Distance-Weighted Frontier Explorer (DWFE)

• The paper introduces DWFE, a geometry-only frontier exploration heuristic that clusters frontier cells with DBSCAN and scores navigable cells based on Euclidean distance.
• It demonstrates improved performance on HM3D with a 61.1% success rate and a 36.0 SPL, highlighting fewer detours and more efficient goal-reaching than baseline methods.
• DWFE operates independently of language semantics, using purely occupancy and geometric cues to direct exploration before goal detection.

Searching arXiv for the primary DWFE paper and closely related frontier-exploration work. Distance-Weighted Frontier Explorer (DWFE) is a geometry-only, training-free frontier exploration heuristic introduced as a replacement for the language- and vision-heavy selector used in InstructNav on HM3D-v1 validation. In that formulation, a frontier is the boundary between already explored free space and unknown space, operationalized as “free cells adjacent to at least one unknown cell.” DWFE uses the occupancy map already maintained by the navigation system, clusters frontier cells into frontier islands with DBSCAN, assigns higher scores to navigable cells that are closer to frontier islands under Euclidean distance, writes those scores into the Action Value Map (AVM), and leaves target selection and motion execution to the unchanged affordance-map and A*-based planning stack (Aghaei et al., 26 Jul 2025).

1. Definition and formal placement

DWFE denotes a specific frontier-selection mechanism rather than a complete autonomous navigation stack. Its role is confined to the exploration phase before the goal object has entered view. Once the goal object is observed, the selector is bypassed and the planner drives directly toward the object’s centroid until stop. Within that division of labor, DWFE governs only where the system prefers to explore next, not how the low-level control policy is executed (Aghaei et al., 26 Jul 2025).

The method consumes an occupancy grid built from RGB-D depth projection, with cells marked free, occupied, or unknown; a current set of frontier points recomputed every step; the set of navigable grid cells; and the robot pose as represented through the map frame and planner. Frontiers are grouped into frontier islands using DBSCAN with parameters $\varepsilon = 1.0$ m and $min\_samples = 1$. The resulting mechanism is explicitly geometry-only and language-independent during exploration. This places DWFE within the frontier-based exploration lineage, but with a deliberately narrow design objective: bias the robot toward nearby frontier structure without semantic reasoning (Aghaei et al., 26 Jul 2025).

A useful way to situate DWFE within prior exploration research is to note that frontier-based methods have long treated the explored/unknown boundary as the principal driver of map expansion. In 2D occupancy-grid exploration, “Exploration frontiers are the regions on the boundary between open space and unexplored space,” and the frontier set is formalized as $\mathcal {F}=\{f_0,...,f_j \}$ with a desired frontier $f^* \in \mathcal{F}$ (Gao et al., 2020). This suggests that DWFE is best understood as a minimalist selector layered atop an already mature frontier-detection and path-planning substrate.

2. Scoring mechanism and integration into the navigation stack

At each time step $t$, DWFE first updates the occupancy map from depth, detects all frontier cells as explored free cells adjacent to unknown cells, and clusters those frontier cells into islands $F_t$. For every frontier island $f \in F_t$, and for every navigable grid cell $x$, it computes a distance-based score using Euclidean distance from $x$ to the closest point in $f$. The scoring function is given as

$min\_samples = 1$0

where $min\_samples = 1$1 is the Euclidean distance from $min\_samples = 1$2 to the closest point in $min\_samples = 1$3, and $min\_samples = 1$4 is the maximum such distance across all navigable cells at the current step (Aghaei et al., 26 Jul 2025).

The paper leaves one implementation detail unspecified: it states that for every frontier island and every navigable grid cell, “these scores are added to the Action Value Map (AVM),” but does not state explicitly whether contributions from multiple islands are summed, max-pooled, or otherwise aggregated. The safest factual reading is therefore that the AVM receives frontier-distance-derived scores over the grid for all current islands, after which the existing affordance and planning stack resolves the final target (Aghaei et al., 26 Jul 2025).

Once the AVM has been updated, the remainder of execution is inherited unchanged from InstructNav. The component maps are summed into an affordance map $min\_samples = 1$5; the grid cell with $min\_samples = 1$6 becomes the target; and the A* planner treats $min\_samples = 1$7 as cost. The selected target is then converted into a path and low-level actions from $min\_samples = 1$8, with $min\_samples = 1$9 equal to $\mathcal {F}=\{f_0,...,f_j \}$0 m and turn increments of $\mathcal {F}=\{f_0,...,f_j \}$1 (Aghaei et al., 26 Jul 2025).

