Pith. sign in

REVIEW 2 major objections 4 minor 33 references

This paper establishes that set-based training on a calibrated low-dimensional bottleneck shrinks the probabilistic reachable-action radius of a visuomotor policy under camera-pose perturbation, while matched behavior-only, consistency, and

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A frozen visual encoder plus a calibrated low-dimensional bottleneck lets set-based training shrink a probabilistic reachable-action radius for visuomotor policies, verified by split conformal prediction.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A genuinely useful scheme—freeze the encoder, propagate a calibrated low-dim bottleneck set through a small policy, and report a conformal action radius—with a real but acknowledged gap between the state-level calibration of the training radius and the group-level coverage it actually gets. the 2 major comments →

arxiv 2608.02545 v1 pith:5UWSACUY submitted 2026-08-03 cs.RO

Probabilistic Reachable-Action Verification of Visuomotor Policies via Set-Based Training

Yanliang Huang , Zhuocheng Zhang , Peng Xie , Zhen Zhang , Wenyuan Wu , Majid Khadiv , Zhuoqi Zeng , Amr Alanwar This is my paper
classification cs.RO
keywords visuomotor policyreachability analysiszonotopeset-based trainingconformal predictioncamera-pose perturbationaction-space verificationmanipulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 tries to establish that a visuomotor policy's actions under camera-pose perturbation can be verified tractably—and tightened during training—if the large visual encoder is frozen and set propagation is confined to a calibrated, low-dimensional bottleneck interface. The downstream policy is trained to minimize both imitation loss and the width of a zonotope enclosure propagated through its own layers, and rollout-level split conformal calibration converts measured action deviations into a probabilistic reachable-action radius with finite-sample coverage. In controlled manipulation experiments, this set-based objective reduced the 99% radius by a median 2.09x over behavior-only fine-tuning while preserving closed-loop task success, and observational-consistency and pointwise-adversarial controls did not reproduce the reduction. The reader should care because the radius is a pre-deployment bound on how far a camera drift can move the commanded action—a safety question for fixed-camera manipulation.

Core claim

The central claim is that the size of the action-set enclosure is trainable: once the frozen encoder and a learned 32-dimensional bottleneck define a fixed interface, the set-based training objective directly contracts a zonotope over-approximation of the actions reachable from the interface set, and this contraction carries over to the physical action deviation induced by sampled camera poses. The paper proves that propagation through the small flow-matching policy encloses the exact action set, and that on the event where the perturbed bottleneck representation stays inside the calibrated interface set, the terminal width upper-bounds the rollout-level action-deviation score. Split conform

What carries the argument

The key object is the calibrated low-dimensional interface: a frozen visual encoder followed by a learned 32-dimensional bottleneck whose representation shift under camera-pose perturbations is calibrated to epsilon = 0.240, forming an isotropic zonotope B_i = <z_i, epsilon I_m>, a centrally symmetric polytope with a center and generator matrix. This set is propagated through the downstream policy—a compact flow-matching (velocity-field) network unrolled into affine and ReLU layers—using affine maps and the single-neuron ReLU relaxation, producing an output zonotope whose largest scaled generator row-sum defines the terminal width rho_i that set-based training minimizes. Rollout-level split

Load-bearing premise

The load-bearing premise is that the frozen visual encoder plus the learned 32-dimensional bottleneck faithfully carries the camera-pose perturbation geometry, so that the calibrated interface set B_epsilon(z) contains the actual bottleneck shifts induced by poses from the prescribed distribution; the conformal guarantee is only over sampled rollout-pose pairs, and the paper notes that a deterministic map from the pose set to the interface set would be needed to upgrade the r

What would settle it

Evaluate the same protocol with a deliberately information-destroying bottleneck (e.g., a random projection) under the same camera-pose distribution: if the calibrated interface set still contains the bottleneck shifts at the claimed rate but the set-based radius no longer beats the behavior-only control, then the bottleneck's information content, not the set objective, is carrying the result. More directly, find any camera pose in the perturbation box whose bottleneck shift exceeds the calibrated epsilon on a held-out state: the Proposition 3 bound then fails on that pose, and the empirical c

