small and violations are rare, because the empirical miscoverage on a finite fold can easily be zero while the true miscoverage is nonzero. The cited references [22], [23] give TV-based bounds, but not the tight pointwise upper bound used here. This is not a peripheral detail; it is the bridge between the local calibration and the fused guarantee.

The simulations do not close the gap. They use K = 2 temporal blocks, no error bars across repeated runs, and a safety factor s = 0.7 with lambda = 0 at the operating point, so they demonstrate that the method is conservative, not that Assumption 3 holds at the nominal threshold. The empirical coverage staying near 0.98 is consistent with the theorem but does not test the premise.

Who is this for? Researchers in multi-robot mapping, conformal prediction, and e-values will find the fusion construction and the conditional analysis useful. It deserves a serious referee, and I would send it to peer review. The authors should either prove a finite-sample conservative epsilon_i^k (for example, via a holdout conformal quantile or a concentration argument that accounts for the temporal dependence), or explicitly state the guarantee as conditional on a validated local calibration and provide an empirical check of that calibration. I’d also ask for more than two blocks and repeated-run variance. With those additions, this becomes a solid contribution; as written, it is a good idea with an unresolved foundational step.

Referee Report

2 major / 4 minor

Summary. The paper addresses multi-robot occupancy mapping with finite-sample coverage guarantees. Each robot trains a local GP, computes conformal p-values, recalibrates an inner threshold to a per-neighborhood miscoverage budget, and constructs e-values that are gated by local observability and attenuated by predictive uncertainty. Agents broadcast only scalar e-values over a shared query set; each receiver averages them and forms a set-valued prediction. Theorem 4.1 claims that the fused map achieves the target coverage 1−α at every agent, for any communication graph and sensor-noise distribution, under Assumption 3 (a local coverage bound). Proposition 2 extends the result to the case where local coverage degrades, using an uncertainty-graded bound. Simulations on ring and mesh topologies report empirical coverage above the nominal level and show that denser graphs reduce the unclassified fraction.

Significance. The e-value fusion mechanism is elegant and practically attractive: it requires only scalar communication, is agnostic to the local likelihood-map estimator, and the proof of Theorem 4.1 is clean and correct conditional on its premise. Proposition 1 is a useful observation that, under a shared measurement kernel, joint total-variation closeness reduces to spatial-marginal closeness. If the local coverage premise can be certified, the paper provides a valuable reduction: recovering a global coverage guarantee from per-agent calibrated e-values. The simulation study is transparent and demonstrates the efficiency benefits of denser topologies. However, the central theoretical claim is only as strong as Assumption 3, and that assumption is asserted rather than derived; the abstract's unconditional phrasing overstates what is actually proven.

major comments (2)

[Section IV, Assumption 3, Eq. (18); Remark 2] The load-bearing premise of Theorem 4.1 is asserted, not established. In Eq. (11), ε_i^k is the miscoverage measured on a held-out fold at the recalibrated threshold a_i^k, and Remark 2 claims that this 'provides' a finite-sample upper bound on the true miscoverage P(p_i^k(X*_k,Y*_k)≤a_i^k). A single empirical miscoverage estimate from a finite fold is a random variable; it is not stochastically guaranteed to dominate the true value, especially when the budget β_i=α/d_i is small (e.g., 0.04 for α=0.2, d_i=5) and miscoverage events are rare. If ε_i^k underestimates the true miscoverage, then E[e_j^k]>1, so the fused e-value is not a valid e-value and the Markov step in Theorem 4.1 fails. The citation to [22],[23] does not supply the required bound: those works give guarantees under weighted exchangeability or spatial-block conditions, not a guarantee that a held-out empirical CDF dominate
[Section IV-B, Proposition 2, Assumption 4] The 'recovery' result does not remove the unverified-premise problem; it only changes its form. Proposition 2 assumes c_i^k(x) ≤ ε_i^0 exp(μ_i σ̂_i^k(x)) for active x, with unknown constants ε_i^0 and μ_i, and requires λ ≥ max_j μ_j. No estimation procedure for μ_i is given, and unlike ε_i^k in the main theorem there is not even a suggested held-out estimator. Thus the claim that attenuation 'restores the target level when a local guarantee fails' is contingent on an assumption that is at least as hard to certify as Assumption 3. If this proposition is intended to make the framework applicable when local bounds are loose, it needs a concrete verification protocol for μ_i or a conservative choice that preserves finite-sample validity.

minor comments (4)

