PaperlayerView exact revision on arXiv

Source-linked AI summary

Dual-Layer Semantic-Spatial Belief Mapping for Aerial Object Goal Navigation

Jianqiang Xiao, Xiang Deng, Yuexuan Sun, Yanjin Wu, Wenbiao Yan, Liqiang Nie

arXiv:2609.08164v1cs.ROcs.AI

TL;DR

Aerial ObjectNav must locate and approach described targets despite noisy, sparse, and transient VLM observations in unknown outdoor environments. AeroBelief converts these observations into persistent dual-layer spatial belief and stable coverage guidance, achieving the best reported overall SR, OSR, and SPL on UAV-ON among compared methods.

  • Problem

    Aerial ObjectNav lacks persistent, reliable spatial handling for noisy VLM judgments that can blur contextual plausibility with target-specific evidence.

  • Method

    AeroBelief combines target-conditioned conservative evidence qualification, dual-layer intuition–evidence mapping with evidence-gated fusion, and temporally committed egocentric regional guidance.

  • Results

    AeroBelief achieves the best reported overall SR, OSR, and SPL among compared methods on UAV-ON, reaching 21.61%, 35.57%, and 10.62, respectively.

  • Takeaways & Limitations

    The results support persistent semantic-spatial belief, conservative evidence qualification, and temporally stable regional guidance for aerial ObjectNav.

  • Takeaways & Limitations

    Evaluation is limited to AirSim simulation, and sim-to-real transfer remains unvalidated; reliable target confirmation and stopping remain challenges.

Abstract

from arXiv · show

Aerial Object Goal Navigation (ObjectNav) requires an unmanned aerial vehicle (UAV) to locate a described target in an unknown outdoor environment using onboard visual observations. Vision-language models (VLMs) can interpret open-ended target descriptions and visual observations, but their frame-level outputs are often noisy, sparse, and spatially transient. We propose AeroBelief, a dual-layer semantic-spatial belief mapping framework that transforms transient VLM observations into persistent spatial guidance. It separates broad contextual plausibility from target-specific evidence: an intuition layer accumulates scene-level semantic cues for exploration, while an evidence layer preserves qualified target-specific observations for approach and confirmation. Evidence-gated fusion combines the two layers into spatial belief hotspots. We further introduce object-conditioned visual reasoning with conservative evidence qualification to improve observation reliability before spatial accumulation. In parallel, egocentric regional guidance converts quadtree coverage into UAV-centered, yaw-aligned directional proposals and stabilizes them through temporal commitment. Its regional scoring is independent of semantic belief values, maintaining exploration pressure and reducing repeated low-gain search. Experiments on the UAV-ON benchmark show that AeroBelief achieves the best reported overall SR, OSR, and SPL among the compared methods, reaching 21.61%, 35.57%, and 10.62, respectively. These results support the effectiveness of persistent semantic-spatial belief, conservative evidence qualification, and temporally stable regional guidance for aerial ObjectNav.

I. Introduction

AeroBelief addresses aerial ObjectNav challenges caused by transient VLM judgments, sparse target visibility, and large unknown outdoor spaces. It combines persistent dual-layer semantic belief with conservative evidence qualification and stable regional exploration guidance.

  • Aerial ObjectNav requires searching and approaching a described target in large, partially unknown outdoor environments with viewpoint changes, occlusion, and sparse visibility.
  • Frame-level VLM judgments may lose spatial persistence and conflate contextual plausibility with target-specific evidence, producing misleading hotspots and repeated irrelevant searches.
  • AeroBelief preserves contextual intuition and qualified target evidence in separate layers, then fuses them into spatial belief hotspots for planner-level decisions.
  • Conservative evidence qualification uses target-conditioned reasoning to restrict insufficiently supported target claims before persistent spatial accumulation.
  • Egocentric regional guidance converts quadtree coverage into yaw-aligned directional bins and uses temporal commitment to reduce oscillation and redundant revisitation.

C. Semantic Mapping for Active Search

The framework interprets target-conditioned observations and projects them into persistent aerial search representations while preserving coverage-oriented exploration. Success requires explicit stopping within a predefined target radius, not merely entering the target vicinity.

  • Aerial ObjectNav uses an unknown outdoor environment, a natural-language target specification, and onboard visual observations without providing the target position.
  • Success requires an explicit stop action within a predefined success radius, distinguishing successful stopping from merely reaching the target vicinity.
  • The framework projects semantic signals into a shared search grid while independently using egocentric, yaw-aligned coverage guidance stabilized by temporal commitment.
  • Object-conditioned reasoning builds a target cue from the specification and initial multi-view observations to guide interpretation rather than action selection.
  • Structured semantic judgments distinguish exact and strong semantic matches from weak contextual matches, which can guide exploration without confirming target presence.

2) Conservative Evidence Qualification:

Conservative evidence qualification revises target claims when visual support is insufficient, preserving useful contextual guidance while preventing unsupported claims from entering the evidence layer.

