acceptodds
Withdrawn from ICLR 2027

Seeing Is Not Searching: Goal-Conditioned Effective Coverage for Aerial Object Goal Navigation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Aerial object goal navigation (ObjectNav) requires an unmanned aerial vehicle (UAV) to actively search for a specified target using onboard visual observations. Existing methods commonly measure exploration progress through geometric coverage, treating observed regions as searched regardless of the sensing conditions under which they are acquired. This assumption becomes problematic in aerial navigation, where altitude simultaneously affects spatial coverage and goal observability: higher altitudes increase the observed area while reducing the visual resolvability of the specified target, causing the same geometric coverage to provide substantially different search value across goals. To capture this goal-dependent search utility, we propose goal-conditioned effective coverage, which redefines exploration progress by weighting newly observed space according to the observability of the specified goal under the current sensing geometry. Because reliable target cues are often unavailable during early exploration, we further derive a goal-conditioned altitude prior from the target footprint and camera geometry, providing a reference sensing altitude before target detection. The prior is incorporated into policy observations and observability-weighted exploration rewards to encourage exploration under goal-appropriate sensing conditions. To mitigate the optimization difficulty arising from jointly learning altitude control, obstacle avoidance, and spatial exploration, we adopt a staged reinforcement learning scheme that progressively integrates these behaviors into a unified aerial search policy. Experiments on UAV-ON achieve 25.17% SR, 41.04% OSR, and 19.10% SPL, outperforming the previous state of the art by 5.67, 11.74, and 8.96 percentage points, respectively.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Object goal Navigation, Aerial navigation, Active exploration, Observability-aware exploration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
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_4axPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies aerial object goal navigation, focusing on the mismatch between geometric coverage and goal observability. The authors argue that increasing flight altitude allows a UAV to observe more space but can make the specified target harder to recognize. As a result, geometric coverage alone may not be a reliable measure of search progress. To address this issue, the paper proposes goal-conditioned effective coverage, which weights newly observed regions according to an altitude-dependent observability gain. A reference altitude is derived from the target's physical footprint and camera geometry, and the resulting prior is incorporated into policy observations and reward shaping. The method also uses staged reinforcement learning to combine altitude regulation, obstacle avoidance, and spatial exploration.

Experiments on UAV-ON report 25.17% SR, 41.04% OSR, and 19.10% SPL, outperforming the best previously reported overall results. The paper further provides ablation studies and analyses of flight altitude, detector responses, and training behavior.

Critical strengths

1. Meaningful Problem Motivation.

The paper identifies a meaningful limitation of geometric coverage as an exploration objective. In aerial ObjectNav, observing more space does not necessarily mean acquiring useful visual evidence of the target, since flight altitude affects both coverage and target resolvability. This distinction is well motivated and highlights an important consideration for aerial search.

2. Simple and Goal-Conditioned Design.

The proposed goal-conditioned effective coverage offers a simple way to account for target observability during exploration. By incorporating goal-dependent sensing considerations into the coverage objective, the method provides a practical approach to addressing the identified mismatch. I consider this the main contribution of the paper.

3. Encouraging Experimental Results and Behavioral Analysis.

The experimental results are also encouraging. The method improves the reported overall performance on UAV-ON, and Table 2 shows a substantial performance drop accompanied by higher flight altitudes when the observability-aware reward terms are removed. The additional analyses in Figures 3–7 help explain the behavior of the learned policy rather than relying solely on navigation scores. These results provide useful support for the paper's motivation.

Critical weaknesses

1. The contribution of effective-coverage weighting is not fully isolated.

The ablation results provide useful evidence for the overall observability-aware reward design. However, the "w/o observability shaping" variant removes the effective-coverage weighting together with the altitude penalty and detection-height gate, as described in Appendix B.2. This makes it difficult to assess the contribution of the coverage weighting itself. Since this is a central part of the proposed method, it would be helpful to understand whether it provides additional benefits beyond conventional altitude regularization. A more focused ablation would make the evidence for the main contribution stronger.

