Pith. sign in

REVIEW 5 major objections 5 minor 43 references

RoboLight contends that linear interpolation of HDR images between synchronized real episodes produces valid lighting conditions for training robotic policies, allowing light-varied data to be scaled arbitrarily without new physical collect

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 18:51 UTC pith:6QY24TI7

load-bearing objection A genuinely useful lighting-variation dataset for manipulation, but the synthetic-data validation is undermined by an internal lux inconsistency that needs fixing. the 5 major comments →

arxiv 2603.04249 v2 pith:6QY24TI7 submitted 2026-03-04 cs.RO

RoboLight: A Dataset with Linearly Composable Illumination for Robotic Manipulation

Shutong Jin , Jin Yang , Muhammad Zahid , Florian T. Pokorny This is my paper
classification cs.RO
keywords RoboLightlighting variationHDR imaginglinear light transportdataset synthesisrobotic manipulationillumination robustnessdata scaling
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 paper introduces RoboLight, a dataset for robotic manipulation in which the same demonstration is replayed under systematically varied lighting inside a controlled enclosure, with all images stored in HDR format. Its central claim is that, because light transport is linear, new lighting conditions can be synthesized by simply interpolating HDR frames of synchronized episodes — no new physical collection needed. This yields 196,000 synthetic episodes from 2,800 real ones, and policies trained on the synthetic data perform comparably to policies trained on real ground-truth lighting. If true, the approach turns illumination into a cheaply scalar dimension for robot learning, and the dataset doubles as a reproducible lighting-robustness benchmark.

Core claim

The paper's central discovery is that a real episode recorded under illumination A and the same episode under illumination B can be averaged frame-by-frame in HDR space to generate an episode under any intermediate illumination. Defining an episode as a synchronized sequence of HDR frames, the authors write Eλ = λE1 + (1−λ)E2 and generate RoboLight-Synthetic by stepping λ in increments of 0.01. Validation against a real 'ground truth' lighting condition (purple, left+right, 700 lux) shows synthesized frames closely match the ground truth by luminance histograms, and policies trained on synthesized data reach success rates close to those trained on real data (0.65 vs 0.70 for RGB stacking, 0.

What carries the argument

The load-bearing object is the HDR image as a linear, radiometrically calibrated representation — here produced as RAW16 frames from a consumer depth camera and processed through a custom calibration pipeline. The second pillar is a record-replay-reset protocol that synchronizes the robot's trajectory across lighting conditions, so that corresponding frames from two episodes show the same scene under different lights. Equation (5) then defines synthesized episodes by weighted interpolation in HDR space; the linearity of light transport (Eq. 3) is what guarantees the interpolated frames are valid images of the scene under intermediate illumination.

Load-bearing premise

The LED bulbs' RGB control values are assumed to map linearly to emitted radiance, so the HDR image under 'purple' equals the sum of the images under 'red' and 'blue' — but the paper's own lux measurements (172 and 72 → 95, not 244) do not obviously support this, and the record-replay-reset protocol assumes object positions remain identical across lighting conditions.

What would settle it

Capture HDR frames of the same static scene under the red, blue, and purple settings and compare the per-pixel sum of the red and blue frames against the actual purple frame; a clear luminance difference would disprove the linear-composition premise. A simpler check: the workspace-center lux for purple should equal red+blue (244), whereas the dataset reports 95.

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

If this is right

  • Light-varied training data can be expanded arbitrarily by refining the interpolation step, removing the need to physically collect every lighting condition.
  • Policies trained on synthesized lighting can substitute for real lighting data when the target condition is hard or impossible to measure directly.
  • The dataset provides a repeatable benchmark for measuring policy robustness to color, direction, and intensity shifts separately.
  • Because all episodes are stored in HDR, global exposure and tone-mapping transformations yield additional visual conditions post hoc, without any new capture.
  • Lighting estimation methods can be paired with the interpolation rule to select or generate training data for a previously unseen environment.

