VLN-Zero: Zero-Shot Vision-Language Navigation

• The paper introduces a novel framework that decomposes navigation into rapid exploration, compact symbolic mapping, and cache-enabled neurosymbolic execution for zero-shot transfer.
• It employs structured prompts for VLM-guided exploration to build a top-down scene graph, enabling efficient reasoning over visual observations and language directives.
• Benchmarks show significant gains with up to 55% fewer VLM calls and doubling success rates compared to state-of-the-art zero-shot navigation models.

Searching arXiv for the specified VLN-Zero paper and closely related zero-shot VLN works. VLN-Zero is a zero-shot vision-language navigation framework for robot navigation in unseen environments that is designed to operate without fine-tuning, without exhaustive exploration, and without repeated expensive VLM querying. Its defining structure is a two-phase pipeline: a rapid exploration phase in which a VLM-guided robot constructs a compact symbolic scene graph or top-down map, followed by a deployment phase in which a neurosymbolic planner reasons over that scene graph, current visual observations, and a natural-language task, while a cache-enabled execution module reuses previously validated task-location trajectories. In this formulation, the central contribution is not merely the use of a vision-LLM for navigation, but the decomposition of navigation into exploration, symbolic map building, and cached neurosymbolic execution (Bhatt et al., 23 Sep 2025).

1. Problem setting and formalization

VLN-Zero is motivated by a practical bottleneck in robot deployment: how to move from an unseen environment to successful goal-directed navigation without retraining. The method is framed against two recurring limitations in prior VLN systems. First, exploration is inefficient: agents either spend too many steps blindly searching or rely on fixed policies that do not generalize well. Second, deployment is expensive or brittle: repeated online planning calls are slow and costly, and many methods do not explicitly reuse prior navigation experience (Bhatt et al., 23 Sep 2025).

The paper formalizes this as two coupled problems. The first is rapid exploration: given a novel environment and user constraints $\Phi$, the goal is to design a prompting strategy for a VLM $M_{VL}$ so that the robot can efficiently explore and construct a usable scene graph $G_S$ within a limited budget. The second is deployment, expressed as

$P(\Phi, o, T_p, G_S) = a,$

where $\Phi$ denotes user constraints in natural language, $o$ is the current visual observation, $T_p$ is the task prompt, $G_S$ is the scene graph from exploration, and $a$ is the action to execute. The planner is required to satisfy constraints, navigate effectively, and adapt to the unseen environment without retraining (Bhatt et al., 23 Sep 2025).

This formalization places VLN-Zero within the broader zero-shot VLN problem class, but with an explicit separation between environment acquisition and task execution. A plausible implication is that the framework treats adaptation as a structured world-model construction problem rather than as direct end-to-end policy transfer.

2. Exploration phase and compact symbolic scene graphs

In the exploration phase, VLN-Zero uses structured prompts to steer the robot toward informative, diverse, and safe trajectories. The exploration prompt $T_e$ asks the VLM to generate simple step-by-step navigation instructions using only four actions: move forward, turn left, turn right, and stop. It also includes explicit constraints: use only allowed actions, return a single action, stop when exploration of the visible environment is complete or unsafe, and avoid revisiting explored areas. Crucially, the prompt includes both the Scene Graph and the Visual Observation, allowing the model to reason over memory and perception jointly (Bhatt et al., 23 Sep 2025).

At each step, the VLM receives the tuple $M_{VL}$0 and outputs an action $M_{VL}$1 together with a partial scene-graph update $M_{VL}$2. The global graph is then updated incrementally as

$M_{VL}$3

Algorithm 1 is described as an iterative loop that initializes $M_{VL}$4, repeatedly reads sensor data $M_{VL}$5, queries the VLM for an action and graph update, executes the action, merges the update into $M_{VL}$6, and returns the final scene graph when the exploration budget is exhausted. Formally, the core query is

$M_{VL}$7

The exploration strategy is explicitly not exhaustive. Instead of sweeping the entire environment, VLN-Zero prefers trajectories that improve coverage and map quality. The prompts and constraints encourage the robot to avoid unsafe areas, avoid revisiting already explored regions, continue toward new visible space, and stop when enough of the accessible region is covered. The stopping condition is not only time-based: the framework also monitors the top-down map for loop closure of the periphery and for the absence of large unexplored empty regions inside the mapped area (Bhatt et al., 23 Sep 2025). This suggests that exploration is judged by structural sufficiency rather than by full reconstruction.

