acceptodds
Under review as a conference paper at ICLR 2027

ReliAct: Policy-Relevant Reliability for Robust Underwater Vision-Language-Action Models

Abstract

Underwater vision-language-action (VLA) models encounter visual degradation and sensor failures in combinations absent from training. We establish an underwater VLA robustness benchmark on U0/USIM with 18 conditions and 3,560 closed-loop episodes, covering visual, sensor, and composite degradation while preserving dynamics and evaluation ground truth. Motivated by state-dependent policy impact and the limitations of fixed visual repair, we introduce ReliAct. It learns policy-relevant reliability from a frozen policy’s response discrepancy under paired clean and corrupted observations. One predicted vector jointly regulates visual repair, multimodal routing, and action generation, without clean references at inference. ReliAct maintains clean performance and raises aggregate corrupted success from 26.7% to 36.0%. With single-source training, it improves success on three unseen composites by 17.1, 14.3, and 13.6 percentage points over U0. Adding 8.6M trainable parameters and 5.9% inference overhead, ReliAct establishes policy-relevant information use as an effective mechanism for robust underwater control across visual degradation, sensor failure, and unseen combinations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, underwater robotics, robustness, policy-relevant reliability, multimodal sensor fusion, observation degradation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_e9xDOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ReliAct, a lightweight framework for robust underwater vision-language-action control. It derives source-specific reliability targets from a frozen policy’s response discrepancy between paired clean and corrupted observations. An estimator predicts these scores from a short observation history, and a shared reliability vector regulates visual feature repair, multimodal routing, and action generation. The paper also introduces an underwater robustness benchmark with 18 conditions and 3,560 closed-loop episodes per model. ReliAct improves aggregate corrupted success from 26.7% to 36.0%, with 5.9% additional inference latency.

Critical strengths
  1. Practical and well-motivated problem. The paper addresses both visual degradation and sensor failures, which are relevant challenges for underwater robotic control.

  2. Coherent method design. Connecting policy-derived reliability to repair, fusion, and action generation provides a clear mechanism for adapting information use to degraded observations.

Critical weaknesses
  1. The reliability definition lacks direct comparison. The paper does not compare independently trained models using alternative signals, such as sensor health, perceptual quality, or an unconstrained learned gate. Replacing the signal at inference with constants or random values does not establish the superiority of the proposed target.

  2. Component contributions are not fully isolated. The incremental ablations also increase trainable capacity, and the RobustVLA baseline implements only part of the original method. Stronger controls would help attribute the gains.

  3. Evaluation is limited to one backbone and simulator. Transfer to other VLA models and real underwater conditions remains untested.

  4. Composite generalization is tested narrowly. Only three unseen combinations and two withheld corruption families are evaluated. Broader simultaneous failures and prolonged sensor outages remain unexplored.

  5. Uncertainty estimates are incomplete. Training uses a single seed, and the aggregate bootstrap does not account for correlations across conditions sharing task seeds. The variability of improvements over the strongest baselines is therefore unclear.

Questions for authors

See Weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer e9xD

Thank you for the comments. We address each concern below.

1. Direct comparison of reliability supervision. We independently retrained the added modules of four variants with the full ReliAct architecture, 8.6M trainable parameters, frozen U0, identical training data and single-source corruption sampling, AdamW, batch size 64, and 20,000 steps. "No Lrel" removes reliability supervision. "Policy-only" additionally removes clean-reliability regularization while retaining clean-policy consistency. The severity proxy uses q=1-s, with s normalized within each corruption family's training range and unaffected sources labeled 1. Severity is used only for training labels.

SupervisionCleanSevere visualSensorThree composites
No Lrel40.027.936.329.2
Policy-only40.027.134.427.5
Severity proxy40.028.837.530.4
ReliAct41.331.743.837.5

Entries are SR (%), on the Table 4 manifest: navigation/grasping/transport/tracking (G/R/T/F), 20 base scenes each; N=80/240/160/240 across columns. Severe visual comprises L3 turbidity/color shift/low-light; Sensor comprises IMU bias/DVL dropout. Methods share initial states, instructions, environment and corruption seeds.

