MapEx: Probabilistic Indoor Exploration

• MapEx is a robotic exploration framework that leverages global map predictions and probabilistic sensor models to navigate and map structured indoor environments.
• It utilizes an ensemble of inpainting networks to estimate occupancy and variance, enabling robust viewpoint selection through joint visibility and uncertainty reasoning.
• Experimental results demonstrate significant gains in coverage, occupancy accuracy, and topological understanding compared to traditional frontier and recent map-prediction methods.

Searching arXiv for the specified MAPEX-related paper and closely related exploration baselines to ground the article in current literature. arxiv_search(query="(Ho et al., 2024) OR \"MapEx: Indoor Structure Exploration with Probabilistic Information Gain from Global Map Predictions\" OR indoor exploration predicted map information gain", max_results=10, sort_by="relevance") MapEx is an exploration framework for robots operating in unknown structured indoor environments, introduced for the setting of a mobile robot equipped with a $2\,\mathrm{D}$ LiDAR that must build an occupancy map within a fixed time or budget $T$. Its central idea is to exploit the predictability of indoor layouts—such as repeated rooms and corridors—without relying on a single deterministic map completion. Instead, MapEx generates multiple global map predictions from partial observations, estimates their variance, and combines that uncertainty with a probabilistic estimate of what the sensor can actually observe from a candidate viewpoint. The resulting information-gain objective is therefore neither coverage-only nor uncertainty-only, but a joint criterion over visibility and predictive variance (Ho et al., 2024).

1. Exploration setting and conceptual contribution

The exploration problem addressed by MapEx assumes an initially unknown indoor environment represented as an occupancy grid. At timestep $t$, the robot maintains an observed map

$O_t \in \{0, 0.5, 1\}^{H \times W},$

where $0$ denotes free space, $0.5$ unknown space, and $1$ occupied space. The environment class of interest is explicitly structured indoor space, in which partial observations often suffice to infer plausible global floorplan structure (Ho et al., 2024).

MapEx is positioned against two established families of methods. Classical frontier-based methods extract the boundary between known and unknown cells and select the next frontier by a heuristic such as “closest first” or distance-weighted gain. In the characterization used by the paper, these methods do not leverage global layout predictability. Recent map-prediction-based methods divide into two contrasting groups. Coverage-only methods, exemplified in the paper by IG-Hector 2019, use a learned inpainting network to complete a local or global map and then score viewpoints by predicted visible area, but ignore prediction uncertainty. Uncertainty-only methods, exemplified by UPEN 2022, use an ensemble of predicted maps and score frontiers by prediction variance, but do not reason over actual sensor visibility. MapEx’s stated novelty is to jointly reason over where the robot’s sensor can see and how uncertain the predicted map is at those locations (Ho et al., 2024).

This design directly addresses two failure modes identified in prior approaches. Coverage-only scoring can be misled by incorrect inpainted walls, while uncertainty-only scoring can prioritize regions that are uncertain yet physically unobservable. MapEx therefore formulates exploration as viewpoint selection under a probabilistic sensor model derived from global map predictions rather than as pure frontier proximity or pure predictive disagreement.

2. Global map prediction and variance estimation

MapEx performs global inpainting on the robot’s current observed map using a fully convolutional inpainting network $G$ based on LaMa 2022. Treating $O_t$ as a single-channel image, the network outputs

$P_t = G(O_t) \in [0,1]^{H \times W},$

where each pixel $T$0 is an estimated occupancy probability. The network is fine-tuned on the KTH floorplan dataset, with $T$1 used for training and $T$2 held out, so that it learns typical indoor layouts including walls, rooms, and corridors (Ho et al., 2024).

To quantify uncertainty, MapEx maintains $T$3 independent LaMa models $T$4, each fine-tuned on a different subset of the training data. Each model produces a predicted map

$T$5

From this ensemble, the framework computes pixel-wise empirical mean and variance: $T$6

$T$7

In practice, the reported system uses $T$8, which the paper describes as a fast but effective uncertainty estimate (Ho et al., 2024).

