acceptodds
Under review as a conference paper at ICLR 2027

Uncertainty-Aware World Model for Aerial Image-Goal Navigation

Abstract

Aerial image-goal navigation requires an unmanned aerial vehicle (UAV) to reach a target location specified by a goal image. Existing world-model-based methods rank candidate trajectories using predicted futures, but typically rely on only one or a few point predictions, which is inadequate for large-scale outdoor environments with substantial future-state uncertainty. To address this limitation, we propose the Uncertainty-Aware Navigation World Model (UA-NWM), an efficient latent world model for aerial image-goal navigation, which formulates trajectory scoring as conditional out-of-distribution detection. UA-NWM represents plausible futures with an uncertainty subspace and decomposes the prediction–goal discrepancy into uncertainty-explainable and unexplainable components. Only the unexplainable residual is used for scoring, enabling robust selection without multiple future samples. Extensive experiments show that UA-NWM consistently outperforms existing navigation world models while maintaining low inference latency. Real-world UAV experiments further validate its practical applicability.

Primary area:
applications to robotics, autonomy, planning
Keywords:
image-goal navigation, world models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_f9RcOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a novel model and a novel dataset for the task of visual aerial image-goal navigation. In their setting, a world-model is used to rank different candidate trajectories (i.e., future actions) given a history of observations and a goal image. The proposed algorithm consist mainly of a deterministic baseline module that predicts a single future image and a projection module (HEP) that estimates the error between the goal and predicted images at multiple scales in order to better compute a score for a candidate trajectory. These ideas are tested on a novel aerial benchmark, showing that their methods excel in navigation and are faster than similar world-based methods. Real world experiments are also provided to validate the practicality and transferability of the method.

The paper overall seems to be of high quality and I commend the authors for their work. I believe it could be of benefit to the community, so my recommendation leans towards acceptance. I do have certain questions and concerns that prevent me from assigning a higher score, as will be expanded later in the review.

Critical strengths

Perhaps the main strength of the submission is that the proposed method seems to be thoroughly validated and performs very well on the benchmarks. This includes the main aerial benchmark (both the novel dataset and how UA-NWM and other methods perform in it), the real world experiments and even 2D navigation tasks presented in the appendix. Though I raise certain perceived weaknesses in much more detail in my review, I still believe the authors provide reasonable and extensive evidence that the proposed ideas have merit and could be valuable to the community.

I also believe the resulting algorithm to be reasonably lightweight compared to similar work, such that it could be a good starting point for future work, particularly under a more limited computational budget and for embedded systems.

Critical weaknesses

In my perception, the paper doesn't analyze nor discuss of the deterministic baseline properly. Some experimental choices may also be inflating the reported performance in relation to previous work

The most impactful component (deterministic baseline) is underexplored

Most of the benefits of the proposed methods in relation to previous work, both in terms of computation time and navigation performance, seem to already be present in the deterministic baseline. The introduction states that the main issue of stochastic world models is that they are computationally expensive (the issue of sampling sensitivity is also raised and is reasonable, but is never quantified or explored further, so it seems to be of lesser importance). One would then expect to observe a tradeoff between performance and computational cost. This, however, seems incongruent with the reported experiments, where the deterministic baseline is already both less expensive and more performant than the other methods presented even without the proposed HEC. So either the introduction doesn't frame the problem correctly or there is a very impactful part of the model that is not present in previous work, yet is never identified.

In the lack of relevant discussion or ablation, it would seem that this key underappreciated component is the delta parametrization. This formulation seems to have come from Delta-World (Kerssies et al., 2026), as the authors of this submission properly cite, and by itself it reduced the FLOP count of other world models by 3 orders of magnitude. That seems to be a very important component that could explain a significant part of the reported performance, perhaps more so than the ideas introduced in UA-NWM. It would seem that not identifying this, or whatever other aspect of the model was most significant, obscures valuable insight and potentially fails to acknowledge the significance of previous work.

To be clear, I am not diminishing the value of the HEC module (which seems valuable by itself and is properly discussed) or other contributions related to the deterministic baseline (e.g., the other components and how to integrate them or how to apply the delta parametrization in this setting). My critique is that the paper focuses on the HEC module, while almost ignoring the baseline that seems essential to their results, even when it seems to contradict what is claimed in the introduction.

Potentially unfair comparisons with previous work

The authors use DINOv3 for their own method, but appear to pick the older and less capable DINOv2 as the backbone for RAE-NWM. There doesn't seem to be any justification for that choice and the natural expectation is that it would unfairly degrade the other method's performance. It could be argued that the authors of RAE-NWM originally used DINOv2, but it could also be counter-argued that the RAE-NWM paper is older and DINOv3 is still relatively new. One-step WM is a newer paper that used the stable diffusion VAE (i.e., it's not in the same family of models), so perhaps the same cannot be said in its case, but it would seem ideal to control for the effect of the encoder choice in the experiments to the extent possible. Otherwise, it's hard to separate the benefits of the paper's own contributions from those resulting from a better encoder choice.

For fairness, I'll acknowledge that adapting other works for a new dataset is an open ended task and many reasonable choices are possible. I'll also acknowledge that the experiments in table 12 do seem to use previous work in their original settings and UA-NWM still outperforms them. Still, that appears to be a weakness of the main experimental results that may also help explain the improved navigation performance of the deterministic baseline as pointed in the previous weakness

Questions for authors

Question (related to assessment):

  • What dominates the computational cost of previous methods? Given that One-step WM used a single diffusion step and was relatively similar to UA-NWM in terms of time per frame, while the other ones took multiple steps and were significantly slower, I am led to believe that using a stochastic world model isn't as much of an issue as the submission seems to claim and ultimately the problem is the cost of multi-step diffusion

Questions and suggestions (not related to assessment):

  • Unless I missed where that was written, I think it would facilitate reading if the other world models tested (NWM, RAE-NWM, etc) were explicitly classified as being stochastic or deterministic, given that is a central point in the text.
  • I believe many readers expect a certain type of content when they see the words "uncertainty". For instance: covariances, ensemble, MC-dropout, aleatoric vs. epistemic uncertainty, calibration. This submission doesn't seem to meet this expectation that well as the connection to uncertainty is relatively lax. For comparison, consider how the works this paper cites (Seo et al., 2025; Mei et al., 2025) do met those expectations and are more rigorous about the topic. So I'd suggest the authors to consider ways to manage this expectation from potential readers earlier in the text. Adding other relevant works that use "uncertainty-aware" and OOD in a similar way would also be helpful
  • Did you try running UA-NWM on the Orin NX rather than the laptop? I believe many practitioners are familiar with the capabilities of all kinds of Jetson GPUs, so understanding how they perform in this task (if they can do it all) seem useful for future work.
7 comments
AuthorsOpenReview ↗

Response to Reviewer f9Rc (1/5)

Thank you for the constructive and perceptive review. We greatly appreciate your positive assessment of the paper's overall quality, extensive validation, and practical efficiency, as well as the concrete suggestions for strengthening our analysis. We have added controlled experiments and clearer discussion, and respond to each point below.