The broader frontier literature contains stronger gain-aware or path-aware formulations than DWFE. For example, in a 2D frontier planner, the heuristic

$\mathcal {F}=\{f_0,...,f_j \}$2

combines immediate information gain, path length, predicted future return, and a cleanup bonus, with the baseline form $\mathcal {F}=\{f_0,...,f_j \}$3 already corresponding to “coverage gain minus weighted travel cost” (Tangri et al., 2021). A plausible implication is that DWFE represents an intentionally stripped-down point in the design space: inverse-distance frontier preference without explicit path-cost, information-gain, or future-return terms.

3. Relationship to frontier-exploration research

DWFE belongs to the classical frontier-exploration family, but it differs from several influential strands in that it scores navigable cells by Euclidean distance to frontier islands rather than selecting frontier centroids or viewpoint candidates using explicit utility-over-cost objectives. In hybrid 2D exploration, a mobile robot may combine a grid-based metric map with a modified GVD-based topological map so that local frontier detection occurs inside a sliding window while global planning handles cases where local information is unreliable (Gao et al., 2020). This suggests a contrast: DWFE uses only local frontier geometry as rasterized value prior, whereas hybrid map methods use an explicit global-local hierarchy.

In submap-based SLAM systems, frontier maintenance itself can be the main systems bottleneck. A local frontier point may be defined as the center of an unobserved occupancy-grid cell adjacent to an unoccupied cell within a submap, while a global frontier point must remain unobserved with respect to all overlapping submaps (Oršulić et al., 2019). That formulation is relevant because DWFE presupposes that frontier points are already available at high frequency; it does not contribute a frontier detector, a loop-closure-robust frontier maintenance scheme, or a global frontier validation mechanism.

In 3D exploration, several planners already combine frontier gain with distance penalties. One multi-resolution OctoMap-based planner scores frontier candidates by

$\mathcal {F}=\{f_0,...,f_j \}$4

where $\mathcal {F}=\{f_0,...,f_j \}$5 is information gain and $\mathcal {F}=\{f_0,...,f_j \}$6 is Euclidean distance, and selects

$\mathcal {F}=\{f_0,...,f_j \}$7

after frontier reduction and mean-shift clustering (Batinović et al., 2020). Another 3D planner uses

$\mathcal {F}=\{f_0,...,f_j \}$8

with roadmap shortest-path length $\mathcal {F}=\{f_0,...,f_j \}$9 rather than raw Euclidean distance (Zhang et al., 28 Feb 2025). Relative to those formulations, DWFE is a narrower mechanism: it injects a frontier-distance prior into an existing action-value grid, rather than performing explicit global optimization over gain, path cost, or clustered candidate viewpoints.

4. Empirical evaluation and reported performance

DWFE was evaluated in Habitat 3.0 on the HM3D-v1 validation split over 2,000 episodes, using RGB and depth at $f^* \in \mathcal{F}$0, depth noise disabled, default physics and navigation mesh, a 500-step episode cap, and a success criterion requiring that the goal be visible and within $f^* \in \mathcal{F}$1 m when stop is issued (Aghaei et al., 26 Jul 2025). The reported metrics are Success and SPL on the full evaluation, with Average Steps additionally reported in the smaller SHF subset study.

On the 2,000-episode validation split, DWFE achieves 61.1% Success and 36.0% SPL, compared with InstructNav’s reported 58.0% Success and 20.9% SPL. The comparison is therefore +3.1 points in Success and +15.1 points in SPL, which the paper describes as a 72% relative increase in SPL (Aghaei et al., 26 Jul 2025).

The paper also reports a comparison against listed training-free baselines:

Method	Success	SPL
ZSON	25.5	12.6
ESC	39.2	22.3
VoroNav	42.0	26.0
L3MVN	50.4	23.1
VLFM	52.5	30.4
InstructNav	58.0	20.9
DWFE	61.1	36.0

The especially large SPL improvement is central to the interpretation offered in the paper: DWFE not only succeeds more often, but reaches goals more efficiently. Qualitatively, the trajectory comparison reports that InstructNav “wanders, makes many detours, and times-out at the 500-step cap,” whereas DWFE “explores a few islands and then reaches the goal” (Aghaei et al., 26 Jul 2025).