The mean map $T$9 and variance map $t$0 serve distinct roles. The former is used to reason about likely geometry and line-of-sight, while the latter is used to quantify where additional observations would reduce predictive uncertainty. This separation is a defining feature of the framework: the predicted map is not used only as a surrogate occupancy map, but as a source of both geometric expectation and uncertainty structure.

3. Probabilistic sensor model and visible-area estimation

The sensor model in MapEx assumes a $t$1 LiDAR at pose $t$2, maximum range $t$3, and full $t$4 coverage. For any candidate viewpoint $t$5, the method computes a binary visibility mask $t$6. Rather than performing a hard raycast on the observed map $t$7 or directly on the mean predicted map $t$8, MapEx uses a probabilistic raycast on $t$9 (Ho et al., 2024).

From $O_t \in \{0, 0.5, 1\}^{H \times W},$0, the algorithm casts $O_t \in \{0, 0.5, 1\}^{H \times W},$1 uniform rays $O_t \in \{0, 0.5, 1\}^{H \times W},$2, each parameterized by a length $O_t \in \{0, 0.5, 1\}^{H \times W},$3. For each ray, it initializes an accumulated occupancy value $O_t \in \{0, 0.5, 1\}^{H \times W},$4. As the ray steps through discrete cells $O_t \in \{0, 0.5, 1\}^{H \times W},$5, it updates

$O_t \in \{0, 0.5, 1\}^{H \times W},$6

If $O_t \in \{0, 0.5, 1\}^{H \times W},$7 exceeds a threshold $O_t \in \{0, 0.5, 1\}^{H \times W},$8, empirically set to $O_t \in \{0, 0.5, 1\}^{H \times W},$9, the ray terminates; otherwise it continues to range $0$0. The endpoints of all rays define a boundary used for a flood-fill in $0$1, yielding a continuous region of “potential visibility.” Cells already observed in $0$2 are then masked out to obtain

$0$3

The resulting mask encodes the probabilistic set of unobserved cells that the LiDAR could see from $0$4, given uncertainty in wall placement. This formulation is central to the method’s interpretation of information gain. A candidate viewpoint is valuable not merely because it lies near unknown space, but because it can expose unknown cells under a visibility model that accounts for uncertain geometry.

4. Information gain and the planning loop

Given the variance map $0$5 and visibility mask $0$6, MapEx defines the information gain of viewpoint $0$7 as

$0$8

In an equivalent cell-indexed notation, if $0$9 is the predicted occupancy at cell $0.5$0, $0.5$1 its variance, and $0.5$2 the visible cell set, then

$0.5$3

with MapEx using $0.5$4. The paper notes that one could add an entropy term

$0.5$5

but reports that the variance sum is sufficient and simpler (Ho et al., 2024).

The planning loop begins with LiDAR acquisition and map update. If there is no current goal, the framework predicts the ensemble $0.5$6, computes $0.5$7, extracts candidate frontiers $0.5$8 from $0.5$9, and evaluates each frontier $1$0 by first computing $1$1, then $1$2, and finally the score

$1$3

which penalizes distant goals. The selected goal is

$1$4

after which the robot uses A$1$5 to plan a collision-free path $1$6 to $1$7. The loop then follows one step of $1$8, updates the pose to $1$9, and repeats until the time budget is exhausted. The output is the final observed map $G$0 and predicted map $G$1 (Ho et al., 2024).

This suggests that MapEx remains compatible with conventional frontier extraction and graph search, but replaces heuristic frontier valuation with a viewpoint criterion grounded in predicted global structure, uncertainty, and sensor reachability.

5. Experimental protocol and quantitative performance

The reported experiments use the KTH floorplan dataset, consisting of 149 real indoor floorplans cleaned and rasterized at $G$2. The split is by building: $G$3 train, approximately 2,367 maps, and $G$4 test, comprising 10 floorplans. The robot is initialized at each of four corners, yielding $G$5 trials, each run for 1,000 time steps. The LiDAR range is $G$6, and each scan uses 2,500 rays (Ho et al., 2024).

Three baselines are reported: Nearest-Frontier, described as classical and distance-only; IG-Hector 2019, described as coverage-only from a single predicted map; and UPEN 2022, described as variance-only from an ensemble without coverage reasoning. Evaluation uses three metrics. Coverage$G$7 is the percentage of non-unknown cells in $G$8. $G$9 is the intersection-over-union between predicted $O_t$0 and the ground-truth map, measured on the occupied class inside the building footprint. Topological Understanding (TU) is computed by planning 100 random A$O_t$1 paths from $O_t$2; TU is the percentage of successful paths that reach goal without collision (Ho et al., 2024).