2. The benefit of the goal-conditioned altitude reference over simpler fixed references is not fully established.

The fixed-anchor ablation in Table 2 provides initial evidence for goal-conditioned altitude adaptation, but it only considers a fixed reference of 15 m. As discussed in Section 4.2, small and medium targets are largely constrained by the 5 m reference-altitude floor. A lower fixed reference might therefore provide a more competitive comparison. Without such a comparison, it remains unclear how much of the improvement comes from adapting the reference to the target rather than choosing a suitable fixed altitude. Evaluating one or two additional fixed references would help clarify the contribution of goal conditioning.

3. Limited evidence of real-world applicability.

The method is evaluated only in simulation. While the results on UAV-ON are encouraging, the proposed observability prior relies on estimated flight altitude and simplified camera geometry, which may be affected by sensing noise, uneven terrain, and viewpoint-dependent detection performance in real environments. It remains unclear whether the learned search behavior would remain effective under these conditions. Although the authors acknowledge this limitation, a real-world evaluation or a targeted robustness analysis would help establish the practical applicability of the approach.

Questions for authors

1. Geometric observability prior and multi-view detection.

Figure 3 provides useful evidence that detector response varies with altitude and target size. However, the proposed observability prior is derived from a downward-looking camera approximation, while the system uses multiple camera views for detection. Could the authors clarify how well the observed detector behavior supports this geometric prior across different viewing directions? This would help assess whether the proposed weighting captures goal-dependent observability beyond a general preference for lower flight altitudes.

2. Comparability of baseline evaluations.

Table 1 reports baseline results taken from previous publications. Could the authors clarify whether the evaluation protocols are directly comparable, particularly regarding test splits, episode settings, and target detection or terminal guidance? This would help assess the reported improvements over existing methods.

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

ECO-Nav targets aerial object-goal navigation, where more coverage does not mean the goal is visible. It sets a reference altitude from the task-given target footprint and camera geometry, with a 5 m floor. The altitude offset is a policy input and scales coverage rewards through an asymmetric gain. A recurrent transformer is trained with PPO, extra altitude rewards and a two-stage curriculum. Scripted rules handle approach and stopping at test time. On UAV-ON it reports 25.17% SR, 41.04% OSR and 19.10% SPL. Ablations and behavior plots link the full design to less over-climbing. I read the code but did not reproduce the results.

Critical strengths
  1. The coverage–resolution trade-off is well known in UAV search, but the paper shows clearly that it matters for learned aerial ObjectNav: with a plain coverage reward, the policy learns to over-climb. The size-based reference is known before detection and cheap to compute.

  2. The ablations are easy to interpret. In Table 2, removing the observability-aware rewards drops SR from 25.17% to 13.27%. Mean search altitude rises from 11.70/13.69/8.40 m to 30.33/31.43/26.11 m (small/medium/large goals). Figures 5 and 7 show the same upward shift, which makes over-climbing a convincing explanation for the failures. The appendix and code document the reward terms and staging in detail.

Critical weaknesses
  1. Beyond classical altitude–resolution geometry, the new part is a goal-conditioned coverage reweighting plus an altitude-offset input. The experiments do not test it alone. The main ablation turns off the coverage gain, altitude penalty and detection-bonus height gate together (Appendix B.2; scripts/run_train.sh; src/tasks/airsim_object_nav_task.py, lines 485–490 and 759–765). So the 11.90-point SR drop and the altitude shifts in Figures 5 and 7 support the package, not the weighting. The only fixed reference is 15 m, yet most small and medium goals sit at the 5 m floor, so goal conditioning is never compared with simply using a fixed 5 m reference for all goals (Q1). The main claim is therefore not tested in isolation.