  • The qualification procedure may revise confidence when available visual support is insufficient and does not strengthen weak observations into target-specific evidence.
  • Qualification examines scene-context conflict, insufficient core verification, and part–whole mismatch for exact or strong semantic claims.
  • Insufficiently supported observations are reclassified as weak contextual, remain eligible for intuition-layer projection, and are excluded from evidence-layer projection.

D. Dual-Layer Semantic-Spatial Belief Mapping

AeroBelief maintains complementary intuition and evidence maps over a shared aerial grid, projecting qualified observations spatially and updating the layers with different persistence rules.

  • Each grid cell stores an intuition value for context-guided exploration and an evidence value for target approach and confirmation.
  • Qualified observations are projected into the search grid using UAV pose, viewing direction, distance band, and observable spatial support.
  • Positive view contributions are transformed into a shared world frame and aggregated per cell by a maximum operation to avoid repeated addition of overlapping observations.
  • The intuition layer admits contextual observations, applies local max-belief updating, and uses temporal forgetting to remove outdated support.
  • The evidence layer admits only qualified target-specific observations above its confidence threshold and preserves them without within-episode decay when visibility is temporarily lost.
  • Evidence-gated fusion integrates the intuition and evidence layers into spatial belief hotspots.

2) Evidence-Gated Fusion and Hotspot Interface:

Evidence-gated fusion combines contextual intuition with qualified target-specific evidence, increasing the evidence contribution as reliable observations accumulate. The fused map is then compressed into directional hotspots for planner use.

  • The fusion weight increases with qualified evidence, keeping intuition dominant when evidence is missing or sparse and strengthening evidence for approach and confirmation.
  • Contextual cues guide exploration through the intuition layer but do not directly enter the evidence term.
  • Without the evidence layer, the fused representation reduces to an intuition-only map.
  • The fused belief map is converted into a compact planner interface by extracting top-K directional belief hotspots.Hotspots summarize target-related spatial hypotheses using direction, bearing, distance, and fused belief score.
  • Direction-binned non-maximum suppression and an explored-region penalty select concise target-relevant candidates from the fused map.

E. Egocentric Regional Guidance

Egocentric regional guidance turns quadtree coverage into UAV-centered directional proposals whose scores remain independent of semantic belief. Temporal commitment stabilizes these proposals while allowing the planner to weigh them against semantic hotspots and execution constraints.

  • Regional scoring evaluates exploration opportunity, reachability, and revisit-aware efficiency independently of semantic belief values.The stream complements fine-grained semantic belief when reliable evidence is sparse.
  • Quadtree coverage is transformed into a 5 × 5 body-frame grid reconstructed from the UAV’s current pose and yaw.Rows follow the forward–backward axis, columns follow the left–right axis, and the center contains the UAV.
  • Each bin represents an egocentric directional region rather than a persistent world-space partition.The same bin index can correspond to different world locations as the UAV moves or rotates.
  • The instantaneous candidate is recomputed from geometrically valid bins as the most promising current exploration direction.
  • Temporal commitment retains a useful direction or updates it when stronger opportunities, low progress, obstacles, or repeated visitation make the current choice stale.Minimum duration and switching hysteresis reduce short-term oscillation.
  • The committed cue expresses an exploration preference, while the planner may depart from it when stronger target evidence or motion constraints support another action.The cue is not an exact metric waypoint or a directly prescribed action.

F. Planner-Level Integration

A shared planner and execution backbone combines semantic observations, belief hotspots, regional guidance, pose–action history, and feasible actions for high-level proposals. On UAV-ON, AeroBelief achieved the best reported overall SR, OSR, and SPL among compared methods, with size-dependent behavior and representative episodes showing progressive belief concentration.

  • The planner receives semantic observations, belief hotspots, committed regional guidance, pose–action history, and feasible-action information to produce a high-level action proposal.
  • The shared execution layer handles stop verification, collision avoidance, boundary and altitude constraints, and fallback recovery across variants.
  • AeroBelief achieves the best reported overall SR, OSR, and SPL, reaching 21.61%, 35.57%, and 10.62, respectively.It improves over the strongest previously reported result by 2.11 percentage points in SR, 6.27 in OSR, and 0.48 in SPL.
  • Small and medium targets rank first across all three metrics, while large targets achieve the second-highest SR and SPL but lower OSR than several prior methods.
  • The largest size-specific OSR–SR gap is 18.44 percentage points for small targets, indicating frequent target-vicinity discovery without consistently successful stopping.
  • Representative successful episodes show intuition maintaining broad contextual belief before qualified evidence concentrates fused belief around the target.The UAV may continue regional search, adjust viewpoint or altitude, and reject ambiguous candidates before sufficient evidence accumulates.

E. Ablation Study

The ablation study shows that combining semantic-spatial belief mapping, object-conditioned reasoning, and regional guidance yields the strongest navigation performance, while persistent evidence and temporal commitment provide additional gains.