2. Component attribution and RobustVLA. We retain the full architecture and all computational branches, fix reliability at only one location to the per-source training-target mean, and independently retrain with the same parameter budget and training steps. Fixing repair/routing/action-conditioning inputs yields composite SRs of 33.3/30.4/32.1%, versus 37.5% with all locations dynamic. Fixing repair lowers severe-visual SR from 31.7% to 27.5%; fixing routing lowers sensor SR from 43.8% to 35.6%. These controls test dynamic regulation at each location within the complete architecture.

RobustVLA-Input adapts input consistency to U0 without the original method's complete output-side components. Claims are limited to this implementation; the matched-architecture controls above provide the attribution evidence.

3. Backbone and simulation scope. Our empirical conclusions concern the evaluated U0/USIM underwater simulation tasks. Transfer to other VLA backbones and real underwater platforms has not been tested.

4. Unseen combinations and corruption families. Standard training uses only single-source degradation. Turbidity+low-light, turbidity+IMU, and low-light+DVL never occur jointly during training; their evaluated visual components are L2. Separately, LOCO excludes the entire motion-blur or DVL-dropout family. SRs are 42.5% and 40.7%, retaining approximately 76% and 78% of the full model's gains over U0.

Composites test combination transfer; LOCO tests withheld-family transfer. DVL outages last 0.5 s every 10 s. Prolonged outages and broader simultaneous failures are not covered by these experiments.

5. Paired uncertainty and training variability. The original condition-wise bootstrap did not account for correlations across conditions sharing a base scene. For the original Table 3 comparisons, we now resample base-scene clusters within each task, jointly retaining all conditions and both methods' outcomes. We use 10,000 replicates and the 2.5th/97.5th percentiles.

ComparisonN11/01/10/00Gain (pp)95% CI (pp)
L3 vs SameAug24056/20/12/1523.33[-1.25, 7.92]
Sensor vs Gated U032097/33/20/1704.06[0.00, 8.44]
Cross-modal vs Gated U032088/36/17/1795.94[1.56, 10.31]

SameAug denotes Fixed IB+SameAug; Gated U0 denotes Reliability-Gated U0. Counts are both succeed/only ReliAct succeeds/only baseline succeeds/both fail. Gains use unrounded counts. Table 3 Sensor/Cross-modal each contain four tasks x 40 base scenes x two conditions. Cross-modal contains only turbidity L2+IMU and low-light L2+DVL; these manifests differ from Table 4.

These CIs describe scene-sampling uncertainty at fixed checkpoints. The L3 CI crosses zero, the Sensor CI touches zero, and the Cross-modal CI is strictly positive. Training variability is reported separately: both methods are independently trained using seeds 0/1/2.

For Table 3, Sensor gains are 4.06/3.44/4.69 pp (mean +/- sample SD: 4.06 +/- 0.63); Cross-modal gains are 5.94/5.31/6.88 (6.04 +/- 0.79). Their G/R/T/F gains are 12.50/-2.50/2.50/3.75 and 5.00/7.50/6.25/5.00, respectively. Thus seed-level gains are positive, with task-level variation.

On Table 4's manifest, mean +/- sample SD of gains is 3.61 +/- 0.24 pp (L3 vs SameAug), 5.42 +/- 0.36 (Sensor vs Gated U0), and 5.97 +/- 0.24 (three composites vs Gated U0). All three seeds show positive gains; composite gains are positive on all four tasks.

Thanks for the authors' rebuttal. I maintain my original positive score.

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

ReliAct is a robust underwater VLA framework that predicts policy-relevant reliability from frozen-policy response discrepancies, jointly regulating visual repair, multimodal routing, and action generation to significantly improve success under visual and sensor degradations.

Critical strengths

Addresses sudden multimodal sensor failures in AUVs/ROVs by proposing a dynamic reliability evaluation and fault-tolerant control mechanism, demonstrating strong application potential.

Combines dynamic reliability estimation with a Flow Matching framework; the proposed RC-IB repair and routing mechanism displays superior robustness over standard baselines under various sensor degradation conditions.