[Section VI] The limitations paragraph says 'temporal mixing (Assumption 1)' but should refer to Assumption 2 (β-mixing). The same sentence duplicates 'Assumption 1' twice.
[Section III-B, Algorithm 1] Line 4 of Algorithm 1 sets β←α/d, while the simulation uses β=sα/d with safety factor s=0.7. The safety factor appears only in Section V-A; please state in Algorithm 1 that β may incorporate a safety factor, or make the two consistent.
[Section IV, after Eq. (18)] The sentence 'The above assumption is in principle the same as the results in [22], [23]' is vague. Please cite the specific theorem in those references and explain how it yields the finite-sample upper bound needed in (18).
[Section V-E, Table I] The text defines an 'empty fraction' (Γ=∅) as the only genuine coverage failure, but Table I reports only coverage, classified fraction, and no-data fraction. Reporting the empty fraction would make the connection between coverage and the set-valued output more direct.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 4.1 is a conditional e-value fusion argument; its unproven local-coverage premise is a correctness risk, not a circular reduction.

full rationale

The derivation chain is explicit: local conformal p-values (8) are recalibrated to a per-agent threshold a_i^k and normalizer epsilon_i^k on a held-out fold (11); local e-values are constructed in (12); a fixed uniform average (15) yields a fused e-value; Lemma 1 (Markov) gives Theorem 4.1. This is a direct conditional construction: under Assumption 3 (P(p_i^k <= a_i^k) <= epsilon_i^k), each local e-value has expectation at most 1, and the weighted average preserves the e-value property, so the fused miscoverage bound follows. The target alpha is not the same object as the local epsilon: the budget beta = alpha/d_i and the 1/d_i averaging connect them, and the gate/attenuation only shrink e-values. The held-out fold is explicitly disjoint from the test folds (Remark 2 and Section V), so the empirical coverage is not evaluated on the fitted fold. No load-bearing self-citation appears: [22] and [23] are external references, and the e-value closure is cited to external work [25], [26]. The main weakness is that Assumption 3 is asserted rather than proved: Remark 2 says the held-out estimate 'provides this under the temporal protocol' without a concentration argument. That is an omitted proof / correctness risk, and the abstract's unconditional phrasing overstates the conditional theorem, but it is not a circular reduction of the conclusion into the inputs. The paper explicitly labels Assumption 3 and Assumption 4 as assumptions and acknowledges the marginal-guarantee limitation. Therefore no significant circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical or mathematical entities are postulated; the gate, attenuation, and budget are algorithmic constructions inside the e-value. The central claim rests on Assumptions 1–4, especially Assumption 3, which supplies the local-coverage upper bound that the fused guarantee is built on. Free parameters include the held-out-estimated normalizer ε_i^k, the recalibrated threshold a_i^k, the safety factor s, gate radius r, attenuation λ, and GP hyperparameters.

free parameters (6)

ε_i^k (local miscoverage normalizer) = not reported (per agent/block)
Denominator of the e-value (12); set to empirical miscoverage of the recalibrated inner set on a held-out fold (Remark 2). Theorem 4.1 requires it to upper-bound the true local miscoverage; finite-sample guarantee not established.
a_i^k (recalibrated inner threshold) = not reported
Chosen as sup{a: empirical miscoverage ≤ β_i} on held-out fold (11); determines the e-value indicator 1{p ≤ a}; central to the budget split.
safety factor s = 0.7
Section V-A: per-neighborhood budget β_i = s·α/d_i; used in all simulations; not in the theory.
gate radius r = 1 m
Equation (9); activates agent only within distance r of training inputs; chosen 'on the order of inter-sample spacing'.
attenuation decay λ = 0 at operating point; swept 0..4
Equation (13); Proposition 2 requires λ ≥ max_j μ_j to recover coverage; at λ=0 the method is already valid in experiments, so attenuation is a safety margin rather than a necessity.
GP hyperparameters (σ_n, lengthscale) = fit per agent
Section II-B: ExactGP with Matérn-3/2 kernel; σ_n fit per agent; affects the predictive variance bound sqrt(1+σ_n^2) used in Remark 3.