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, set-based training gives a practical way to verify visuomotor policies under camera perturbation without propagating sets through a large visual encoder end to end.
  • The trained terminal width doubles as a deterministic over-approximation of the action set, and the paper's Proposition 3 makes that width a conformal upper bound on the sampled physical deviation whenever the bottleneck shift stays inside the calibrated interface set.
  • A policy trained with the set objective can be deployed with a pre-computed radius q such that, with probability at least 1-alpha over a new rollout-pose draw, the largest action deviation stays within q.
  • The same protocol yields a 2.03x median radius reduction across ten tasks under two-dimensional image augmentation, so the effect is not specific to one task.
  • Because the directly calibrated radius q is the primary guarantee and stays below the conservative propagated bound on all twelve checkpoints, the verification claim does not depend on the looseness of the zonotope enclosure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural extension: if the same calibrated-interface idea is applied to other sensor perturbations (lighting, occlusion, extrinsics of additional cameras), the bottleneck must be re-calibrated per threat, and the radius guarantee would transfer only to the extent the bottleneck preserves that perturbation geometry.
  • The paper's own closing note implies a route to a stronger guarantee: replacing the probabilistic sampled-pose statement with a deterministic map from the pose set to the interface set would upgrade the result from marginal conformal coverage to an end-to-end worst-case bound; the current evidence leaves that as future work.
  • One testable implication of the mechanism: the roughly 16x tighter zonotope enclosure for set-based training suggests that if the same objective were applied to a larger-capacity policy or a higher-dimensional interface, the radius reduction should persist only while the enclosure remains informative; otherwise the training signal decouples from the physical score.
  • Another testable implication: because the training radius epsilon is calibrated on 180 held-out states, the reported coverage is sensitive to the choice of epsilon; re-calibrating epsilon per seed produced a behavior-invalid seed, so a robust recipe would need an epsilon-selection rule that trades behavior and tightness.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes a method for probabilistic reachable-action verification of visuomotor policies under camera-pose perturbation. The approach freezes a visual encoder and uses a learned 32-dimensional bottleneck as a low-dimensional interface; an interface set is calibrated from held-out state-level bottleneck residuals, propagated through a compact flow-matching policy with zonotopes, and the output enclosure width is used as a training objective. At evaluation, one camera pose per rollout initialization is drawn from a bounded uniform distribution, rollout-level action deviations are computed, and split conformal calibration produces a finite-sample probabilistic radius q. Controlled experiments on LIBERO-10 Task 3 report that set-based training reduces the 99% conformal radius by a median per-seed factor of 2.09 over behavior-only fine-tuning, with empirical coverage 0.99–1.00 on 100 held-out initializations; observational-consistency and PGD adversarial-training controls do not reproduce the reduction.

Significance. If the result holds, the paper makes a useful practical contribution: it makes set-based training tractable for visuomotor policies by confining propagation to a low-dimensional interface, and it connects the trained enclosure to a finite-sample probabilistic verification statement. The direct split-conformal calibration of the rollout-level radius is methodologically clean and does not inherit the looseness of the zonotope abstraction. The controlled comparison is carefully designed: all configurations share the same frozen encoder, bottleneck, data, checkpoint schedule, and evaluation poses, so the observed q reduction is attributable to the training signal rather than to augmentation or architecture. The paper is also transparent about coverage and violation counts (Table S5), reports a deterministic-enclosure tightness study, and states that code and the pose manifest are released. The main weakness is that the mechanistic link between the training objective and the physical score is only conditional on an event whose group-level coverage is substantially below the nominal 95% in some seeds.

