Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

A frozen vision-language-action model can see temperature: adding a thermal camera channel and a control-barrier-function safety filter lets a robot follow 'pick the hot bottle' in the dark, find a buried warm object, and ignore mirror refl

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 ↗

T0 review · deepseek-v4-flash

2026-08-02 18:40 UTC pith:JFL3EHAC

load-bearing objection Frozen VLM + pseudo-color thermal + CBF filter is a real, incremental result; the causal grounding claim outruns the evidence. the 4 major comments →

arxiv 2603.05754 v2 pith:JFL3EHAC submitted 2026-03-05 cs.RO

Safe-Night VLA: Seeing the Unseen via Thermal-Perceptive Vision-Language-Action Models for Safety-Critical Manipulation

Dian Yu , Qingchuan Zhou , Bingkun Huang , Majid Khadiv , Zewen Yang This is my paper
classification cs.RO
keywords vision-language-action modelsthermal perceptionlong-wave infraredcontrol barrier functionssafe manipulationsubsurface localizationmirror reflection disambiguationfrozen backbone adaptation
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.

Current vision-language-action (VLA) models are blind to temperature and hidden physical states, and their generative action heads have no safety guarantees. Safe-Night VLA proposes to feed a long-wave infrared thermal image, rendered as a pseudo-color picture, into a frozen pretrained VLA while a runtime control-barrier-function quadratic-program filter bounds the executed motion. The paper reports that thermal-equipped variants outperform RGB-only and RGB-D models on hot/cold bottle selection (64% vs 0% success in dim light), subsurface localization (72% vs 2%), and mirror-reflection rejection (17/20 with the full model under dim light with safety filtering). The claim is that a pretrained RGB encoder can transfer its shape and intensity bias to the thermal domain without fine-tuning, and that safety can be decoupled from semantics as a post-hoc geometric filter.

Core claim

The paper claims that a frozen pretrained vision-language-action model can gain thermal awareness by feeding it a long-wave infrared pseudo-color image through the same vision encoder, training only the action head. It further claims that the policy grounds semantic tokens like 'hot' in thermal gradients rather than dataset layout: attention ablation shows masking thermal input raises attention entropy from 0.052 to 0.228 and drops target mass from 53.5% to 16.8%. A control-barrier-function QP filter then converts the policy's Cartesian intent into safe joint displacements, enforcing boundary constraints and blocking extrapolative actions, so that perception and safety are decoupled.

What carries the argument

The central mechanism is a thermal perception channel: long-wave infrared images are formatted as 3-channel pseudo-color pictures and passed through the frozen vision encoder alongside RGB and depth, while only the action head is trained. The runtime safety layer is a control-barrier-function QP that, at each control step, solves for a joint displacement closest to the policy's desired Cartesian delta while keeping a collision-distance function above its safe boundary. The transfer mechanism is claimed to be the pretrained encoder's shape and intensity bias, which maps thermal intensity patterns onto semantic tokens without catastrophic forgetting.

Load-bearing premise

The claim stands on two premises: that the policy's success comes from grounding 'hot' in thermal gradients rather than in fixed positions or dataset layout, and that the CBF filter has accurate workspace boundaries and state estimates; if either fails, the reported advantages and safety guarantees do not generalize.

What would settle it

Run Scenario I with bottle positions randomized each episode and RGB/depth zero-masked, so only thermal distinguishes the target. If success does not track the thermal signature (e.g., the model picks the cold bottle when the hot bottle's position is swapped, or it ignores a decoy that is warm but not the instructed object), the thermal-grounding claim is false. A complementary check: in Scenario II, bury a room-temperature object and a hot object at different, randomized locations; if the policy digs where the hot object normally sits rather than where the thermal bloom appears, the spatial-b

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

If this is right

  • Thermal-aware manipulation becomes practical on existing VLA stacks: only the action head is trained, so the approach is parameter-efficient and preserves pretrained knowledge.
  • Safety can be added as a post-hoc geometric filter, decoupling semantic correctness from collision avoidance even when the policy hallucinates out-of-distribution motions.
  • Under degraded illumination, combining thermal and depth produces a policy that keeps working where RGB-only and RGB-D collapse: thermal supplies semantic discrimination and depth supplies geometric stabilization.
  • Mirror ambiguity, a hard case for RGB, can be resolved because common mirrors do not emit long-wave infrared; a policy with thermal input sees the reflected object as absent.
  • The three diagnostic scenarios provide a reusable evaluation paradigm for testing whether a VLA grounds physical state rather than appearance.