assumptions (5)

domain assumption Assumption 1 (Local Stationarity): within-block TV distance of the joint law over block I_cal^k is at most δ_T.
Section II-A; needed for the claimed local coverage property and for the validity of split conformal scores.
domain assumption Assumption 2 (β-mixing): each agent's process {(X_i_n, Y_i_n)} is β_i-mixing with coefficient β_i(l) → 0.
Section II-A; ensures calibration data approximates i.i.d. within blocks.
domain assumption Assumption 3 (Local Coverage): ε_i^k upper-bounds the miscoverage of the recalibrated local conformal set at the test point.
Section IV, Eq. (18). Paper states it is 'in principle the same' as [22,23] but provides no proof that the held-out fold estimate (Remark 2) gives a finite-sample upper bound. Load-bearing for Theorem 4.1.
ad hoc to paper Assumption 4 (Uncertainty-graded miscoverage): c_i^k(x) ≤ ε_0 exp(μ σ_hat(x)) for active x.
Section IV-B; introduced for Proposition 2; the paper calls it 'a modeling assumption, not a derived fact'.
domain assumption Shared measurement kernel for Proposition 1: L(Y|X=x) depends only on x, not on time or agent.
Used to certify joint TV closeness from spatial marginals; plausible for the sensing model (deterministic occupancy + homoscedastic noise), but it is stated, not derived.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Finite-Sample Conformal Coverage Recovery via Fusion under Degraded Local Guarantees in Occupancy Map Estimation." pith.science (2026). https://pith.science/paper/7P2OVBNU

@misc{pith2026260714906,
  author       = {Pith},
  title        = {Pith review of: Finite-Sample Conformal Coverage Recovery via Fusion under Degraded Local Guarantees in Occupancy Map Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7P2OVBNU}},
  note         = {Machine review of arXiv:2607.14906}
}

read the original abstract

Accurate and reliable environmental mapping is a fundamental requirement for multi-robot autonomy. While continuous mapping techniques like Gaussian Process Occupancy Mapping (GPOM) provide rich spatial correlation and uncertainty estimates, they lack formal, finite-sample guarantees on their predictive reliability. Conformal prediction can equip each robot's local map with a distribution-free coverage guarantee, but this local guarantee degrades in practice: temporal correlation along a robot's trajectory breaks the exchangeability on which conformal calibration relies, and each robot observes only a spatially limited, non-uniform portion of the environment. Taking these degraded per-agent guarantees as given, we develop a distributed fusion algorithm that recovers the desired coverage across the team. Robots exchange only lightweight scalar e-values with their neighbors, and a receiver fuses them using a per-neighborhood miscoverage budget and an uncertainty-attenuated fusion operator. We prove that the fused set-valued map recovers the target user-specified coverage level regardless of the communication graph topology or the underlying sensor noise distribution. However, a drawback is that wherever the fused evidence is insufficient, the map declines to commit and returns both labels (free and occupied), leaving a significant fraction of the domain unclassified rather than thresholded into a single decision. Simulated multi-agent mapping experiments demonstrate that the fused predictor reliably meets its theoretical coverage bounds, and illustrate that denser communication topologies significantly enhance map efficiency by shrinking this unclassified fraction.

Figures

Figures reproduced from arXiv: 2607.14906 by the authors.

**Figure 2.** Occupancy probability of a single agent’s pre-fusion [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Occupied-class F1 versus classification threshold [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: The five agents’ trajectories over the workspace [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Per-agent local maps (a)–(e), each at the agent’s own level [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Empirical fused coverage versus nominal coverage [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: Ring vs. mesh at α = 0.2 (test set, mean over agents). The mesh determines more of the map (higher classification, lower uncertainty, and no-data fraction) at equal coverage and accuracy. 0 1 2 3 4 attenuation decay λ 0.75 0.80 0.85 0.90 0.95 1.00 empirical coverage Va…

**Figure 8.** Figure 8: Increasing the attenuation decay λ raises coverage toward 1 while shrinking the classified fraction toward 0 (sets grow to {−1, +1}). VI. CONCLUSION This paper presented a distributed fusion algorithm for multi-robot environmental mapping that combines Gaussian Process…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 3 linked inside Pith

[22]

Con- formal prediction beyond exchangeability,

R. F. Barber, E. J. Candes, A. Ramdas, and R. J. Tibshirani, “Con- formal prediction beyond exchangeability,”The Annals of Statistics, vol. 51, no. 2, pp. 816–845, 2023

2023
[23]

Valid model-free spatial prediction,

H. Mao, R. Martin, and B. J. Reich, “Valid model-free spatial prediction,”Journal of the American Statistical Association, vol. 119, no. 546, pp. 904–914, 2024

2024
[1]

Using occupancy grids for mobile robot perception and navigation,

A. Elfes, “Using occupancy grids for mobile robot perception and navigation,”Computer, vol. 22, no. 6, pp. 46–57, 1989

1989
[2]

