Source-linked AI summary
CAVE-NAV: VLM-Based Autonomous 3D Navigation in Underwater Cave Environments
Zhenqi Wu, Yuanjie Lu, Yisheng Zhang, Miao Yu, Xuesu Xiao, Jaejeong Shin, Xiaomin Lin
TL;DR
Underwater cave navigation is difficult because visual degradation makes feature-based localization unreliable in confined environments. CAVE-NAV uses a VLM with Chain-of-Thought reasoning over RGB, depth, and vertical-clearance observations to infer safe 3D directions. In high-fidelity simulations, it completes all evaluated end-to-end traversals without collisions while maintaining clearance from cave boundaries.
Problem
Visual degradation in underwater caves makes dense-feature localization unreliable, motivating alternative navigation for confined environments.
Method
CAVE-NAV uses a VLM with Chain-of-Thought reasoning to interpret RGB, depth, and vertical-clearance observations and generate navigation commands.
Results
The framework completes all evaluated end-to-end traversals without collisions while maintaining safe clearance from cave boundaries.
Takeaways & Limitations
The results support semantic reasoning combined with geometric perception and clearance-aware planning as a viable approach in feature-denied underwater caves.
Abstract
from arXiv · showhide
Autonomous navigation in underwater cave environments is essential for search-and-rescue operations, scientific exploration, and emergency egress. Traditional navigation systems commonly depend on dense visual features for localization and mapping. In underwater caves, however, visual degradation can undermine feature-based localization, sonar-based mapping may yield overly conservative obstacle representations, and communication constraints preclude real-time human guidance. To address these limitations, we propose an autonomous underwater cave navigation framework that leverages a vision-language model (VLM) with Chain-of-Thought (CoT) reasoning to infer navigable directions from environmental cues, including light intensity gradients, passage morphology, and geometric complexity, captured through multimodal inputs comprising RGB imagery, depth maps, and sonar-based vertical-clearance measurements, thereby supporting safe 3D navigation through confined cave passages. High-fidelity simulations across multiple cave topologies demonstrate that the proposed framework completes all evaluated end-to-end traversals without collisions while maintaining safe clearance from cave boundaries.
I. INTRODUCTION
CAVE-NAV addresses underwater cave navigation challenges by using VLM-based semantic reasoning over multimodal observations instead of relying solely on dense visual features. The framework targets safe 3D traversal and reports collision-free evaluation across diverse cave topologies.
- I. INTRODUCTION: Underwater caves challenge autonomous navigation through visual degradation, communication constraints, and difficult operating conditions relevant to rescue and scientific exploration.Suspended sediment and low-texture rock surfaces make reliable perception especially difficult where navigation errors are hazardous.
- I. INTRODUCTION: Feature-based localization becomes unreliable because suspended sediment, low-texture surfaces, and vehicle illumination undermine feature correspondences and tracking.These conditions can produce tracking failures and accumulated drift.
- I. INTRODUCTION: CAVE-NAV uses a VLM to infer navigable directions from passage morphology, illumination patterns, and structural complexity using RGB, depth, and vertical-clearance observations.The framework is proposed for previously unexplored underwater caves.
- I. INTRODUCTION: The prompt encodes clearance requirements with zero-shot Chain-of-Thought reasoning so the model evaluates active constraints before producing a 3D motion command.The reasoning process is paired with direct motion-command generation.
- I. INTRODUCTION: The framework demonstrates collision-free traversal across five cave topologies and transfers the same prompt to real imagery captured at Blue Grotto, Florida.The tested topologies include constrictions, vertical undulations, sharp bends, and cluttered geometry.
II. RELATED WORK
Related underwater navigation work improves geometric estimation or uses predefined semantics, while foundation-model approaches broaden visual-language interpretation. CAVE-NAV is positioned around environment-centric reasoning about admissible directions in unmapped caves.
- II. RELATED WORK: Underwater SLAM integrates visual, inertial, sonar, and pressure measurements, but remains focused on geometric state estimation rather than identifying navigable structures.These systems improve metric estimation under degraded sensing without changing the central estimation objective.
- II. RELATED WORK: Cave-specific methods address reconstruction or semantic navigation, but remain constrained by illumination geometry, predefined categories, or pre-installed cavelines.The cited cave systems broaden navigation beyond pure reconstruction while retaining infrastructure or category dependencies.
- II. RELATED WORK: Contrastive models such as CLIP and BLIP provide visual-semantic embeddings, similarity scores, or captions rather than task-specified deliberation over actions.Generative multimodal models can instead combine mission, sensor, and safety requirements in a single query.
- II. RELATED WORK: Prior navigation applications span perception and control layers, but many target terrestrial landmarks or open water where vehicles can retreat and hold station.The paper frames unmapped caves as requiring direction-level admissibility reasoning under tighter geometric constraints.
III. APPROACH
The approach uses a closed-loop perception–planning–execution cycle built from RGB-D imagery and vertical-clearance sensing. Multimodal observations are rendered and processed by a VLM before each direct vehicle-pose update.
- III. APPROACH: At each step, the simulator renders visual and depth channels, a VLM interprets them with a fixed prompt, and the resulting motion command is applied directly to the vehicle pose.The next observation is then rendered for the following cycle.
- III. APPROACH: The perception module combines a forward-facing RGB-D camera with two single-beam echo sounders measuring distances to the ceiling and floor.The vertical measurements support avoidance of vertical collisions and adequate clearance in narrow passages.
- III. APPROACH: Raw depth values are normalized to [0, 1] and encoded as an 8-bit PNG image for VLM input.The depth image represents per-pixel distance to scene points.
B. Planning Module
The planning module supplies multimodal observations and a task-specific prompt to a VLM, which reasons with Chain-of-Thought and returns an action plus an explanation.
