acceptodds
Under review as a conference paper at ICLR 2027

AeroEVA: Unified World Modeling and VLM Policy Learning for Embodied Aerial Search

Abstract

Efficient aerial search for human-specified goals requires anticipating what each move will reveal. Content next to the current view is often predictable even when distant regions are not, yet existing agents remain limited in turning such local foresight into better actions. We introduce AeroEVA (Envision, Verify, and Act), a vision-language model (VLM) agent that uses transition prediction in two roles: as a training signal, and as evidence checked before each move. To envision, its vision tower learns to predict features of the nearest patches in the neighboring region under each action, and textual transition targets supervise the decoder alongside actions. To verify, AeroEVA reuses this predictor to score each candidate move against an aerial-image goal. To act, the policy judges when these one-step scores are relevant and combines them with observations and history. Imitation learning alone surpasses prior search agents, and multi-turn reinforcement learning widens the margin to nearly 20 points in success rate. Without retraining, AeroEVA also leads prior methods in unseen regions, with ground-photo and text goals, under disaster-induced appearance change, and on finer grids. We further introduce AeroTask, 696 search tasks across 13 application scenarios whose language goals state task requirements rather than target appearance. Code, model weights, and the AeroTask dataset will be released upon acceptance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Geolocalization, Agentic RL, World modelling, Embodied navigation, Aerial Visual Search, Remote Sensing

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

The paper proposes AeroEVA, a Qwen3-VL-4B agent for active geo-localization: navigating aerial grids with limited moves and local observations to reach image- or text-specified goals. Its Envision–Verify–Act framework predicts upcoming views, scores candidate moves by similarity to the goal image, and uses these scores to guide actions. Training combines joint SFT on shortest paths with multi-turn GRPO. On MASA, AeroEVA achieves 71.8% success with SFT and 86.1% with GRPO, outperforming retrained GOMAA-Geo (65.0%) and same-backbone GeoExplorer (47.9%). The checkpoint transfers to unseen regions, post-disaster imagery, larger grids, and ground-photo and text goals. The paper also introduces AeroTask, comprising 696 LLM-generated task-style language goals.

Critical strengths
  1. Well-supported design choices. Diagnostics support joint supervision and nearest-row prediction, while ablations show that visual prediction helps only when paired with text targets and verification.
  2. Strong gains and broad transfer. AeroEVA outperforms same-backbone controls (77.0 vs. 56.4) and generalizes across regions, time, goal types, and grid sizes, reaching 60.9% on 10×10 grids versus at most 12.9% for baselines.
Critical weaknesses
  1. The "unified" part is not isolated. The verifier is a separate frozen snapshot, so it could exist without visual prediction in the policy itself. Table 2 has no "Action SFT + text + tool" row without visual-latent supervision, which is needed to show that the policy's own prediction matters.
  2. The verifier's share of the gain is unknown. I could not find MASA results for the final model with the tool turned off on aerial-image goals. Tool-off results on other goal types are encouraging but not a controlled comparison.
  3. No training-seed variance. All results come from one training run; the reported ± values reflect only decoding randomness. Multi-turn GRPO is often sensitive to seeds.
  4. Training uses privileged supervision. AeroEVA learns from shortest-path action labels and text targets built from offline land-cover maps, which the RL baselines do not use. The paper should state which gains depend on the map-derived labels.
  5. AeroTask is weakly validated. The same model (Qwen3-VL-32B) generates and judges the tasks on segmentation pseudo-labels, with no human check of whether each instruction uniquely identifies the target cell. With 268 evaluation records, label noise could affect the conclusions.
Questions for authors
  1. Can you add an "Action SFT + text targets + frozen verifier" variant without visual-latent loss on the policy?
  2. How does the final SFT+GRPO model perform on MASA aerial-image goals with the tool turned off?
  3. Can you report at least 3 training seeds for AeroEVA and the main same-backbone baselines?
  4. How does a variant perform whose text targets come from the image alone, without land-cover maps?
  5. Was any part of AeroTask checked by humans for target uniqueness, and what was the agreement rate?
Reviewer_wyKDOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces AeroEVA, a VLM agent for embodied aerial search. The method is organized around three stages: Envision–Verify–Act. Envision trains the VLM's vision tower to predict the DINOv3 features of the nearest row of patches in the neighboring cell under each candidate action, teaching the model "what the next view will look like if I move this way." Verify computes cosine similarity between the predicted features for each direction and the goal features, yielding a goal-compatibility score for each candidate move. Act: the policy combines these verification scores with the current observation and interaction history to select a move; after each executed move the agent receives a real observation and repeats the process, and multi-turn GRPO further refines decision-making. AeroEVA reaches 86.1% average success on MASA, surpassing prior methods by roughly 20 points, and a single checkpoint transfers zero-shot to new regions, disaster-induced temporal changes, ground-photo/text goals, and finer grids, leading prior methods in all settings.

Critical strengths
  1. The method delivers large and consistent performance gains, together with a breadth of zero-shot generalization that is rare in this area: it surpasses the strongest prior method on MASA by roughly 20 points (86.1% vs. 66.4%), and a single checkpoint—without any adaptation—remains ahead across regions, time (disaster-induced appearance change), goal modalities, and grid resolutions. This shows that unifying world-model foresight with VLM policy learning brings substantial benefits to embodied aerial search.

  2. The design is principled and the analysis thorough. Through careful experiments, the paper answers questions such as which prediction target best preserves action sensitivity (the nearest row of patches, rather than the whole next cell or pooled features), and its component ablations verify the individual contributions of visual prediction, textual transition supervision, and verifier evidence—showing that predictive features pay off only when connected to language reasoning and decision-time evidence.