Critical weaknesses

The main text lacks self-containment, leaving multiple formulas and symbols undefined in context. Subscripts like r^t,e,r^t,w,r^t,imu,r^t,dvl\hat{r}_{t,e}, \hat{r}_{t,w}, \hat{r}_{t,\text{imu}}, \hat{r}_{t,\text{dvl}} appear abruptly without an upfront definition of the sensor set M\mathcal{M}, causing notational inconsistency with earlier expressions like v∈{e,w}v \in \{e, w\}.

A gap exists between motivation and empirical proof. While the motivation emphasizes that ground-truth r∗r^* exhibits strong task phase-dependency, the experiments rely solely on indirect attribution via improved closed-loop success rates, lacking direct empirical evidence that inference-time predictions (r^\hat{r}) reproduce this phase-dependent behavior during closed-loop execution.

The proposed "policy-relevant reliability" supervision objective lacks rigorous validation. Merely swapping reliability inputs in a pre-trained model introduces severe distribution shift; a fair baseline requires a head-to-head comparison where alternative targets (e.g., image quality, sensor health, feature entropy) are trained from scratch under identical network architecture, parameter count, and training budget.

Questions for authors

see Weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer qPzh

Thank you for the feedback. We address the three concerns below.

1. Observation sources and notation. We use M={e,w,imu,dvl}\mathcal{M}=\{e,w,\mathrm{imu},\mathrm{dvl}\} for all observation sources, with m∈Mm\in\mathcal{M}, while v∈{e,w}v\in\{e,w\} refers only to visual sources. Here e/w denote the ego/wrist cameras, t is the control timestep, and τ\tau is flow-matching time.

rt,m∗r^*_{t,m} is the supervision target constructed from paired policy responses; r^t,m\hat r_{t,m} is the deployment prediction. XX denotes visual input features, ZZ repaired visual or encoded sensor features, Z~\tilde Z routed features, and NN a learnable neutral representation. PηP_\eta projects the reliability vector into an action condition. φ~t\tilde\varphi_t and S~t\tilde S_t are constructed from routed visual and state features. These definitions clarify Fig. 3 and Eqs. (5)-(9).

2. Closed-loop phase analysis under fixed degradation. We compare deployment predictions with policy-derived targets at the same states along closed-loop trajectories. Each setting uses 20 base scenes. Grasping and transport examine the wrist source under 15% occlusion, with the mask unchanged within each episode; navigation examines the ego source under fixed L2 low-light.

Predictions are logged during execution, and clean observations are saved at the same simulator states to compute targets offline. Each paired reference-policy query shares the action chunk, flow time, and noise. Clean observations are used only offline; the policy receives corrupted observations.

The later phase starts when the relevant distance first reaches:

  • Grasping: end effector to target <=0.30 m.
  • Transport: carried object to placement target <=0.50 m.
  • Navigation: vehicle to goal <=2.00 m.

Samples before/after this boundary define the two phases. Trajectories that do not reach the boundary or lack valid records in either phase are excluded from paired analysis but remain in the total-scene denominator. We first compute phase means within each trajectory, then average equally over valid paired trajectories. Phase differences are earlier minus later. Direction agreement requires both predicted and target differences to be nonzero and have the same sign; zero differences count as disagreements.

Setting and phasesPaired/totalTarget: early -> latePredicted: early -> lateDirection agreement
Grasping: approach -> final alignment18/200.69 -> 0.420.66 -> 0.4416/18 (88.9%)
Transport: cruise -> placement alignment17/200.71 -> 0.450.68 -> 0.4715/17 (88.2%)
Navigation: cruise -> target approach19/200.73 -> 0.540.70 -> 0.5517/19 (89.5%)

Predicted reliability decreases by 0.22/0.21/0.15 in the later phases, matching the target's direction. This directly supports phase-dependent deployment predictions under the tested fixed degradations. This table reports multi-scene statistics; original Table 1 reports representative cases.

3. Matched-architecture supervision controls. We independently retrained the added modules of four variants with the full ReliAct architecture, 8.6M trainable parameters, frozen U0, identical training data and single-source corruption sampling, AdamW, batch size 64, and 20,000 steps.

