acceptodds
Under review as a conference paper at ICLR 2027

Action-Stable HDR Interfaces for Frozen Vision-Language-Action Models

Abstract

Vision-language-action (VLA) policies fail when an exposure change or a bright light patch clips the camera image, even though the task is unchanged: an EV+3 shift takes OpenVLA-OFT, OpenVLA and from 82/90 to 0/90 successful episodes on LIBERO. We ask which camera interface preserves the actions of a frozen VLA and derive a minimum-intervention answer from the observation model. A common task weight cancels from the per-pixel choice of source, and clipping makes single-image recovery non-identifiable. With a second, shorter exposure under an ideal noise-free, unquantized two-level lighting model, four image operations follow: threshold the near-saturated pixels, close them to the smallest compatible lit region, dilate by one pixel to cover a one-pixel localization error, and write the recorded protected values inside. The resulting training-free interface, HDR-Guard, restores 82/90 under the EV+3 shift, where five single-image corrections reach at most 1/45. Under a sensor-aware simulation of a hard-edged light patch brighter than the room, it reaches 76/90 successes, against 70/90 for a soft saturation blend, 57/90 for full-frame Mertens fusion and 21/90 for the raw camera; with 500 episodes per rule, OpenVLA reaches 335/500 with HDR-Guard against 156/500 with Mertens fusion. Against the nominal frame, the lit-region error of HDR-Guard is 6.1 gray levels, against 15.7 for the soft blend; the 1.1% of the frame that the closing adds accounts for 75% of the reconstruction gain, and over 1,100 paired states, writing protected values there has an estimated +3.4-point equal-suite success effect (95% task-cluster interval [-0.4, +7.7]). In a simulation of trigger-free acquisition with the protected frame 50 ms old, OpenVLA reaches 362/500, against 336/500 without delay. On tripod-bracketed DSLR RAW frames, valid protected samples normalized by one calibrated global gain have a 4.8% median relative error against an independent short-exposure reference, and in 178,200 frames of real firefighting footage, 49% of frames with a person show a person with over 5% of their pixels near saturation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action, HDR imaging, embodied AI, robustness, robotics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jiYtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

For the the robustness issue of frozen vision-language-action models under exposure changes or bright light patch clips. The authors argue that existing HDR pipelines are optimized for image reconstruction quality rather than preserving the input distribution required by downstream foundation models. To address this issue, this paper proposes HDR-Guard, a training-free preprocessing module that selectively replaces corrupted regions using a short-exposure protected stream. Experiments on LIBERO with several frozen VLA policies demonstrate that HDR-Guard substantially improves task success rates under severe exposure shifts and localized illumination changes.

Critical strengths
  1. The paper introduces an interesting perspective by formulating HDR processing as an action-preserving sensing interface problem for frozen VLA models, rather than conventional image reconstruction.

  2. The analysis of clipping-induced non-identifiability provides a principled explanation for why single-image correction is insufficient.

  3. HDR-Guard is training-free, interpretable, and computationally efficient, making it attractive for deployment with frozen foundation models.

Critical weaknesses
  1. Limited algorithmic novelty. The proposed HDR-Guard mainly relies on classical image processing operations, and the main contribution lies in problem formulation rather than algorithm design.

  2. The evaluation is mainly conducted on LIBERO with simulated illumination changes, lacking real-world robot experiments and more diverse lighting conditions.

  3. The evaluation mainly compares HDR-Guard with classical HDR methods such as Mertens and Reinhard. Recent learning-based HDR approaches (e.g., HyHDRNet and HDRFlow) are not considered. Evaluating these methods as preprocessing modules for frozen VLA policies would better demonstrate the advantage of the proposed minimum-intervention strategy over modern HDR reconstruction approaches.

  4. The theoretical analysis relies on a strong assumption, a two-level illumination assumption, which may not fully capture complex real-world lighting variations.

Questions for authors
  1. The paper mainly compares with classical HDR approaches. Have the authors evaluated recent learning-based HDR methods as preprocessing modules for frozen VLA policies? If not, could the authors clarify why the proposed minimum-intervention strategy is expected to be more suitable?

  2. The paper motivates HDR-Guard as preserving the nominal input expected by frozen VLA policies. Beyond action consistency, have the authors analyzed whether the restored images or intermediate visual features are closer to the nominal inputs?