major comments (2)
  1. [§3.3, §4.3, Table S5] The interface radius ε=0.240 is calibrated as the 95th percentile of per-state bottleneck L∞ residuals over only 180 held-out states (Appendix B), but the verification score Sθ(g,v) of Eq. (11) is a max over the 16 states of a rollout group. Consequently the event E_r(g,v) in Proposition 3 holds for only 76–100% of test groups across seeds (Table S5: seed s622 coverage 0.76). On the ~24% uncovered groups at that seed, the propagated zonotope does not contain the perturbed action, so the terminal width ρ_i of Eq. (18) is not minimizing a valid upper bound on the physical deviation Sθ. The sample-adaptive bound eUθ(g,v) of Eq. (27) holds unconditionally and is reduced by set-based training, but it is not the quantity optimized during training. This is load-bearing for the paper's mechanistic claim that the training contracts an upper bound on the physical action deviation. The authors shou
  2. [§4.3] The claim that '95% paired bootstrap intervals exclude one' for the PGD-versus-set-based comparison is based on only n=2 behavior-valid PGD seeds (s622 and s623, after excluding s621). With two paired observations, the 95% percentile bootstrap interval is the observed range by construction, so this does not provide meaningful statistical evidence. The raw comparison on all three seeds (q_PGD > q_set on every seed) is clear, but the behavior-matched subset has no statistical power. Please remove the bootstrap claim or replace it with an appropriate statement of the limitation.
minor comments (4)
  1. [Table 1] The reported 'median per-seed ratio of 2.09×' does not exactly match the per-seed ratios implied by the table (0.231/0.111≈2.08, 0.181/0.113≈1.60, 0.191/0.077≈2.48; median=2.08). Please check the rounding or the computation.
  2. [§4.3] The sentence 'The fixed training radius covers 0.76 to 1.00 of the test groups across seeds' is important but under-explained in the main text; the distinction between the state-level calibration of ε and the group-level event E_r should appear earlier, since it directly affects how readers interpret Proposition 3.
  3. [Figure 3 / §4.4] The all-task median ratio of 2.03 is acknowledged to include behavior-invalid tasks, which is good, but the figure would be clearer if behavior-invalid tasks were visually distinguished from the capable subset rather than color-coded only by which method passes the gate.
  4. [Appendix D] The statement that the reported behavior success is not an independent held-out estimate is candid and the re-evaluation is helpful. Consider reporting the binomial confidence intervals for the 50-rollout estimates so the reader can see the precision of the behavior-gate numbers.

Circularity Check

0 steps flagged

No significant circularity: the reported radius q is directly conformally calibrated from sampled action deviations, and the width-to-score bound is a conditional theorem with acknowledged coverage limitations.

full rationale

The paper's central quantitative claim is the rollout-level conformal radius qθ, obtained by split conformal calibration directly from measured action-deviation scores Sθ(g,v) on calibration initialization groups (Eq. 13; Algorithm S1), with separate held-out empirical coverage. The training objective minimizes a zonotope terminal width ρ_i (Eq. 18) for an interface set B_i = ⟨z_i, ε I_m⟩, where ε = 0.240 is a state-level bottleneck-residual quantile. The derivation from width to physical score is Proposition 3: on the event E_r(g,v) that the perturbed bottleneck representation lies in the interface box, Sθ(g,v) ≤ Uθ(g;r), and using the sample-specific r_a(v) yields a pointwise bound. This is a conditional inequality, not an equivalence: qθ is not defined in terms of ε, ρ, or U, and the paper explicitly calls the propagated bound a 'surrogate' and qθ the 'tighter primary result.' The empirical support in Table S5 reports fixed-radius coverage between 0.76 and 1.00 across seeds, with zero violations on covered groups, and Section 4.3 openly states this coverage range. The low group-level coverage on some seeds is a real validity/calibration limitation, but it is not circularity: it does not make the measured qθ an artifact of the fitted ε, because qθ is independently conformally calibrated. The conclusion's caveat that a deterministic pose-set-to-interface-set map would be needed to upgrade toward an end-to-end guarantee further confirms that the paper does not claim a first-principles derivation of qθ from ε. Self-citations in Related Work are descriptive and not load-bearing for the main results. No step reduces a prediction to its own input by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim depends on the calibrated interface-set radius epsilon=0.240, which is a fitted parameter. The other free parameters are hyperparameters of the training procedure. No new physical or conceptual entities are introduced.

