acceptodds
Withdrawn from ICLR 2027

SearchWorld: Spatial Value-Grounded Imagination for UAV Object Search via World Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Autonomous unmanned aerial vehicle (UAV) object search involves a closed loop of perception, decision-making, and action under partial observability. Urban environments pose several challenges: large search areas and narrow egocentric views limit coverage, dense 3D geometry constrains safe motion, and open-world instructions require identifying a specific target among distractors. Many existing methods mitigate partial observability through explicit maps or memory representations, yet remain largely reactive, reasoning over past observations without explicitly predicting future states. World models enable prospective reasoning through imagined rollouts. However, image-generating world models can incur high inference latency, while spatially grounded planning remains challenging for latent world models. We propose SearchWorld, a recurrent state-space world model that connects explicit spatial memory with value-guided imagination. The model maintains bird’s-eye-view exploration and obstacle memory and decodes a task-aware spatial value layer to guide search. A cognition–action network uses this learned spatial value prior to improve the policy through imagined rollouts, without training a separate scalar critic. Training progresses from world-model learning to expert imitation and imagination-based exploration refinement. On UAV-ON, SearchWorld improves the success rate to 23.8% (19.5% for the strongest published agent) and raises oracle success to 35.5%, while remaining robust on unseen scenes (19.9% success rate). By grounding imagination in explicit spatial representations, SearchWorld enables UAV agents to plan prospectively rather than react.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Object Search, World Model, Object-goal Navigation, City Space

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_3PrhOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

SearchWorld combine spatial memory and RSSM(Recurrent State Space Model) for UAV object goal search. The architecture uses 3 BEV (Bird’s-Eye View) layers, which represent the search area as a top-down 2D grid where each cell corresponds to a physical location. This research focuses on connecting explicit spatial memory and future state prediction. Authors argue that these two parts are not well connected in prior works. Exploration and obstacle layers are constructed from depth and pose. Spatial value layer is decoded from the latent state with the target instruction. For each imagined movement action, the method pools predicted values over its sensing footprint to score the action. These scores guide policy refinement without a separate scalar critic. The value map uses exploration, obstacle, and target position as training supervision. Target position is not given at inference. Training has 3 stages: random rollout world model learning, expert imitation, and value-guided imagination refinement. Stage 3 changes movement behavior while keeping the BC stopping probability.

Critical strengths
  • Suggest explicit usage of world model’s memory about environment and link WM research to Aerial object search task. Replacing the scalar critic with decoded spatial value layer over an inspectable BEV memory. I think this is an interesting direction for aerial object search. My main concern is whether the experiments clearly show why this combination works.

  • The proposed method is efficient at inference. It runs without LLM or VLA at execution time. Compare with LLM / VLM baseline like WMNav and PRPSearcher, SearchWorld's inference efficiency is much faster than both prior works. Lower inference time is useful because search requires many sequential decisions. But Table 3 mixes timings measured by the authors and timings from prior papers. So, the comparison may not use the same measurement conditions.

Critical weaknesses
  • (Learned future state prediction's effect is not disentangled in ablation) Maintain same memory, supervision, policy refinement objective, there's lack of compare results about current state decoding without learned future-state rollout. So, in stage 3's metric improvement, it's hard to distinguish the role of future prediction and dense spatial value guidence's contribution. Removing Stage 3 removes the whole imagination-based, value-guided refinement stage. This does not separately test future prediction. The Dreamer-style comparison also changes the critic and learning objective. Does this baseline receive the same dense supervision? Without this information, it is hard to tell which part explains the performance gap.

  • (Direct validation of world model predictions is missing) In the paper, there's lack of explanation about prediction accuracy of BEV and value maps which are decoded from imagined future states or evaluate whether the action utilities reliably rank candidate actions. It makes difficult to assess whether imagination provides reliable information for policy refinement. Authors could report prediction errors at different rollout depths and compare the imagined action rankings with simulator-based reference rankings. Checking reconstruction from posterior states alone would not show whether prior-only future predictions are accurate. Stage 3 uses these predictions to compare actions, so this validation is important.

Questions for authors
  • In section 4.1, there’s 1000 test episodes in the experiment. In Table 1 and 2, are each methods evaluated by the same episode set or not? Appendix G says the full test set is used. But Appendix D says some results are reproduced and others are taken from prior papers.

  • In related works, the paper mentions WMNav’s online curiosity value map and says prior work’s env memory is implicit. I read that WMNav has explicit spatial memory isn’t it? What is the main difference between WMNav’s Map and SearchWorld’s spatial value map?