For clarity, our response is organized as follows:

  1. What actually dominates the computational cost? (W1&Q1)
  2. What trade-offs and components explain the strong deterministic baseline? (W1&Q1)
  3. Encoder fairness (W2)
  4. Explicit classification of evaluated world models (Q2)
  5. Mathematical and empirical meaning of “uncertainty” (Q3)
  6. Jetson Orin NX deployment (Q4)

1. What actually dominates the computational cost? (W1&Q1)

Thank you for asking us to distinguish the computational cost of stochastic world models from that of multi-step diffusion. We agree that the two should not be treated as equivalent. The extra computational cost of existing stochastic world models mainly comes from two sources.

  • First, many stochastic world models use multi-step diffusion because they place considerable emphasis on the visual quality and details of generated observations. Iterative denoising is therefore a major source of computational cost for most existing stochastic world models. However, this is not an inherent requirement of every stochastic world model. One-Step WM, as you note, is an important exception that performs stochastic generation in a single step.

  • Second, repeated sampling can introduce extra cost. Stochastic world models can represent multiple plausible future states, but a single forward pass produces only one sample from that distribution. The resulting trajectory score can therefore be sensitive to which future happens to be sampled. The official NWM paper addresses this issue by running NWM three times for each action sequence and averaging the three LPIPS scores between the predicted final state and the goal. This existing strategy shows that repeated sampling is an established practical approach to stabilizing trajectory scoring. As stated in our Introduction: “Stochastic models can capture diverse plausible futures, but determining whether the goal belongs to this distribution requires sufficient samples to approximate it, incurring high computational cost.”

We further explore the issue of sampling sensitivity using One-Step WM. In the main experiment, each candidate trajectory receives one predicted future (N=1N=1), which is compared with the goal image to score the trajectory. We additionally evaluate N=3N=3. For every candidate, we perform three forward passes and score it using the sampled future most similar to the goal image.

One-Step WM samples per candidateRanking-8 ATE / RPE ↓Ranking-16 ATE / RPE ↓Ranking-32 ATE / RPE ↓Planning ATE / RPE ↓
N=1N=11.52 / 0.411.48 / 0.401.44 / 0.391.53 / 0.44
N=3N=31.50 / 0.411.46 / 0.401.41 / 0.391.52 / 0.44

Increasing the number of samples consistently improves ATE, while RPE remains tied at the reported precision. This experiment shows that additional sampling can partly reduce the sampling sensitivity of the same stochastic world model at the cost of additional computation.

We nevertheless agree that “high computational cost” is too absolute as a description of the entire category. Even with N=3N=3, One-Step WM remains among the faster evaluated world models. We have therefore revised the manuscript to describe the general drawback as extra computational cost, which may arise from iterative generation, repeated sampling, or both. We have also added Appendix A.9.1, which formally defines the multi-sample scoring strategy, reports the complete One-Step WM N=1N=1 versus N=3N=3 results, and discusses the trade-off between reduced sampling sensitivity and additional world-model evaluations.

2. What trade-offs and components explain the strong deterministic baseline? (W1&Q1)

Thank you for asking us to examine both the expected trade-offs and the components underlying the strong deterministic baseline.

Expected trade-offs

You note that the efficiency of our deterministic baseline would normally be expected to come with a trade-off. Our results reveal two relevant trade-offs, which should be distinguished from a universal trade-off between inference cost and navigation performance.

  • Within stochastic world models, there is a trade-off between sampling sensitivity and inference cost. Increasing the number of samples can partly reduce sampling sensitivity and improve navigation performance, but it also increases inference cost and cannot completely remove sampling randomness, as already discussed above.
AuthorsOpenReview ↗

Response to Reviewer f9Rc (2/5)

  • Across many existing stochastic models and our deterministic baseline, there is also a trade-off between visual detail and inference cost. Here, visual detail refers only to image sharpness and fine-grained appearance, not to semantic, physical, or navigation accuracy. As shown in Fig. 5 and Appendix Figs. 21--23, stochastic models such as NWM and RAE-NWM produce images with finer visual detail than our deterministic baseline. This difference partly reflects the tendency of deterministic prediction to average multiple possible futures, and it is also strongly influenced by the iterative denoising used by many stochastic models to improve visual detail.

However, finer visual detail does not necessarily imply better navigation performance. Our new NWM diffusion-step comparison provides a direct control for this distinction because it changes the inference denoising budget while keeping the checkpoint, test cases, candidate sets, and task protocol unchanged. The official 250-step setting performs substantially more denoising and would generally be expected to produce finer visual detail than the 25-step setting. Nevertheless, 250 steps do not improve navigation performance, while 25 steps achieve equal or lower errors in every comparison. This result shows that visual detail and navigation performance are not equivalent. We added this diffusion-step analysis to Appendix A.9.2.

NWM inferenceRanking-8 ATE / RPE ↓\downarrowRanking-16 ATE / RPE ↓\downarrowRanking-32 ATE / RPE ↓\downarrowPlanning ATE / RPE ↓\downarrow
25 steps1.47 / 0.401.40 / 0.381.37 / 0.371.40 / 0.39
250 steps1.48 / 0.401.43 / 0.391.40 / 0.381.42 / 0.39

We do not claim a universal trade-off between inference cost and navigation performance for the stochastic and deterministic categories. The Introduction, Related Work, and Fig. 2 describe the respective strengths and limitations of both categories without assuming that either must navigate better. In our experiments, stochastic world models are often outperformed by the deterministic approach, indicating that sampling sensitivity can outweigh the benefit of finer visual detail for navigation performance.

Ablation and discussion for the deterministic baseline

The original submission did not sufficiently isolate and analyze the deterministic baseline. We have therefore moved the staged-training and delta-representation ablation from the appendix to Sec. 5.4 and expanded it to evaluate both the deterministic baseline and UA-NWM. The updated results are shown below.

