Language-goal aerial navigation requires a UAV to reach a destination described in natural language using observations gathered along its flight. A central challenge is that observations and motion commands depend on the UAV's changing pose, whereas environmental landmarks provide fixed spatial references for long-range reasoning. To reconcile these reference frames, we propose a vision-language-action model for aerial city navigation, named CityNav-VLA. Its dual-frame design retains accumulated observations in allocentric spatial memory, with metric coordinates anchoring each map token to a fixed physical location. A navigation query combines instruction-conditioned perception with the UAV's state and motion history to retrieve relevant evidence from this memory and predict a waypoint. The waypoint remains in the map frame until execution, when a controller converts it into body-frame motion commands. Early goal commitment then limits the influence of later prediction drift by fixing the execution waypoint to the median of initial estimates. On CityNav Test Unseen, CityNav-VLA reduces navigation error by 30.6 m and increases success rate by 4.6 percentage points relative to the best competing result for each metric. Cumulative ablations show that navigation-map features provide the largest gains and that metric coordinates further improve endpoint accuracy. With a 4B-scale backbone, CityNav-VLA averages 0.68 s of model inference per decision on single GPU, reducing at least 47.8% GPU memory and 84.0% decision latency with the comparable LLM-based methods.
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.Reviewer_GGaAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary
This paper studies language-goal aerial navigation, where a UAV must reach a destination described by natural language using observations collected during flight. To reconcile pose-dependent observations and motion commands with pose-independent environmental landmarks, the authors propose CityNav-VLA, a dual-frame vision-language-action model. The method stores accumulated observations in an allocentric spatial memory with metric coordinate encodings, while combining instruction-conditioned perception, UAV state, and motion history to predict map-referenced waypoints. A deterministic controller converts these waypoints into body-frame motion commands, and an early goal commitment mechanism reduces later prediction drift by fixing the execution waypoint based on the median of early predictions. On CityNav Test Unseen, CityNav-VLA improves navigation accuracy and success rate over prior methods while reporting lower GPU memory usage and decision latency than comparable LLM-based approaches.
Critical strengths
Novel dual-frame design. The separation between persistent allocentric spatial memory and dynamic egocentric interaction provides a clear and relatively novel formulation for city-scale aerial navigation. Predicting waypoints in a stable map frame and converting them into body-frame commands only during execution offers a principled way to combine long-range spatial consistency with pose-dependent control.
Strong experimental performance. CityNav-VLA achieves consistent improvements on CityNav, particularly on the Test Unseen split, where it outperforms prior methods across NE, SR, OSR, and SPL. The ablation studies provide useful evidence for the contributions of navigation-map features, metric coordinate encoding, and early goal commitment. The reported inference latency and GPU memory comparisons also strengthen the practical value of the proposed framework.
Critical weaknesses
Unclear motivation. The paper does not clearly explain how the reference-frame mismatch affects the learning objective or why a dual-frame model is necessary beyond standard coordinate transformation.
Questionable VLA formulation. The model predicts a map-referenced waypoint rather than an action, while a deterministic controller separately generates motion commands. The method therefore appears closer to multimodal waypoint prediction than a VLA policy that directly learns actions.
Limited generalizability and reliance on privileged spatial information. The framework depends on an accurately georeferenced global map, metric coordinates, landmark locations, and UAV states. Its robustness to missing or noisy spatial information and its generalization beyond CityNav remain unclear.
Limited closed-loop adaptation. Early goal commitment ignores subsequent waypoint predictions after the first K steps, which appears inconsistent with the claimed closed-loop replanning capability.