The resulting scene graph $M_{VL}$8 is a top-down map representing traversable floor, obstacles or walls, landmarks, and semantic room labels. Objects are grouped into meaningful areas such as living room, kitchen, and bedroom. The graph is therefore not only geometric but also symbolic and semantically tagged. The paper emphasizes that it is built incrementally during exploration using odometry and sensor input, and that the output is a compact representation sufficient for later navigation (Bhatt et al., 23 Sep 2025).

3. Neurosymbolic deployment and executable planning

Once exploration ends, deployment begins. The deployment prompt $M_{VL}$9 includes constraints, the scene graph, the visual observation, and navigation rules. The planner is instructed to use the top-down scene graph to decide how to move from the start or current location to the goal. In this phase, the planner reasons over feasible traversable corridors, obstacles and walls, current position and heading, and likely routes to task-relevant landmarks (Bhatt et al., 23 Sep 2025).

The method is described as neurosymbolic because the VLM combines symbolic structure from the scene graph, visual perception from the camera, and language understanding of the task prompt. For example, for the task “deliver an item to the kitchen,” the planner uses $G_S$0 to identify a feasible route to the kitchen while avoiding non-traversable areas. The planner generates one action at a time. If a cached plan is available, the action is retrieved directly; otherwise, the planner reasons from scratch over the current map and observations (Bhatt et al., 23 Sep 2025).

A recurrent misconception in discussions of zero-shot VLN is that a VLM is being used as a monolithic end-to-end navigator. VLN-Zero is explicitly organized otherwise. Its conceptual shift is to use the VLM first as an explorer or map-builder, then as a symbolic reasoning engine, and finally as a retrieval-augmented executor. This architecture is intended to move deployment away from raw perceptual control and toward planning over a structured world model (Bhatt et al., 23 Sep 2025).

4. Cache-enabled execution and hierarchical trajectory reuse

Cache-enabled execution is one of the framework’s main efficiency mechanisms. The cache $G_S$1 stores validated trajectories indexed by task and location:

$G_S$2

A trajectory is defined as a sequence of consecutive points in the scene graph. The cache is hierarchical rather than limited to full task trajectories. It stores task-level trajectories $G_S$3, subtask or location-level trajectories $G_S$4, and reusable fragments such as room-to-room and room-to-object segments (Bhatt et al., 23 Sep 2025).

Algorithm 2 follows a staged retrieval procedure. It first checks whether the full task $G_S$5 is already in cache; if so, the trajectory is retrieved and the next action is executed. If not, it checks whether a subtask or location pair $G_S$6 is cached; if so, that segment is used. Otherwise, the system calls the planner $G_S$7, computes a new trajectory, stores it in cache, and merges it into higher-level task memory (Bhatt et al., 23 Sep 2025). The cache therefore functions as a trajectory memory that reuses verified navigation behavior instead of recomputing it.

The paper argues that VLMs do not have explicit memory, and that caching compensates for this limitation. The intended effects are reduced redundant VLM calls, lower per-step latency, shorter total episode time, and lower query cost. This is also the mechanism behind the framework’s claim that it reaches goal locations in half the time with 55% fewer VLM calls on average compared to state-of-the-art models across diverse environments (Bhatt et al., 23 Sep 2025).

5. Benchmarks, quantitative performance, and real-robot validation

VLN-Zero is evaluated in Habitat Simulator and the VLN-CE environment on R2R Val-Unseen and RxR Val-Unseen. The reported metrics are Navigation Error (NE), Oracle Success Rate (OS), Success Rate (SR), and SPL, with success defined as stopping within 3 meters of the goal (Bhatt et al., 23 Sep 2025).

On R2R Val-Unseen, VLN-Zero achieves SR 42.4%, OS 51.6%, SPL 26.3%, and NE 5.97. The strongest zero-shot baseline, CA-Nav, is reported at SR 25.3%, OS 48.0%, and SPL 10.8%, so VLN-Zero improves SR by roughly 17 percentage points over the best zero-shot baseline. On RxR Val-Unseen, VLN-Zero achieves SR 30.8%, OS 37.5%, SPL 19.0%, and NE 9.13, outperforming the second-best zero-shot baseline, AO-Planner, by 8.4% SR. The paper summarizes these outcomes as 2x higher success rate compared to state-of-the-art zero-shot models and states that the framework outperforms most fine-tuned baselines (Bhatt et al., 23 Sep 2025).