MethodStaged trainingDelta representationRanking-8 ATE / RPE ↓\downarrowRanking-16 ATE / RPE ↓\downarrowRanking-32 ATE / RPE ↓\downarrow
Deterministic Baseline×\times×\times1.424 / 0.3931.411 / 0.3891.423 / 0.392
Deterministic Baseline×\times✓\checkmark1.420 / 0.3881.351 / 0.3701.281 / 0.351
Deterministic Baseline✓\checkmark×\times1.423 / 0.3921.389 / 0.3831.403 / 0.387
Deterministic Baseline✓\checkmark✓\checkmark1.391 / 0.3801.325 / 0.3641.267 / 0.348
UA-NWM×\times×\times1.383 / 0.3801.321 / 0.3631.297 / 0.357
UA-NWM×\times✓\checkmark1.309 / 0.3601.222 / 0.3381.189 / 0.330
UA-NWM✓\checkmark×\times1.370 / 0.3751.309 / 0.3591.263 / 0.347
UA-NWM✓\checkmark✓\checkmark1.256 / 0.3471.174 / 0.3261.091 / 0.304

For both the deterministic baseline and UA-NWM, the delta representation provides the largest gain, followed by staged training. Predicting inter-frame changes reduces the visual redundancy shared by adjacent observations and directs model capacity toward dynamic variations. Staged pretraining provides a stable initialization for the compressor, fuser, and delta decoder before future prediction is learned.

Beyond the demonstrated contributions of staged training and delta representation, an additional plausible explanation for the strong baseline is deterministic prediction itself, which produces a repeatable, conservative future estimate without inference-time sampling variance. This property may contribute to the observed gap over stochastic world models when sampling sensitivity outweighs the benefit of generating sharper visual details.

AuthorsOpenReview ↗

Response to Reviewer f9Rc (3/5)

3. Encoder fairness (W2)

Thank you for highlighting the visual encoder as a potential source of the observed performance differences. We agree that the visual encoder is an important part of a world model because it determines the representation space in which future prediction is learned. In the original comparison, we followed the official configurations as closely as possible, including DINOv2 for RAE-NWM and the VAE for One-Step WM. To separate the effect of the encoder from that of the world-model design, we have now conducted controlled comparisons in which competing methods use the same visual encoder.

Shared DINOv3 encoder.

MethodRanking-8 ATE / RPE ↓\downarrowRanking-16 ATE / RPE ↓\downarrowRanking-32 ATE / RPE ↓\downarrowPlanning ATE / RPE ↓\downarrow
RAE-NWM (DINOv3)1.50 / 0.411.42 / 0.391.42 / 0.391.37 / 0.39
One-Step WM (DINOv3)1.55 / 0.421.51 / 0.411.50 / 0.411.63 / 0.47
InJepa (DINOv3)1.43 / 0.391.34 / 0.371.27 / 0.351.42 / 0.42
Deterministic Baseline (Ours, DINOv3)1.39 / 0.381.33 / 0.361.27 / 0.351.26 / 0.35
UA-NWM (Ours, DINOv3)1.26 / 0.351.17 / 0.331.09 / 0.301.22 / 0.33

Shared DINOv2 encoder.

MethodRanking-8 ATE / RPE ↓\downarrowRanking-16 ATE / RPE ↓\downarrowRanking-32 ATE / RPE ↓\downarrowPlanning ATE / RPE ↓\downarrow
RAE-NWM (DINOv2)1.54 / 0.421.54 / 0.421.50 / 0.411.46 / 0.43
Deterministic Baseline (Ours, DINOv2)1.38 / 0.381.28 / 0.351.23 / 0.341.30 / 0.36
UA-NWM (Ours, DINOv2)1.35 / 0.371.26 / 0.351.21 / 0.331.28 / 0.35

Shared V-JEPA 2.1 encoder.

MethodRanking-8 ATE / RPE ↓\downarrowRanking-16 ATE / RPE ↓\downarrowRanking-32 ATE / RPE ↓\downarrowPlanning ATE / RPE ↓\downarrow
InJepa (V-JEPA 2.1)1.48 / 0.401.41 / 0.391.33 / 0.371.54 / 0.45
Deterministic Baseline (Ours, V-JEPA 2.1)1.34 / 0.371.26 / 0.351.19 / 0.331.29 / 0.36
UA-NWM (Ours, V-JEPA 2.1)1.33 / 0.371.25 / 0.351.18 / 0.331.27 / 0.36

Shared VAE encoder.

MethodRanking-8 ATE / RPE ↓\downarrowRanking-16 ATE / RPE ↓\downarrowRanking-32 ATE / RPE ↓\downarrowPlanning ATE / RPE ↓\downarrow
NWM (VAE)1.47 / 0.401.40 / 0.381.37 / 0.371.40 / 0.39
One-Step WM (VAE)1.52 / 0.411.48 / 0.401.44 / 0.391.53 / 0.44
MWM (VAE)1.58 / 0.431.60 / 0.431.59 / 0.431.54 / 0.43
InJepa (VAE)1.66 / 0.461.70 / 0.461.74 / 0.471.93 / 0.58
Deterministic Baseline (Ours, VAE)1.58 / 0.431.59 / 0.431.62 / 0.441.62 / 0.47
UA-NWM (Ours, VAE)1.60 / 0.441.61 / 0.441.65 / 0.451.77 / 0.52

These results support three conclusions.

  • First, the encoder does matter. For example, RAE-NWM improves when DINOv2 is replaced by DINOv3, and InJepa similarly improves when V-JEPA 2.1 is replaced by DINOv3.

  • Second, the advantage of UA-NWM is not explained by using DINOv3 alone. Under the controlled DINOv3 comparison, UA-NWM outperforms RAE-NWM, One-Step WM, and InJepa in all four tasks. It also remains the best method in the corresponding DINOv2 and V-JEPA 2.1 comparisons. Moreover, HEP consistently improves over its matched deterministic baseline with these three feature encoders, showing that its benefit transfers across multiple representation spaces, although the gain varies with the encoder.

  • Third, the VAE control exposes an important limitation rather than favoring our method. Our deterministic baseline performs poorly with this encoder, and adding HEP further degrades the results. InJepa also deteriorates in every task when its native V-JEPA 2.1 representation is replaced with a VAE, with a particularly large planning degradation. Conversely, replacing One-Step WM's native VAE representation with DINOv3 features also reduces its performance. Thus, the controlled experiments do not support the conclusion that one encoder is universally better or that our advantage over stochastic models comes merely from replacing a VAE with DINOv3. Instead, they indicate an interaction between the representation space and the prediction formulation. A VAE latent is optimized for probabilistic image reconstruction and is naturally paired with a generative objective that models its distribution. Directly regressing its next latent with a deterministic causal Transformer can yield an inaccurate central prediction. Because HEP learns local residual directions around that prediction, an unreliable center can in turn produce an overly permissive subspace that absorbs discrepancies rather than preserving useful trajectory discrimination.