Where Pith is reading between the lines

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

  • The method's reach is strongest inside a calibrated enclosure where illumination is the only varying factor; outside such a setup, trajectory synchronization errors will likely dominate over radiometric errors.
  • The reported workspace-center lux values (red 172, blue 72, purple 95) do not satisfy the linear-sum expectation of 244, suggesting the 'purple' ground truth may not be a strict linear combination of 'red' and 'blue' under the current hardware calibration — so the validation in Fig. 6 may underestimate the true residual error.
  • Because interpolation is performed on RAW16 frames that have already undergone denoising and white-balance processing, any non-linearities in the camera pipeline could bias the result; a cleaner test would interpolate purely linear RAW data before white balance.
  • The same linear-blending idea could extend beyond LEDs to any controllable light source and could serve as a physics-grounded augmentation for sim-to-real transfer, where interpolating between real and rendered lighting might close the domain gap.

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

5 major / 5 minor

Summary. The paper introduces RoboLight, a real-world robotic manipulation dataset collected in a custom 'Light Cube' with eight programmable RGB LED lights. It contains RoboLight-Real (2,800 episodes across 14 systematically varied lighting conditions) and RoboLight-Synthetic (claimed 196,000 episodes) generated by linear interpolation in HDR space between synchronized episodes, leveraging the linearity of light transport. The dataset is validated through visual comparisons, Diffusion Policy roll-outs, and three use cases (lighting robustness benchmark, lighting estimation, and HDR-based visual condition scaling).

Significance. If the linear composability claim holds, the dataset is a potentially valuable community resource: it addresses an underrepresented axis (lighting) in manipulation datasets and proposes a method to scale light-varied training data without additional real collection. The release of hardware design, HDR acquisition pipeline, and synchronized episodes would support reproducibility. The paper's strengths include a real, calibrated light-control system, HDR imaging, and a forward validation strategy that does not fit free parameters to the targets. However, the significance is conditional on resolving internal inconsistencies in the additivity evidence and the synthetic episode count.

major comments (5)
  1. [Sec. V-C / VI-C / Table II] The validation of RoboLight-Synthetic for the color subset assumes that 'Purple' is a linear combination of 'Red' and 'Blue' in HDR space. Table II reports center illuminance values of 172 lux (Red), 72 lux (Blue), and 95 lux (Purple). Under additive light transport, Purple should be approximately 244 lux if the control values (255,0,255) are the sum of (255,0,0) and (0,0,255), or at least between 72 and 172 lux if a convex combination with weights summing to 1 is intended. The measured 95 lux is inconsistent with both. This is load-bearing because the Synthetic-vs-Ground-Truth comparison in Fig. 6 and Table IV for RGB Stacking is the primary evidence that interpolated HDR episodes reproduce real lighting. Please provide a direct radiometric calibration of the LED system (e.g., per-channel HDR images and the combined condition) or re-run the validation on conditions for which Eq. (3) is
  2. [Sec. V-C] The total count of 196,000 synthetic episodes is not derivable from the stated protocol. With 14 conditions, 200 episodes per condition, and λ in steps of 0.01, a naive count using all condition pairs and all episodes gives C(14,2)×99×200 = 1,801,800 episodes; other interpretations (e.g., only pairs within a task, or one episode per pair) give different numbers. The manuscript should specify the exact set of condition pairs, the λ values, and the episode pairing rule used to arrive at 196,000, or correct the count. This number appears in the abstract, Sec. I, and Sec. V-C, so the arithmetic should be verifiable.
  3. [Sec. VI, Tables III, IV, Fig. 7] All validation numbers are point estimates from 20 roll-outs without error bars. For example, Table IV reports 0.95 vs 0.80 for Donut Hanging; with n=20, the 95% Wilson intervals overlap substantially, so the claim that synthetic training is 'comparable' to ground-truth training is not statistically supported. Please report confidence intervals, perform more roll-outs, or use a paired design that exploits the synchronized episodes to reduce variance.
  4. [Sec. VI-C, Fig. 6] The visual fidelity comparison is qualitative. The luminance histograms are not quantified, and the statement that synthesized frames 'closely match' ground truth is a subjective judgment. Because the entire synthetic-data claim rests on radiometric accuracy, please provide a quantitative image metric (e.g., PSNR, SSIM, or mean absolute error in RAW16/HDR space) between the synthesized and ground-truth frames for the three validation conditions.
  5. [Sec. IV-B and V-C] The HDR pipeline includes bilateral denoising (nonlinear), lens shading correction, white balance, color correction, and gamma correction. The paper states that interpolation is performed on RAW16 data, but it is not clear whether the RAW16 frames are the raw sensor values or the denoised/lens-corrected values. If any nonlinear operation is applied before the interpolation in Eq. (5), the linearity assumption no longer holds. Please clarify the exact processing stage at which interpolation is performed and verify that all preceding operations are linear in radiance.
minor comments (5)
  1. [References] Reference [12] contains a formatting artifact: 'Cacti: A 578 framework...' should be cleaned up.
  2. [Eq. (4)] The symbol T is used both for the number of frames in Eq. (4) and for the light transport operator in Eq. (1). Consider renaming one of them to avoid ambiguity.
  3. [Sec. IV-C] The record-replay-reset protocol mentions 'dataset synchronization checks' but does not specify how desynchronization is detected or what tolerance is used. Please elaborate.
  4. [Sec. II-B / Abstract] The claim of being the 'first real-world robotic manipulation dataset capturing synchronized episodes under systematically varied lighting conditions' should more explicitly discuss prior lighting-controlled datasets (e.g., RePLab [17], OlatVerse [22]) to contextualize the novelty.
  5. [Fig. 5] The scatter plot of object positions is informative, but the axes are in millimeters with no coordinate frame indicator. Adding a reference to Fig. 3a would improve readability.