  • Introduction mentions that world model’s implicit spatial representation and scalar critic centric design limit effective spatial planning. What is the experimental evidence of this diagnosis? Mentioned architecture(design) is the real bottleneck in aerial object search? How do the experiments separate the effect of the spatial value representation from the effect of dense supervision using training-time target positions?

  • Table 1, 2 numbers may have errors. In paper, PRPSearcher’s OSR(25.65%) is higher than seen & unseen results (23.32% & 21.97%) why?

Reviewer_EC7jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SearchWorld combines an RSSM with explicit BEV memory and a supervised spatial value layer for UAV object search. Its three-stage training uses imitation and value-guided imagination to refine movement while preserving the stopping policy. The paper reports 23.8% success on UAV-ON, with component ablations and seen/unseen evaluations.

Critical strengths

The spatial memory and value-guided policy update form a coherent design. The reported ablations show that removing explicit memory reduces success from 23.80% to 15.96%, while removing imagination-based refinement reduces it to 19.71%. These results support the usefulness of the combined system.

Critical weaknesses
  1. The main comparison uses unequal action budgets: SearchWorld allows 150 steps, while the imported OctMem-Agent results use 100. This confounds the reported 4.3-point success gain.
  2. The reported aggregates are inconsistent. PRPSearcher’s overall OSR of 25.65% exceeds both its seen and unseen values, 23.32% and 21.97%. SearchWorld’s totals also imply different split weights across SR and OSR.
  3. The Dreamer comparison does not clearly match the target-derived supervision and policy-update objective. Its performance gap therefore does not isolate the contribution of spatial value grounding.
  4. The inference procedure is unclear: the main text describes policy distillation and one RSSM transition, whereas Appendix C describes online H-step rollouts. The rollout horizon, guidance temperature, and utility aggregation are unspecified.
Questions for authors
  1. How does SearchWorld compare with OctMem-Agent under the same action budget?
  2. Can you reconcile the overall and seen/unseen results, providing episode counts and aggregation rules?
  3. Does the advantage remain with scalar or current-map value guidance using equivalent supervision and policy refinement?
  4. Is imagination performed during deployment? Please specify the rollout configuration and what the reported timing includes.
Reviewer_hhQbOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes SearchWorld: a recurrent state-space world model for UAV target search that integrates explicit spatial memory with value-guided imagination. To address challenges in partially observable urban environments, including large search areas, complex 3D geometry, and the need to identify specific targets from distractors under open-ended instructions, the model maintains BEV exploration and obstacle memory, and decodes a task-aware spatial value map to guide search. The cognition-action network leverages the learned spatial value prior to refine the policy through imaginative rollouts without training a separate scalar critic. The training pipeline proceeds sequentially from world model learning and expert imitation to imagination-based exploration refinement. On the UAV-ON benchmark, SearchWorld boosts the success rate to 23.8 while retaining real-time inference speed. The work features a clear motivation and comprehensive ablation studies.

Critical strengths

1.Combine the RSSM world model with an explicit three-layer BEV spatial memory (exploration layer/obstacle layer/learnable value layer), and decode the "spatial value layer" to replace the scalar critic for imagined planning 2.The design of the cognition-action network (the value layer only performs action reweighting, and the stop probability is fixed to the BC policy) is a targeted engineering innovation that avoids the failure of critic bootstrapping under sparse rewards.

Critical weaknesses

None

Questions for authors

Why does SearchWorld replace the Dreamer-style scalar critic with a decoded spatial value layer? What is the failure mechanism of scalar critics under sparse rewards, and how does spatial value supervision alleviate this issue?

Reviewer_rMowOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SearchWorld addresses instruction-conditioned UAV object search under partial observability. It combines an RSSM with allocentric BEV exploration and obstacle maps constructed from depth and pose. A decoder predicts a task-aware spatial utility map, supervised by a mixture of unexplored-area preference, obstacle masking, and a Gaussian centered at the training-time ground-truth target position. The target coordinate is not supplied at inference. Training proceeds through random-interaction world-model learning, behavior cloning from a privileged rule-based expert, and imagination-based movement-policy refinement with the world model frozen. Candidate actions are scored by pooling predicted utility over their sensing footprints; a KL-regularized reweighting of the behavior-cloned policy is distilled into the action head, while the conditional stop probability is preserved. As the paper acknowledges, this last stage optimizes supervised utility rather than environment returns. On UAV-ON, the paper reports 23.80% SR and 35.52% OSR, compared with 19.50% SR and 29.30% OSR for OctMem-Agent. Ablations report 19.71% SR without Stage 3 and 15.96% without explicit memory. These results suggest useful gains, subject to the reporting inconsistencies discussed below.