This paper studies language-goal-driven aerial navigation at city scale and proposes CityNav-VLA. The method adopts a dual-frame design that combines dynamic navigation context, comprising the UAV’s current observations, state, and motion history, with global spatial memory represented in a fixed map coordinate frame. Global visual features, navigation-map features, positional embeddings, and metric coordinate encodings are integrated into a unified spatial representation. A cross-attention decoder predicts waypoints in the global coordinate frame, which are then converted by a deterministic controller into body-frame motion commands. The authors further introduce Early Goal Commitment (EGC), which fixes the execution target to the coordinate-wise median of early waypoint predictions to mitigate prediction drift during sequential decision-making. On CityNav Test Unseen, the method achieves a navigation error (NE) of 37.91 m, a success rate (SR) of 30.53%, an Oracle Success Rate (OSR) of 43.08%, and a Success weighted by Path Length (SPL) of 29.32%. The paper also conducts ablation studies on spatial-memory components and reports low inference latency and GPU memory usage.
Critical strengths
The method achieves competitive performance on CityNav Test Unseen. It obtains an NE of 37.91 m, an SR of 30.53%, and an SPL of 29.32%, outperforming the comparison methods reported in the paper on these metrics. These results indicate advantages in target localization, task success, and path efficiency.
The model offers relatively efficient inference and has potential for resource-conscious deployment. It contains approximately 4.46B parameters, of which only 4.67M are trainable. On an NVIDIA A100 GPU, its per-decision inference time is 679.7 ms and peak GPU memory usage is 8.56 GB, suggesting that freezing the multimodal backbone and training lightweight navigation modules can help control computational costs.
Critical weaknesses
The ablation study is relatively simplistic and does not sufficiently validate the independent contributions of individual components. The paper primarily evaluates the spatial memory module by progressively incorporating navigation-map features, positional embeddings, and metric coordinate encodings, but lacks an in-depth analysis of the specific role of each component and its impact on overall navigation performance. Therefore, the current ablation results are insufficient to fully validate the necessity of each design choice or clearly identify which components are primarily responsible for the observed performance improvements.
EGC may lock in early localization errors and limit online correction. The method fixes the execution target after the first 24 prediction steps, preventing subsequent observations from updating the committed waypoint. In a failure case reported in the paper, an inaccurate initial target commitment leads to a final localization error of 128.70 m, whereas the model without EGC reaches the vicinity of the target. This suggests that, although EGC can suppress prediction drift, its effectiveness depends heavily on the reliability of early predictions. The paper does not sufficiently analyze its applicability under different levels of prediction uncertainty or compare it with a confidence-aware commitment strategy.
The independent novelty of the dual-frame design needs more direct experimental support. The framework combines global spatial memory, navigation-map features, positional embeddings, metric coordinate encodings, and cross-attention, but the experiments do not clearly establish how the dual-frame design differs from existing map-enhanced navigation and spatial-memory methods. It remains unclear whether the gains primarily arise from global spatial memory, explicit geometric encoding, or interactions among the components. More targeted comparisons with structurally related methods would help establish the necessity and distinct contribution of the proposed design.
Questions for authors
What are the key technical differences between the proposed dual-frame design and existing map-enhanced navigation or spatial-memory methods? Could the authors compare alternative reference-frame designs while keeping the remaining modules and training settings unchanged, in order to isolate the benefit of the dual-frame design?
Could the authors further evaluate the independent effect of metric coordinate encoding across different scenes and target distances, and analyze its interaction with navigation-map features?
Given that the current method lacks an explicit stop action, have the authors considered adding stop prediction or a distance-based termination condition? Could such a mechanism reduce cases in which the UAV enters the success region but continues moving, while avoiding premature termination?
This paper studies language-goal aerial navigation at city scale and proposes CityNav-VLA, which combines spatial memory maintained in a fixed map frame with visual, state, and history information that changes as the UAV moves, in order to predict the navigation goal.