Reviewer_KJK9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work focus on aerial image-goal navigation. Different from world-model-based methods that using predicted futures from a few samples, this work purpose Uncertainty-Aware Navigation World Model. The proposed method outperforms existing navigation world models with low inference latency. This work is a good attempt for replacing mainstream method that predict single data point.

Critical strengths
  1. Apply world model in Navigation is a straightforward transfer and this is an important direction to research.

  2. The key contribution of this work is to try challenge existing world model for aerial navigation paradigm. Previous works treat the future as a point. This work try a different way.

  3. The main story is interesting. What this work truly challenges is the prevailing paradigm in world models for aerial navigation—specifically, treating the future as a point-based estimate. Typically, candidate trajectories are scored by predicting what will be seen after executing the trajectory and matching it against the goal image. Existing methods either use deterministic approaches (collapsing the future into a single point) or stochastic approaches (sampling multiple points). The problem is that real-world navigation futures exhibit a multi-modal distribution, making both approaches prone to misjudging viable trajectories. Instead of asking 'what does the future look like?', this paper subtly shifts the question to 'does the target lie within this trajectory's future distribution?' By framing trajectory scoring as conditional OOD detection, the paper convincingly demonstrates that explicit uncertainty modeling yields significant performance gains."

Critical weaknesses
  1. The main story is appealing but this work is somewhat far away from addressing it. The core claim was actually framed too broadly, the current design does not fully cover it.

  2. Although this work reframes trajectory scoring as evaluating whether a goal belongs to the potential future generated by that trajectory, its implementation still follows a two-step process: it first predicts a deterministic future, and then constructs a tolerance space for variation around it based on the prediction errors learned during training. It primarily describes which discrepancies can be tolerated, but fails to fully capture the probabilities of various futures occurring. This means that when an unseen scenario during training introduces novel prediction errors, the model may struggle to distinguish whether these differences represent reasonable future variations or indicate an inherent flaw in the trajectory itself.

Questions for authors

Pls discuss the weakness point-by-point.

4 comments
AuthorsOpenReview ↗

Response to Reviewer KJK9 (1/3)

Thank you for the thoughtful, encouraging, and insightful review. We especially appreciate your recognition that moving beyond point-prediction trajectory scoring is an important problem and that the central idea of UA-NWM is promising. We address both of your concerns below and have strengthened the revised manuscript with direct evidence and additional evaluation.

1. What supports the paper's central story and core claim? (W1)

Thank you for asking whether the current method and evidence substantiate the paper's central story. The main story consists of a connected sequence: multiple future visual states can be compatible with the same context and action, HEP represents their structured, candidate-conditioned feature variations, its remaining residual distinguishes compatible from incompatible goals, and this distinction improves trajectory selection. The revised manuscript provides complementary evidence for every part of this sequence.

Does HEP represent multiple admissible futures beyond one prediction? The qualitative results in Figure 5 and Figure 19 of the revised manuscript show that the same observation context and action sequence can support different tree coverage, roof appearance, and local building details. Samples generated along the learned directions express these alternatives while retaining the overall scene layout. The goal-oriented visualization in Figure 17 of the revised manuscript provides a stronger controlled test. HEP can reconstruct edited goals that remain compatible with the original trajectory, whereas a goal image taken from the endpoint of a different trajectory leaves a large unexplained discrepancy. Figure 13 of the revised manuscript further shows that the learned directions produce structured variations in separately collected unseen environments. These results show that the representation captures multiple admissible feature variations rather than collapsing the future into one prediction.

Are the learned variations candidate-conditioned rather than a fixed global error model? HEP predicts its local multi-scale bases from the context and action-conditioned deterministic prediction. Appendix A.8.3 explains why its training objective favors recurring residual directions under similar context--action conditions. Empirically, Tables 11 and 12 of the revised manuscript show that HEP consistently outperforms global diagonal Mahalanobis and PPCA formulations, heteroscedastic regression, and PPCA sampling with the same frozen predictor. Together, these results support context--action conditioning rather than a fixed residual correction rule.

Does the remaining residual distinguish compatible from incompatible goals? In the controlled compatible-versus-hard-incompatible evaluation in Table 28 of the revised manuscript, HEP improves AUROC, pairwise ordering, and R@1 over the shared-backbone deterministic score for every difficulty level. The task-aligned Pair Accuracy reaches 87.75%, 85.25%, and 82.02% for 4, 8, and 16 incompatible candidates. The separate score-shift analysis in Figure 16 and Table 32 also shows a progression from the ground-truth trajectory to the lowest-ATE proposal and then to the highest-ATE proposal. Because navigation ranks candidates within the same case, this evidence directly evaluates the role for which the score is used.

Do these mechanisms translate into navigation gains? UA-NWM improves over its shared-backbone deterministic baseline in all ranking and planning settings, and every paired ATE and RPE confidence interval in Table 30 of the revised manuscript excludes zero. Table 13 shows that HEP also outperforms matched-capacity supervised pairwise-ranking and ATE-regression heads that use the same frozen backbone. Its advantage therefore cannot be explained only by additional parameters, a stronger visual predictor, or a generic learned metric. The same pattern extends across offline ranking and CEM planning in Tables 1 and 2, closed-loop simulation in Table 3, 2D navigation benchmarks in Table 21, unseen environments in Tables 22 and 23, different visual encoders in Table 20, and real-world UAV flights in Table 7.

AuthorsOpenReview ↗

Response to Reviewer KJK9 (2/3)

Together, these results establish the complete empirical chain underlying the paper's main story: HEP represents multiple structured future-feature variations, conditions them on the candidate, retains incompatible discrepancies in its residual, improves compatible-versus-incompatible ordering, and consequently improves navigation. The evidence spans representation analysis, controlled discrimination tests, matched-capacity comparisons, multiple environments, and real-world deployment rather than relying on downstream aggregate performance alone. What remains open is how its benefits scale with broader data, model capacity, environmental diversity, and deployment conditions. We view this as an important empirical and systems-scaling direction, rather than a missing link in the conceptual contribution evaluated here.

2. Can UA-NWM distinguish admissible variation from novel errors in unseen environments? (W2)

Thank you for raising the important question of whether HEP can remain useful under environment shift without absorbing novel prediction errors. To evaluate whether UA-NWM remains effective in unseen environments, we collected an additional test set of 500 trajectories from simulation environments that do not overlap with any environment used by the AirGoal-10k training, validation, or original test splits. No model is fine-tuned on these scenes, and this unseen-scene split will be publicly released with AirGoal-10k.

