Wanderland: Grounded Exploration in Embodied AI

• Wanderland is a framework and research idiom that blends photorealistic urban simulation with open-ended, contextually constrained exploration in embodied AI.
• It employs a real-to-sim pipeline integrating 3D LiDAR, IMU, RTK-GNSS, and fisheye imaging to create globally consistent, geometrically accurate virtual environments.
• The framework underpins diverse applications—from few-shot learning benchmarks to assistive navigation and geolocation systems—offering practical insights into closed-loop evaluation and policy learning.

Wanderland denotes, most concretely, a real-to-sim framework and dataset for photorealistic, geometrically grounded simulation of large open-world urban environments for embodied AI (Liu et al., 25 Nov 2025). In adjacent literatures, the same label, or a closely related “wandering within a world” formulation, is used for online contextualized few-shot learning (Ren et al., 2020), walkability-aware urban discovery (Amendola et al., 4 Dec 2025), actionable geolocation in navigable panorama graphs (Zheng et al., 11 Mar 2026), and map-less robotic exploration support for blind users (Kuribayashi et al., 13 Feb 2025). This suggests that Wanderland functions both as a named system and as a broader research idiom for grounded exploration: movement is open-ended, but it is never unstructured, because geometry, context, narrative, or environmental feedback continuously constrain what counts as valid wandering.

1. Terminological scope and research uses

Across the cited literature, “Wanderland” is not a single uniformly defined method. It appears as an explicit framework in embodied AI, as a benchmark-oriented learning setting in few-shot recognition, and as an interpretive label for systems that support exploration, discovery, or wandering-like behavior.

Research context	Status of “Wanderland”	Core object
Embodied AI simulation (Liu et al., 25 Nov 2025)	Explicit framework and dataset	Real-to-sim, photorealistic, geometrically grounded urban simulation
Online contextualized few-shot learning (Ren et al., 2020)	More naturalistic benchmark and learning setting	Agent wandering within a world
Urban discovery, geolocation, accessibility (Amendola et al., 4 Dec 2025, Zheng et al., 11 Mar 2026, Kuribayashi et al., 13 Feb 2025)	Exploration-centered usage	Walkability-aware itineraries, actionable geolocation, map-less wandering
Generative, behavioral, and mathematical wandering (Inch et al., 1 Nov 2025, Utley et al., 23 Jun 2026, Dimitrov, 2024)	Adjacent or analogical usage	Novelty search, movement continua, wanderer line ensembles

The most stable common denominator is not nomenclature but problem structure. Wanderland-type research treats exploration as a sequential process in which the system must update beliefs, retrieve grounded information, or navigate a structured space without collapsing everything into one-shot prediction. In some papers, the term itself is absent and only the underlying logic of wandering is present. In others, especially the embodied-AI framework, Wanderland is the formal name of the artifact.

2. Wanderland as a geometrically grounded embodied-AI framework

In its primary technical sense, Wanderland is a real-to-sim pipeline designed for reproducible closed-loop evaluation in embodied AI, especially navigation in mixed indoor-outdoor urban environments (Liu et al., 25 Nov 2025). Its premise is that open-world embodied benchmarking requires two properties simultaneously: photorealistic sensor rendering and reliable metric geometry. The framework was introduced in response to the claim that classic RGB-D mesh pipelines scale poorly to open urban spaces, while recent video-3DGS methods remain unsuitable for benchmarking because of large visual and geometric sim-to-real gaps.

The capture stack is built around MetaCam Air, a handheld commercial 3D scanner comprising a Livox Mid-360 non-repetitive LiDAR, a built-in IMU, an RTK-GNSS antenna, and two synchronized 4K fisheye cameras with over 180° field of view (Liu et al., 25 Nov 2025). Scenes are selected to be roughly 5,000–10,000 m² each. RGB images are not collected at a fixed frame rate; capture is triggered when the device moves by a fixed distance or rotates by a fixed angle, which produces more uniformly distributed views. Data collection spans New York City and Jersey City and includes residential buildings, business districts, public streets, plazas, and university campuses under varied time-of-day and weather conditions.