  3. How robust is HDR-Guard under realistic deployment conditions?

Reviewer_1s7fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates camera-side exposure correction for frozen vision-language-action policies. It proposes HDR-Guard, a training-free interface that combines a policy-camera image with a shorter-exposure image. The method detects near-saturated pixels, applies morphological closing and one-pixel dilation, and substitutes protected-exposure values within the resulting region while preserving the policy stream elsewhere. Conditional theoretical results motivate the construction and decompose its reconstruction error. Experiments evaluate three frozen VLAs on LIBERO, supplemented by action-consistency measurements, regional interventions, acquisition-delay simulations, and real-camera photometric analysis.

Critical strengths
  1. The problem is practically relevant, and the camera-interface perspective is useful. Exposure clipping can remove task-relevant evidence before a policy processes it. Addressing this through sensing and preprocessing offers a practical alternative to policy finetuning. Evaluating restored inputs through actions and closed-loop success is appropriate for robotic deployment.

  2. The method is simple, training-free, and inexpensive. HDR-Guard requires neither policy modification nor gradient access. Its standard image operations incur only 0.27 ms processing overhead per 256×256 frame on an RTX 4060 Ti, making the interface computationally attractive.

Critical weaknesses
  1. The method works best under restrictive lighting assumptions. It relies on the illumination gain matching the exposure bracket. Under a soft spotlight, HDR-Guard achieves 45/90 successes, substantially below Mertens fusion’s 76/90.

  2. Important restoration baselines are missing. The single-image baselines are simple brightness corrections. Learned restoration methods, particularly those using the same exposure pair, would help separate the benefit of additional sensing from the proposed fusion rule.

  3. The advantage over simple local fusion is inconsistent. HDR-Guard performs similarly to or worse than the soft blend on several suites and policies. Larger experiments mainly focus on OpenVLA, leaving its benefits across different VLA families insufficiently established.

  4. There is no real-world closed-loop robot evaluation. The real-camera experiments examine image statistics and photometric accuracy. They do not show whether HDR-Guard improves robot success under actual lighting and acquisition conditions.

  5. The methodological novelty is limited. The method combines classical saturation-aware merging with morphological operations. Its theoretical results explain this construction under specific assumptions but provide limited guarantees for actual policy behavior.

Questions for authors

See weaknesses.

Reviewer_rjGQOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses image clipping / over-exposure problem for frozen VLA robot policies. Authors point out single-image repair cannot fix clipped observation theoretically, then derive a training-free image interface called HDR-Guard. It uses two bracketed exposures, applies morphological closing+dilation to identify bright regions and selectively fills in content from the short-exposure frame. They run extensive simulation experiments on LIBERO benchmark, compare against classic HDR fusion pipelines, measure closed-loop success rate, action consistency and do various stress tests. Results show HDR-Guard greatly improves robot task success under global exposure shift and local bright light patches.

Critical strengths

1.Clear motivation. Frozen VLA policies completely break under exposure clipping, they even give theoretical proposition to prove single-image recovery is non-identifiable. 2.Extensive experiments. Closed-loop robot rollouts, paired action consistency check, ablation studies, mis-registration / time-delay stress tests, real raw-data sanity check are all present. 3.Good results. The proposed HDR-Guard works well within its assumed setup. Under matched hard-edge lighting patch condition, it outperforms traditional full-frame HDR fusion baselines.

Critical weaknesses

1.This work is a very specific engineering adaptation for frozen robot VLA policies. The scenario assumption is restrictive, audience is limited. It lacks general scientific insight transferable to broader ML / robot learning community. 2.Heavily engineering-oriented rather than fundamental methodological innovation. A lot of design choices are tied tightly to their two-level lighting model and morphological image processing tricks. The heavy case-specific tuning makes it hard to extract general-purpose methodology for other tasks. 3.Abstract puts excessive emphasis on numerical comparison against prior methods. While empirical performance is good, the paper has limited conceptual novelty; good benchmark numbers cannot compensate limited general insight for ICLR.

Questions for authors

This paper is hard to follow, as it contains too much techinical details. I suggest the authors to rewrite abstract: cut down excessive experimental comparison numbers, spend more space on conceptual contributions and limitations of the underlying assumptions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.