  2. The coverage code seems to conflict with Appendix B.1, which claims a geometric union with no gain on revisits. _update_quadtree_with_polygon (src/utils/explore_map_new.py, line 701) sets c_new = 1 − (1 − c_old)(1 − a), where a is the cell's overlap with the current footprint. So re-observing a partly covered cell adds coverage again. get_explored_area (lines 987–991) returns the root's total coverage, so this goes straight into the reward. The coverage reward, and the coverage values reported in Fig. 7, rest on this computation (see Q2). I found this by reading the code, not by running it.

  3. The paper's own data do not support the altitude story behind its main gains. In Fig. 3(b–c), first-detection rates for small and medium goals are roughly flat or non-monotonic in altitude, so the altitude-dependent part of the prior is supported mainly by large goals. For most small and medium goals, h* sits at the 5 m floor, so for them the "goal-conditioned" reference is just a constant 5 m. Section 5.2 presents the largest gain (+18.42 SR on large goals) as consistent with the policy adapting its altitude once h* rises above the floor. But Fig. 5(a) shows that successful large-goal episodes search at 5.23 m on average, i.e. near the 5 m floor, and Appendix C.2 explains this by early detection. Fig. 3(d) also shows the large-goal detection rate is highest at the lowest altitudes. So the large-goal gain does not seem to come from the adaptive regime that Section 5.2 credits; staying low and detecting early, which a fixed 5 m reference would also encourage, explains it equally well. On small goals, where by the paper's own geometric argument flying too high should hurt detection most, ECO-Nav's SR (17.90) is even below OctMem-Agent's (18.91). Section 5.2 blames the 5 m floor, but this means the method does not help where it should help most.

  4. The Table 1 comparison is not like for like. ECO-Nav also uses GroundingDINO detection, CLIP verification, Bayesian fusion and a scripted approach/stop rule (stop when estimated distance plus half the goal diagonal is below 15 m; success radius 20 m; Appendix A.3). The Table 1 baselines do not share this detector and terminal module, and no experiment removes it from ECO-Nav or gives the baselines the same module. In Fig. 6, a large goal is detected at step 1 at 3.1 m, with almost no search. So I cannot tell how much of the Table 1 margin (e.g., 42.02 vs. 23.60 SR on large goals) comes from search rather than from perception and the terminal pipeline. A single benchmark and a single run per variant add to this uncertainty.

Minor points, to help the presentation and not part of my assessment: the Fig. 2 terminal-guidance panels look unlike the simulator frames; please mark them as illustrative if they are not simulator output. In Fig. 4, the stage-1 marker sits visibly before the 10k steps stated in Appendix C.1 (please clarify the unit), and the curves stop at 24,449 of the 30k training steps, before the evaluated checkpoint; since joint training has not plateaued there, the full curves would help. The peak-altitude ECDFs in Fig. 5(c) only reach 1 near the right end of the axis, well above 100 m, while Appendix C.1 describes a ~100 m ceiling; please clarify whether the ceiling is a hard limit and how peak altitude is measured. In Fig. 7, the coverage maps in the ablation rows fill the whole square although coverage is normalized over the 70 m disk.

Questions for authors
  1. Could you run a φ-only ablation that keeps the altitude penalty and detection gate, and a fixed 5 m or validation-selected reference? Since small and medium goals mostly have h* at the 5 m floor anyway, a fixed 5 m reference is the natural control. These would show whether the coverage weighting and goal dependence add anything beyond the other altitude terms. This could change my rating.

  2. Did Tables 1–2 use the supplied fractional coverage update? How does it fit Appendix B.1's union / no-repeat claim? Please quantify how repeated observations, or an exact union, change revisit credit, the coverage reward and the key comparisons.

  3. The Table 1 caption says the baseline numbers come from published papers under their own protocols, so how were ECO-Nav's evaluation settings and test episodes aligned with each baseline's, especially APEX's seen/unseen split? APEX's own Table 1 percentages (seen 12.36/19.10, unseen 16.13/22.58 SR/OSR) suggest 120 episodes (89 seen / 31 unseen; e.g., 11/89 = 12.36%, 5/31 = 16.13%), not the full UAV-ON test set. How many episodes and environments did ECO-Nav use? Please qualify the +8.96 SPL claim to match.