- B. Planning Module: At each decision step, the VLM receives the current RGB image, normalized depth image, vertical-clearance measurements, and task-specific navigation prompt.These inputs are fused for high-level navigation-action generation.
- B. Planning Module: The VLM outputs a discrete navigation action and textual reasoning over motion primitives including forward, turning, ascending, and descending.The explanation allows human operators to verify the robot’s decision logic.
C. Vision-Language Model Prompting
The prompting framework combines multimodal sensor interpretation with safety-first and goal-directed strategies to guide VLM navigation through confined underwater caves.
- C. Vision-Language Model Prompting: The prompt combines mission specification, sensor interpretation guidance, strategic guidance, and reasoning scaffolding.
- 1) Mission Specification and Input Interpretation:: The VLM receives the navigation objective alongside front-facing RGB, depth, and vertical-clearance measurements.
- 1) Mission Specification and Input Interpretation:: RGB provides scene context, depth encodes nearby versus distant surfaces, and ceiling-floor distances indicate vertical clearance.
- 2) Strategic Decision-Making:: The prompt prioritizes collision avoidance over efficiency, accepting slower progress to preserve safety margins.
- 2) Strategic Decision-Making:: The obstacle protocol targets a 2.0 m standoff and uses shorter, turning, lateral, or vertical commands when clearance decreases.
- 2) Strategic Decision-Making:: Below nominal safety distance, forward motion is capped at 0.1–0.2 m and paired with turning or vertical adjustment.
- 2) Strategic Decision-Making:: For exploration, the VLM favors locally visible open directions that extend the passage and avoids immediate reversal of recent motion.
3) Chain-of-Thought Reasoning:
Chain-of-Thought prompting structures each navigation decision as explicit assessment, constraint analysis, and action selection, then converts the result into a bounded motion command.
- 3) Chain-of-Thought Reasoning:: Zero-shot Chain-of-Thought prompting asks the VLM to reason step by step before selecting a navigation action.
- 3) Chain-of-Thought Reasoning:: The reasoning sequence assesses the situation, identifies active constraints, and proposes and justifies a navigation command.
- 1) Output format:: The output is a single dictionary key–value pair containing an action string and a short justification for deterministic parsing and auditability.
- 1) Output format:: Parsing the action string yields a rigid-body increment with forward, rightward, vertical, and yaw components.
- 1) Output format:: The first three action components use discrete option sets, while the vertical increment is selected from a continuous range.
- 1) Output format:: Invalid or out-of-range responses trigger a re-query, while persistent parsing failures terminate the run.
2) Pose update:
The simulator applies accepted actions directly to the vehicle pose, updating heading and lateral displacement relative to a defined world-frame reference.
- 2) Pose update:: Accepted actions are applied directly to the vehicle pose in the simulator.
- 2) Pose update:: The reference heading sets ψ = 0 along the world frame’s −y axis, with Δrt defining rightward displacement from that heading.
IV. EXPERIMENTS AND RESULTS
The experiments use five high-fidelity Blender cave scenarios designed to test distinct geometric and navigational challenges, with trajectories visualized from top-down and side views.
- 1) Simulation Environment:: Five Blender-generated cave scenarios test narrow constrictions, vertical undulations, obstacles, a 90-degree turn, and complex topology.
- 1) Simulation Environment:: The scenarios collectively evaluate confined-space navigation, vertical maneuvering, sharp turns, and cluttered environments.
- 1) Simulation Environment:: Figure 3 uses top-down and side views across scenarios S1–S5, with orange lines marking AUV trajectories.
2) AUV Configuration:
CAVE-NAV is evaluated across five simulated cave scenarios and a real Blue Grotto image, producing collision-free trajectories with appropriate clearance and geometry-aware commands.
- Scenario results: At the sharp S4 turn, the VLM reads the exit direction before the bend and issues turning commands through the curved section.
- Trajectory visualization: The orange trajectories remain near cave centers in top-down and side views, maintaining clearance from walls, ceiling, and floor.These complementary views summarize three-dimensional obstacle avoidance and trajectory placement across the five scenarios.
- Scenario results: In narrow passages S1 and S2, the AUV follows central trajectories while preserving clearance from surrounding cave boundaries.
- Scenario results: In complex topology S5, the vehicle traverses the full cave while avoiding every obstacle and preserving the prescribed standoff.
- Real-image transfer: On real Blue Grotto imagery, the model combined ascent, right yaw, and a short forward step after identifying an overhead opening and nearby left wall.
- Evaluation: 100% task completion with zero collisions was achieved across all five simulated cave environments.The evaluated topologies include constrictions, vertical undulations, obstacles, sharp turns, and complex geometry.
V. CONCLUSION
CAVE-NAV uses multimodal VLM semantic reasoning and clearance-aware planning to navigate feature-denied underwater caves. Simulations across five diverse topologies achieved complete traversals, while field deployment and onboard efficiency remain future requirements.
- Framework: CAVE-NAV interprets passage geometry, illumination gradients, and structural complexity using VLMs, CoT reasoning, and multimodal perception.
- Framework: The framework avoids reliance on dense visual features, pre-built maps, and human-placed infrastructure for underwater cave navigation.
- Results: Across five diverse cave topologies, the system completed end-to-end traversals while maintaining central trajectories that maximize clearance from cave walls.
- Future work: The authors identify field validation under turbidity, backscatter, variable currents, and non-uniform illumination as a requirement for deployment.
- Future work: Developing compressed VLMs for resource-constrained hardware is proposed to support onboard deployment, lower communication dependencies, and improve response latency.
- Implication: The work takes a step toward autonomous cave exploration by showing that VLMs can reason about environmental structure and clearance margins.