Reconstruction is performed by MetaCam Studio, which uses LiDAR-inertial-visual-GNSS fusion to produce a globally consistent dense metric point cloud and accurate camera poses in a shared world coordinate frame (Liu et al., 25 Nov 2025). This coordinate frame is reused for 3D Gaussian Splatting, mesh extraction, and simulator integration, so rendering and interaction operate in the same metric system. The paper presents this shared metric frame as the key mechanism by which physics, collision checking, shortest-path planning, and photorealistic rendering remain mutually consistent.

The image-processing stage applies a two-stage masking procedure: Egoblur masks faces and license plates, and an object detector masks dynamic objects such as people, animals, and vehicles (Liu et al., 25 Nov 2025). Raw fisheye images are cropped to 120° and undistorted into perspective views for 3DGS training. The neural rendering component is built on gsplat, initialized from the LiDAR point cloud. Depth supervision is derived from the initialized Gaussians projected into each camera, rather than from monocular depth prediction. For interaction, geometry is not extracted from 3DGS opacity fields; instead, the point cloud is filtered, voxelized, converted to a triangle mesh with Marching Cubes, and simplified into a collision layer.

The final simulator asset combines a collision mesh and a 3DGS renderer inside a single USD scene loaded into Isaac Sim (Liu et al., 25 Nov 2025). The collision mesh supplies traversability and physics, while 3DGS supplies the primary photorealistic observations. In the paper’s formulation, this division is what makes the framework “geometrically grounded.”

3. Dataset, benchmark structure, and empirical findings

The accompanying Wanderland dataset currently contains 530 distinct scenes, 420,000+ frames, 100+ hours of recording, and 3.8+ million square meters of total area, with a stated expansion target of 1,000+ scenes (Liu et al., 25 Nov 2025). Each scene includes synchronized RGB fisheye images, intrinsic calibrations, globally consistent camera poses, colorized metric point clouds, optimized 3D Gaussian Splatting models, extracted collision meshes, and USD simulator scenes. The raw point cloud spacing is reported as 5–10 mm, with roughly 10–50 million points per scene before downsampling to about 5 million points per scene for 3DGS initialization.

Wanderland supports point-goal navigation, image-goal navigation, and vision-language navigation (Liu et al., 25 Nov 2025). For expert trajectories, the collision mesh is imported into Unity and converted into a NavMesh; start and goal points are sampled near capture cameras, and shortest collision-free paths are generated on the navigable surface. For VLN, the expert route is replayed in simulation, rendered as egocentric video, described by Gemini 2.5 Flash, and then verified by humans. The benchmark therefore couples simulator assets with executable task definitions rather than only releasing raw reconstruction outputs.

The paper’s empirical program is organized around three claims. First, image-only 3D reconstruction is not yet as good as LIV-SLAM for geometric grounding. Even an optimistic “best of all” configuration over the compared vision-only methods yields T-ATE$^S$ = 0.30 m, R-ATE = 5.0°, T-RTE = 3.9 m, and AUC@30 = 0.83, which the authors still regard as too inaccurate for reliable closed-loop benchmarking (Liu et al., 25 Nov 2025). Second, photorealistic simulation quality improves when geometry is accurate. On interpolated views, Wanderland reports PSNR 20.37, SSIM 0.688, and LPIPS 0.327, outperforming the listed baselines; on extrapolated views, it reports PSNR 17.92, SSIM 0.591, and LPIPS 0.445 (Liu et al., 25 Nov 2025). Third, policy learning is sensitive to simulator fidelity. RL post-training inside Vid2Sim-built environments generally hurts or fails to help, whereas RL post-training inside Wanderland improves the tested policies’ success rate or intervention rate. The paper states that no method exceeds 50% success rate, and that outdoor navigation remains significantly harder than indoor navigation (Liu et al., 25 Nov 2025).

The navigation metric emphasized by the paper is SPL:

$$\text{SPL} = \frac{1}{N} \sum_{i=1}^N S_i \frac{l_i}{\max(p_i, l_i)},$$

where $S_i$ is the success indicator, $l_i$ is the optimal path length, and $p_i$ is the actual path length (Liu et al., 25 Nov 2025). The framework’s broader claim is that such closed-loop evaluation is only trustworthy when geometry and rendering are jointly grounded.