Circularity Check

0 steps flagged

No significant circularity: interpolation is a forward, parameter-free operation with independent ground-truth checks; the Table II additivity discrepancy is a validity concern, not a circular reduction.

full rationale

The central derivation is Eq. (5), Eλ = λE1 + (1−λ)E2, a convex combination of synchronized HDR episodes. This is computed directly from recorded images; no parameter is fitted to the validation targets, and the validation in Sec. VI-C and Table IV compares against separately collected real conditions (Purple, Left+Right, 700 lux). Even if the lux values in Table II (172+72≠95) indicate that the LED RGB settings may not combine additively, that is an empirical-validity threat to the stated premise, not a circular reduction: the synthetic frames are not defined in terms of the ground-truth frames, nor is the ground truth defined as the output of Eq. (5). The linearity premise is supported by standard external citations (Debevec et al. 2000; Haeberli 1992), not by the authors' own prior work. No self-citation, uniqueness theorem, or ansatz-via-self-citation is load-bearing. The paper's synchronization caveat in the Conclusion is a data-quality limitation, not a circular step. The derivation is self-contained for the purpose of circularity analysis.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard light-transport linearity plus two implementation-level assumptions: RAW16 is linear and LED RGB controls are linear. The latter is contradicted by the paper's own lux table, which is the main source of correctness risk. No new physical entities are introduced.

free parameters (3)
  • Interpolation step Δλ = 0.01
    Discretization chosen by hand; controls the 196,000-episode count and is arbitrary.
  • Object placement ranges = not specified numerically
    Set through iterative task difficulty tests (Sec. VI-B) to avoid extreme settings; affects distribution diversity but not the physical claim.
  • Color correction matrix / white balance / exposure = not provided
    Calibrated using a color board (Sec. IV-D); needed for the HDR pipeline, but parameters are not released.
axioms (4)
  • domain assumption Light transport is linear: T(L1+L2)=T(L1)+T(L2) (Eq. 2)
    Standard physical assumption for fixed geometry and incoherent light; experimentally validated in cited graphics work [33,37].
  • domain assumption RAW16 output from RealSense D435i is a linear radiometric measurement
    Stated in Sec. IV-B; not independently verified in the paper and critical for HDR interpolation.
  • ad hoc to paper LED RGB control values map linearly to radiance
    Implicit in Eq. 5; contradicted by Table II color-subset lux measurements (Red 172 + Blue 72 ≠ Purple 95).
  • domain assumption Record-replay-reset maintains synchronized scene geometry across lighting conditions
    Sec. IV-C; acknowledged as fragile in Sec. VIII, with manual inspection required.

pith-pipeline@v1.3.0-alltime-deepseek · 11371 in / 16855 out tokens · 164485 ms · 2026-08-02T18:51:29.779516+00:00 · methodology