free parameters (6)
  • interface set half-width epsilon = 0.240 (in standardized bottleneck-representation units)
    Fixed before training as the 95% held-out state-level marginal quantile of bottleneck-representation shifts induced by camera perturbations. Used as the training interface set; coverage of this radius on evaluation groups varies from 0.76 to 1.00 across seeds.
  • conformal rank k = k = ceil((n+1)(1-alpha))
    Determined by the user-specified alpha and n=200 calibration groups. Not fitted to data, but the choice of the 99% level as the primary result is a reporting choice.
  • target radius r_target = 0.25 (small tier), 1.0 (medium/large tier)
    Chosen by hand for each capacity tier.
  • dual step eta_lambda and clamp lambda_max = eta_lambda=0.01, lambda_max=100
    Chosen by hand in the adaptive Lagrangian.
  • bottleneck dimension m = 32
    Chosen from a sweep over 16, 32, 64 based on clean and image-jitter behavior.
  • other hyperparameters (lr, batch size, behavior gate 0.85, etc.) = varies
    Standard hyperparameter choices, used for all methods; not fitted to the target result.
axioms (4)
  • domain assumption Exchangeability of rollout-initialization/camera-pose pairs for split conformal calibration
    Section 2: the guarantee Pr[S <= q] >= 1-alpha uses exchangeability of the sampled pairs. This is a standard conformal assumption, reasonable in the protocol, but not verified.
  • domain assumption The image of any perturbed camera pose lies within the calibrated bottleneck box with probability approximately 0.95
    The training interface set inherits its size from held-out state-level residual quantiles; evaluation coverage of this box varies across seeds (0.76-1.00), so the assumption is only partially validated.
  • domain assumption The downstream policy has an architecture whose zonotope abstraction is exactly affine plus ReLU relaxations
    To make the set propagation valid, GELU is replaced with ReLU, LayerNorm is replaced with BatchNorm, and BatchNorm is frozen/folded. These modifications could alter the policy behavior, though the imitation loss is intended to preserve it.
  • domain assumption Closed-loop behavior in the simulator is a valid proxy for the physical setting
    Numerous claims (behavior gate, open-loop action deviations) rely on the LIBERO/robosuite simulator.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Probabilistic Reachable-Action Verification of Visuomotor Policies via Set-Based Training." pith.science (2026). https://pith.science/paper/5UWSACUY

@misc{pith2026260802545,
  author       = {Pith},
  title        = {Pith review of: Probabilistic Reachable-Action Verification of Visuomotor Policies via Set-Based Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5UWSACUY}},
  note         = {Machine review of arXiv:2608.02545}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Reachability analysis for visuomotor policies is difficult because large visual encoders make end-to-end set propagation computationally expensive and excessively conservative. We therefore freeze the visual encoder and confine set propagation to a low-dimensional interface between it and the downstream policy, with the interface set calibrated from held-out camera-pose perturbations. Propagating this set through the policy with zonotopes yields a terminal output-enclosure width that set-based training optimizes directly. During evaluation, camera-pose perturbations are sampled from the prescribed distribution, and rollout-level split conformal calibration converts the resulting action-deviation scores into a probabilistic reachable-action radius with finite-sample coverage. In controlled manipulation experiments, set-based training reduces this radius while preserving closed-loop task capability, and matched behavior-only, observational-consistency, and pointwise-adversarial controls all leave a larger radius.

Figures

Figures reproduced from arXiv: 2608.02545 by Amr Alanwar, Majid Khadiv, Peng Xie, Wenyuan Wu, Yanliang Huang, Zhen Zhang, Zhuocheng Zhang, Zhuoqi Zeng.

Figure 1
Figure 1. Figure 1: Overview of the proposed pipeline. (1) Set-based training: an observation is mapped through the frozen vision encoder and bottleneck to a representation z. The calibrated bottleneck set Bϵ(z) = ⟨z, ϵIm⟩ is propagated through the Vision￾to-Action Flow Matching Policy (VITA) (Gao et al. 2026) variant to obtain an output zonotope enclosing the actions reachable from that set. The objective reduces the enclosu… view at source ↗
Figure 2
Figure 2. Figure 2: Rollout-initialization-level 99% conformal reachable-action radius q on the held-out Task 3 eval￾uation set, comparing two augmentation protocols, two￾dimensional image jitter and six-dimensional camera￾extrinsic perturbation, each without and with set-based train￾ing. Bars are three-seed means with individual seeds as dots, lower is better. The annotated factors are the set-versus-no￾set reductions of thr… view at source ↗
Figure 3
Figure 3. Figure 3: Ten-task breadth of set-based training. The per [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 10 linked inside Pith

  1. [1]

    Proceedings of the 35th International Conference on Machine Learning (ICML) , series =

    Differentiable Abstract Interpretation for Provably Robust Neural Networks , author =. Proceedings of the 35th International Conference on Machine Learning (ICML) , series =. 2018 , publisher =

    2018

  2. [2]