4. Wanderland as online contextualized few-shot learning

A second important usage presents Wanderland as a more naturalistic benchmark and learning setting for few-shot recognition, instantiated as online contextualized few-shot learning (OC-FSL) and the RoamingRooms dataset (Ren et al., 2020). The central departure from standard episodic few-shot learning is the removal of the internal support/query split. Prediction is formulated as

$$\hat{y}_t = f(x_t; (x_1,\tilde{y}_1), \dots, (x_{t-1},\tilde{y}_{t-1})), \qquad t=1,\dots,T,$$

where $\tilde y_t$ is either the true label $y_t$ or $-1$ if the item is unlabeled (Ren et al., 2020). Every time step is simultaneously a learning step and an evaluation step.

The “wandering within a world” aspect is literal in the data-generation process. RoamingRooms is built from Matterport3D, comprising 90 indoor worlds, split into 60 worlds for training, 10 for validation, and 20 for testing in the main text, with about 1.22M frames, 6,971 sequences, and about 7.0K unique instance classes (Ren et al., 2020). Episodes are produced by a random walking agent: MatterSim samples a path through discrete panoramic viewpoints, and HabitatSim renders aligned RGB frames and instance-segmentation maps. Each input highlights one target object instance inside a broader room-level visual context, so the learner is exposed to both object identity and environmental co-occurrence structure.

The sequential statistics are designed to make context materially useful. The paper reports that 30% of the time the same instance appears in the next frame, and that a 100-image sequence has about 3 different viewpoints on average (Ren et al., 2020). These properties induce soft task boundaries and recurring latent environments rather than i.i.d. examples. The proposed Contextual Prototypical Memory (CPM) combines a ResNet-12-based 512-dimensional visual feature, a 256-dimensional LSTM, context-modulated metric scaling, and prototype memory. On RoamingRooms, CPM attains AP 89.14 in the supervised setting and AP 84.12 in the semi-supervised setting, outperforming the listed online baselines on the main AP criterion (Ren et al., 2020).

In this literature, Wanderland names a benchmark philosophy more than a single platform. The environment is not merely a source of nuisance variation; it is a latent contextual process that determines what classes are likely to recur and when prior knowledge should be retrieved.

5. Urban discovery, geolocation, accessibility, and travel planning

In applied exploration systems, Wanderland denotes or implies a family of architectures that treat movement through urban or indoor space as a dialogue between spatial grounding and user intent. WalkRAG is a spatial RAG framework with a conversational interface for recommending walkable itineraries and supporting follow-up questions about the route and points of interest (Amendola et al., 4 Dec 2025). It is organized around QUAG (Query Understanding and Answer Generation), a Spatial component, and an Information Retrieval component. The walkability model uses four indicators—sidewalk/pedestrian footway availability, air pollution levels, presence of green areas, and accessibility for individuals with disabilities—and scores routes by

$WS = \frac{\sum_i w_i \, c_i}{\tau},$

with $$\text{SPL} = \frac{1}{N} \sum_{i=1}^N S_i \frac{l_i}{\max(p_i, l_i)},$$0 and default weights $$\text{SPL} = \frac{1}{N} \sum_{i=1}^N S_i \frac{l_i}{\max(p_i, l_i)},$$1 when no preferences are supplied (Amendola et al., 4 Dec 2025). The system correctly classified all 40 queries in the evaluation and improved route-grounded answer quality relative to a closed-book baseline.

WanderBench and GeoAoT extend this exploration logic to geolocation (Zheng et al., 11 Mar 2026). WanderBench is presented as the first open-access global geolocation benchmark designed for actionable geolocation reasoning in embodied scenarios, containing 1,047 unique locations, 32,741 panoramas, and 39,442 edges across six continents. Environments are panorama graphs satisfying

$$\text{SPL} = \frac{1}{N} \sum_{i=1}^N S_i \frac{l_i}{\max(p_i, l_i)},$$2

so the agent can explore at least ten steps from any starting node before reaching a boundary (Zheng et al., 11 Mar 2026). GeoAoT, operating purely at inference time without additional training, replaces textual chain-of-thought with action-conditioned reasoning: instead of only verbalizing uncertainty, the model can rotate, move, and gather disambiguating evidence. The paper states that GeoAoT improves every tested model.