0 comments
read the original abstract

In this paper, we introduce RoboLight, the first real-world robotic manipulation dataset capturing synchronized episodes under systematically varied lighting conditions. RoboLight consists of two components. (a) RoboLight-Real contains 2,800 real-world episodes collected in our custom Light Cube setup, a calibrated system equipped with eight programmable RGB LED lights. It includes structured illumination variation along three independently controlled dimensions: color, direction, and intensity. Each dimension is paired with a dedicated task featuring objects of diverse geometries and materials to induce perceptual challenges. All image data are recorded in high-dynamic-range (HDR) format to preserve radiometric accuracy. Leveraging the linearity of light transport, we introduce (b) RoboLight-Synthetic, comprising 196,000 episodes synthesized through interpolation in the HDR image space of RoboLight-Real. In principle, RoboLight-Synthetic can be arbitrarily expanded by refining the interpolation granularity. We further verify the dataset quality through qualitative analysis and real-world policy roll-outs, analyzing task difficulty, distributional diversity, and the effectiveness of synthesized data. We additionally demonstrate three representative use cases of the proposed dataset. The full dataset, along with the system software and hardware design, will be released as open-source to support continued research.

Figures

Figures reproduced from arXiv: 2603.04249 by Florian T. Pokorny, Jin Yang, Muhammad Zahid, Shutong Jin.

Figure 1
Figure 1. Figure 1: The Light Cube system developed for controlled and repeatable robotic lighting data curation. We use this system to collect RoboLight, the first real-world robotic manipulation dataset capturing synchronized episodes under systematically varied light￾ing conditions. collection often spans long time horizons, during which uncontrolled light sources, such as light shining through windows or display monitors,… view at source ↗
Figure 2
Figure 2. Figure 2: This pipeline is custom-developed based on es￾tablished HDR imaging and radiometric calibration proce￾dures [39] and is not included in the official RealSense software development kit (SDK). Parameters for steps (5) and (6) are determined through the calibration procedure described in Sec. IV-D. (a) (b) (c) (d) (e) (f) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) Top-down schematic of the Light Cube. Overview of the RoboLight-Real dataset: (b) the color subset, with lighting conditions from left to right: white, red, green, blue, and purple; (c) the direction subset, including all, front, rear, left, right, and left+right directional illumination; and (d) the intensity subset, featuring lighting levels of 1400 lux, 700 lux, and 140 lux. All images are cropped a… view at source ↗
Figure 4
Figure 4. Figure 4: Examples from RoboLight-Synthetic. New episodes are synthesized by interpolating between HDR images from RoboLight-Real. In each row, only the first and last images are real; all intermediate images are synthesized. TABLE II: Dataset overview for RoboLight-Real. Task Color Direction Intensity RGB Stacking White All 700 lux Red All 172 lux Green All 335 lux Blue All 72 lux Purple All 95 lux Donut Hanging Wh… view at source ↗
Figure 5
Figure 5. Figure 5: Initial object position distributions for the three tasks, each computed from 200 episodes. Coordinates are in millimeters and referenced to the robot base frame. Point density varies across tasks due to differences in object count. Within each task, object positions are randomized across placement patches. frame from RoboLight-Synthetic, and frames generated by averaging PNG images within the color, direc… view at source ↗
Figure 6
Figure 6. Figure 6: Visual fidelity verification. From left to right, we compare the ground-truth frame, the synthesized frame from RoboLight￾Synthetic, and frames generated by averaging PNG images across (a) color, (b) direction, and (c) intensity. For each comparison, the luminance histogram is included to provide a quantitative view of the luminance distribution. We further evaluate the effectiveness of RoboLight￾Synthetic… view at source ↗
Figure 9
Figure 9. Figure 9: By adjusting global exposure and tone mapping of the HDR images, we generate (a) ambient white lighting, (b) high-exposure conditions resembling industrial environ￾ments, and (c) dimmer, yellow-toned illumination. These global transformations can be consistently applied to all HDR episodes in RoboLight, enabling scalable augmentation of visual conditions through post-hoc adjustments. (a) (b) (c) [PITH_FUL… view at source ↗
Figure 8
Figure 8. Figure 8: Lighting estimation results produced by DiffusionLight [43] on Donut Hanging (Right lighting). (a) DiffusionLight estimates scene illumination by inpainting a virtual mirror ball at the image center, which reflects the surrounding environment and lighting. (b) The unwrapped environment map derived from the mirror ball, visualizing the predicted lighting color, direction, and intensity; the bright region in… 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

43 extracted references · 12 linked inside Pith

  1. [1]

    “data will solve robotics and automation: True or false?