In the 200-episode subset used for the SHF add-on, DWFE obtains 57.5% Success, 34.1% SPL, and 284.6 average steps, while SHF improves this to 59.5% Success, 35.0% SPL, and 279.8 steps. This suggests that the incremental language contribution reported there is +2.0 Success, +0.9 SPL, and about 5 fewer steps on average, whereas the primary performance shift is already present in geometry-only DWFE (Aghaei et al., 26 Jul 2025).

5. Interpretation, significance, and controversy

The principal interpretation attached to DWFE is polemical only in the limited sense that it was introduced to test whether reported gains in instruction-guided object navigation were actually due to LLM reasoning. The paper explicitly states that it strips InstructNav of its Dynamic Chain-of-Navigation prompt, GLEE detector, and Intuition saliency map, and replaces them with DWFE, then argues that frontier geometry rather than emergent LLM reasoning drives most reported gains (Aghaei et al., 26 Jul 2025).

The qualitative rationale supplied is straightforward: nearby frontier islands tend to continue the current route into adjacent unseen space, while distant islands often require backtracking or long detours. Since the original prompt-based selector does not include coordinates or metric relations, GPT-4 cannot exploit this structure. DWFE can, despite being much simpler (Aghaei et al., 26 Jul 2025).

This interpretation intersects with a longstanding concern in frontier exploration: Euclidean proximity is often an imperfect proxy for true travel cost. In office-like 2D exploration, existing methods were criticized for low exploration efficiency in complex environments “due to the lack of a systematical way to determine and assign optimal exploration command,” and specifically for frontier selection by Euclidean distance instead of actual travel distance (Gao et al., 2020). A plausible implication is that DWFE’s success should not be read as establishing the optimality of Euclidean frontier weighting; rather, it demonstrates that even a simple geometric prior can outperform a semantically elaborate selector when that selector lacks metric grounding.

The controversy, then, is not whether language can ever help, but where the burden of explanation lies. The same paper notes that SHF helps most in kitchens and bathrooms where object co-occurrence priors are informative, and helps little in long corridors or sparse lofts where semantics provide weak adjacency cues (Aghaei et al., 26 Jul 2025). This suggests that DWFE is best interpreted as an argument for geometry-first baselines and metric-aware evaluation, not as a general denial of semantic utility.

6. Limitations, underspecification, and subsequent directions

DWFE is deliberately simple, and several limitations are explicit. It uses Euclidean distance from each navigable cell to the nearest point on a frontier island; it does not state that geodesic distance, planner cost, travel time, or obstacle-aware path length are used in scoring (Aghaei et al., 26 Jul 2025). It also leaves unspecified the exact multi-island score aggregation mechanism, the occupancy-map resolution, map size, inflation radius, exact frontier detection implementation details beyond free/unknown adjacency, and exact AVM initialization (Aghaei et al., 26 Jul 2025).

Because of those omissions, DWFE should not be conflated with broader path-aware or information-aware frontier selection. In 3D occupancy exploration, for example, a frontier-based automatic-differentiable information gain framework defines a sampled viewpoint score

$f^* \in \mathcal{F}$2

and a path-level objective

$f^* \in \mathcal{F}$3

thereby combining frontier utility, safety, and path length in a differentiable optimization framework (Deng et al., 2020). Compared with such methods, DWFE is only a frontier-distance prior written into an action-value raster.

Other extensions in the frontier literature emphasize dynamic obstacles, bounded frontier-processing complexity, or better global travel models. Dynamic-aware exploration distinguishes ordinary frontiers from dynamic-aware frontiers in populated environments and defers those associated with temporary blockage (Cavinato et al., 2021). OctoMap-based 3D exploration with approximate frontier detection and Bayesian information gain claims frontier-processing complexity $f^* \in \mathcal{F}$4 and reports up to a 54.4% exploration-time improvement over a standard deterministic frontier baseline in one forest setting (Lewis et al., 3 Apr 2026). Roadmap-based 3D planners evaluate frontier-associated candidates with path-length-aware utility $f^* \in \mathcal{F}$5 rather than raw Euclidean proximity (Zhang et al., 28 Feb 2025). These comparisons indicate that DWFE occupies the minimalist end of the frontier-exploration spectrum.

The future directions explicitly suggested in the DWFE paper are “metric-aware prompts” and “offline semantic graphs” (Aghaei et al., 26 Jul 2025). This suggests that, if language is to be credited for navigation gains, it should receive structured spatial information rather than vague object lists. A plausible implication is that the enduring contribution of DWFE is not a new exploration theory, but a sharp baseline lesson: before attributing performance to semantic reasoning, one must first control for straightforward frontier geometry.