For assistive robotics, WanderGuide addresses indoor exploration by blind users without requiring a prebuilt map (Kuribayashi et al., 13 Feb 2025). The system emerged from a formative study with 10 blind participants and was later evaluated with 5 blind participants. It provides three levels of detail for environmental descriptions, supports question answering, and implements a “Take-Me-There” revisitation function. Although described as map-less, it performs online mapping with Cartographer and detects candidate waypoints by skeletonizing the cost map, applying a kernel-based corner detection algorithm, clustering with DBSCAN, and falling back to points 3 meters in front, behind, left, and right when intersections are absent (Kuribayashi et al., 13 Feb 2025). The paper reports that the system could provide blind people with the enjoyable experience of wandering around without a specific destination in their minds.

Adjacent travel-planning work makes the same shift from one-shot recommendation to exploratory orchestration. Vaiage is a graph-structured multi-agent framework built around TravelGraph, with Chat, Information, Recommendation, Route, Strategy, and Communication agents, and reports an average score of 8.5/10, compared with 7.2 for a no-strategy variant and 6.8 for a no-external-API variant (Liu et al., 16 May 2025). NarrativeGuide models narrative-driven travel planning as an optimization problem over itinerary smoothness, travel time, and attraction popularity:

$$\text{SPL} = \frac{1}{N} \sum_{i=1}^N S_i \frac{l_i}{\max(p_i, l_i)},$$3

thereby making storytelling part of the planning objective itself (Zhang et al., 20 Feb 2025).

6. Generative, behavioral, and resilient interpretations of wandering

In generative AI, the term is closely associated with open-ended search rather than optimization toward a single best output. WANDER, introduced in “Evolve to Inspire: Novelty Search for Diverse Image Generation,” is a black-box evolutionary framework for producing diverse sets of images from a single text prompt (Inch et al., 1 Nov 2025). The paper states that “Wanderland” does not name a separate method there; the actual system name is Wander. The method mutates or crossovers prompts with an LLM, evaluates resulting images in CLIP embedding space, and uses human-written emitters such as “Completely change the composition” or “Suggest a novel color scheme.” Novelty is measured by average cosine distance to $$\text{SPL} = \frac{1}{N} \sum_{i=1}^N S_i \frac{l_i}{\max(p_i, l_i)},$$4-nearest neighbors in the current pool:

$$\text{SPL} = \frac{1}{N} \sum_{i=1}^N S_i \frac{l_i}{\max(p_i, l_i)},$$5

On the main benchmark, Wander achieves Vendi $$\text{SPL} = \frac{1}{N} \sum_{i=1}^N S_i \frac{l_i}{\max(p_i, l_i)},$$6 and LPIPS $$\text{SPL} = \frac{1}{N} \sum_{i=1}^N S_i \frac{l_i}{\max(p_i, l_i)},$$7, compared with Lluminate at Vendi $$\text{SPL} = \frac{1}{N} \sum_{i=1}^N S_i \frac{l_i}{\max(p_i, l_i)},$$8 and LPIPS $$\text{SPL} = \frac{1}{N} \sum_{i=1}^N S_i \frac{l_i}{\max(p_i, l_i)},$$9, while using 24,347 \pm 649 tokens versus 175,902 \pm 9,390 (Inch et al., 1 Nov 2025). In this context, Wanderland is best understood as a prompt-and-image space to be traversed rather than solved.