Critical strengths

1 Coherent use of spatial structure for search The method links persistent exploration and obstacle memory to a task-conditioned utility decoder and a compact latent dynamics model. This provides a concrete way to guide imagined movement using physically indexed spatial information. Combining an imitated task policy with constrained movement refinement is a sensible design for a task where both exploration and stopping matter. 2 Promising gains with relevant component ablations The reported 4.30-point SR improvement over the strongest published comparison is meaningful for this difficult benchmark. The Stage-3 and explicit-memory ablations provide preliminary evidence that both components contribute within the proposed system. The evaluation also includes held-out scenes, and the paper explicitly acknowledges its simulation and sensing assumptions.

Critical weaknesses

1 The main results need reconciliation Tables 1 and 2 contain quantities that cannot be reconciled under the stated per-episode averaging protocol and a common exhaustive seen/unseen partition. For PRPSearcher, overall OSR is 25.65%, whereas seen and unseen OSR are 23.32% and 21.97%; overall SPL is 5.61%, below both subgroup values, 6.26% and 6.48%. A pooled mean cannot lie outside both subgroup means. There is also an inconsistency for SearchWorld. Its overall SR of 23.80% and seen/unseen SR of 25.14%/19.92% imply a seen proportion of approximately 74.3%. Applying that proportion to OSR of 37.88%/33.12% gives approximately 36.66%, rather than the reported 35.52%. Ordinary rounding does not explain this difference. Different runs, subsets, or aggregation rules could account for it, but must be specified. This is consequential because the main performance and generalization claims rely on these tables. I regard it as a reporting issue requiring resolution, not evidence of misconduct. 2 The evidence does not yet isolate the central mechanism The comparison with the Dreamer-style scalar critic is described as isolating spatial value grounding, but the proposed spatial map receives direct target-coordinate and exploration/obstacle supervision. The scalar-critic baseline is described as learning imagined returns, without sufficiently specifying its reward signal or whether it receives comparable privileged supervision. Thus, the reported gap can reflect supervision and objective differences as well as the spatial representation. The current results support the effectiveness of the combined design more strongly than the claim that spatial value representation itself explains the gain. The role of imagination at deployment is also ambiguous. Sections 3.5 and 4.2.2 emphasize policy distillation and a lightweight execution path, whereas Appendix C describes H-step prior rollouts used for evaluation-time decisions. The horizon, continuation policy, and multi-step utility aggregation are not specified. These distinctions affect what was actually evaluated and how to interpret the planning and latency claims. A comparison retaining the same spatial supervision while removing future-dynamics rollout would help establish the additional contribution of imagination. This is a targeted test of the central claim, rather than a request for broad additional benchmarking.

Questions for authors

Clarifications that could change the recommendation

  1. Can you reconcile Tables 1 and 2 and provide the test-episode counts and aggregation rules for each split? Please identify any differences in evaluated tasks, model checkpoints, or runs, and update the affected overall and subgroup metrics.
  2. At evaluation, does the agent directly execute the distilled policy, perform online H-step planning, or combine both? Please specify H, the rollout continuation policy, how per-step utility is accumulated, and whether Table 3 includes the complete computation used for action selection.
  3. What reward and supervision are used by the Dreamer-style baseline, given that Section 3.1 says benchmark success is not used as a training signal? Does it share the imitation initialization and privileged information used by SearchWorld? A matched-supervision comparison, or a direct spatial-utility policy without imagined dynamics, would materially strengthen the attribution of the gain. If such evidence is unavailable, would you narrow the conclusion to the effectiveness of the combined system?
Reviewer_VMtTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses UAV object search in urban environments: given RGB-D observations, the UAV pose, and a text instruction describing a target instance, the agent must locate and reach the target over an area far larger than its field of view, without any external route. The authors identify two limitations in existing approaches. Map-based agents use explicit spatial memory to mitigate partial observability, but their reasoning over past observations is reactive and does not predict the future states induced by their own actions. World-model agents can perform prospective rollouts, yet they compress the spatial information that governs search efficiency — visited, impassable, and unexplored regions — into a flat latent vector, which actor–critic training then collapses into a scalar value, discarding spatial structure. The authors propose SearchWorld, a recurrent state-space world model (RSSM) grounded in an explicit spatial-cognition memory stored as a bird's-eye-view (BEV) grid. Its exploration and obstacle layers are constructed deterministically from pose and depth, while its value layer is decoded from the world-model latent state; the supervision for this value layer is built at training time from exploration status, obstacle constraints, and a Gaussian prior around the target, and target coordinates are not used at inference. Visual observations and instructions are encoded with frozen DINOv2 and SigLIP. A cognition–action network replaces the Dreamer-style scalar critic with the decoded spatial value layer: for each candidate movement action, it decodes the value layer of the imagined successor state, pools it over the action's motion footprint, accumulates an H-step action utility, and uses it to exponentially reweight the behavior-cloned policy, distilled through a KL divergence. Only movement actions are updated; the stop decision remains that of the expert-imitation policy. Training proceeds in three stages: world-model pretraining, expert behavior cloning, and imagination-based value-guided refinement. On UAV-ON, the method attains 23.80% SR and 35.52% OSR, outperforming the strongest published agent, OctMem-Agent (19.50% / 29.30%), and retains 19.92% SR on unseen scenes. Ablations examine the world-model architecture, the two memory layers, and the three-stage curriculum.