The method uses a frozen CityBot-4B backbone to extract vision-language features. Global-map visual features, auxiliary navigation-map channels, learned positional embeddings, and explicit metric coordinate encodings are fused into a spatial memory on a 14×14 grid. The current multimodal context, the UAV state, and the recent motion history form a navigation query, which regresses the goal position in the map frame via cross-attention; a deterministic controller then converts this position into motion commands in the body frame. During training, the final destination coordinate serves as the supervision target at every step. At inference, an early goal commitment (EGC) mechanism is applied: after the first K=24 decisions, the coordinate-wise median of these predictions is fixed as the execution goal to suppress later prediction drift.
On CityNav Test Unseen, the paper reports NE 37.91 m, SR 30.53%, OSR 43.08%, and SPL 29.32%, together with ablations on spatial-memory components, reference frames, backbone adaptation, and EGC. The model has about 4.67M trainable parameters, and per-decision model inference takes 679.7 ms on an A100.
Critical strengths
1. Clear motivation and a well-grounded division of reference frames.
City-scale UAV navigation must handle both long-term, stable environmental structure and observations and controls that change with motion. The paper stores environmental evidence in a fixed map frame and converts it to the body frame according to the current heading only at execution time, which matches the geometry of the task. Eq. (1) encodes map-cell positions metrically, and §3.5 converts normalized coordinate residuals to physical units when computing the loss, allowing the model to distinguish the physical meaning of the same normalized displacement across maps of different scales. These design choices have a clear physical interpretation and provide a targeted spatial inductive bias.
2. A large reduction in endpoint error that persists without EGC and across training seeds.
With EGC disabled, Test Unseen NE is 40.14 m (Table 5); for the 10K-step checkpoints across three training seeds, it is 43.24±2.62 m (Table 6). Both are well below the best competing result of 68.5 m in Table 1, indicating that the localization ability comes mainly from the model itself rather than from EGC or a particular training seed. Note, however, that this comparison is made against published results on the original CityNav, so its magnitude depends on whether the dataset version and metric definition are consistent (see Weakness 1); the SR advantage, by contrast, comes mainly from EGC.
3. Several controlled experiments are well designed and supported by statistical tests.
The EGC ablation keeps the checkpoint and all other components fixed, compares enabling and disabling commitment on all three splits, and uses episode-paired bootstrap and McNemar tests, providing direct evidence for the effect of EGC in the current system (Appendix A.1, Table 5). In addition, the authors use only Validation Seen to select the main checkpoint and the commitment step, and use fixed-step checkpoints for the representation comparisons, which helps control the influence of model selection on the conclusions.
Critical weaknesses
1. The evaluation sets differ from the official CityNav splits, and the cited baselines may come from a different dataset version and metric definition.
I consider this the issue with the greatest impact on the main conclusions.
(a) Size of the evaluation sets. Both §4.1 and Appendix B state that evaluation covers the "complete" splits, yet the reported sample counts are smaller than those of the official splits in the original CityNav paper:
Validation Seen
Validation Unseen
Test Unseen
Official CityNav splits
2,498
2,826
5,311
This paper (Appendix B)
2,470
2,697
5,281
The three splits are short by 28, 129, and 30 episodes, respectively (about 4.6% of Validation Unseen), and the paper does not state which episodes were excluded or why. The three difficulty groups listed in §4.2.1 sum to exactly 5,311, matching the official Test Unseen split, which further indicates that the paper's evaluation set differs from the official one.
(b) Dataset version. HTNav[1] reports a revised version of CityNav, with roughly 800 annotation corrections and 311 trajectories lacking landmark descriptions removed. HTNav's results on this version (NE/SR/OSR/SPL) are 37.5/31.05/51.82/27.48 on Validation Seen, 49.8/17.69/37.82/15.44 on Validation Unseen, and 40.3/25.49/48.82/21.56 on Test Unseen, far better than the original-version numbers cited in Table 1 (Test Unseen NE 68.5 m). I cannot determine which version this paper uses; but if it is a filtered or revised version, Table 1 effectively compares against baseline numbers obtained on the original version. Taking HTNav on the revised version as the reference, the NE advantage on Test Unseen would shrink to about 2.4 m, and OSR would be 5.7 percentage points (pp) lower (SR and SPL would still be higher by 5.0 and 7.8 pp, respectively); on Validation Seen and Validation Unseen, the proposed method would not lead in SR or OSR.