Where Pith is reading between the lines

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

  • Beyond the paper: a decisive test the paper does not run is randomizing object positions per trial while holding thermal signatures fixed; if a position heuristic drives the results, RGB-Only success would not collapse in dim light once positions vary, while if thermal grounding is real, RGB-T should track the correct target across position changes.
  • Beyond the paper: the frozen-backbone transfer mechanism suggests the same recipe may work for other non-visible, intensity-based modalities (near-infrared, ultrasound, gas concentration) whenever the signal can be rendered as a pseudo-color image, since the model appears to learn intensity-to-semantic correlation.
  • Beyond the paper: the paper's residual-failure analysis implies that adding a recovery policy — lifting and re-centering when the CBF blocks an approach — could convert many 'safely stopped' failures into successes, a testable extension beyond the current static-filter design.
  • Beyond the paper: replacing the static CBF model with real-time thermal point clouds could turn the safety layer from geometric-only into temperature-aware obstacle avoidance, an extension the paper names as future work rather than a demonstrated result.

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

4 major / 5 minor

Summary. The paper proposes Safe-Night VLA, a framework that augments a frozen pretrained vision-language-action model (GR00T-N1.5-3B) with thermal (LWIR) and depth perception, formatted as pseudo-color images, and couples it with a control-barrier-function quadratic program (CBF-QP) as a runtime safety filter. It is evaluated on three real-robot scenarios designed to expose RGB failure modes: temperature-conditioned selection of visually identical bottles, localization of a buried heated object under granular media, and rejection of mirror-induced false targets. Four modality configurations (RGB-Only, RGB-D, RGB-T, and RGB-T-D) are compared under normal and dim/night illumination, with and without the safety filter. The paper claims that thermal input is the dominant semantic cue, that depth and safety filtering mainly improve execution robustness, and that a cross-modal attention ablation suggests the policy grounds semantic tokens in thermal gradients rather than dataset-induced spatial biases.

Significance. If the claims hold, the paper would make a useful empirical contribution: it demonstrates a parameter-efficient way to inject a non-visible physical modality into a frozen VLA backbone and shows a plausible role for runtime safety filtering. The strengths are the real-robot evaluation across three complementary failure modes, the separate training of modality variants, the presence of held-out runtime modifications (an unseen hot obstacle, mirrors absent in training), and the standard, self-contained CBF-QP formulation. The paper is also candid about its limitations: the tasks are diagnostic rather than a general benchmark, scaling to larger models is not tested, and the CBF filter depends on accurate workspace models and state estimation. However, the evidence is not yet sufficient to establish the strongest mechanistic claim about thermal grounding, and the statistical reporting is below the standard expected for a journal.