    Advances in Neural Information Processing Systems 31 (NeurIPS 2018) , pages =

    Fast and Effective Robustness Certification , author =. Advances in Neural Information Processing Systems 31 (NeurIPS 2018) , pages =. 2018 , url =

    2018

  3. [3]

    Proceedings of the ACM on Programming Languages , volume =

    An Abstract Domain for Certifying Neural Networks , author =. Proceedings of the ACM on Programming Languages , volume =. 2019 , doi =

    2019

  4. [4]

    Proceedings of the 35th International Conference on Machine Learning (ICML) , series =

    Provable Defenses against Adversarial Examples via the Convex Outer Adversarial Polytope , author =. Proceedings of the 35th International Conference on Machine Learning (ICML) , series =. 2018 , publisher =

    2018

  5. [5]

    Advances in Neural Information Processing Systems 31 (NeurIPS 2018) , pages =

    Efficient Neural Network Robustness Certification with General Activation Functions , author =. Advances in Neural Information Processing Systems 31 (NeurIPS 2018) , pages =. 2018 , url =

    2018

  6. [6]

    Advances in Neural Information Processing Systems 33 (NeurIPS 2020) , year =

    Automatic Perturbation Analysis for Scalable Certified Robustness and Beyond , author =. Advances in Neural Information Processing Systems 33 (NeurIPS 2020) , year =

    2020

  7. [7]

    The Sixth International Conference on Learning Representations (ICLR) , year =

    Towards Deep Learning Models Resistant to Adversarial Attacks , author =. The Sixth International Conference on Learning Representations (ICLR) , year =

  8. [8]

    Proceedings of the 3rd Conference on Learning for Dynamics and Control , series =

    Data-Driven Reachability Analysis Using Matrix Zonotopes , author =. Proceedings of the 3rd Conference on Learning for Dynamics and Control , series =. 2021 , publisher =

    2021

  9. [9]

    Transactions on Machine Learning Research , year =

    Set-Based Training for Neural Network Verification , author =. Transactions on Machine Learning Research , year =. 2401.14961 , archivePrefix =

    Pith/arXiv arXiv

  10. [10]

    Transactions on Machine Learning Research , year =

    Training Verifiably Robust Agents Using Set-Based Reinforcement Learning , author =. Transactions on Machine Learning Research , year =. 2408.09112 , archivePrefix =

    Pith/arXiv arXiv

  11. [11]

    2026 , eprint =

    Set-Based Training of Neural Barrier Certificates for Safety Verification of Dynamical Systems , author =. 2026 , eprint =

    2026

  12. [12]

    Proceedings of the 6th Conference on Robot Learning (CoRL) , series =

    Robustness Certification of Visual Perception Models via Camera Motion Smoothing , author =. Proceedings of the 6th Conference on Robot Learning (CoRL) , series =. 2023 , publisher =. 2210.04625 , archivePrefix =

    Pith/arXiv arXiv 2023

  13. [13]

    Proceedings of the 27th International Conference on Artificial Intelligence and Statistics (AISTATS) , series =

    Pixel-wise Smoothing for Certified Robustness against Camera Motion Perturbations , author =. Proceedings of the 27th International Conference on Artificial Intelligence and Statistics (AISTATS) , series =. 2024 , publisher =. 2309.13150 , archivePrefix =

    Pith/arXiv arXiv 2024

  14. [14]

    2022 , url =

    Dong, Yinpeng and Ruan, Shouwei and Su, Hang and Kang, Caixin and Wei, Xingxing and Zhu, Jun , booktitle =. 2022 , url =. 2210.03895 , archivePrefix =

    Pith/arXiv arXiv 2022

  15. [15]

    Computer Aided Verification (CAV) , year =

    Perception with Guarantees: Certified Pose Estimation via Reachability Analysis , author =. Computer Aided Verification (CAV) , year =. 2602.10032 , archivePrefix =

    Pith/arXiv arXiv

  16. [16]

    The Thirteenth International Conference on Learning Representations (ICLR) , year =

    Conformalized Interactive Imitation Learning: Handling Expert Shift and Intermittent Feedback , author =. The Thirteenth International Conference on Learning Representations (ICLR) , year =. 2410.08852 , archivePrefix =

    Pith/arXiv arXiv

  17. [17]