    N. M. Amato, S. Hutchinson, A. Garg, A. Billard, D. Rus, R. Tedrake, F. Park, and K. Goldberg, ““data will solve robotics and automation: True or false?”: A debate,”Science Robotics, vol. 10, no. 105, p. eaea7897, 2025

    2025

  2. [2]

    Open x-embodiment: Robotic learning datasets and rt-x models: Open x- embodiment collaboration 0,

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jainet al., “Open x-embodiment: Robotic learning datasets and rt-x models: Open x- embodiment collaboration 0,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 6892–6903

    2024

  3. [3]

    Droid: A large-scale in-the-wild robot manipulation dataset,

    A. Khazatsky, K. Pertsch, S. Nair, A. Balakrishna, S. Dasari, S. Karam- cheti, S. Nasiriany, M. K. Srirama, L. Y . Chen, K. Elliset al., “Droid: A large-scale in-the-wild robot manipulation dataset,”arXiv preprint arXiv:2403.12945, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    pi 0: A vision- language-action flow model for general robot control,

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichteret al., “pi 0: A vision- language-action flow model for general robot control,”arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Open- vla: An open-source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketiet al., “Open- vla: An open-source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Cosmos world foundation model platform for physical ai,

    N. Agarwal, A. Ali, M. Bala, Y . Balaji, E. Barker, T. Cai, P. Chattopad- hyay, Y . Chen, Y . Cui, Y . Dinget al., “Cosmos world foundation model platform for physical ai,”arXiv preprint arXiv:2501.03575, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    V-jepa 2: Self- supervised video models enable understanding, prediction and plan- ning,

    M. Assran, A. Bardes, D. Fan, Q. Garrido, R. Howes, M. Muckley, A. Rizvi, C. Roberts, K. Sinha, A. Zholuset al., “V-jepa 2: Self- supervised video models enable understanding, prediction and plan- ning,”arXiv preprint arXiv:2506.09985, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Armbench: An object-centric benchmark dataset for robotic manipulation,

    C. Mitash, F. Wang, S. Lu, V . Terhuja, T. Garaas, F. Polido, and M. Nambi, “Armbench: An object-centric benchmark dataset for robotic manipulation,”arXiv preprint arXiv:2303.16382, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Furniturebench: Reproducible real-world benchmark for long-horizon complex manipulation,

    M. Heo, Y . Lee, D. Lee, and J. J. Lim, “Furniturebench: Reproducible real-world benchmark for long-horizon complex manipulation,” in Robotics: Science and Systems, 2023

    2023

  10. [10]

    Decomposing the generalization gap in imitation learning for visual robotic manipulation,

    A. Xie, L. Lee, T. Xiao, and C. Finn, “Decomposing the generalization gap in imitation learning for visual robotic manipulation,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 3153–3160

    2024

  11. [11]

    Efficient data collection for robotic manipulation via compositional generalization,

    J. Gao, A. Xie, T. Xiao, C. Finn, and D. Sadigh, “Efficient data collection for robotic manipulation via compositional generalization,” arXiv preprint arXiv:2403.05110, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Cacti: A 578 framework for scalable multi-task multi- scene visual imitation learning. arxiv preprint 579,

    Z. Mandi, H. Bharadhwaj, V . Moens, S. Song, A. Rajeswaran, and V . Kumar, “Cacti: A 578 framework for scalable multi-task multi- scene visual imitation learning. arxiv preprint 579,”arXiv preprint arXiv:2212.05711, vol. 580, 2022

    Pith/arXiv arXiv 2022

  13. [13]

    Differentiable robot rendering,

    R. Liu, A. Canberk, S. Song, and C. V ondrick, “Differentiable robot rendering,”arXiv preprint arXiv:2410.13851, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    The ycb object and model set: Towards common benchmarks for manipulation research,

    B. Calli, A. Singh, A. Walsman, S. Srinivasa, P. Abbeel, and A. M. Dollar, “The ycb object and model set: Towards common benchmarks for manipulation research,” in2015 international conference on ad- vanced robotics (ICAR). IEEE, 2015, pp. 510–517

    2015

  15. [15]

    More than a million ways to be pushed. a high-fidelity experimental dataset of planar pushing,