(c) Metric definition. Appendix B states that the method uses planar distance and planar path length, whereas the original CityNav paper defines success as stopping within a "20-meter spherical radius" of the goal, i.e., using 3D distance. In CityNav, the UAV starts at an altitude of 100–150 m; if it does not descend sufficiently, the vertical error alone can exceed 20 m. Since
d2D=Δx2+Δy2≤Δx2+Δy2+Δz2=d3D,
planar NE, SR, and OSR (and SPL computed with planar path length) are systematically more lenient. If the cited baselines use the original definition, the two sets of numbers are not directly comparable.
(d) Baseline provenance. The baselines in Table 1 are listed only by method name and venue, without citations, and FLAME, HETT, and Seq2Seq have no corresponding entries in the reference list. DroneNav is cited as Xu et al. (2026c) in Table 4, but the corresponding reference entry is titled AerialVLA; whether these are the same work should be clarified. This makes it difficult for readers to verify the source and evaluation setting of each baseline number.
I suggest that the authors state the dataset version used and the excluded episodes, compare with the baselines under a consistent dataset version and metric definition (if necessary, against same-version results such as HTNav on the revised version), and complete the baseline citations.
[1]. Fan C, Pan C, Liu Z, et al. HTNav: A Hybrid Navigation Framework with Tiered Structure for Urban Aerial Vision-and-Language Navigation[J]. arXiv preprint arXiv:2604.08883, 2026.
2. Insufficient description of the execution conditions and the evaluation budget.
§3.6 gives the map-to-body rotation R(−ψt) (Eq. (7)) and states that the controller outputs bounded motion commands, but it does not specify the per-decision translation and altitude-change limits, how the heading is updated, the execution duration of each action, or the control behavior near the predicted goal. §3.6 states that the controller parameters are provided in the implementation details, while the implementation details in Appendix B refer back to Section 3; because of this circular reference, the parameters are not actually given anywhere in the paper.
These details determine how model decisions translate into actual trajectories, and also what the 120-step budget actually means. CityNav's primitive actions move forward 5 m, turn 30∘, or change altitude by 2 m, and human demonstrations average about 240 actions; GeoNav uses a 200-step limit. If a single decision in this paper can turn, translate, and change altitude simultaneously, 120 decisions do not correspond to the same physical motion as the budget of discrete-action baselines. In addition, the method has no explicit stop action, so how episodes terminate and whether the UAV hovers after reaching the goal also affect NE and SPL. I suggest providing the full controller parameters and stating how many CityNav primitive actions one decision corresponds to, so that readers can judge whether the improvement comes from goal prediction, the execution policy, or both.
3. Limited robustness of the SR advantage: reliance on a single training run and on EGC.
The main results come from a single training run at the 17K-step checkpoint, with no across-seed variance reported. Table 6 shows that the same architecture at 10K steps with EGC disabled achieves a Test Unseen SR of 22.40±3.08% across three training seeds: the across-seed standard deviation is of the same order as the 4.6 pp SR advantage over the best baseline in Table 1, and the mean is also below GeoNav's 25.91%.
Moreover, with EGC disabled, the main model's Test Unseen SR is 26.66%, only 0.75 pp above GeoNav, and its Validation Unseen SR of 14.76% is below several baselines, including HETT, GeoNav, and HTNav. In other words, about 3.86 pp of the 4.6 pp SR advantage comes from EGC. EGC is an inference-time execution strategy that does not depend on the paper's representation design and could, in principle, be combined with other methods that output goal estimates; it is therefore necessary to separate the contribution of the model itself from that of the execution strategy.