    Journal of Machine Learning Research , volume =

    Zono-Conformal Prediction: Zonotope-Based Uncertainty Quantification for Regression and Classification Tasks , author =. Journal of Machine Learning Research , volume =. 2025 , url =. 2508.11025 , archivePrefix =

    arXiv 2025

  18. [18]

    IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , year =

    Robust Operational Space Control with Conformal Disturbance Bounds for Safe Redundant Manipulation , author =. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , year =. 2607.00424 , archivePrefix =

    Pith/arXiv arXiv

  19. [19]

    Proceedings of Robotics: Science and Systems (RSS) , year =

    Diffusion Policy: Visuomotor Policy Learning via Action Diffusion , author =. Proceedings of Robotics: Science and Systems (RSS) , year =

  20. [20]

    2605.19678 , archivePrefix =

    Luo, Jingzhou and Wen, Yifan and Bai, Yongjie and Song, Xinshuai and Liu, Yang and Lin, Liang , year =. 2605.19678 , archivePrefix =

    Pith/arXiv arXiv

  21. [21]

    2026 , url =

    Gao, Dechen and Zhao, Boqi and Lee, Andrew and Chuang, Ian and Zhou, Hanchu and Wang, Hang and Zhao, Zhe and Zhang, Junshan and Soltani, Iman , booktitle =. 2026 , url =. 2507.13231 , archivePrefix =

    arXiv 2026

  22. [22]

    2026 , eprint =

    Fei, Senyu and Wang, Siyin and Shi, Junhao and Dai, Zihao and Cai, Jikun and Qian, Pengfang and Ji, Li and He, Xinzhe and Zhang, Shiduo and Fei, Zhaoye and Fu, Jinlan and Gong, Jingjing and Qiu, Xipeng , booktitle =. 2026 , eprint =

    2026

  23. [23]

    Concrete Problems in

    Amodei, Dario and Olah, Chris and Steinhardt, Jacob and Christiano, Paul and Schulman, John and Man. Concrete Problems in. 2016 , eprint =

    2016

  24. [24]

    Unsolved Problems in

    Hendrycks, Dan and Carlini, Nicholas and Schulman, John and Steinhardt, Jacob , year =. Unsolved Problems in. 2109.13916 , archivePrefix =

    Pith/arXiv arXiv

  25. [25]

    2023 , eprint =

    Liu, Bo and Zhu, Yifeng and Gao, Chongkai and Feng, Yihao and Liu, Qiang and Zhu, Yuke and Stone, Peter , booktitle =. 2023 , eprint =

    2023

  26. [26]

    2020 IEEE International Conference on Robotics and Automation (ICRA) , pages =

    Camera-to-Robot Pose Estimation from a Single Image , author =. 2020 IEEE International Conference on Robotics and Automation (ICRA) , pages =. 2020 , publisher =. doi:10.1109/ICRA40945.2020.9196596 , eprint =

    arXiv 2020

  27. [27]

    , booktitle =

    Nobre, Fernando and Heckman, Christoffer R. , booktitle =. 2020 , publisher =

    2020

  28. [28]

    Sensors , volume =

    Methodology for Designing an Optimal Test Stand for Camera Thermal Drift Measurements and Its Stability Verification , author =. Sensors , volume =. 2022 , publisher =

    2022

  29. [29]

    2026 , eprint =

    From Points to Sets: Set-Based Safety Verification in the Latent Space , author =. 2026 , eprint =

    2026

  30. [30]

    Conformalized Data-Driven Reachability Analysis with

    Huang, Yanliang and Zhang, Zhen and Xie, Peng and Zeng, Zhuoqi and Alanwar, Amr , year =. Conformalized Data-Driven Reachability Analysis with. 2603.12220 , archivePrefix =

    arXiv

  31. [31]

    Journal of the American Statistical Association , volume =

    Distribution-Free Predictive Inference for Regression , author =. Journal of the American Statistical Association , volume =. 2018 , publisher =

    2018

  32. [32]

    2026 , eprint =

    Certified Set Convergence for Piecewise Affine Systems via Neural Lyapunov Functions , author =. 2026 , eprint =

    2026

  33. [33]

    Rigorously computed orbits of dynamical systems without the wrapping effect , volume =

    K. Rigorously computed orbits of dynamical systems without the wrapping effect , volume =. Computing , number =. 1998 , doi =

    1998

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.