acceptodds
Withdrawn from ICLR 2027

Self-Evolving Grounding World Model for Dynamic Long-Horizon Aerial Object Goal Navigation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

14 views

Abstract

Object-goal aerial navigation has witnessed rapid emergence with the development of vision–language–action models. However, dynamic long-horizon object-goal aerial navigation remains challenging: with goals arriving dynamically one after another, the model's current decision basis is repeatedly invalidated and a fresh direction must be determined entirely from its own observations, and as the chain grows long, the scene state and the value of earlier observations dynamically shift along the progressively revealed sequence and must be maintained rather than reset. Existing methods follow transient per-step observations with no direction signal and discard observation memory across goals, and therefore fail to handle the changing goals and to sustain success across the chain. To address these, we propose the Self-Evolving Grounding World Model (SE-GWM), a unified framework that couples goal grounding with memory evolution. Specifically, we introduce three collaborative modules: i) a **multi-view video grounding** module that performs *visual sequence grounding* over the current observations and remembered keyframes, converting each described object into an actionable 3-D goal estimate together with target, exploration, and avoidance judgments; ii) an **exploitation–exploration navigation** module that turns the grounded state into discrete flight actions through an evidence-bound planner and a goal-conditioned policy; and iii) **priority-gated memory evolution**, which preserves grounded evidence under a decaying admission threshold and feeds remembered keyframes back for re-grounding, letting the world model self-evolve with its own flight experience. Extensive experiments on our newly constructed UAV-ON-MG benchmark (1,000 episodes, 5,744 goals, 14 open-world scenes) show that SE-GWM consistently outperforms state-of-the-art methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
World Model, Aerial Object Goal Navigation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_6ZAnOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SE-GWM, a unified frame work for dynamic long-horizon object-goal aerial navigation that decomposes the task into grounded goal estimation, evidence-bound flight, and gated memory evolution. The paper further construct UAV-ON-MG, a 5,744-goal benchmark stratified by cross-goal search context.

Critical strengths

The paper presents a complete navigation system that integrates multi-view goal grounding, exploration-exploitation navigation, and cross-goal memory within a unified pipeline. It also constructs UAV-ON-MG, a new benchmark for evaluating dynamic long-horizon aerial object-goal navigation across chained goals and different memory contexts. Together, the system and benchmark provide a useful foundation for studying cross-goal information reuse in aerial navigation.

Critical weaknesses

Insufficient motivation and comparisons with existing memory-based methods. The claim that existing approaches rely on transient observations and discard long-term memory is overly broad. The paper itself cites prior work on navigation history, persistent embodied memory, sequential-horizon navigation, and self-evolving memory, including HETT, SeqWalker, EvoMemNav, HIMM, and MEM. However, none of these memory-enabled frameworks is included in the experiments. Most reported baselines are VLN/VLA models that require directional instructions and are not originally designed for category-only object-goal search. The paper acknowledges that several of them collapse to nearly constant forward actions without an explicit goal-direction signal. Consequently, the comparison may primarily reflect a task-interface mismatch rather than demonstrating the superiority of the proposed memory mechanism.

Insufficient validation of the UAV-ON-MG benchmark. Although the paper reports basic statistics for chain length, object size, and search context, it does not convincingly establish that the benchmark measures cross-goal memory as intended. A goal being visible from an earlier viewpoint does not necessarily mean that its success depends on retaining that observation, yet the paper directly claims that 59% of goals are memory-dependent. More importantly, the reported results contradict the expected monotonic advantage: blind goals achieve the highest success rate, whereas memory-guided goals achieve the lowest. The authors acknowledge that this result is confounded by chain position and that the position-controlled analysis is still ongoing. The reliability of the search-context labels, chain construction procedure, and claimed memory dependency therefore remains insufficiently demonstrated.

Incomplete and insufficient experimental evaluation. The experiments do not contain a clean memory-free SE-GWM baseline or a complete removal of PAME. Table 3 only changes the write threshold and keyframe sampling density, which cannot isolate the contribution of cross-goal memory. The paper also explicitly states that the position-controlled analysis and the planner, grounding-frequency, recovery, and ungrounded-state ablations are still ongoing or being rerun. Moreover, the LLM branch of the evidence-bound planner is not invoked during evaluation; only the deterministic fallback is used, leaving one of the claimed modules experimentally unvalidated. Finally, the improvement over the strongest baseline is relatively limited in overall SR (27.5% versus 25.9%), while no multi-seed variance, confidence interval, or statistical significance is reported.

Questions for authors

See Weaknesses.

Reviewer_9oEQOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SE-GWM for dynamic long-horizon object-goal aerial navigation, where a UAV must find a chain of goals that arrive sequentially.

Critical strengths

The research is well-motivated. Long-horizon navigation is an important problem when embodied agents deployed into the real world.

Critical weaknesses
  1. Section 5.1 names chain SR and the average number of goals found (tf) as the primary metrics, yet neither is reported in the main results (Tables 1 and 2).

  2. The method description is inconsistent with the actual implementation. Section 3.3 describes the LLM-based planner in detail, but Appendix A states that the LLM branch was not invoked in evaluation.

  3. In the search-context results, memory-guided goals are the hardest (22.4%) and blind goals the easiest (32.1%). This contradicts the paper's core claim about self-evolving memory, and the authors only attribute it to a confound that is still under analysis.

  4. The paper is not easy to follow.

Questions for authors

Please see the weaknesses above.

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

This paper studies UAV navigation to multiple object goals that arrive sequentially. It proposes SE-GWM, which uses Multi-view Video Grounding to convert the current four-view images and historical keyframes into target positions plus exploration and obstacle-avoidance information. A planner and a goal-conditioned policy then execute actions. Priority-gated Memory Evolution retains and reuses observations across different goals. The authors also build UAV-ON-MG, which contains 14 test scenes, 1,000 episodes, and 5,744 goals. Each task chain includes 3–8 goals.

Critical strengths
  1. UAV-ON-MG explicitly labels whether subsequent goals were seen during earlier searches as a task attribute, providing a concrete data foundation for studying cross-goal memory.
  2. The method provides a clear spatial anchoring and control interface for historical visual evidence, allowing previously observed goals to serve as directional signals for later navigation.
  3. The local ablations of memory gating and keyframe sampling hold the VLM invocation budget constant, supplying evidence for the contribution of how memory is used.
Critical weaknesses
  1. The paper describes three key mechanisms—LLM-driven planning, sub-goal embedding conditioning, and curriculum training—but Appendix A reveals that the LLM branch was not invoked in evaluation, the sub-goal embedding was held at a fixed index, and no curriculum was used.
  2. The paper states that chain SR and average goals found are primary metrics (§5.1, L375–376), but Tables 1–4 report only per-goal SR, OSR, SPL, and DCA. Without chain SR, average goals found.
  3. Removing the goal-direction input eliminates a key control signal for the policy, making it impossible to isolate the contribution of cross-goal historical memory
  4. The paper attributes the 11.1 percentage-point gap between OSR (38.6%) and SR (27.5%) to the fixed 25 m depth prior used in reprojection (Eq. 7, Appendix A, L686–687; §5.2, L427–430), but provides no direct depth-error measurement or ablation to verify that depth is the primary failure cause.
Questions for authors

see weakness

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

Related papers

Discussion (0)

Sign in to comment.