In field robotics, wandering becomes a resilient navigation primitive. Blind-Wayfarer is a probing-driven navigation framework for perception-degraded environments that relies primarily on a compass, an IMU, and anomaly detection rather than vision, LiDAR, GPS, or precise localization (2503.07492). Its Pledge-inspired recovery rule maintains a cumulative turn counter $S_i$0, turning angle $S_i$1, and turning direction $S_i$2, so that obstacle encounters trigger systematic reverse-turn-probe cycles rather than random dithering. The paper reports a 99.7% success rate in 1,000 simulated forest experiments, all 20 trials successful in real-world tests, and a long-distance escape from 45 m inside a dense woodland to a paved pathway (2503.07492).

In movement analytics, wandering is a behavioral continuum rather than a crisp class. The theme-park study “Do Waders, Swimmers, and Divers Exist?” clusters visitors by eight interpretable GPS-derived features and concludes that behavioral groups recur reliably but without sharp boundaries (Utley et al., 23 Jun 2026). At Knott’s Berry Farm, the recovered groups have sizes waders $S_i$3, swimmers $S_i$4, and divers $S_i$5, but no site reaches silhouette $S_i$6, and the authors argue for a continuum rather than discrete natural kinds. Self-report is a weak proxy for observed behavior, and the sign of some movement features reverses across parks, so behavioral parameters cannot be transferred unchanged between sites (Utley et al., 23 Jun 2026). This use of wandering is descriptive rather than prescriptive: it identifies a style of movement whose empirical signature depends on destination choice and sequencing more than on locomotion mechanics alone.

7. Theoretical and mathematical formulations of wandering

At the theoretical end of the spectrum, wandering is formalized either as motivational arbitration or as a structural deformation of stochastic line ensembles. Wanderer, in “Should I Stay or Should I Go,” is a computational model in which exploration is determined by two competing subsystems: an excitatory subsystem representing the need to acquire energy and an inhibitory subsystem representing the need to avoid predators (Gabora, 2013). Exploration is a continuous output,

$S_i$7

with food reducing exploratory drive via satiation and predator encounters both suppressing exploration immediately and increasing future cautiousness (Gabora, 2013). Neutral cues become predictive through a simple associative rule,

$S_i$8

so wandering is shaped by both immediate stimuli and learned environmental regularities.

A mathematically distinct usage appears in the theory of Airy wanderer line ensembles (Dimitrov, 2024). These are infinite-parameter generalizations of the classical Airy line ensemble, defined via a determinantal kernel built from

$S_i$9

and, in a substantial subregime, lifted to continuous Brownian Gibbsian line ensembles (Dimitrov, 2024). The point is that, after parabolic centering, selected top curves acquire nonzero linear asymptotic slopes rather than remaining of order one. The later paper on their structural properties proves continuous dependence on the parameters, multiple monotone couplings, and extremality in the space of Brownian Gibbsian line ensembles (Dimitrov, 23 Dec 2025). The asymptotic slope theorem identifies the right- and left-side drifts through the parameter sequences $l_i$0 and $l_i$1, making the “wanderer” terminology precise at the level of limiting curve geometry.

Taken together, these theoretical constructions show that wandering is not restricted to colloquial exploration. It can be written as a logistic motivational controller, a determinantal process, a Brownian Gibbs state, or an asymptotic slope deformation, depending on the domain.

Wanderland therefore occupies a distinctive place across contemporary research. In its narrowest sense, it is a geometrically grounded, photorealistic simulation framework for open-world embodied AI (Liu et al., 25 Nov 2025). In its broader sense, it names a recurring technical agenda in which exploration is sequential, context-dependent, and environment-grounded: few-shot learners acquire labels while moving through latent worlds (Ren et al., 2020); route planners couple walking, retrieval, and dialogue (Amendola et al., 4 Dec 2025); multimodal models geolocate by acting in panorama graphs (Zheng et al., 11 Mar 2026); blind users wander through indoor spaces with robotic guidance (Kuribayashi et al., 13 Feb 2025); generative systems search prompt and image space for novelty (Inch et al., 1 Nov 2025); and even visitor movement or stochastic line ensembles are analyzed through wandering-based structure (Utley et al., 23 Jun 2026, Dimitrov, 23 Dec 2025). The unifying theme is not aimlessness, but constrained open-endedness: the system wanders, yet it remains accountable to geometry, history, motivation, cultural structure, or measurable behavior.