The main results are reported as area under the curve over $O_t$3:

Metric	vs. Nearest	vs. UPEN	vs. IG-Hector
Coverage	+25.4%	+16.2%	+15.5%
$O_t$4	+22.9%	+9.9%	+12.3%
TU	+20.1%	+9.6%	+11.1%

The abstract additionally summarizes the outcome as an average $O_t$5 improvement over a representative map-prediction-based exploration method and a $O_t$6 improvement over the nearest frontier approach (Ho et al., 2024).

The evaluation is notable for going beyond raw area coverage. By reporting $O_t$7 and TU, the study treats exploration as both map completion and structural understanding. This is consistent with the framework’s use of global predictions rather than only local frontier expansion.

6. Ablations, limitations, and extension paths

The ablation study isolates the contribution of each major component of the information-gain formulation. Removing visibility reasoning and summing $O_t$8 within $O_t$9 yields poor coverage and IoU, with a reported $P_t = G(O_t) \in [0,1]^{H \times W},$0 change. Replacing the probabilistic raycast with a deterministic raycast using hard walls on $P_t = G(O_t) \in [0,1]^{H \times W},$1 reduces coverage by $P_t = G(O_t) \in [0,1]^{H \times W},$2. Removing variance and summing only the number of cells in the visibility mask reduces coverage by $P_t = G(O_t) \in [0,1]^{H \times W},$3. Using only the observed map $P_t = G(O_t) \in [0,1]^{H \times W},$4 for casting and summing visible-cell count reduces coverage by $P_t = G(O_t) \in [0,1]^{H \times W},$5. The paper concludes from these ablations that both probabilistic visibility and variance-weighted gain are critical (Ho et al., 2024).

The paper also states three key insights. First, jointly reasoning over where the robot could see and how uncertain the prediction is yields more robust and efficient exploration than either criterion alone. Second, global inpainting allows planning long-horizon motions over the entire building rather than only local next-step heuristics. Third, probabilistic raycasting prevents overconfident coverage estimates in areas where predicted walls may block the view (Ho et al., 2024).

Several limitations are explicitly identified. Performance depends on the quality of the inpainting network $P_t = G(O_t) \in [0,1]^{H \times W},$6, so very novel layouts may be mispredicted and can penalize performance. The method incurs computational overhead because it runs an ensemble of inpainting networks and probabilistic raycasts during replanning. It also assumes perfect robot localization and a static environment. The extension paths proposed in the paper include multi-robot MapEx with shared predictions and variances, incorporation of semantic predictions such as room labels for task-driven exploration, adaptive ensemble size or online Bayesian updating to reduce compute, and integration with learning-based planners such as RL using $P_t = G(O_t) \in [0,1]^{H \times W},$7 as reward (Ho et al., 2024).

7. Terminological ambiguity of “MAPEX”

The name “MAPEX” is not unique in recent arXiv literature. In autonomous driving, MapEX denotes an online HDMap estimation framework that accounts for existing maps by encoding map elements into query tokens and refining matching in DETR-style estimators (Sun et al., 2023). In multi-objective reinforcement learning, MAPEX denotes Mixed Advantage Pareto Extraction, an offline method for post hoc Pareto-front construction from single-objective specialists, critics, and replay buffers (Thakar et al., 3 Mar 2026). In remote sensing, MAPEX denotes a mixture-of-modality expert foundation model with modality-aware pruning for downstream tasks (Hanna et al., 10 Jul 2025). In natural language processing, MAPEX denotes a multi-agent pipeline for keyphrase extraction with expert recruitment, candidate extraction, topic guidance, knowledge augmentation, and post-processing (Zhang et al., 23 Sep 2025).

Within robotics exploration, however, “MapEx” specifically refers to the framework for indoor structure exploration with probabilistic information gain from global map predictions (Ho et al., 2024). The distinction matters because the acronym alone is insufficient to identify a unique method across fields.