major comments (4)
  1. [V-A, Table II] The quantitative comparison lacks any measure of uncertainty. Each reported success rate/count comes from 20–50 trials with no confidence intervals, no per-seed variance, and no significance tests. Several differences between thermal-equipped variants are small (e.g., RGB-T 86% vs. Ours 82% in Scenario I Normal w/ filter; RGB-T 10/20 vs. Ours 9/20 in Scenario IIIa Normal w/o filter), so claims about the relative contribution of depth are not supported. The large thermal-vs-RGB margins are encouraging, but without error bars or a multi-seed evaluation the reader cannot assess whether the differences are reproducible. Please add confidence intervals or standard deviations across at least 3 seeds, and report statistical tests for key pairwise comparisons.
  2. [V-C and Contribution 3] The claim that the policy 'actively grounds semantic tokens in thermal gradients rather than relying on dataset-induced spatial biases' is not established by the provided evidence. The attention ablation is correlational and based on only 10 episodes; the reported metrics (normalized entropy 0.228→0.052, Pearson r −0.064→+0.081, attention mass on hot object 16.8%→53.5%) are equally consistent with the policy attending to the fixed spatial location of the hot bottle in the training setup. A causal manipulation is needed: for example, swapping thermal patches between the hot and cold bottles at inference, randomizing object locations while holding the thermal layout constant, or masking the thermal channel only at test time and showing that behavior tracks the thermal cue. Without such a control, the mechanistic conclusion is unsupported.
  3. [IV-B, IV-C and Table II, Scenarios I and II] The scenario designs confound thermal evidence with fixed layout. In Scenario II, the buried target is at a fixed position inside the same box across all 200 demonstrations, and the expert trajectories consistently avoid the box boundary; a policy could succeed with a 'dig in the same place' trajectory, with thermal input as a correlated but non-causal signal. In Scenario I, bottle positions are fixed, and the added 'unseen hot obstacle' is not a controlled intervention on the thermal channel itself. To support the attribution of success to thermal semantics rather than positional heuristics, please add control experiments that randomize object/box positions and independently vary thermal signatures across trials.
  4. [IV-D and Table II, Scenario III] The Scenario III evaluation is a 1-step open-loop action prediction, and the safety filter intercepts actions that violate a modeled wall constraint. This conflates semantic target selection with safety-layer intervention: a proposal toward the true target that also contains an unsafe backward component is counted as a failure, while a proposal toward the mirror reflection that happens to satisfy the wall constraint could be counted as a success. Please report the semantic selection accuracy (whether the intended direction corresponds to the true target) separately from the safety-filter-passed execution. This separation is necessary to support the 'mirror rejection' claim as a perceptual result rather than a filter artifact.
minor comments (5)
  1. [III-A] The citation to the language model is missing: 'Qwen3-1.7B large language model [?]' should be replaced with the proper reference.
  2. [III-B] Typo: 'mathemetically' should be 'mathematically'.
  3. [Eq. (1)] The objective uses J(q)Δq − u_vla, where u_vla is described as a 6-DoF desired Cartesian EE delta pose. Please clarify the relative weighting between translation and rotation components, and how the axis-angle representation enters the norm.
  4. [IV-A] The paper states each episode lasts approximately 200 state-action pairs; clarify whether this is a fixed horizon or a nominal duration, since it affects interpretation of the action-chunking setup.
  5. [V-B] The observation that being safely stopped by the CBF filter and physically colliding with the boundary can both lead to task failure is important; it suggests the current recovery behavior is a bottleneck. A brief discussion of potential recovery mechanisms would strengthen the paper.

Circularity Check

0 steps flagged

No load-bearing circularity; central claims are empirical and the only self-citations are contextual.

full rationale

The paper's derivation chain is not circular. The CBF-QP safety layer in Eq. (1) is a standard convex program: the objective and constraints are stated explicitly (tracking Jacobian-projected delta poses subject to collision and joint-limit constraints), and no parameter is fitted to the success metrics and then reported as a prediction. The main empirical claims in Table II are based on four independently trained modality variants and are evaluated on held-out or runtime-shifted conditions: an unseen heated obstacle in Scenario I, mirror configurations absent from the Scenario III training set, and programmatic dim-light attenuation. These are genuine out-of-distribution checks rather than re-statements of training inputs. The mechanistic claim in Sec. V-C -- that the policy grounds semantic tokens in thermal gradients rather than spatial biases -- is supported only by correlational attention metrics on 10 episodes, and the paper itself cautions that 'these findings do not imply that the VLA performs complex thermodynamic reasoning.' This is an evidence-quality limitation (confounded by possible layout shortcuts), not a circular reduction. The self-citations [27], [28] appear only in related-work positioning and are not load-bearing; no uniqueness theorem is imported from the authors' prior work, and the frozen-backbone strategy is motivated by external citations [20], [21] as well as by empirical results. The acknowledged limitations in the Conclusion (targeted diagnostic scenarios, CBF dependence on accurate workspace models and state estimation) are stated transparently and do not mask circular reasoning. Overall, I find no step in which a claimed result reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced. The assumptions are standard for the domain: frozen-backbone transferability, thermal physics of buried objects and mirrors, and accurate geometric models for CBF. Free parameters are mostly standard hyperparameters, though γ and λ are unspecified, which is a reproducibility gap.