"No Lrel" removes reliability supervision. "Policy-only" additionally removes clean-reliability regularization while retaining clean-policy consistency. The severity proxy replaces the policy-derived target with q=1-s, where s is normalized within each corruption family's original training range; unaffected sources have label 1. Severity is used only for training labels. Labels decrease as degradation increases: lower transmission, color gain or brightness, or greater occlusion, blur, IMU bias or DVL outage duration.

SupervisionCleanSevere visualSensorThree composites
No Lrel40.027.936.329.2
Policy-only40.027.134.427.5
Severity proxy40.028.837.530.4
ReliAct41.331.743.837.5

Entries are SR (%), using the original Table 4 manifest: navigation/grasping/transport/tracking, 20 base scenes each. Column sizes are 80/240/160/240 episodes. Methods share initial states, instructions, environment and corruption seeds. Severe visual comprises L3 turbidity/color shift/low-light; Sensor comprises IMU bias/DVL dropout. The composites are turbidity+low-light, turbidity+IMU and low-light+DVL, with L2 visual components.

Policy-derived supervision improves severe-visual/sensor/composite success by 2.9/6.3/7.1 percentage points over the severity proxy and 4.6/9.4/10.0 over policy-only learning. The phase analysis tests prediction behavior; the retrained controls support downstream utility relative to the compared alternatives.

Hi,

please avoid irresponsible review. current quality is not quite acceptable.

AC

Thank you for the detailed rebuttal. Two concerns remain.

  1. The Introduction motivates ReliAct against fixed processing such as feature filtering and consistency training (StableVLA, RobustVLA). But underwater image enhancement and restoration is a mature field, and a restored image can be fed directly into a VLA. The paper cites Akkaynak & Treibitz and Li et al. (UIEB) only for degradation modeling. The Related Work does not discuss enhancement or restoration methods, and no baseline uses them. The Introduction therefore does not state the gap relative to the underwater literature, and I think it needs further revision before publication.

  2. Sec. 4.3 state that training uses single-source corruption only, and that the targets do not explicitly model nonlinear interactions between simultaneous degradations. Real underwater failures are often concurrent. This train/deploy mismatch remains unaddressed.

Given these remaining concerns, especially since the introduction needs further revision, I maintain my current rating.

AuthorsOpenReview ↗

Follow-up on Underwater Restoration and Composite Degradation

Thank you for the follow-up. We have made focused revisions to clarify ReliAct’s relationship to underwater restoration and the interpretation of its composite-degradation results. The corresponding changes are marked in blue in the revised manuscript.

1. Positioning relative to underwater enhancement and restoration.

The revised Introduction explicitly recognizes underwater enhancement and restoration as established approaches whose outputs can serve as visual inputs to a VLA policy. Related Work now discusses Sea-thru and UIEB/Water-Net.

We clarify the complementary question addressed by ReliAct: how should a policy regulate visual and navigation inputs when their usefulness changes with task state? ReliAct connects this decision to policy-relevant reliability, using one shared signal to regulate feature repair, multimodal routing, and action generation. Section 3 now explicitly identifies the evaluated fixed-repair baseline as StableVLA’s feature-space IB operator. Section 6 identifies direct closed-loop comparison with dedicated underwater restoration pipelines as a remaining evaluation direction.

2. Single-source training and concurrent failures.

Single-source supervision isolates each source’s effect on the reference policy, while withholding composite conditions allows us to evaluate transfer to concurrent degradation. At inference, ReliAct estimates reliability from the available observations and conditions action generation on the complete predicted vector.

The existing closed-loop evaluation covers three withheld combinations: turbidity with low-light, turbidity with IMU bias, and low-light with DVL dropout. Table 2 reports improvements of 17.1, 14.3, and 13.6 percentage points over U0, respectively. These results provide evidence of transfer from individual corruptions to their co-occurrence.

We have clarified this protocol in Section 4.3 and aligned Sections 5.2 and 6 with that evidence. The revised text distinguishes the demonstrated transfer to these combinations from explicit supervision of nonlinear failure interactions, which remains an open aspect of the single-source training design.

