ReasonNavi: Zero-Shot Navigation Framework

• ReasonNavi is a zero-shot embodied navigation framework that leverages multimodal language models and deterministic planners to convert top-down maps into discrete reasoning spaces.
• It explicitly separates global semantic reasoning from local motion control by segmenting maps into rooms and nodes, enabling efficient, interpretable navigation decisions.
• The framework achieves competitive performance on benchmarks like HM3D for ObjectNav, ImageNav, and TextNav without any fine-tuning of its large language models.

ReasonNavi is a zero-shot embodied navigation framework that operationalizes a human-inspired “reason-then-act” paradigm by coupling multimodal LLMs with deterministic planners. Rather than asking an MLLM to emit continuous coordinates or low-level control, it converts a global top-down 2D map into a discrete reasoning space, uses the MLLM to select a plausible room and waypoint for an object-, image-, or text-specified goal, and then executes motion through classical planning on an online occupancy map. The result is a unified navigation system for ObjectNav, instance ImageNav, and TextNav that requires no MLLM fine-tuning and explicitly separates global semantic reasoning from local motion control (Ao et al., 26 Jan 2026).

1. Problem formulation and design principle

ReasonNavi addresses goal-oriented indoor navigation in an environment $E$ where an agent starts at an arbitrary pose $p_0 \in E$, receives egocentric RGB-D observations $o_t$, has access to a global 2D map $M$, and is localized on that map at pose $p_t$. The action space is discrete,

$$\mathcal{A} = \{\texttt{move\_forward},\ \texttt{turn\_left},\ \texttt{turn\_right},\ \texttt{stop}\},$$

and an episode is successful if the agent executes stop within distance $d$ meters of the true target instance $O$ within $T$ steps (Ao et al., 26 Jan 2026).

The framework is defined over three tasks. In Object-goal Navigation, the goal $g$ is an object category $p_0 \in E$0, such as “mug” or “bed.” In Instance Image-goal Navigation, $p_0 \in E$1 is an image containing a target instance present in the scene. In Text-goal Navigation, $p_0 \in E$2 is a free-form textual description such as “the TV in the living room.” In every case, the objective is to navigate to an instance $p_0 \in E$3 of category $p_0 \in E$4 or described by $p_0 \in E$5 (Ao et al., 26 Jan 2026).

Its central design principle is an explicit separation between global semantic reasoning—“where in the building should I go?”—and local motion control—“how do I move my body safely to that place?” ReasonNavi assumes a global top-down 2D map from a CAD floorplan, a reconstruction system such as VGGT, or similar. It then turns navigation into a discrete map reasoning problem over rooms and sampled nodes, while delegating trajectory execution to deterministic planning over an online occupancy map. This structure is presented as a response to common failure modes of RL-style or reactive embodied agents: lack of global foresight, over-exploration, meandering paths, and heavy dependence on training (Ao et al., 26 Jan 2026).

2. Discrete global reasoning over maps

ReasonNavi begins by converting the global map into a discrete reasoning space. From the top-down map it derives a binary wall mask $p_0 \in E$6, applies Euclidean Distance Transform, extracts seeds by thresholding the distance field with a distance threshold $p_0 \in E$7, runs Watershed segmentation constrained by walls, and then merges tiny regions to avoid over-segmentation. The resulting room masks are

$p_0 \in E$8

with $p_0 \in E$9, and these are overlaid as labels on the map to produce an annotated map $o_t$0 for room-level reasoning (Ao et al., 26 Jan 2026).

To avoid continuous coordinate regression, the free space is discretized into candidate nodes by Poisson Disk Sampling with radius $o_t$1 over navigable cells after wall padding. This produces a global node set $o_t$2. The paper emphasizes that this sidesteps a core weakness of MLLMs: precise spatial regression. The MLLM is asked to solve only structured discrete selection problems (Ao et al., 26 Jan 2026).

Global reasoning is hierarchical and multi-stage. At the room level, the MLLM is asked which segmented room most likely contains the goal, using the segmented floor plan and the goal specification. Formally,

$o_t$3

The prompt instructs the model to look for the object directly if visible in the top-down map and otherwise infer the most plausible room type, such as mugs in kitchens or beds in bedrooms (Ao et al., 26 Jan 2026).

After selecting $o_t$4, the system filters nodes to those inside that room,

$o_t$5

creates a cropped room map $o_t$6 and a node-annotated version $o_t$7, and asks the MLLM to choose the node most consistent with the goal: $o_t$8 The chosen node is mapped to coordinates as the global target $o_t$9. The prompt directs the model to choose a node directly on the object if visible, otherwise the nearest node, otherwise a good vantage point. The paper reports that this room-then-node decomposition substantially outperforms both direct coordinate prediction and single-stage node selection (Ao et al., 26 Jan 2026).