free parameters (4)
  • CBF constraint parameter gamma (γ) = not reported
    The CBF constraint is ∇h^T Δq + γ h(q) ≥ 0; the rate constant γ controls how aggressively safety is enforced and is not specified. It is a tuning choice that affects filter behavior.
  • DLS regularizer lambda (λ) = not reported
    The QP objective includes λ||Δq||^2; the value is not given, affecting how strongly the filter regularizes joint velocities.
  • Depth clipping range (5 meters) and Turbo colormap mapping = 5 m
    Depth normalization is hand-chosen; mild, but part of the input encoding.
  • Training timesteps (5000), LR (3e-5), batch (128), weight decay = 5000 steps, 3e-5, 128, 1e-4
    Chosen without reported sweeps; affects which variant checkpoints are compared.
axioms (4)
  • domain assumption Frozen SigLIP/CLIP-style vision encoders retain sufficient shape/intensity bias to transfer to pseudo-color thermal and depth domains without LoRA or architectural changes.
    Section III.A states this as the basis for the parameter-efficient strategy; empirically supported only by the paper's own ablations, not by a mechanistic proof.
  • domain assumption The thermal bloom of a ~50°C object buried under 0.5 cm of granular media is detectable by the Topdon TC001 at the employed camera position and persists during deployment.
    Section IV.C relies on heat transfer through granular media; no thermal imagery with labeled surface temperatures is provided.
  • domain assumption Glass/silver mirrors are near-opaque in the LWIR band and remain near room temperature in this setup.
    Section IV.D cites [34], [35]; a plausible physics assumption, but no measured LWIR image of the mirror is shown.
  • domain assumption The CBF collision function h(q) is continuously differentiable and computed from known geometric models of robot, box, table, and wall.
    Section III.C relies on h(q) as minimum distance to collision spheres; implementation (e.g., FCL, custom meshes) is not described.

pith-pipeline@v1.3.0-alltime-deepseek · 10993 in / 7419 out tokens · 56228 ms · 2026-08-02T18:40:04.263002+00:00 · methodology

0 comments
read the original abstract

Current Vision-Language-Action (VLA) models rely primarily on RGB perception, preventing them from capturing modalities such as thermal signals that are imperceptible to conventional visual sensors. Moreover, end-to-end generative policies lack explicit safety constraints, making them fragile when encountering obstacles and novel scenarios outside the training distribution. To address these limitations, we propose Safe-Night VLA, a multimodal manipulation framework that enables robots to see the unseen while enforcing rigorous safety constraints for thermal-aware manipulation in unstructured environments. Specifically, Safe-Night VLA integrates long-wave infrared thermal perception into a pre-trained vision-language backbone, enabling semantic reasoning grounded in thermodynamic properties. To ensure safe execution under out-of-distribution conditions, we incorporate a safety filter via control barrier functions, which provide deterministic workspace constraint enforcement during policy execution. We validate our framework through real-world experiments on a Franka manipulator, introducing a novel evaluation paradigm featuring temperature-conditioned manipulation, subsurface target localization, and reflection disambiguation, while maintaining constrained execution at inference time. Results demonstrate that Safe-Night VLA outperforms RGB-only baselines and provide empirical evidence that foundation models can effectively leverage non-visible physical modalities for robust manipulation.

Figures

Figures reproduced from arXiv: 2603.05754 by Bingkun Huang, Dian Yu, Majid Khadiv, Qingchuan Zhou, Zewen Yang.

Figure 1
Figure 1. Figure 1: Multimodal perception comparison in downstream tasks. LWIR [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: System Architecture of Safe-Night VLA. Rather than designing a multimodal architecture from scratch, we adapt the foundation model to process non-RGB modalities via a simple parameter-efficient strategy. The VLM backbone, comprising the vision encoder and language model, is strictly frozen to preserve pre-trained semantic representations. Training is restricted exclusively to the action head components (th… view at source ↗
Figure 3
Figure 3. Figure 3: Temperature-Conditioned manipulation. Training employs the AdamW optimizer (β1 = 0.95, β2 = 0.999, ϵ = 10−8 ) with a weight decay of 10−4 . We use a learning rate of 3 × 10−5 with a cosine decay schedule and a 5% linear warmup. The per-GPU batch size is set to 16 with 8 gradient accumulation steps, resulting in an effective batch size of 128. Training is conducted in bfloat16 mixed precision with TF32 enab… view at source ↗
Figure 5
Figure 5. Figure 5: Cross-modal disambiguation under mirror-induced ambiguity. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Attention ablation in Scenario I. VI. CONCLUSION In this work, we presented Safe-Night VLA, a multimodal manipulation framework that addresses the fundamental lim￾itations of RGB-only policies by integrating thermal per￾ception with safety guarantees. Rather than treating thermal imaging merely as a low-light heuristic, we demonstrated its capability to ground unobservable physical states, enabling [PITH_… view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Model