I suggest reporting multi-seed means and standard deviations for the main configuration (17K-step checkpoint with EGC) as well as with EGC disabled. By contrast, the NE advantage is relatively stable across seeds (see Strength 2).
4. The main-table results cannot be reconciled with the difficulty-level breakdown.
§4.2.1 splits Test Unseen into Easy, Medium, and Hard groups containing 2,096, 1,686, and 1,529 episodes, respectively. If these groups partition the same evaluation set, the overall value of each metric must be the sample-weighted mean of the group values:
mˉ=∑knk∑knkmk,minkmk≤mˉ≤maxkmk,
where nk and mk denote the size and the metric value of group k∈{Easy,Medium,Hard}. Weighting the results of the proposed method in Table 2 by the listed group sizes and comparing them with Tables 1 and 5 gives:
Metric
Table 2, weighted by listed group sizes
Table 1, overall
Table 5, EGC disabled
NE (↓)
40.85 m
37.91 m
40.14 m
SR (↑)
26.53%
30.53%
26.66%
OSR (↑)
42.25%
43.08%
42.40%
SPL (↑)
24.72%
29.32%
24.84%
The highest SR among the three difficulty groups in Table 2 is 27.58% and the highest SPL is 26.22%, whereas the overall SR and SPL in Table 1 are 30.53% and 29.32%. If these were the group-wise and overall results of the same system under the same setting on the full test set, the overall mean could not exceed every group value. This contradiction does not depend on the group weights and therefore cannot be explained by the difference between 5,281 and 5,311 episodes.
Numerically, Table 2 closely matches the Test Unseen results with EGC disabled in Table 5: the success counts implied by the listed group sizes are 578+459+372=1,409, while 26.66%×5,281≈1,408. This suggests that Table 2 may report results with EGC disabled, but the paper does not say so.
It should also be noted that a similar phenomenon appears for the baselines: in Table 1, HTNav's overall SR is 22.23%, higher than all three of its difficulty-group values in Table 2 (19.99%, 20.40%, and 21.91%), and GeoNav's overall NE (73.5 m) and OSR (41.68%) likewise fall outside the range of its group values. The same phenomenon already exists in the original HTNav and GeoNav papers, suggesting that the difficulty-level and overall evaluations in prior work may have used different episode sets or protocols. In addition, I could not find a definition of Easy/Medium/Hard in the original CityNav paper, and GeoNav and HTNav state only that the levels are defined by the start-to-goal distance, without giving thresholds. Since the paper calls these the "official" partitions, it should state their source and thresholds.
Therefore, this issue more likely stems from unclear configuration labeling and from protocols inherited from prior work, and does not necessarily mean that the results themselves are wrong. Until the authors clarify, however, the cross-difficulty performance and the overall main-table performance cannot both be verified.
5. The evidence for the core reference-frame claim remains thin.
In Table 3, the comparison between V+M+P and V+M+P+C supports the benefit of the coordinate branch in the current configuration (NE decreases from 57.01 m to 52.42 m), and the authors appropriately limit this conclusion to the improvement in endpoint error. However, Table 7, which bears most directly on the core claim that the allocentric encoding outperforms the egocentric one, compares the variants on only 200 Validation Unseen episodes with a single training seed, reports no paired confidence intervals, and does not explain how the 200 episodes were selected.
The existing numbers suggest that this subset may exaggerate the effect. Tables 7 and 3 share the same seed, training schedule, and checkpoint step, so the no-coordinate and allocentric rows of Table 7 most likely correspond to F2 and F3 in Table 3. On the full set of 2,697 episodes, the two differ by 4.59 m in NE and 1.45 pp in SR; on the 200-episode subset, they differ by 8.99 m and 3.0 pp, roughly twice as much. Moreover, the SR difference in Table 7 corresponds to 32 versus 25 successes. Let b and c denote the numbers of episodes on which only the allocentric variant or only the egocentric variant succeeds, respectively, so that b−c=7. The exact McNemar test reaches p<0.05 only if c≤1 (e.g., (b,c)=(8,1) gives p≈0.039, whereas (b,c)=(9,2) gives p≈0.065), so the SR difference is most likely not significant. Since full-split results already exist for the other two variants, evaluating the egocentric variant on the complete Validation Unseen split would be inexpensive.