  • AeroBelief achieves the highest SR, OSR, and SPL and the lowest collision rate across the ablation variants.Removing any complete proposed pathway causes a larger degradation than removing only its paired subcomponent.
  • The results support semantic-spatial belief mapping, object-conditioned reasoning, and regional guidance, with evidence qualification and temporal commitment adding further gains.
  • 25.00% SR versus 11.00% for Vanilla accompanies a collision-termination reduction from 39.50% to 28.00%.
  • Removing the evidence layer increases MaxCov but lowers SR and SPL, showing that broader coverage alone does not ensure target confirmation.Persistent target-specific evidence helps convert exploration into successful approach and stopping.

Appendix A Shared Autonomous UAV Backbone

The shared UAV backbone converts multi-view RGB-D observations and state information into structured coverage and navigation inputs. Its incremental quadtree represents geometric coverage, while common execution and verification procedures constrain actions consistently across variants.

  • At each timestep, the UAV receives four RGB-D views, pose and flight state, pose–action history, and a target specification for structured planning.The shared interpretation stage processes raw camera observations into view summaries.
  • The incremental quadtree records explored and unexplored spatial structure and supplies common coverage information to all evaluated variants.It is maintained independently of the semantic belief map.
  • Action space, motion increments, action budget, search radius, and success criteria remain identical across AeroBelief, the vanilla backbone, and ablations.Stop verification is also shared across variants.
  • The shared execution layer rejects collision-risk or boundary-violating motions, protects minimum altitude, regulates altitude by target size, and provides recovery behavior.
  • Downward observations directly cover ground areas, whereas horizontal views cover larger but more geometrically uncertain projected regions.The coverage update therefore treats horizontal visibility conservatively relative to direct downward observation.
  • Adaptive subdivision refines partially intersected leaves, while homogeneous regions remain compact and internal-node statistics are updated from child nodes.

B. Coverage-Guided Regional Reasoning

Coverage-guided regional reasoning converts quadtree leaves into UAV-centered directional bins and selects exploration proposals using geometry and coverage rather than semantic belief. Temporal commitment stabilizes these proposals across steps.

  • Quadtree leaves are transformed into the UAV body frame and aggregated into a 5 × 5 yaw-aligned grid of relative directional bins.Rows follow the forward–backward axis, columns follow the left–right axis, and the center contains the UAV.
  • Each bin receives a coverage-oriented score based on exploration opportunity, reachability, and revisit-aware spatial penalties.The score considers unknown area, frontiers, heading alignment, travel distance, boundaries, spatial support, and visitation history, but not semantic belief.
  • The instantaneous candidate is selected from geometrically valid bins under the current search boundary.
  • Temporal commitment retains a useful direction until a stronger candidate appears or the current one becomes stale.Minimum commitment duration and switching hysteresis reduce short-term oscillation.
  • The committed relative bin is reinterpreted in the current body-frame grid after movement or rotation and supplied as coarse guidance rather than an exact waypoint.

Appendix C Prompt Design and Structured Reasoning Interface

The prompt design uses target-conditioned cues and structured per-view reasoning to support conservative, non-action-oriented visual interpretation. The planner receives organized summaries and rendered maps instead of raw camera data.

  • A. Target-Conditioned Visual Reasoning: The episode-level cue-generation prompt identifies discriminative target attributes and part–whole verification cues from the target specification and initial views.
  • A. Target-Conditioned Visual Reasoning: The cue also summarizes range-dependent appearance and supportive or suppressive scene contexts for later interpretation.
  • A. Target-Conditioned Visual Reasoning: The resulting cue is reused throughout the episode rather than regenerated at each semantic update step.
  • A. Target-Conditioned Visual Reasoning: Per-view reasoning estimates target-match type, confidence, distance band, contextual and target-specific support, visible content, and stop readiness.
  • A. Structured Reasoning Interface: The navigation planner receives rendered spatial maps and organized textual context rather than raw RGB-D observations.
  • A. Target-Conditioned Visual Reasoning: Structured observations pass through conservative evidence qualification before entering the semantic-spatial belief map.
  • A. Target-Conditioned Visual Reasoning: The cue supports visual interpretation but neither prescribes navigation actions nor enters the planner as a standalone instruction.

B. Structured Planner Interface

The structured planner combines fused semantic belief and quadtree coverage with organized textual context, then passes high-level proposals through shared execution safeguards. Failure analysis shows that incomplete exploration and termination miscalibration remain important error sources, especially for large targets.

  • Structured Planner Interface: The planner receives rendered fused semantic-belief and quadtree-coverage maps rather than raw RGB-D observations.
  • Structured Planner Interface: Its textual context includes task progress, UAV pose and motion, qualified visual interpretations, semantic hotspots, and coverage statistics.
  • Structured Planner Interface: High-level action proposals pass through shared stop verification, collision avoidance, boundary and altitude handling, and fallback recovery.
  • Failure Mode Analysis: Explicit stops account for 51.03% of terminations, while 27.50% reach the action limit and 21.48% end through collision.
  • Target-Size Sensitivity: Large targets can be recognized from farther away, but visual scale makes proximity estimation ambiguous and can trigger premature stopping beyond the success radius.
Loading 2609.08164v1…