Comparison in unseen environments

MethodRanking-8 ATE / RPE ↓\downarrowRanking-16 ATE / RPE ↓\downarrowRanking-32 ATE / RPE ↓\downarrow
Random Selection1.96 / 0.531.97 / 0.541.99 / 0.54
NWM1.75 / 0.481.76 / 0.481.74 / 0.47
One-Step WM1.73 / 0.471.73 / 0.471.74 / 0.48
MWM1.74 / 0.471.72 / 0.471.71 / 0.47
RAE-NWM1.82 / 0.491.80 / 0.491.79 / 0.48
InJepa (V-JEPA 2.1)1.69 / 0.461.60 / 0.441.56 / 0.43
Deterministic Baseline (Ours)1.63 / 0.451.56 / 0.431.56 / 0.43
UA-NWM (Ours)1.57 / 0.431.49 / 0.411.47 / 0.40
Oracle Selection1.02 / 0.290.84 / 0.250.70 / 0.21
MethodUnseen-scene planning ATE ↓\downarrowUnseen-scene planning RPE ↓\downarrow
GNM1.450.39
ViNT1.660.46
NoMaD1.990.54
FlowNav1.900.52
NaviBridgerCVAE\mathrm{NaviBridger}_{\mathrm{CVAE}}1.780.49
NWM1.610.45
One-Step WM1.700.50
MWM1.610.45
RAE-NWM1.670.49
InJepa (V-JEPA 2.1)1.590.46
Deterministic Baseline (Ours)1.460.40
UA-NWM (Ours)1.430.39

UA-NWM remains the best learned method for all three unseen-scene ranking settings and obtains the lowest planning ATE. Its planning RPE is also lower than every world-model-based baseline. These results show that UA-NWM's relative advantage is preserved in the evaluated unseen environments without fine-tuning. We added the complete unseen-environment evaluation and discussion to Appendix A.7.

We have also added a qualitative comparison on the unseen split in Figure 13 of the revised manuscript. NWM accumulates errors and does not preserve the global scene structure. RAE-NWM produces sharper images, but its sampled future can depict a visually plausible outcome whose content and layout differ substantially from the goal. Both failure modes can distort trajectory scores. UA-NWM instead retains a deterministic prediction as its structural reference. Although this prediction contains less visual detail, HEP uses its learned variation directions to construct xgPx_g^P. Across the displayed unseen scenes, xgPx_g^P better aligns with the goal's global layout, while the additional samples visualize other plausible variations around the same prediction.

Why UA-NWM outperforms others in unseen environments

The unseen-setting concern is whether a novel prediction error could be mistaken for reasonable future variation. HEP does not automatically tolerate every discrepancy that appears in a new environment. It removes only the components that agree with its learned low-rank, action-conditioned variation directions. A novel error outside these directions remains in the residual and is penalized. The low rank and local multi-scale structure therefore provide a conservative inductive bias. They allow recurring structured variations, such as viewpoint- and occlusion-related feature changes, to transfer across environments without giving HEP unrestricted capacity to absorb arbitrary scene-specific errors.

AuthorsOpenReview ↗

Response to Reviewer KJK9 (3/3)

This distinction also explains the relative advantage over methods that score against individual predicted futures. A deterministic scorer penalizes every deviation from its single prediction, including plausible variations that remain valid in an unseen scene. A sampling-based scorer can avoid that limitation only when one of its generated futures happens to realize a mode sufficiently close to the goal. HEP instead discounts transferable, candidate-conditioned variation directions without requiring one sampled image to reproduce the goal. At the same time, discrepancies inconsistent with those directions continue to distinguish flawed trajectories. Moreover, navigation compares candidates within the same case. A scene-dependent shift in absolute score can affect all candidates without requiring scores from different unseen scenes to share a calibrated scale. What matters is whether the goal is more compatible with the variation directions induced by one candidate than by the others.

HEP models directions of plausible future variation rather than their relative probabilities, so it cannot guarantee perfect separation between every novel error and plausible variation. As already acknowledged in our appendix, “the uncertainty subspace remains only a low-dimensional, local approximation to the conditional future-state distribution.” The evidence here instead supports relative robustness under environment shift. UA-NWM improves over its own deterministic baseline on the unseen split, while the qualitative examples show that its HEP-based reconstruction better preserves goal-relevant layout than matching against one deterministic or sampled future. We state this scope explicitly in Appendix A.7.

AuthorsOpenReview ↗

Follow-up on Our Response to Your Review

Dear Reviewer KJK9,

Thank you for your careful review and constructive feedback. We greatly appreciate your time and effort in helping us improve our work.

While the discussion period is still ongoing, we would appreciate hearing whether our rebuttal has addressed your concerns or if any questions remain.

We welcome further feedback and would be happy to provide additional clarification. Thank you again for your valuable input.

Best regards,
The Authors

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

The paper targets image-goal navigation for UAVs in large outdoor scenes, where a world model ranks candidate trajectories by predicting their terminal observation and comparing it to the goal image. The authors argue that comparing the goal against a single point prediction is brittle, since occlusion and long-horizon drift make many futures plausible. Their fix is to treat scoring as conditional out-of-distribution detection and to approximate the conditional future distribution with a learned low-rank "uncertainty subspace" around the deterministic prediction. A module called HEP builds this subspace on a coarse-to-fine grid (1, 2, 7, 14 cells) over the 14×14 DINOv3 patch field, projects the prediction-goal discrepancy onto a per-cell rank-2 basis, subtracts it, and scores only the residual that survives. The backbone is a 115M deterministic latent world model that rolls out eight steps with delta tokens; HEP adds about 7M parameters and is trained on top of the frozen backbone with a normalized residual loss. The authors also build AirGoal-10k from AirSim and report offline ranking, standalone planning, closed-loop simulation, and a real quadrotor deployment.

Critical strengths
  1. The diagnosis is right and the cost is low. Point-wise comparison of a blurry mean prediction really is fragile in occluded outdoor scenes, and the proposed fix stays at one forward pass. At 8.47 ms per frame against 10.44 ms for One-Step WM and 438.67 ms for NWM, with 7M extra parameters, the trade-off is credible for onboard use. That engineering result is worth having regardless of how the theory is framed.