3. Deterministic local planning and target verification

Once $M$0 is selected, the MLLM is no longer used for movement. Local navigation is formulated as a deterministic robotics problem. The agent maintains an online occupancy map $M$1, initialized with the wall mask and updated from RGB-D observations to mark free, occupied, and unknown cells. Every $M$2 steps, A* search is run on $M$3 from the current pose to $M$4, with an EDF-based cost augmentation that biases paths away from walls. A short-term waypoint $M$5 at distance $M$6 along the A* path is then passed to VFH* for local collision avoidance and steering. If $M$7 or $M$8 becomes occupied under updated mapping, it is snapped to the nearest valid free cell (Ao et al., 26 Jan 2026).

This local-control stage is explicitly modular. The global plan is fixed by the MLLM once per episode, while motion uses hybrid A* + VFH* over an evolving occupancy map. The paper argues that this circumvents the brittleness of RL-based policies, avoids per-step MLLM inference, and yields interpretable behavior because both the selected room and waypoint remain inspectable (Ao et al., 26 Jan 2026).

When the agent reaches a proximity radius $M$9 around $p_t$0, the system switches to target verification. At each step it runs a pretrained object detector on the RGB image. If an object of target category $p_t$1 is detected with high confidence, MobileSAM is used to obtain a precise segmentation mask, mask pixels are back-projected to 3D with depth, and the centroid of the resulting point cloud yields the object location $p_t$2. The agent then uses VFH* to move to $p_t$3 and issues stop. If the detector fails near the waypoint, the agent approaches a tighter threshold $p_t$4; if necessary it performs a 360° in-place scan, running the detector at each yaw. If no detection occurs after the full scan, the agent stops and the episode fails (Ao et al., 26 Jan 2026).

This separation of concerns is fundamental to the framework. Semantic reasoning is used only where it is comparatively strong, whereas coordinate execution, obstacle avoidance, and final positioning are handled by classical planning and geometry-aware vision modules (Ao et al., 26 Jan 2026).

4. Foundation models, ensemble design, and zero-shot operation

ReasonNavi uses several MLLMs as black-box vision-language reasoners for global map reasoning. The reported reasoning models are Qwen2.5-7B-VL, Seed-1.6-Thinking, and Gemini-2.5-Pro, while GPT-5 serves as a discriminator in an ensemble configuration. These models are not fine-tuned for navigation. They receive top-down map images, segmented maps, node annotations, and goal text and/or goal images, and return discrete room indices, node identifiers, or choices between candidate points (Ao et al., 26 Jan 2026).

The ensemble mechanism runs two global reasoning units, $p_t$5 and $p_t$6, each using a different base MLLM, producing two candidate waypoints $p_t$7 and $p_t$8. A discriminator MLLM then receives map crops with both candidate points marked and chooses which point is closer to the plausible target location and more reachable in navigable space. The final choice is

$p_t$9

The paper characterizes this ensemble as plug-and-play and argues that it exploits complementary strengths and biases across MLLMs (Ao et al., 26 Jan 2026).

The zero-shot claim is precise. No supervised loss over navigation trajectories is defined, and no navigation policy is trained. Object detectors, segmenters, and MLLMs are used off the shelf. All navigation-specific scoring is implicit in the MLLM’s discrete choices over rooms and nodes rather than in an end-to-end learned controller (Ao et al., 26 Jan 2026).

5. Benchmarks, quantitative results, and ablations

ReasonNavi is evaluated on HM3D benchmarks for ObjectNav, ImageNav, and TextNav, against both training-based and training-free baselines. The compared methods include OVRL-v2, OVRL-v2-IIN, IEVE, PSL, GOAT, ESC, OpenFMNav, Mod-IIN, SG-Nav, VLFM, TriHelper, and UniGoal (Ao et al., 26 Jan 2026).

The main reported results are as follows:

Task	ReasonNavi	Selected comparison
ObjectNav (HM3D)	SR 57.9, SPL 31.4	TriHelper 56.5 / 25.3; VLFM 52.5 / 30.4
ImageNav (HM3D)	SR 47.8, SPL 30.4	Mod-IIN 56.1 / 23.3; IEVE 70.2 / 25.2; UniGoal 60.2 / 23.7
TextNav	SR 38.8, SPL 24.3	UniGoal 20.2 / 11.4; GOAT 17.0 / 8.8; PSL 16.5 / 7.5

On HM3D ObjectNav, the framework achieves the highest SPL among all methods, which the paper interprets as evidence of the shortest and most efficient paths. On ImageNav, its SR is lower than specialized methods such as IEVE and Mod-IIN, but its SPL is higher than all baselines because it relies on generic verification rather than instance-specific similarity matching. On TextNav, it exceeds UniGoal, GOAT, and PSL in both SR and SPL by a large margin, which the paper attributes to strong map-aware language-conditioned reasoning (Ao et al., 26 Jan 2026).