    K.-T. Yu, M. Bauza, N. Fazeli, and A. Rodriguez, “More than a million ways to be pushed. a high-fidelity experimental dataset of planar pushing,” in2016 IEEE/RSJ international conference on intelligent robots and systems (IROS). IEEE, 2016, pp. 30–37

    2016

  16. [16]

    Omnipush: accurate, diverse, real-world dataset of pushing dynamics with rgb-d video,

    M. Bauza, F. Alet, Y .-C. Lin, T. Lozano-P´erez, L. P. Kaelbling, P. Isola, and A. Rodriguez, “Omnipush: accurate, diverse, real-world dataset of pushing dynamics with rgb-d video,” in2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2019, pp. 4265–4272

    2019

  17. [17]

    Replab: A reproducible low-cost arm benchmark platform for robotic learning,

    B. Yang, J. Zhang, V . Pong, S. Levine, and D. Jayaraman, “Replab: A reproducible low-cost arm benchmark platform for robotic learning,” arXiv preprint arXiv:1905.07447, 2019

    Pith/arXiv arXiv 1905

  18. [18]

    Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations,

    S. Song, A. Zeng, J. Lee, and T. Funkhouser, “Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations,” IEEE Robotics and Automation Letters, vol. 5, no. 3, pp. 4978–4985, 2020

    2020

  19. [19]

    Reassemble: A multimodal dataset for contact-rich robotic assembly and disassembly,

    D. Sliwowski, S. Jadav, S. Stanovcic, J. Orbik, J. Heidersberger, and D. Lee, “Reassemble: A multimodal dataset for contact-rich robotic assembly and disassembly,”arXiv preprint arXiv:2502.05086, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    Train offline, test online: A real robot learning benchmark,

    G. Zhou, V . Dean, M. K. Srirama, A. Rajeswaran, J. Pari, K. Hatch, A. Jain, T. Yu, P. Abbeel, L. Pinto, C. Finn, and A. Gupta, “Train offline, test online: A real robot learning benchmark,” in2023 IEEE International Conference on Robotics and Automation (ICRA), 2023

    2023

  21. [21]

    Domain randomization for transferring deep neural networks from simulation to the real world,

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, “Domain randomization for transferring deep neural networks from simulation to the real world,” in2017 IEEE/RSJ international con- ference on intelligent robots and systems (IROS). IEEE, 2017, pp. 23–30

    2017

  22. [22]

    Olatverse: A large-scale real-world object dataset with precise lighting control,

    X. Zhou, J. Chen, P. Rao, T. Teufel, L. Lyu, T. Minasian, O. Sot- nychenko, X.-X. Long, M. Habermann, and C. Theobalt, “Olatverse: A large-scale real-world object dataset with precise lighting control,” arXiv preprint arXiv:2511.02483, 2025

    arXiv 2025

  23. [23]

    Getting to know low-light images with the exclusively dark dataset,

    Y . P. Loh and C. S. Chan, “Getting to know low-light images with the exclusively dark dataset,”Computer vision and image understanding, vol. 178, pp. 30–42, 2019

    2019

  24. [24]

    See clearer at night: towards robust nighttime semantic segmentation through day-night image conversion,

    L. Sun, K. Wang, K. Yang, and K. Xiang, “See clearer at night: towards robust nighttime semantic segmentation through day-night image conversion,” inArtificial Intelligence and Machine Learning in Defense Applications, vol. 11169. SPIE, 2019, pp. 77–89

    2019

  25. [25]

    A study on data selection for object detection in various lighting conditions for autonomous vehicles,

    H. Lin, A. Parsi, D. Mullins, J. Horgan, E. Ward, C. Eising, P. Denny, B. Deegan, M. Glavin, and E. Jones, “A study on data selection for object detection in various lighting conditions for autonomous vehicles,”Journal of Imaging, vol. 10, no. 7, p. 153, 2024

    2024

  26. [26]

    Quantifying how lighting and focus affect face recognition performance,

    J. R. Beveridge, D. S. Bolme, B. A. Draper, G. H. Givens, Y . M. Lui, and P. J. Phillips, “Quantifying how lighting and focus affect face recognition performance,” in2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition-Workshops. IEEE, 2010, pp. 74–81

    2010

  27. [27]

    Hand gesture recognition with symmetric pattern under diverse illu- minated conditions using artificial neural network,