Critical strengths
  1. The most substantive design choice is a spatial value layer, decoded from the latent state and aligned with the BEV grid, which supplies action utility during imagination in place of a separately trained scalar critic. At training time this layer is densely supervised by exploration status, obstacle constraints, and a Gaussian prior around the target; at deployment it relies only on observations, pose, and instructions. Because the task provides only a sparse terminal success signal and training uses no return-based objective, the layer offers a dense signal for action evaluation that does not require bootstrapping from returns, mitigating the difficulty of training a scalar critic under sparse rewards. Value guidance is applied to movement actions only, while the stop probability is fixed to that of the behavior-cloned policy, so value optimization does not directly alter the stopping probability.
  2. All ablation variants are trained and evaluated within a consistent setup. Direct imitation, X-Mobility-style, Dreamer-style, and the variants that remove the explicit memory, the individual exploration and obstacle layers, or the individual training stages are all retrained by the authors on identical data and in the same action space, rather than quoted from other papers. Removing explicit memory or Stage 2 lowers the success rate by about 7.8 and 12.8 points respectively, and the full model holds a 4.7-point advantage over the Dreamer-style scalar-critic variant. The sources of the reported gain can therefore be examined within a single experimental setting.
Critical weaknesses
  1. The causal attribution of the core contribution is not isolated. The paper argues that the gains come from deciding on a spatial map rather than on the latent state, but the ablation supporting this claim varies several factors at once. In Table 4, the Dreamer-style variant reaches 19.07% SR and the full model 23.80%, while the two differ along three dimensions: the value representation (spatial map vs. scalar), the learning signal for that value (dense supervision constructed from the target location vs. learning from imagined returns), and the policy optimization procedure (value-guided distillation vs. actor-critic training). The 4.7-point gap therefore cannot be attributed to spatial structure alone. The minimal controlled comparison needed to test the claim — holding the dense supervision and the policy optimization procedure fixed while replacing only the spatial value representation with a scalar one — is absent. This bears on how independent the contribution is: the current results show that the full system outperforms the Dreamer-style baseline as implemented by the authors, but they do not yet support the causal reading that deciding on the map is the main source of the improvement.
  2. The reported baseline results and the evaluation protocol require clarification. The paper states that the OctMem-Agent results are quoted from its original paper, yet Table 2 reports its seen results as 20.17 / 29.55 / 6.79, whereas the original paper reports 22.72 / 30.48 / 8.40; the three unseen-scene metrics are identical. Since these values were not reproduced by the authors, the discrepancy needs an explanation. In addition, Appendix B states that the action magnitudes are fixed by the benchmark (5 m translation, 2 m vertical, 15° rotation), whereas the UAV-ON paper defines the action space as continuous, parameterized controls for translation, rotation, and stopping and lists fixed-step motion as one supported strategy; OctMem-Agent, in contrast, uses six discrete actions with 3 units forward, 3 m vertical, and 30° rotation. This does not by itself invalidate the comparison, but the authors should clarify whether results obtained under different action configurations are directly comparable, and justify attributing the fixed magnitudes to the benchmark. Taken together, these two points leave me unable to confirm that the cross-method comparisons in the tables were obtained under the same execution conditions.
Questions for authors
  1. Table 2 reports the seen-scene results for OctMem-Agent as 20.17 / 29.55 / 6.79, whereas the original paper reports 22.72 / 30.48 / 8.40, while the three unseen-scene values are identical. Please clarify the provenance of these numbers: were they taken from a particular version or split, or obtained by re-running the method?
  2. Were the baselines re-run under the eight-action configuration used in this paper? If their original configurations were retained, please specify the action magnitudes and discretization used when reproducing WMNav and PRPSearcher.
  3. A controlled variant that keeps the dense goal supervision and the policy optimization procedure unchanged and replaces only the spatial value representation with a scalar one would indicate how much of the gain is attributable to spatial structure; correspondingly, results with λ_goal set to zero would show how much of the Stage 3 gain derives from goal-distance supervision itself. These results would materially affect my assessment of the core contribution.

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

Related papers

Discussion (0)

Sign in to comment.