Thrun, W

S. Thrun, W. Burgard, and D. Fox,Probabilistic Robotics. MIT Press, 2005

2005
[3]

Gaussian process occupancy maps,

S. T. O’Callaghan and F. T. Ramos, “Gaussian process occupancy maps,”The International Journal of Robotics Research, vol. 31, no. 14, pp. 1542–1570, 2012

2012
[4]

Gaussian processes au- tonomous mapping and exploration for range-sensing mobile robots,

M. G. Jadidi, J. V . Miro, and G. Dissanayake, “Gaussian processes au- tonomous mapping and exploration for range-sensing mobile robots,” Autonomous Robots, vol. 42, no. 2, pp. 273–290, 2018

2018
[5]

Hilbert maps: Scalable continuous occupancy mapping with stochastic gradient descent,

F. Ramos and L. Ott, “Hilbert maps: Scalable continuous occupancy mapping with stochastic gradient descent,”The International Journal of Robotics Research, vol. 35, no. 14, pp. 1717–1730, 2016

2016
[6]

Bayesian hilbert maps for dynamic continuous occupancy mapping,

R. Senanayake and F. Ramos, “Bayesian hilbert maps for dynamic continuous occupancy mapping,” inProceedings of the 1st Annual Conference on Robot Learning (CoRL), vol. 78 ofProceedings of Machine Learning Research, pp. 458–471, PMLR, 2017

2017

Show all 30 references

[7]

Building continuous occupancy maps with moving robots,

R. Senanayake and F. Ramos, “Building continuous occupancy maps with moving robots,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 32, 2018

2018
[8]

Mr-gmmapping: Communication efficient multi-robot mapping sys- tem via gaussian mixture model,

H. Dong, J. Yu, Y . Xu, Z. Xu, Z. Shen, J. Tang, Y . Shen, and Y . Wang, “Mr-gmmapping: Communication efficient multi-robot mapping sys- tem via gaussian mixture model,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 3294–3301, 2022

2022
[9]

Distributed gaussian process mapping for robot teams with time-varying communication,

J. Di, E. Zobeidi, A. Koppel, and N. Atanasov, “Distributed gaussian process mapping for robot teams with time-varying communication,” in2022 American Control Conference (ACC), pp. 4458–4464, 2022

2022
[10]

iMESA: Incremental distributed opti- mization for collaborative simultaneous localization and mapping,

D. McGann and M. Kaess, “iMESA: Incremental distributed opti- mization for collaborative simultaneous localization and mapping,” in Proceedings of Robotics: Science and Systems (RSS), 2024

2024
[11]

V ovk, A

V . V ovk, A. Gammerman, and G. Shafer,Algorithmic Learning in a Random World. Springer, 2005

2005
[12]

Induc- tive confidence machines for regression,

H. Papadopoulos, K. Proedrou, V . V ovk, and A. Gammerman, “Induc- tive confidence machines for regression,” inMachine Learning: ECML 2002, vol. 2430 ofLecture Notes in Computer Science, pp. 345–356, Springer Berlin Heidelberg, 2002

2002
[13]

Adaptive conformal inference under distribution shift,

I. Gibbs and E. J. Cand `es, “Adaptive conformal inference under distribution shift,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 34, pp. 1660–1672, 2021

2021
[14]

Label noise robustness of conformal prediction,

B.-S. Einbinder, S. Feldman, S. Bates, A. N. Angelopoulos, A. Gendler, and Y . Romano, “Label noise robustness of conformal prediction,”Journal of Machine Learning Research, vol. 25, no. 328, pp. 1–66, 2024

2024
[15]

Formal verification and control with conformal prediction,

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

2024 arXiv
[16]

Distributed conformal prediction via message passing,

H. Wen, H. Xing, and O. Simeone, “Distributed conformal prediction via message passing,”arXiv preprint arXiv:2501.14544, 2025

2025 arXiv
[17]

α-occ: Uncertainty-aware camera-based 3d semantic occupancy prediction,

S. Su, N. Chen, C. Lin, F. Juefei-Xu, C. Feng, and F. Miao, “α-occ: Uncertainty-aware camera-based 3d semantic occupancy prediction,” 2025

2025
[18]

Safe planning in unknown environments using con- formalized semantic maps,

D. S. Sundarsingh, Y . Li, T. Tang, G. J. Pappas, N. Atanasov, and Y . Kantaros, “Safe planning in unknown environments using con- formalized semantic maps,”IEEE Robotics and Automation Letters, vol. 11, no. 4, pp. 4921–4928, 2026

2026
[19]