  2. The ablation discipline is better than most papers in this area. Table 9 and Table 10 compare against four alternative uncertainty formulations fit on the same frozen backbone, and every paired bootstrap interval excludes zero. The α sweep in Table 11 is monotone and moves in the direction the design predicts. That is the evidence that convinced me the projection, and not just the switch to a patch-wise distance, is doing the work.

  3. The appendix does not oversell. A.7.2 states that the decomposition is not orthogonal and not a single global projection, and A.8 admits that samples drawn from S look like a shuffled jigsaw. I would rather read that than a clean story.

Critical weaknesses

W1. "Uncertainty" is never defined, and the OOD framing does not survive the paper's own test. There is no scalar uncertainty, no variance, no threshold, no coverage statement. S is a linear subspace, so the "in-distribution region" is an unbounded slab rather than a region, and since no cutoff is ever set, scoring is a ranking, not a detection. Table 13 is the paper's own conditional-compatibility experiment, and it reports AUROC 0.684 / 0.666 / 0.642 against 0.675 / 0.657 / 0.634 for the deterministic baseline, a difference of about 0.01. For a paper whose central claim is the OOD formulation, that is the number that matters, and it is not discussed. The symbol μ compounds the problem: it denotes the terminal prediction (Eq. 3), the one-step prediction (Eq. 14, 36), the "conditional mean of all plausible futures" (Fig. 2), and the centre of the generative model in A.7.1. Because e is computed after channel-wise L2 normalization, norm(μ) ≠ E[norm(y)], so e is not a residual around any well-defined mean even when the goal equals the true future. The OOD story depends on the reading that Eq. 3 does not license.

W2. The training loss is scale-invariant, so HEP learns where the error points but never how large it is. L_HEP reduces to 1 − ρ² with ρ² = ‖e∥‖²/‖e‖², a dimensionless quantity. The backbone is frozen, so cos(x̂g, μ̂) is constant with respect to HEP's parameters and minimizing the ratio cannot be read as maximizing goal-prediction similarity. Inference then scores the unnormalized residual energy, which means HEP is fit on a purely directional quantity and applied to magnitudes that never entered its objective. The 1/‖e‖² denominator also down-weights exactly the large-error samples that dominate selection, so training and inference bias the weighting in opposite directions. A.7.1 makes the gap explicit: it derives eᵀΣ⁻¹e = ‖e∥‖²/(τ²+σ²) + ‖e⊥‖²/σ², where both terms should be penalized with weights set by τ²/σ². HEP drops the first term, which amounts to asserting τ² ≫ σ². That ratio is never estimated, only tuned after the fact through α.

W3. The score is a cosine with one learned gate, and the gate is goal-driven. Writing it out, s = (2/N) Σ (1 − c_i)(1 − ρ_i²), so the only new factor relative to the deterministic scorer is (1 − ρ²). Since ρ depends on e and e depends on x_g, the score is the distance between x_g and a repaired prediction μ + P(e). The repair moves μ toward whichever goal is being tested, provided the displacement lies in S, so a distant goal displaced along a learned direction receives a near-zero cost. This is the mechanism behind the weak AUROC above. Nothing bounds how much S can absorb: no penalty on ‖e∥‖, no constraint on ‖U‖, no control over the effective volume of S. The only guard is capacity, and Table 5 (R = 5 worse than R = 2) is as consistent with the failure mode as with a defense against it.

W4. HEP is trained and tested on different residual distributions, and there is no supervised baseline. Training uses rollouts conditioned on dataset actions; at test time HEP scores NoMaD or CEM candidates, many of which sit far from the training action distribution, so the basis is queried about an error cone it never saw. The degradation in Table 13 is consistent with this: R@1 drops from 79.0 to 68.6 to 56.3 as K grows, and AUROC from 0.684 to 0.642. More importantly, every comparison in the paper is against another unsupervised distance form. The obvious competitor is a lightweight supervised head on the same frozen backbone, trained with a pairwise ranking loss and using ATE-to-GT as free labels on AirGoal-10k. If that matches HEP, the uncertainty argument becomes unnecessary.

W5. The baselines are handicapped and the benchmark belongs to the authors. NWM runs at 25 diffusion steps instead of the official 250, and One-Step WM is trained from scratch because no checkpoint was released. Both are then retrained on AirGoal-10k, a new dataset with no external users to validate it. There are no seeds and no run-to-run variance; the paired bootstrap in Tables 10 and 14 covers case variance only. A.6 also does not say how UA-NWM and the baselines were trained for RECON and GO Stanford, and the GO Stanford jump from 2.11 to 1.51 needs that clarification.

W6. The real-world section is qualitative, and "low latency" means two different things. Five tasks are reported with no success rate, no distance error, no baseline and no failure count. Onboard planning is about 9 s per trajectory under CEM with 32 samples and three iterations, measured on a MacBook Air over a phone hotspot, while the headline number is 8.47 ms. That is three orders of magnitude apart and the two are never reconciled.

Questions for authors

Q1. Please define the uncertainty. What is the scalar quantity, what does it estimate (variance, support, coverage), and what threshold turns "in-distribution" into a decision rather than a ranking? How do you reconcile the OOD formulation with AUROC 0.642–0.684 in Table 13? A split-conformal quantile of the residual under the same frozen predictor would give the coverage guarantee S lacks at comparable cost, so why is it not compared?

Q2. Please report the distribution of ρ² across candidates, and in particular whether high ρ² is assigned systematically to bad candidates. Also report the held-out residual ratio and Table 13's R@1 as R grows well past 2. Given that L_HEP has no penalty on ‖e∥‖ and no constraint on ‖U‖, what keeps S from expanding until the score stops discriminating?

Q3. Why is the loss normalized during training while the score is unnormalized at inference? Please ablate training with the unnormalized residual energy, and report the per-sample weighting that the current 1/‖e‖² denominator induces.

Q4. Add the supervised baseline: same frozen backbone, a small head, pairwise ranking or direct ATE regression, labels free from AirGoal-10k. Does HEP still win, and if so is the margin explained by the subspace or by the extra parameters?

Q5. Retrain HEP with candidate actions sampled from the proposal policy or from CEM instead of dataset actions, and quantify the score distribution shift between GT-conditioned and candidate-conditioned residuals.

Q6. Report NWM at its official 250-step setting on a subset and confirm that all baselines received the same tuning budget. On RECON and GO Stanford, were all methods trained under identical protocols and on the same data?

Q7. For the real-world flights: number of trials per task, success rate, terminal distance error, failure cases, and a latency breakdown showing what runs onboard the Jetson versus on the laptop.