Critical weaknesses
  1. The "+tool" ablation bundles verifier score inputs with goal-adjacency supervision, so the gain from text to full (36.3%→59.8%) may come largely from the adjacency-belief supervision rather than from decision-time verification scores. A control that removes only the verifier inputs on top of full joint training is missing, leaving the causal contribution of Verify as decision evidence unclear.
  2. Verify scores each candidate move using only the edge-aligned single row/column (14 tokens) of the goal features, and the paper does not report the final search accuracy when scoring with full-grid similarity (or more position-tolerant matching).
Questions for authors
  1. Could you provide an ablation that removes only the verifier score inputs (while keeping the goal-adjacency supervision) on top of the full joint training? In other words, how much of the success rate is contributed by the decision-time verification scores themselves?
  2. If Verify instead scored candidate moves using full-grid similarity against the goal features (or more position-tolerant matching such as max-over-positions), what would the final search success rate be?
Reviewer_7Et8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies embodied aerial search through active geo-localization, where an agent uses local aerial views to find an image- or language-specified target on an image grid. AeroEVA combines neighboring-region feature prediction and textual transition supervision with policy learning. For aerial-image goals, it also uses predicted feature-to-goal matches as evidence for action selection, followed by reinforcement learning to refine decisions. The paper introduces AeroTask, whose language goals describe operational requirements rather than target appearance alone. The experiments show that visual prediction by itself does not improve search, whereas combining it with textual supervision and verifier evidence improves performance on the evaluated tasks. Reinforcement learning further improves goal-adjacency judgments and handling of misleading evidence. Reported transfer results across regions, goal modalities, appearance changes, and grid settings support the method's applicability within the tested image-grid search environments.

Critical strengths

S1. Analysis-driven design with matched internal ablations. Sec. 5.2 first uses an inverse-dynamics probe to show that action-related information decays in deeper decoder layers (Fig. 4; App. B.1 states that Action SFT and Joint SFT use the same 1,125 updates, data and trainable parameters, with scene-level bootstrap intervals within ±1.7 points, p. 19), and then uses Fig. 5 to choose the nearest row of the next cell as the prediction target. Table 2 (p. 8) clearly shows that visual prediction alone is below Action SFT (29.7 vs. 33.3), reaches 59.8 when combined with textual targets and tool evidence, and 77.0 after GRPO, vs. 56.4 for Action SFT + GRPO.

S2. Strong results over a broad evaluation. On the MASA test set, 86.1% vs. 66.4% (Table 1; 4,475 fixed test episodes, App. A.2, p. 13); the same checkpoint transfers zero-shot and leads prior methods on SwissView100, MM-GAG, xBD, 10×10 grids, and ground-photo and text goals (Tables 4–6, p. 9). Decoding randomness (Table 8: std ≤ 0.6 points, p. 17) and failure cases (Fig. 10, p. 20) are also reported.

Critical weaknesses

W1. Neither the test-time role of Verify nor the source of the transfer gains is isolated

  • Issue and location: The core claim is that transition prediction plays two roles—training signal and test-time evidence (Abstract; Sec. 6)—and Sec. 5.5 states that "AeroEVA’s latent prediction abstracts away pixel-level details to learn transferable predictive representations" (p. 8). There are three gaps. (i) No experiment compares tool-on vs. tool-off for the same checkpoint with the same aerial-image goals; "tool off" only appears with ground-photo, text and AeroTask goals, which also change the goal modality (p. 13). (ii) Transfer experiments compare only the full policy against prior methods with different backbones (GOMAA-Geo and GeoExplorer use Sat2Cap + Falcon-7B; App. A.1, p. 13), and App. A.2 itself notes that "it does not isolate robustness gains from latent transition prediction or tool use" (p. 13). (iii) The only same-backbone RL control without spatial prediction—"RL w/o spatial prediction" in Table 9 (p. 17)—is undefined in the main text and evaluated only on SwissView100, where its C4–C8 average (78.0) is higher than AeroEVA's (75.2) while its C4–C6 average (63.3) is lower (AeroEVA: 71.4).

W2. Goal use is not verified for language and ground-photo goals (including AeroTask), and AeroTask quality is checked only by the VLM that generated it

  • Issue and location: App. A.4 shows that fixing the goal descriptor changes the ground-photo and text-goal success of GOMAA-Geo and GeoExplorer by at most 3 points, i.e., these baselines perform largely goal-independent search, and states that "These controls do not cover the full AeroEVA checkpoint" (p. 17). The AeroTask gap in Table 7 (55.4% vs. 43.8%/41.0%) may therefore partly reflect a better generic search strategy rather than reasoning from task requirements. Moreover, AeroTask instructions are generated by Qwen3-VL-32B and judged by the same model (target fit, blind localization, rival-ambiguity checks); the paper acknowledges that "these model-based quality criteria do not constitute independent-model validation" (p. 15), and there is no human audit.
Questions for authors
  1. What is the success rate of the same AeroEVA checkpoint on the MASA test set (aerial-image goals) with tool evidence turned off? And on SwissView100 or xBD-disaster?
  2. What training recipe does "RL w/o spatial prediction" in Table 9 use—is it the same as Action SFT + GRPO or cold-start GRPO in Table 2? Why is it evaluated only on SwissView100, and what are its results on MM-GAG and xBD?
  3. On AeroTask and on MM-GAG text/ground-photo goals, how much does AeroEVA's success drop when its goal is replaced by another task's goal from the same dataset (or by a fixed goal)?
  4. In Table 2, what happens if only the adjacency supervision is added without the tool scores (or vice versa)?
  5. Has the AeroTask evaluation set been checked by humans? What fraction of tasks have requirements that uniquely identify the target cell?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.