The definition of "egocentric metric displacement" in Table 7 is also incomplete. The allocentric XYR encoding in Eq. (1) is defined relative to the fixed map center pˉ:
Is the egocentric variant obtained simply by replacing pˉ with the current UAV position pt, or is the displacement additionally rotated into the body frame by R(−ψt)? Does it keep the same dimensionality, radial term, and physical scale sc? These details determine whether the comparison truly isolates the effect of the reference frame.
Regarding novelty, maintaining memory in a map frame while acting in robot-relative coordinates has clear precedents. For example, Active Neural SLAM (Chaplot et al., ICLR 2020) maintains spatial information in a global map anchored at the episode start and converts short-term goals into relative distance and angle from the agent's current location for its local policy; VLMaps (Huang et al., ICRA 2023, already cited in §2.2) fuses vision-language features into a metric map for language-goal localization and navigation. The paper's added value should therefore be located more precisely in its specific coordinate encoding, query–memory interaction, and goal-regression mechanism. Clearly defining the reference-frame comparison and strengthening its statistical evidence would be more convincing than the general "dual-frame" narrative.
6. The construction of the auxiliary map channels and the training inputs is insufficiently specified for reproduction.
The authors state that the internal map is updated from accumulated observations, that the urban scenes come from SensatUrban, and that landmark locations and contours come from CityRefer; they also give the main network modules, the optimizer, and the learning rate. However, the availability and update rules of the auxiliary map channels during navigation remain unclear:
Are the terrain-height and landmark channels available only after the corresponding regions have been observed, or are some of them provided as a prior from the start?
How are unobserved regions encoded, and how do the exploration and visibility masks act on the different channels?
How is depth or elevation information obtained, and how are first-person observations projected and registered onto the fixed map?
Are the maps and state histories used in training generated from expert trajectories, policy rollouts, or another trajectory-generation procedure?
These questions bear directly on what information the model actually receives and on the consistency of the input distribution between training and testing. In Table 3, F0, which uses only global visual tokens, has an NE of 207.34 m, close to the random policy (223.0 m on Validation Unseen), and adding the navigation-map channels reduces NE to 58.81 m, showing that performance relies heavily on these auxiliary channels. Given also that 4B backbone adaptation yields no stable benefit in Table 6, I suggest adding an ablation that removes the visual tokens (M+P+C only) to clarify the actual contribution of the frozen VLM; this also bears on whether the "VLA" framing is justified. In addition, the values of δ, λz, and sz in the loss, as well as the specific layers used for feature extraction and the aggregation of language features, should be reported.
7. Two supplementary comments whose scope should be qualified.
Late-stage adaptivity of EGC. The authors have demonstrated the effect of EGC in the current system through ablation and have shown failures caused by incorrect commitment, so this is an acknowledged design trade-off. A valuable additional analysis would examine how committing to an early goal performs when the target is not visible before commitment or when later viewpoints are needed for disambiguation, and whether fixing an early estimate still offers additional value over simple temporal stabilization methods such as a sliding median or exponential smoothing. In addition, the offline-replay "No commitment" baseline in Figure 8 is about 27% (read from the figure), whereas the closed-loop SR with EGC disabled on Validation Seen in Table 5 is 28.14%; this discrepancy deserves an explanation. These questions do not negate the evidence already provided in Table 5.