12 comments
AuthorsOpenReview ↗

Response to Reviewer XQcS (1/11)

Thank you for the exceptionally detailed, mathematically rigorous, and technically insightful review. We greatly appreciate your recognition of the practical value of single-pass trajectory scoring, our extensive controlled ablations, and our explicit discussion of HEP's limitations. We address each of your concerns below and have incorporated the corresponding clarifications and evidence into the revised manuscript.

For clarity, our response is organized as follows:

  1. What HEP calls “uncertainty” and what its score measures (W1&Q1)
  2. Does HEP systematically absorb genuine trajectory errors? (W3&Q2)
  3. Why an unbounded linear span does not erase trajectory discrimination (W3&Q2)
  4. Why training normalizes residuals while inference uses absolute residual energy (W2&Q3)
  5. Supervised same-backbone compatibility baseline (W4&Q4)
  6. Does training on dataset actions generalize to NoMaD and CEM proposals? (W4&Q5)
  7. Baseline fairness, diffusion steps, training protocols, and variance (W5&Q6)
  8. Real-world quantitative evaluation and latency accounting (W6&Q7)

1. What HEP calls “uncertainty” and what its score measures (W1&Q1)

Thank you for pointing out that our presentation did not make sufficiently clear how the terminology relates to the quantity computed by HEP and its actual use in navigation. We clarify the relevant definitions, decision rule, and empirical interpretation as follows.

Definition of future-state uncertainty

In this work, future-state uncertainty refers to the variation among multiple future visual features that may be admissible under the same observation context and candidate action sequence. HEP does not attempt to estimate the probability of each possible future. Instead, uncertainty-aware means that the scorer learns action-conditioned directions of such admissible feature variation and does not penalize every displacement along these directions as a trajectory error. We have added this definition where UA-NWM and HEP are first introduced in Sec. 1.

The scalar quantity and its operational meaning

For an observation context OO, candidate action sequence AA, and goal ogo_g, the scalar used at inference is

s(A,og∣O)=1N∑i=1N∥ei⊥∥22.s(A,o_g\mid O)=\frac{1}{N}\sum_{i=1}^{N}\|e_i^\perp\|_2^2.

It is a nonnegative, action-conditioned trajectory--goal incompatibility energy: it measures the part of the normalized goal--prediction discrepancy that the learned future-variation directions do not explain. A lower value indicates greater compatibility. It does not estimate predictive variance, conditional likelihood, coverage probability, or an aleatoric--epistemic decomposition. The deployed decision rule is A∗=arg⁡min⁡A∈As(A,og∣O)A^*=\arg\min_{A\in\mathcal A}s(A,o_g\mid O) over the proposals generated for the current case. We do not apply a universal raw-score threshold, and we do not claim a reject-all decision based on such a threshold. We have therefore revised the manuscript to describe HEP directly as a conditional compatibility scorer and made the scalar score and decision rule explicit in Appendix A.8.1. We no longer name the task or formulation as an OOD problem.

How OOD reasoning motivates the score

Our idea draws on the OOD principle of assessing distributional support. OOD scoring does not require the score itself to be a calibrated predictive probability. Established approaches use scalar energy, distance from a learned normal-data region, reconstruction residual, or compatibility with learned feature subspaces to distinguish supported from unsupported inputs. HEP adapts this score-based principle to a conditional navigation setting. Its unexplained residual measures whether a goal is supported by the future-feature variations associated with a candidate trajectory. We have added the relevant energy-based, reconstruction-based, and subspace-based works in Appendix A.1.2. These precedents explain the design intuition behind using an unexplained residual to measure support under candidate-conditioned future variations. We now state this connection only as motivation in the Introduction.

What the learned subspace represents

You are correct that a linear span is theoretically unbounded. Our intended use of S\mathcal S is to represent local directions along which future features may plausibly vary, not a bounded probability support or confidence region. The score is determined by the discrepancy that remains after hierarchical fitting along these directions. The revised manuscript therefore describes S\mathcal S as local variation directions rather than equating the whole span with the support of a conditional distribution. The practical consequence of its theoretical unboundedness is discussed in response 3. Why an unbounded linear span does not erase trajectory discrimination (W3&Q2).

AuthorsOpenReview ↗

Response to Reviewer XQcS (2/11)

Why the pooled AUROC is moderate

Table 28 of the revised manuscript (Table 13 of the initial submission) pools scores from all 1,000 test cases when computing AUROC. This evaluates whether a raw score from one scene is comparable with a raw score from another scene. Our navigation decision is different: candidate trajectories are ranked only within the same case, where the observation and goal are fixed. Different cases can have different intrinsic difficulty and score scales, so their raw scores need not be globally comparable. We neither use nor claim a universal threshold across cases. The absolute AUROCs of 0.684, 0.666, and 0.642 therefore show the limitation of interpreting the score as a global detector, but they do not directly measure the within-case decision for which HEP is designed. Moreover, the minimum- and maximum-ATE proposals provide controlled compatibility proxies rather than ground-truth labels for a universal conditional OOD event.

For convenience, we reproduce Table 28 below. OOD denotes a hard-incompatible proposal. Parentheses report 95% case-bootstrap confidence intervals.

KKMethodAUROCAUPR-OODAUPR-IDPair Acc.R@1
4UA-NWM0.684 (0.673, 0.697)0.8920.33487.75% (86.08, 89.33)79.0% (76.5, 81.3)
4Deterministic Baseline0.675 (0.663, 0.687)0.8880.32585.53% (83.63, 87.28)75.0% (72.2, 77.6)
8UA-NWM0.666 (0.655, 0.678)0.9380.19085.25% (83.50, 86.98)68.6% (65.7, 71.5)
8Deterministic Baseline0.657 (0.646, 0.669)0.9350.18582.83% (80.99, 84.69)63.7% (60.7, 66.7)
16UA-NWM0.642 (0.632, 0.653)0.9640.09682.02% (80.28, 83.78)56.3% (53.1, 59.4)
16Deterministic Baseline0.634 (0.623, 0.645)0.9630.09379.56% (77.64, 81.48)51.3% (48.2, 54.3)

The task-aligned metrics in the same table compare candidates under the same condition. HEP assigns a higher score to a hard-incompatible proposal than to the compatible proposal in 87.75%, 85.25%, and 82.02% of within-case comparisons for K=4,8,16K=4,8,16, respectively, and ranks the compatible proposal first in 79.0%, 68.6%, and 56.3% of cases. The corresponding deterministic results are 85.53%, 82.83%, and 79.56% for pairwise accuracy, and 75.0%, 63.7%, and 51.3% for R@1. Moreover, the paired improvements in AUROC, pairwise accuracy, and R@1 all have 95% case-level bootstrap intervals excluding zero. Thus, Table 28 supports the claim needed by our method: HEP improves conditional trajectory ranking over its shared deterministic predictor. We have revised the surrounding discussion so that the moderate pooled AUROC is not presented as establishing a globally thresholded OOD detector. This evaluation and its interpretation are included in Appendix A.8.6.