    M. Haroon, S. Altaf, S. Ahmad, M. Zaindin, S. Huda, and S. Iqbal, “Hand gesture recognition with symmetric pattern under diverse illu- minated conditions using artificial neural network,”Symmetry, vol. 14, no. 10, p. 2045, 2022

    2045

  28. [28]

    Openillumination: A multi- illumination dataset for inverse rendering evaluation on real objects,

    I. Liu, L. Chen, Z. Fu, L. Wu, H. Jin, Z. Li, C. M. R. Wong, Y . Xu, R. Ramamoorthi, Z. Xuet al., “Openillumination: A multi- illumination dataset for inverse rendering evaluation on real objects,” Advances in Neural Information Processing Systems, vol. 36, pp. 36 951–36 962, 2023

    2023

  29. [29]

    Acquiring linear subspaces for face recognition under variable lighting,

    K.-C. Lee, J. Ho, and D. J. Kriegman, “Acquiring linear subspaces for face recognition under variable lighting,”IEEE Transactions on pattern analysis and machine intelligence, vol. 27, no. 5, pp. 684– 698, 2005

    2005

  30. [30]

    The cmu pose, illumination, and expression (pie) database,

    T. Sim, S. Baker, and M. Bsat, “The cmu pose, illumination, and expression (pie) database,” inProceedings of fifth IEEE international conference on automatic face gesture recognition. IEEE, 2002, pp. 53–58

    2002

  31. [31]

    What is the set of images of an object under all possible illumination conditions?

    P. N. Belhumeur and D. J. Kriegman, “What is the set of images of an object under all possible illumination conditions?”International journal of computer vision, vol. 28, no. 3, pp. 245–260, 1998

    1998

  32. [32]

    An efficient representation for irradiance environment maps,

    R. Ramamoorthi and P. Hanrahan, “An efficient representation for irradiance environment maps,” inProceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001, pp. 497–500

    2001

  33. [33]

    Acquiring the reflectance field of a human face,

    P. Debevec, T. Hawkins, C. Tchou, H.-P. Duiker, W. Sarokin, and M. Sagar, “Acquiring the reflectance field of a human face,” in Proceedings of the 27th annual conference on Computer graphics and interactive techniques, 2000, pp. 145–156

    2000

  34. [34]

    The light stages and their applications to photoreal digital actors,

    P. Debevec, “The light stages and their applications to photoreal digital actors,” 2012

    2012

  35. [35]

    The rendering equation,

    J. T. Kajiya, “The rendering equation,” inProceedings of the 13th annual conference on Computer graphics and interactive techniques, 1986, pp. 143–150

    1986

  36. [36]

    I. W. Busbridge,The mathematics of radiative transfer. Cambridge University Press, 1960

    1960

  37. [37]

    Synthetic lighting for photography,

    P. Haeberli, “Synthetic lighting for photography,”http://www. grafi- caobscura. com/synth/index. html, 1992

    1992

  38. [38]

    Blender,

    Blender, “Blender,” https://www.blender.org/, 2026

    2026

  39. [39]

    High dynamic range imaging,

    E. Reinhard, G. Ward, S. Pattanaik, and P. Debevec, “High dynamic range imaging,” 2006

    2006

  40. [40]

    A flexible new technique for camera calibration,

    Z. Zhang, “A flexible new technique for camera calibration,”IEEE Transactions on pattern analysis and machine intelligence, vol. 22, no. 11, pp. 1330–1334, 2002

    2002

  41. [41]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,”The International Journal of Robotics Research, vol. 44, no. 10-11, pp. 1684–1704, 2025

    2025

  42. [42]

    What matters in learning from large-scale datasets for robot manipulation,

    V . Saxena, M. Bronars, N. R. Arachchige, K. Wang, W. C. Shin, S. Nasiriany, A. Mandlekar, and D. Xu, “What matters in learning from large-scale datasets for robot manipulation,”arXiv preprint arXiv:2506.13536, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    Diffusionlight: Light probes for free by painting a chrome ball,

    P. Phongthawee, W. Chinchuthakun, N. Sinsunthithet, V . Jampani, A. Raj, P. Khungurn, and S. Suwajanakorn, “Diffusionlight: Light probes for free by painting a chrome ball,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 98–108

    2024