    cs.RO 2026-06 unverdicted novelty 6.0

    SafeDojo is a new world model-based safe RL framework for VLA that outperforms baselines on SafeLIBERO and real robot tasks.

  2. Vision-Language-Action Safety: Threats, Challenges, Evaluations, and Mechanisms

    cs.RO 2026-04 accept novelty 4.0

    A literature survey that unifies fragmented work on attacks, defenses, evaluations, and deployment challenges for Vision-Language-Action models in robotics.

Reference graph

Works this paper leans on

39 extracted references · 4 linked inside Pith · cited by 2 Pith papers

  1. [1]

    GR00T N1: An open foundation model for generalist humanoid robots,

    NVIDIA, “GR00T N1: An open foundation model for generalist humanoid robots,” 2025. [Online]. Available: https://arxiv.org/abs/ 2503.14734

    Pith/arXiv arXiv 2025

  2. [2]

    OpenVLA: An Open-Source Vision-Language-Action Model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn, “OpenVLA: An Open-Source Vision-Language-Action Model,” in 8th Annual Conference on Robot Learning, 2024. [Online]. Available: https://openreview.net/f...

    2024

  3. [3]

    RT-2: Vision-language-action models transfer web knowledge to robotic control,

    B. Zitkovichet al., “RT-2: Vision-language-action models transfer web knowledge to robotic control,” in7th Annual Conference on Robot Learning, 2023. [Online]. Available: https://openreview.net/ forum?id=XMQgwiJ7KSX

    2023

  4. [4]

    π ∗ 0.6: a VLA That Learns From Experience,

    P. Intelligenceet al., “π ∗ 0.6: a VLA That Learns From Experience,”

  5. [5]

    In- depth robustness analysis for vision-language-action models,

    S. Fei, S. Wang, J. Shi, Z. Dai, J. Cai, P. Qian, L. Ji, X. He, S. Zhang, Z. Fei, J. Fu, J. Gong, and X. Qiu, “In- depth robustness analysis for vision-language-action models,” 2025. [Online]. Available: https://openreview.net/forum?id=6mEfYoMRpF

    2025

  6. [6]

    3D-VLA: A 3D vision-language-action generative world model,

    H. Zhenet al., “3D-VLA: A 3D vision-language-action generative world model,” inProceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 235. PMLR, 21–27 Jul 2024, pp. 61 229–61 245. [Online]. Available: https://proceedings.mlr.press/v235/zhen24a.html

    2024

  7. [7]

    Roboagent: Generalization and efficiency in robot ma- nipulation via semantic augmentations and action chunking,

    H. Bharadhwaj, J. Vakil, M. Sharma, A. Gupta, S. Tulsiani, and V . Kumar, “Roboagent: Generalization and efficiency in robot ma- nipulation via semantic augmentations and action chunking,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 4788–4795

    2024

  8. [8]

    Audio-Visual Efficient Conformer for Robust Speech Recognition,

    M. Burchi and V . Vielzeuf, “Audio-Visual Efficient Conformer for Robust Speech Recognition,” inIEEE WACV, 2021

    2021

  9. [9]

    Control barrier function based quadratic programs for safety critical systems,

    A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs for safety critical systems,”IEEE Transactions on Automatic Control, vol. 62, no. 8, pp. 3861–3876, 2017

    2017

  10. [10]

    Octo: An Open-Source Generalist Robot Policy,

    O. M. Teamet al., “Octo: An Open-Source Generalist Robot Policy,”

  11. [11]