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

ReliAct addresses underwater vision-language-action robustness under visual degradation and sensor failure. The paper introduces an 18-condition benchmark on U0/USIM, comprising 3,560 closed-loop episodes per model. Its central mechanism is source-specific reliability supervised by the discrepancy between a frozen policy’s responses to paired clean and single-source-corrupted observations. At deployment, a lightweight estimator predicts reliability without clean references, and the resulting vector jointly regulates visual repair, multimodal routing, and action generation. Experiments report improved corrupted-condition success, transfer to three unseen corruption combinations, and modest adaptation overhead.

Critical strengths

The principal results measure closed-loop task success. Across 3,320 corrupted episodes, success increases from 26.7% to 36.0%, with a reported paired-bootstrap confidence interval of approximately +7.7 to +10.8 percentage points. This provides meaningful evidence that the complete adaptation improves behavior under the tested conditions.

The evaluation includes relevant alternatives to the unadapted policy. Fixed IB + SameAug addresses the contribution of corruption augmentation, while Reliability-Gated U0 provides a simpler reliability-based approach. Single-source training followed by unseen-composite evaluation also directly addresses the motivation for handling combinations absent from training.

The adaptation is lightweight, adding 8.6M trainable parameters and 5.9% mean inference overhead. The paper also appropriately limits the interpretation of its reliability target: preserving a reference policy’s response is neither objective sensor accuracy nor a calibrated probability of task success.

Critical weaknesses

The distinctive value of policy-derived reliability supervision is not sufficiently isolated. The results support the complete system, but lack a comparison that preserves its architecture and training budget while changing only the reliability supervision. Replacing the signal with random or all-one values after training introduces distribution shift and does not establish superiority over alternative learned gating signals.

The component ablations are confounded by model capacity. Incrementally adding repair, routing, and action conditioning also changes the trainable architecture. The fixed-capacity joint-versus-frozen-estimator comparison is informative, but primarily tests optimization strategy rather than isolating the contribution of each component.

The matching conditions for the key gating baseline are unclear. Reliability-Gated U0 is an important test of whether ReliAct’s coordinated design is necessary. However, it is not clear whether the baseline shares the same reliability target, estimator inputs, supervision, and training procedure, leaving the source of the performance difference ambiguous.

The gains over the strongest adapted baselines are less firmly established than the improvement over U0. Table 3 reports margins of 3.4 percentage points over Fixed IB + SameAug under severe visual corruption and 4.0/6.0 points over Reliability-Gated U0 under sensor/composite degradation. The strongest statistical evidence concerns comparison with U0; the stability of these more consequential margins across tasks and adaptation seeds remains insufficiently characterized.

Questions for authors

Is there a comparison that preserves the full architecture and training budget while changing the reliability supervision—for example, learning the routing signal through the policy loss alone? Such a comparison would address the proposed target more directly than replacing its output only at inference.

Please clarify what is matched between Reliability-Gated U0 and ReliAct, particularly the reliability target, estimator inputs, supervision, and training procedure. This is necessary to determine whether their difference isolates coordinated information use or also reflects differences in estimator quality.

For the key Table 3 comparisons, please provide paired outcome counts and uncertainty estimates, particularly against Fixed IB + SameAug under severe visual corruption and Reliability-Gated U0 under sensor/composite degradation. How were the evaluation manifests selected, and are the gains consistent across tasks or any available adaptation seeds?

1 comment
AuthorsOpenReview ↗

Response to Reviewer deHJ

Thank you for the constructive comments. We address the supervision target, component controls, baseline matching, and statistical stability below.

1. Direct comparison of reliability supervision. We independently retrained four variants with the full ReliAct architecture, 8.6M trainable parameters, frozen U0, identical training data and single-source corruption sampling, AdamW, batch size 64, and 20,000 steps. "No Lrel" removes the reliability-supervision loss. "Policy-only" also removes clean-reliability regularization while retaining clean-policy consistency. The severity proxy uses q=1-s, where s is normalized within each corruption family's training range; unaffected sources have label 1. Severity is used only to construct training labels.