Ablations isolate the contribution of discrete hierarchical reasoning. On HM3D ObjectNav with Seed-1.6-Thinking as the reasoning MLLM, direct coordinate prediction yields SR 12.3 and SPL 6.13; single-stage selection yields 44.5 and 23.1; multi-stage selection yields 55.1 and 29.6. The stated conclusion is that MLLMs are poor continuous spatial regressors, that discrete selection already helps, and that the proposed room-then-node decomposition is crucial (Ao et al., 26 Jan 2026).

A second ablation examines the reasoning backbone itself. Qwen2.5-7B-VL gives 41.3 / 21.2 on ObjectNav, 22.8 / 14.3 on ImageNav, and 17.0 / 10.1 on TextNav; Seed-1.6-Thinking yields 55.1 / 29.6, 40.2 / 25.2, and 30.1 / 19.0; Gemini-2.5-Pro yields 55.8 / 29.1, 40.7 / 25.6, and 37.2 / 22.8; the Seed + Gemini ensemble with discriminator reaches 57.9 / 31.4, 47.8 / 30.4, and 38.8 / 24.3. The paper describes this as a clear monotonic improvement with stronger reasoning models, with Gemini-2.5-Pro notably better on TextNav and the ensemble improving all metrics further (Ao et al., 26 Jan 2026).

The qualitative analysis also notes multi-floor navigation through chain-of-thought decomposition—such as “go to stairs on current floor → go upstairs → walk to target room”—and an advantage in multi-agent settings because local control depends on depth-based occupancy rather than dense static semantic scene graphs (Ao et al., 26 Jan 2026).

6. Limitations, interpretation, and relation to broader reasoning-centric navigation

ReasonNavi has three explicit limitations. First, it depends on a high-quality global map; although this may come from CAD drawings or tools like VGGT, map availability remains a prerequisite. Second, semantic reasoning is essentially one-shot at the beginning, apart from ensemble evaluation; the system does not re-invoke the MLLM when local observations contradict the initial global hypothesis. Third, some objects have ambiguous or tiny top-down footprints, such as clothes or a thin TV screen, which can make room-level or node-level map reasoning unreliable even with multi-stage zoom-in reasoning (Ao et al., 26 Jan 2026).

The paper’s stated future directions are diverse map modalities, a global–local reasoning loop, and active exploration under uncertainty. This suggests a broader research trajectory in which ReasonNavi’s modular “reason-then-act” design is extended with iterative semantic feedback rather than relying on a single initial global hypothesis (Ao et al., 26 Jan 2026).

Within the wider literature, ReasonNavi belongs to a larger class of modular systems that separate semantic deliberation from low-level execution. “ReasonNav” for large buildings integrates SLAM, landmark memory, sign reading, and asking humans for directions, using a VLM to choose the next landmark rather than directly control motion (Chandaka et al., 25 Sep 2025). “Nav-R1” pursues explicit > ... reasoning trained with GRPO and a Fast-in-Slow dual system for long-horizon semantic reasoning plus low-latency control (Liu et al., 13 Sep 2025). “P2DNav” decomposes zero-shot VLN into panoramic direction selection, downview local grounding, sliding-window dialogue memory, and reflective reorientation (Sheng et al., 19 May 2026). “ReMemNav” similarly combines panoramic semantic priors, episodic memory, and an adaptive dual-modal rethinking mechanism for zero-shot ObjectNav (Wu et al., 25 Mar 2026).

Related work also shows the breadth of the modular reasoning pattern. “GeoNav” builds a schematic cognitive map and hierarchical scene graph for language-goal aerial navigation (Xu et al., 13 Apr 2025). “NavA$$\mathcal{A} = \{\texttt{move\_forward},\ \texttt{turn\_left},\ \texttt{turn\_right},\ \texttt{stop}\},$$0” uses a global Reasoning-VLM and a local Pointing-VLM trained on a 1.0 million-sample spatial-aware affordance dataset (Zhang et al., 6 Aug 2025). “CoNav” lets a 3D-text model guide an image-text navigation agent through textual spatial hypotheses and Cross-Modal Belief Alignment (Hao et al., 22 May 2025). “NaviDriveVLM” decouples a large frozen Navigator from a lightweight trainable Driver in autonomous driving (Tao et al., 9 Mar 2026). Earlier VLN work such as HSPR builds hierarchical spatial proximity knowledge over regions and objects for multi-step reasoning on semantic topological maps (Xu et al., 2024).

Taken together, these neighboring systems reinforce the interpretation of ReasonNavi as part of a shift away from monolithic end-to-end policies and toward explicit interfaces between high-level reasoning and low-level planning. In that landscape, ReasonNavi’s distinguishing contribution is to make the global plan itself discrete, map-grounded, and inspectable, while reserving deterministic planners and geometry-aware verification for execution (Ao et al., 26 Jan 2026).