The authors also report consistent behavior across different foundation models, specifically GPT-4.1 and GPT-5, which is presented as evidence that the framework is not brittle to one particular VLM backend. In repeated episodes in the same explored scene, caching reduces VLM calls by up to 78.6% and elapsed time by up to 78.8%. The episode-level examples are explicit: for episode 1475, VLM calls drop from 47 to 32, time per step from 2.161s to 1.352s, total time from 101.551s to 72.996s, and cost from \$G_S$80.136; for episode 1691, VLM calls drop from 84 to 18, time per step from 1.995s to 0.888s, total time from 167.605s to 35.500s, and cost from \$G_S$90.077 (Bhatt et al., 23 Sep 2025).

A real-robot demonstration is performed on a Unitree Go2 quadruped in an apartment environment. Exploration consists of a single exploration pass, under 10 minutes, with about 30 m traveled in an approximately 30 m$P(\Phi, o, T_p, G_S) = a,$0 apartment, and produces a usable scene graph without revisiting. For the deployment task “Go to the bookshelf to find a book and then return to the coffee table in the living room,” the planner decomposes the goal into subtasks and retrieves cached segments such as “go to the living room” and “go to the kitchen” directly. The robot completes the task in about 3 minutes (Bhatt et al., 23 Sep 2025).

6. Position within zero-shot VLN research

VLN-Zero belongs to a rapidly expanding family of zero-shot or training-free VLN systems in continuous environments, but its mechanism is distinct. SmartWay combines an Occupancy-aware Waypoint Predictor with an MLLM-based Navigator that adds history-aware reasoning and adaptive backtracking; on R2R-CE it reports OSR 51%, SR 29%, SPL 22.46%, and NE 7.01, and it is explicitly described as a zero-shot VLN-CE or VLN-Zero system (Shi et al., 13 Mar 2025). Another 2025 framework also called VLN-Zero couples an abstract obstacle map-based waypoint predictor with TopoGraph-and-VisitInfo-aware prompting, reporting SR 41 and SPL 25.4 on R2R-CE and SR 35.7 and SPL 21.7 on RxR-CE (Li et al., 24 Sep 2025). These systems share the zero-shot objective but differ in what constitutes the primary structured intermediate representation: waypoints and visitation records in one case, symbolic scene graphs and hierarchical cached trajectories in the other.

Other adjacent approaches expose additional fault lines in the field. GC-VLN formulates navigation guidance as graph constraint optimization by decomposing instructions into explicit spatial constraints and solving waypoint coordinates with a constraint solver, presenting itself as training-free in continuous environments (Yin et al., 12 Sep 2025). Spatial-VLN argues that the main bottleneck of LLM-based agents is insufficient explicit spatial perception and introduces Spatial Perception Enhancement together with Explored Multi-expert Reasoning and a query-and-explore mechanism (Yue et al., 19 Jan 2026). A separate analysis of VLM-LLM navigation reports a perception saturation phenomenon, according to which navigation success saturates once semantic retention and bounding-box localization are “good enough,” suggesting that navigation-relevant core vocabularies and accurate bounding box proportions may matter more than strict pixel-level precision (Xia et al., 14 May 2026).

Historically, earlier zero-shot VLN in continuous environments was often much weaker in absolute terms. VLN-CM, for example, uses Attention Spot Predictor, View Selector, Progress Monitor, and Open Map Predictor modules, and reports SR 0.11 and SPL 0.02 on VLN-CE unseen validation data, while emphasizing that collision mitigation is critical (Jeong et al., 2024). Relative to that line, VLN-Zero represents a later neurosymbolic turn in which exploration, symbolic world modeling, and reusable trajectory memory become the central organizing abstractions.

Within that broader landscape, the distinctive claim of VLN-Zero is that robust zero-shot transfer in unseen environments can be obtained by rapid exploration, symbolic reasoning, and cache-enabled execution in a single pipeline. This suggests a conception of zero-shot VLN in which the dominant challenge is not only action selection under language grounding, but also efficient acquisition and reuse of structured environment knowledge (Bhatt et al., 23 Sep 2025).