The meaning of μ\mu and ee

There is an idealized reason for motivating a deterministic prediction as an averaged future. Under a pure squared-error population objective, the Bayes-optimal predictor is the conditional mean, f∗(O,A)=E[X∣O,A]f^*(O,A)=\mathbb E[X\mid O,A]. Deterministic future prediction has consequently been observed to average multiple possible outcomes. Our deterministic predictor is trained with an MSE-dominated objective, so the squared-error result motivates its tendency toward a central, averaged future representation.

However, we agree that the actual network output should not be identified with an exact conditional mean. Its practical objective also contains cosine and latent-space terms, it has finite capacity, and it is evaluated through autoregressive rollout. Furthermore, channel-wise normalization is nonlinear, so even if μ\mu were the mean in the unnormalized feature space, in general norm⁡(μ)≠E[norm⁡(X)∣O,A]\operatorname{norm}(\mu)\neq \mathbb E[\operatorname{norm}(X)\mid O,A]. We have therefore defined μ\mu throughout as the deterministic feature prediction, and ee as the normalized goal--prediction feature discrepancy, rather than a zero-mean probabilistic residual. The pure-MSE result above explains the motivation behind our original averaging interpretation. In the revised manuscript, however, we define μ\mu solely as the deterministic feature prediction and do not rely on its being the exact conditional mean.

AuthorsOpenReview ↗

Response to Reviewer XQcS (3/11)

This clarification does not alter the reasoning on which HEP depends. HEP requires a conditional deterministic anchor, recurring directions in the discrepancies around that anchor, and an unexplained component that discriminates candidate trajectories. Neither its computation nor its ranking interpretation requires μ\mu to equal the exact conditional mean or ee to have zero expectation. We have revised the idealized Gaussian discussion to make this distinction explicit while preserving it as intuition for residual-energy compatibility scoring. These clarifications are included in Appendix A.8.1.

Split-conformal calibration

Our original evaluation focused on within-case candidate ranking because this is the decision rule used by our navigation method. To additionally examine the threshold-and-coverage interpretation raised in your question, we have now added a split-conformal control using the same frozen predictor and no model retraining. The complete calibration analysis is included in Appendix A.8.7.

We partition the original 1,000 test cases into 500 calibration cases and 500 disjoint evaluation cases. For the deterministic cosine score and HEP score separately, the conformal threshold is computed from the ground-truth action and its true future in the calibration cases. The observed true-future coverage is the fraction of evaluation cases whose corresponding score does not exceed this threshold. We also report the acceptance rate obtained by hypothetically applying the same threshold to the minimum-ATE proposal and the highest-ATE K∈{4,8,16}K\in\{4,8,16\} proposals.

These acceptance rates come from a diagnostic global-threshold protocol that differs from our navigation setting, where candidates are compared only within each case and the lowest-cost trajectory is selected. They therefore characterize hypothetical rejection behavior rather than performance under the actual navigation protocol. Our inference procedure does not use these thresholds for acceptance or rejection.

Nominal coverageScoreThresholdTrue-future coverageMin-ATE acceptanceMax-ATE-4 acceptanceMax-ATE-8 acceptanceMax-ATE-16 acceptance
90%Deterministic cosine0.28982450/500 (90.0%)445/500 (89.0%)1444/2000 (72.2%)2984/4000 (74.6%)6270/8000 (78.38%)
90%HEP0.46702448/500 (89.6%)449/500 (89.8%)1444/2000 (72.2%)2976/4000 (74.4%)6250/8000 (78.13%)
95%Deterministic cosine0.33582468/500 (93.6%)477/500 (95.4%)1658/2000 (82.9%)3415/4000 (85.38%)7014/8000 (87.68%)
95%HEP0.53138468/500 (93.6%)477/500 (95.4%)1611/2000 (80.55%)3334/4000 (83.35%)6886/8000 (86.08%)

This control shows that split-conformal calibration can attach a threshold and a marginal coverage guarantee to either fixed score. The guarantee is supplied by the calibration procedure, not by the geometry of S\mathcal S. At 90% nominal coverage, the Max-ATE-16 acceptance-rate difference between HEP and deterministic scoring is −0.25-0.25 percentage points, with a paired 95% interval of [−0.95,+0.41][-0.95,+0.41]. At 95% nominal coverage, it is −1.60-1.60 points, with a paired interval of [−2.56,−0.73][-2.56,-0.73]. Thus, HEP provides a small reduction in the latter setting, but both calibrated scores still accept most high-ATE proposals. Here, high ATE provides a controlled proxy for trajectory--goal incompatibility. It is not a direct annotation of whether a goal lies outside the true conditional future-state distribution.

A nominal coverage level is a long-run guarantee over exchangeable calibration and future samples, rather than a requirement that every finite evaluation set achieve exactly that percentage. Therefore, observing 468 covered cases out of 500 at the 95% level does not contradict the guarantee. Because the thresholds are calibrated using ground-truth actions and true futures, this guarantee does not directly apply to proposals generated by NoMaD or CEM. We report their acceptance rates only as a diagnostic analysis. The calibration and evaluation cases are two disjoint subsets of the original test set, not a newly collected validation set.

Finally, conformal calibration does not alter the ordering induced by a fixed score. On the same 500 evaluation cases, HEP obtains R@1 of 79.2%, 68.2%, and 55.2% for K=4,8,16K=4,8,16, compared with 72.8%, 61.8%, and 50.4% for deterministic scoring. We therefore report marginal coverage and within-candidate-set ranking as distinct properties: conformal calibration supplies the former, while HEP's demonstrated contribution in our navigation system is the latter.

Comments on the paper

1 comment

We sincerely thank all reviewers for their careful reading and constructive feedback. We appreciate their recognition of our novel problem perspective (Reviewers KJK9 and f9Rc), efficient and lightweight design (Reviewers XQcS and f9Rc), strong empirical validation (Reviewers XQcS and f9Rc), and value to the community (Reviewer f9Rc).

We addressed every concern point by point, conducted extensive new controlled experiments and analyses, and revised the manuscript accordingly. All added or modified text is highlighted in blue.

We believe the new evidence and revisions further strengthen the rigor, clarity, and completeness of the work. We again thank all reviewers and the Area Chair for their time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.