A safety-critical framework for ugvs in complex environments: A data-driven discrepancy-aware approach,

S. X. Wei, L. Gan, and J. W. Burdick, “A safety-critical framework for ugvs in complex environments: A data-driven discrepancy-aware approach,” 2024

2024
[20]

Conformal prediction under covariate shift,

R. J. Tibshirani, R. Foygel Barber, E. Candes, and A. Ramdas, “Conformal prediction under covariate shift,”Advances in neural information processing systems, vol. 32, 2019

2019
[21]

Adaptive conformal predictions for time series,

M. Zaffran, O. F ´eron, Y . Goude, J. Josse, and A. Dieuleveut, “Adaptive conformal predictions for time series,” inInternational conference on machine learning, pp. 25834–25866, PMLR, 2022

2022
[24]

R. M. Dudley,Real analysis and probability. Chapman and Hall/CRC, 2018

2018
[25]

E-values: Calibration, combination and appli- cations,

V . V ovk and R. Wang, “E-values: Calibration, combination and appli- cations,”Annals of Statistics, vol. 49, no. 3, pp. 1736–1754, 2021

2021
[26]

The only admissible way of merging arbitrary e-values,

R. Wang, “The only admissible way of merging arbitrary e-values,” Biometrika, vol. 112, no. 2, p. asaf020, 2025

2025
[27]

E-values expand the scope of conformal prediction,

E. Gauthier, F. Bach, and M. I. Jordan, “E-values expand the scope of conformal prediction,”arXiv preprint arXiv:2503.13050, 2025

2025 arXiv
[28]

Conformal e-prediction,

V . V ovk, “Conformal e-prediction,”Pattern Recognition, vol. 166, p. 111674, 2025

2025
[29]

Flykites: Human-centric interactive exploration and assistance under limited communication,

Y . Zhang, Z. Tian, J. Wei, and M. Guo, “Flykites: Human-centric interactive exploration and assistance under limited communication,” in2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 10222–10228, IEEE, 2025

2025
[30]

Mrgs: A multi-robot slam framework for ros with efficient information sharing,

G. S. Martins, D. Portugal, and R. P. Rocha, “Mrgs: A multi-robot slam framework for ros with efficient information sharing,” inRobot Operating System (ROS) The Complete Reference (Volume 5), pp. 45– 75, Springer, 2020

2020

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [22]

Con- formal prediction beyond exchangeability,

R. F. Barber, E. J. Candes, A. Ramdas, and R. J. Tibshirani, “Con- formal prediction beyond exchangeability,”The Annals of Statistics, vol. 51, no. 2, pp. 816–845, 2023

2023

[2] [23]

Valid model-free spatial prediction,

H. Mao, R. Martin, and B. J. Reich, “Valid model-free spatial prediction,”Journal of the American Statistical Association, vol. 119, no. 546, pp. 904–914, 2024

2024

[3] [1]

Using occupancy grids for mobile robot perception and navigation,

A. Elfes, “Using occupancy grids for mobile robot perception and navigation,”Computer, vol. 22, no. 6, pp. 46–57, 1989

1989

[4] [2]

Thrun, W

S. Thrun, W. Burgard, and D. Fox,Probabilistic Robotics. MIT Press, 2005

2005

[5] [3]

Gaussian process occupancy maps,

S. T. O’Callaghan and F. T. Ramos, “Gaussian process occupancy maps,”The International Journal of Robotics Research, vol. 31, no. 14, pp. 1542–1570, 2012

2012

[6] [4]

Gaussian processes au- tonomous mapping and exploration for range-sensing mobile robots,

M. G. Jadidi, J. V . Miro, and G. Dissanayake, “Gaussian processes au- tonomous mapping and exploration for range-sensing mobile robots,” Autonomous Robots, vol. 42, no. 2, pp. 273–290, 2018

2018

[7] [5]

Hilbert maps: Scalable continuous occupancy mapping with stochastic gradient descent,

F. Ramos and L. Ott, “Hilbert maps: Scalable continuous occupancy mapping with stochastic gradient descent,”The International Journal of Robotics Research, vol. 35, no. 14, pp. 1717–1730, 2016

2016

[8] [6]

Bayesian hilbert maps for dynamic continuous occupancy mapping,

R. Senanayake and F. Ramos, “Bayesian hilbert maps for dynamic continuous occupancy mapping,” inProceedings of the 1st Annual Conference on Robot Learning (CoRL), vol. 78 ofProceedings of Machine Learning Research, pp. 458–471, PMLR, 2017

2017