Conditions of the efficiency comparison. §4.4 describes the hardware and timing scope of the proposed method and notes that the comparison numbers are taken from the literature, so the current results support an inference-cost comparison relative to the reported configurations. However, without measurements under a common hardware platform, numerical precision, and input budget, and given that §4.4 states the comparison is against DroneNav's "second stage", the specific reduction ratios remain conditional. I suggest that the 47.8% and 84.0% reductions in the abstract carry the same qualification as in the main text. End-to-end measurements and onboard experiments would strengthen the evidence for applicability, but they are not necessary additions for the current benchmark study.
Minor issues
§4.2.1 states that HTNav's SR is "2.4 to 6.8 points" below that of the proposed method across difficulty groups, but Table 2 gives gaps of 27.58−19.99=7.59, 27.22−20.40=6.82, and 24.33−21.91=2.42 pp, so the range should be 2.4–7.6.
Appendix A.1 refers to additional analysis on Validation Unseen, but the corresponding results are not shown.
Questions for authors
Dataset version and evaluation sets. Does the paper use the original CityNav, the revised version from HTNav, or another filtered version? Compared with the official splits, Validation Seen, Validation Unseen, and Test Unseen are short by 28, 129, and 30 episodes, respectively; on what basis were these episodes excluded? Can the authors provide comparisons on the same dataset version as the cited baselines? If the revised version is used, please compare against same-version results such as HTNav on the revised version.
Metric definition and execution budget. Do the baselines in Table 1 use planar distance or the original 3D (spherical-radius) definition of CityNav? Can the authors also report the proposed method's results under the 3D definition? Please provide the full controller parameters (per-decision translation and altitude-change limits, heading updates, execution duration, and behavior near the goal), state how many CityNav primitive actions one decision corresponds to, and explain how the 120-step budget corresponds to that of the cited results.
Multi-seed results. Can the authors report multi-seed means and standard deviations for the main configuration (17K-step checkpoint with EGC) and with EGC disabled?
Correspondence between Tables 1 and 2. Are the results of the proposed method in Table 2 obtained with EGC disabled? Do the two tables use the same checkpoint, episode set, and aggregation? Please state the source and distance thresholds of the "official" difficulty partitions, and explain the difference between 5,311 and 5,281.
Reference-frame comparison. Please give the complete definition of the egocentric displacement: its origin (e.g., whether pˉ in Eq. (1) is replaced by pt), whether it is rotated by R(−ψt), and its encoding dimensionality, radial term, and normalization scale. How were the 200 episodes selected? Are the no-coordinate and allocentric rows of Table 7 identical to F2 and F3 in Table 3? Can the authors report the egocentric variant on the complete Validation Unseen split, together with paired confidence intervals or success/failure flip counts?
Map and training-input construction. Please specify the initialization and update rules of the auxiliary channels (elevation, landmarks, visibility, and exploration), distinguishing priors available at the start from online observations; describe the trajectory source used to generate the training maps and motion histories; and report the loss scales and key feature-extraction settings. Can the authors provide results without the visual tokens (M+P+C only)?
Real-world validation. Are there any preliminary results from real flights, onboard deployment, or real camera imagery? Can the authors inject position noise σxy and heading noise σψ in simulation and report how the allocentric and egocentric variants degrade with the noise magnitude? Can they report the per-decision latency on an embedded platform (or with ground-station inference over a communication link)? Does the controller include obstacle avoidance or other safety constraints?
Scope of EGC. For episodes in which the target is not yet visible within the first K=24 decisions, or in which later observations are needed for disambiguation, how does the method perform with and without EGC? Does committing to an early goal still have an advantage over simple alternatives such as a sliding median or exponential smoothing? Why does the offline-replay "No commitment" baseline in Figure 8 differ from the closed-loop result in Table 5?
**Efficiency comparison. ** Can the authors list the available hardware, precision, and timing boundaries for each baseline in Table 4, including what DroneNav's "second stage" specifically refers to? If unified re-measurement is not possible, I suggest stating in the abstract that the relative reductions refer to the configurations reported in the literature.
From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.