    STARE-VLA: Progressive Stage-Aware Reinforcement for Fine-Tuning Vision-Language-Action Models,

    F. Xu, G. Zhai, X. Kong, T. Fu, D. F. N. Gordon, X. An, and B. Busam, “STARE-VLA: Progressive Stage-Aware Reinforcement for Fine-Tuning Vision-Language-Action Models,” 2025. [Online]. Available: https://arxiv.org/abs/2512.05107

    arXiv 2025

  12. [12]

    ReVLA: Reverting Visual Domain Limitation of Robotic Foundation Models,

    S. Dey, J.-N. Zaech, N. Nikolov, L. V . Gool, and D. P. Paudel, “ReVLA: Reverting Visual Domain Limitation of Robotic Foundation Models,” 2025. [Online]. Available: https://arxiv.org/abs/2409.15250

    Pith/arXiv arXiv 2025

  13. [13]

    OmniVLA: An omni-modal vision-language-action model for robot navigation,

    N. Hirose, C. Glossop, D. Shah, and S. Levine, “OmniVLA: An omni-modal vision-language-action model for robot navigation,”arXiv preprint arXiv:2509.19480, 2025

    arXiv 2025

  14. [14]

    OmniVLA: Physically-Grounded Multimodal VLA with Unified Multi-Sensor Perception for Robotic Manipulation,

    H. Guoet al., “OmniVLA: Physically-Grounded Multimodal VLA with Unified Multi-Sensor Perception for Robotic Manipulation,”

  15. [15]

    ContactDB: Analyzing and Predicting Grasp Contact via Thermal Imaging,

    S. Brahmbhattet al., “ContactDB: Analyzing and Predicting Grasp Contact via Thermal Imaging,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  16. [16]

    A Novel Visible-Depth-Thermal Image Dataset of Salient Object Detection for Robotic Visual Perception,

    K. Songet al., “A Novel Visible-Depth-Thermal Image Dataset of Salient Object Detection for Robotic Visual Perception,”IEEE/ASME Transactions on Mechatronics, vol. 28, no. 3, pp. 1558–1569, 2023

    2023

  17. [17]

    Available: https://arxiv.org/abs/2511.01210

    [Online]. Available: https://arxiv.org/abs/2511.01210

    arXiv

  18. [18]

    Robotic perception of transparent objects: A review,

    J. Jiang, G. Cao, J. Deng, T.-T. Do, and S. Luo, “Robotic perception of transparent objects: A review,”IEEE Transactions on Artificial Intelligence, vol. 5, no. 6, pp. 2547–2567, 2024

    2024

  19. [19]

    TRansPose: Large-Scale Multispectral Dataset for Transparent Object,

    A. Kimet al., “TRansPose: Large-Scale Multispectral Dataset for Transparent Object,”The International Journal of Robotics Research, 2023

    2023

  20. [20]

    Multispectral RGB-LWIR Fusion with YOLO for Robust Detection,

    J. Shreyaset al., “Multispectral RGB-LWIR Fusion with YOLO for Robust Detection,”Journal of Sensors and Systems Research, 2025

    2025

  21. [21]

    Enhancing RGB-IR object detection: a frozen backbone approach with multi-receptive field attention, author=Lu, Bingyu and others,

    “Enhancing RGB-IR object detection: a frozen backbone approach with multi-receptive field attention, author=Lu, Bingyu and others,” Research Square / arXiv preprint, 2026

    2026

  22. [22]

    Shapeclipper: Scalable 3d shape learning from single-view images via geometric and clip-based consistency,

    Z. Huang, V . Jampani, A. Thai, Y . Li, S. Stojanov, and J. M. Rehg, “Shapeclipper: Scalable 3d shape learning from single-view images via geometric and clip-based consistency,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 12 912–12 922

    2023

  23. [23]

    ViPT: Visual Prompt Multi-Modal Tracking,

    J. Zhuet al., “ViPT: Visual Prompt Multi-Modal Tracking,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  24. [24]

    AutoRT: Embodied Foundation Models with Large Scale Safety Alignment,

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakrishnan, K. Hausmanet al., “AutoRT: Embodied Foundation Models with Large Scale Safety Alignment,” International Conference on Robotics and Automation (ICRA), 2024