  4. In the snapshot, use_fixed_height can reach RuledModel, and the actor also calls the rules on near-identical consecutive frames and at very low height, even when use_ruled is off. Which of these paths were active in the reported runs, and what was kept fixed across variants? Do the key full vs. fixed-reference results hold across independent seeds?

  5. Can you report ECO-Nav without the scripted terminal guidance, or the baselines with the same detector and terminal module? Since first-detection rates for small and medium goals are nearly flat in altitude (Fig. 3(b–c)), please also explain why these goals need the altitude prior at all.

  6. Scoring observations by altitude-dependent detection quality is not new in UAV target search. Meera et al. (ICRA 2019, "Obstacle-aware Adaptive Informative Path Planning for UAV-based Target Search") measure how a detector's accuracy changes with altitude and build this into the planning objective of a UAV target-search planner. Lanča et al. (Drones 9(9):595, 2025) define "effective coverage" as the area coverage rate multiplied by detection confidence, both modeled as functions of flight speed and GSD. Section 4.2 says the contribution is using this classical relation as a stand-in for observability during search. These works already weight search by altitude-dependent detection quality, although they use measured detector curves instead of a closed-form formula from the goal footprint. Please cite and discuss them. What is new beyond turning this idea into a goal-conditioned RL shaping term for object-goal navigation? I am not asking for a comparison, since neither work targets UAV-ON.

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

The paper tackles an object navigation task that requires a UAV to actively search for a specified target using onboard visual observations. Prior work treats observed regions as searched, which is a problematic assumption. Because altitude affects both spatial coverage and goal observability, higher altitudes increase the observed area while reducing the visual resolvability of the target, and the same geometric coverage can provide substantially different search value across goals. To address this, the authors propose goal-conditioned effective coverage, which weights newly observed space by the observability of the specified goal under the current sensing geometry. Specifically, the prior incorporates policy observations and observability-weighted exploration rewards, and a staged reinforcement learning scheme progressively integrates altitude control, obstacle avoidance, and spatial exploration into a unified aerial search policy. Experimental results also support the proposed approach.

Critical strengths

The paper addresses an important practical problem.

The empirical results are encouraging.

Critical weaknesses

The technical contribution is limited. The proposed framework is incremental and a composition of existing ideas.

Prior work on ObjectNav has not been discussed at all that tackles the task of actively searching for a specified target using onboard visual observations, while treating observed regions as searched. For example: GOMAA-Geo: GOal Modality Agnostic Active Geo-localization; Aerial View Goal Localization with Reinforcement Learning; GeoExplorer: Active Geo-Localization with Curiosity-Driven Exploration. It is important to see how the prior approaches perform in the proposed setting.

Experimental analysis is limited to a single benchmark. It is important to test the generalizability of the proposed approach in other environments and different settings, where some of the constraints are relaxed. Moreover, the evaluation is restricted to the UAV-ON benchmark. It remains uncertain how robust the staged RL framework is when deployed in highly dense urban settings, heavily occluded areas like forest canopies, and dynamic wind disturbances that challenge precise altitude regulation.

The proposed method might be sensitive to initial target footprints. I am not sure how robust the proposed approach is when the UAV is tasked with locating rare, unusually small, or irregularly shaped objects for which reliable target geometry priors cannot be easily established beforehand.

Questions for authors

How sensitive is the policy to sensor noise or drift in real-world altimeters and onboard cameras during active search and navigation?

It is not clear what the reasons for failure are when the agent fails. As there are many components in the proposed framework, thus it is important to analyse the failure cases. Are failures mostly driven by timeout limits, obstacle collisions, or might it be because the UAV gets trapped at sub-optimal altitudes where targets remain unresolvable?

How does the altitude prior adapt when target dimensions are unknown before the search?

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

Related papers

Discussion (0)

Sign in to comment.