SupervisionCleanL3 visualSensorComposite
No Lrel40.027.936.329.2
Policy-only40.027.134.427.5
Severity proxy40.028.837.530.4
ReliAct41.331.743.837.5

Entries are SR (%). These controls use the Table 4 manifest: navigation/grasping/transport/tracking (G/R/T/F), 20 base scenes per task; N=80/240/160/240 across columns. L3 visual covers turbidity, color shift, and low-light; Sensor covers IMU/DVL; Composite covers turbidity+low-light, turbidity+IMU, and low-light+DVL, with L2 visual components. Initial states, instructions, environment and corruption seeds are shared. The manifest retains the first 20 valid initializations per task starting from seeds 0-19, fixed before evaluation without using success outcomes.

ReliAct improves over the severity proxy by 2.9/6.3/7.1 points and over policy-only learning by 4.6/9.4/10.0 points on the three corrupted groups.

2. Dynamic regulation at fixed architecture and parameter budget. We retain every computational branch and independently retrain each variant, fixing reliability at only one location to the per-source training-target mean determined before evaluation. Other locations remain dynamic; data and training steps are matched.

Fixed reliability inputL3 visualSensorComposite
Repair27.542.533.3
Routing30.835.630.4
Action conditioning31.340.632.1
None (full ReliAct)31.743.837.5

This tests each location's dynamic use of reliability within the complete architecture.

3. Matching Reliability-Gated U0. Both methods share frozen U0, the policy-derived target, three-frame multimodal inputs, the 1536-1024-512-4 estimator, reliability supervision, and corruption sampling. Both use AdamW, learning rate 2e-4, weight decay 0.01, batch size 64, and 20,000 steps. Total trainable parameters are 2.6M/8.6M for Gated U0/ReliAct.

In a separate diagnostic sharing and freezing the same estimator checkpoint, direct gating obtains 24.6/37.5/29.2% on L3 visual/Sensor/Composite, versus 30.4/40.6/33.8% for the complete downstream design. This controls estimator differences; the fixed-budget controls above isolate dynamic regulation at each location.

4. Paired statistics for the original Table 3. These use the original Table 3 manifests. L3 uses 240 episodes; Sensor and Cross-modal each use four tasks x 40 base scenes x two conditions (N=320). Cross-modal contains only turbidity L2+IMU and low-light L2+DVL. These sensor/composite manifests differ from Table 4.

ComparisonN11/01/10/00Gain (pp)95% CI (pp)
L3 vs SameAug24056/20/12/1523.33[-1.25, 7.92]
Sensor vs Gated U032097/33/20/1704.06[0.00, 8.44]
Cross-modal vs Gated U032088/36/17/1795.94[1.56, 10.31]

SameAug denotes Fixed IB+SameAug. Counts are both succeed/only ReliAct succeeds/only baseline succeeds/both fail. Gains use unrounded counts. We resample base-scene clusters within each task, jointly retaining all conditions and both methods' outcomes, for 10,000 replicates; CIs use the 2.5th/97.5th percentiles. They describe scene-sampling uncertainty at fixed checkpoints, separately from training variability. The L3 interval crosses zero, Sensor touches zero, and Cross-modal is strictly positive.

With both methods independently trained using seeds 0/1/2, Table 3 Sensor gains are 4.06/3.44/4.69 pp (mean +/- sample SD: 4.06 +/- 0.63); Cross-modal gains are 5.94/5.31/6.88 (6.04 +/- 0.79). Per-task G/R/T/F gains are 12.50/-2.50/2.50/3.75 for Sensor and 5.00/7.50/6.25/5.00 for Cross-modal. Thus gains are positive across training seeds, with task-level variation.

On the Table 4 manifest, paired gains over three training seeds (mean +/- sample SD) are 3.61 +/- 0.24 (L3 vs SameAug), 5.42 +/- 0.36 (Sensor vs Gated U0), and 5.97 +/- 0.24 pp (Composite vs Gated U0); all three seeds give positive gains.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.