    2024

  25. [25]

    SafeDiffuser: Safe Planning with Diffusion Probabilistic Models,

    W. Ni, Y . Zhang, A. Stooke, and A. Gupta, “SafeDiffuser: Safe Planning with Diffusion Probabilistic Models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  26. [26]

    Toward a holistic evaluation of robustness in clip models,

    W. Tu, W. Deng, and T. Gedeon, “Toward a holistic evaluation of robustness in clip models,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 9, pp. 8280–8296, 2025

    2025

  27. [27]

    SafeFlow: Safe Robot Motion Planning with Flow Matching via Control Barrier Functions,

    X. Dai, Z. Yang, D. Yu, F. Liu, H. Sadeghian, S. Haddadin, and S. Hirche, “SafeFlow: Safe Robot Motion Planning with Flow Matching via Control Barrier Functions,” 2025. [Online]. Available: https://arxiv.org/abs/2504.08661

    arXiv 2025

  28. [28]

    UniConFlow: A Unified Constrained Flow-Matching Framework for Certified Motion Planning,

    Z. Yang, X. Dai, D. Yu, Z. Li, M. Khadiv, S. Hirche, and S. Haddadin, “UniConFlow: A Unified Constrained Flow-Matching Framework for Certified Motion Planning,” 2026. [Online]. Available: https://arxiv.org/abs/2506.02955

    arXiv 2026

  29. [29]

    SafeVLA: Towards safety alignment of vision-language-action model via constrained learning,

    B. Zhang, Y . Zhang, J. Ji, Y . Lei, J. Dai, Y . Chen, and Y . Yang, “SafeVLA: Towards safety alignment of vision-language-action model via constrained learning,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. [Online]. Available: https://openreview.net/forum?id=dt940loCBT

    2025

  30. [30]

    Scalable Diffusion Models with Transform- ers,

    W. Peebles and S. Xie, “Scalable Diffusion Models with Transform- ers,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 4172–4182

    2023

  31. [31]

    Turbo, An Improved Rainbow Col- ormap for Visualization,

    A. Mikhailov, “Turbo, An Improved Rainbow Col- ormap for Visualization,” Google Research Blog,

  32. [32]

    Eagle 2.5: Boosting Long-Context Post- Training for Frontier Vision-Language Models,

    G. Chen, Z. Li, S. Wang, J. Jiang, Y . Liu, L. Lu, D.-A. Huang, W. Byeon, M. Le, M. Ehrlich, T. Lu, L. Wang, B. Catanzaro, J. Kautz, A. Tao, Z. Yu, and G. Liu, “Eagle 2.5: Boosting Long-Context Post- Training for Frontier Vision-Language Models,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. [Online]. Available: http...

    2025

  33. [33]

    Kaviany,Principles of Heat Transfer in Porous Media, 2nd ed

    M. Kaviany,Principles of Heat Transfer in Porous Media, 2nd ed. Springer Science & Business Media, 2012

    2012

  34. [34]

    Infrared detectors: an overview,

    A. Rogalski, “Infrared detectors: an overview,”Infrared physics & technology, vol. 43, no. 3-5, pp. 187–210, 2002

    2002

  35. [35]

    V ollmer and K.-P

    M. V ollmer and K.-P. M ¨ollmann,Infrared Thermal Imaging: Fun- damentals, Research and Applications, 2nd ed. John Wiley & Sons, 2017, explains emissivity and transmission properties of glass in LWIR

    2017

  36. [36]

    D. A. Nield and A. Bejan,Convection in Porous Media, 3rd ed. New York: Springer, 2006

    2006

  37. [2019]

    Available: https://research.google/blog/ turbo-an-improved-rainbow-colormap-for-visualization/

    [Online]. Available: https://research.google/blog/ turbo-an-improved-rainbow-colormap-for-visualization/

  38. [2024]

    Available: https://arxiv.org/abs/2405.12213

    [Online]. Available: https://arxiv.org/abs/2405.12213

    Pith/arXiv arXiv

  39. [2025]

    Available: https://arxiv.org/abs/2511.14759

    [Online]. Available: https://arxiv.org/abs/2511.14759

    Pith/arXiv arXiv