Technical Summary: G-DRAGON

Problem Statement

Autonomous ground robots operating in large-scale outdoor environments face a dual challenge: executing robust long-range navigation across complex terrains and performing fine-grained "last-mile" exploration to locate specific, open-vocabulary targets. Existing Visual-Language Navigation (VLN) and Vision-Language-Action (VLA) methods excel in short-horizon tasks but lack geospatial grounding, leading to failures in long-distance missions due to visual drift and an inability to generate reliable global plans. Conversely, OpenStreetMap (OSM)-based approaches often rely on cloud-based Large Language Models (LLMs), which introduce high latency, risk factual hallucinations, and prevent fully offline operation. Furthermore, current "last-mile" exploration methods struggle to bridge the gap between open-vocabulary semantic search and the computational constraints of onboard edge devices in dynamic outdoor settings.

Methodology

The authors propose G-DRAGON (Geospatial Reasoning and Dynamic Planning for Retrieval-Augmented Outdoor Navigation), a unified framework designed to decouple high-level reasoning from low-level control while maintaining full offline capability. The system comprises three core modules:

1. GeoQA Module (Geospatial Question Answering)

This module addresses the grounding of natural language commands into precise geodetic coordinates.

• Generative Retrieval: Instead of relying on dense retrievers or unconstrained generation, the system employs a "zero-shot generative retrieval" framework using a locally deployed LLM.
• Hierarchical ID Construction: Each OSM entity is assigned a unique, discriminative natural-language ID (e.g., "Teaching Area-Building A") to reduce ambiguity.
• Constrained Decoding: The retrieval process utilizes a two-stage approach:
  1. Chain-of-Thought (CoT): The LLM extracts intermediate semantic cues (category, modifiers) to narrow the candidate set.
  2. Trie-Constrained Beam Search: A prefix tree (Trie) of valid entity IDs constrains the decoder, ensuring the generated output corresponds strictly to an existing OSM entity, thereby eliminating hallucinations and guaranteeing valid coordinates for downstream planning.

2. RAPPER Module (Retrieval-Augmented Planner with Programmatic and Executable Reasoning)

RAPPER serves as the central reasoning orchestrator, built upon a locally deployed Qwen2.5-14B LLM.

• Task Parsing: It parses natural language instructions into a logical sequence of sub-tasks (navigation and exploration).
• Behavior Tree (BT) Synthesis: The LLM dynamically constructs a hierarchical Behavior Tree in a single shot, mapping sub-tasks to a predefined skill pool (e.g., GlobalPlanning, FollowPath, Explore). The output is formatted as structured XML to ensure robot-executable logic.
• Global Path Generation: For navigation tasks, RAPPER queries GeoQA for coordinates, projects the robot's position onto the target polygon's contour, and utilizes a local routing engine (OSRM) to generate a global topological path.
• Dynamic Monitoring: The system continuously monitors the robot's state. If execution stalls (e.g., due to blockages), the BT structure triggers recovery mechanisms to re-initialize path search without human intervention.

3. NæVIS Module (Navigation and Exploration System)

NæVIS handles the execution of long-range navigation and the transition to "last-mile" exploration.

• Global-to-Local Alignment: It bridges the domain gap between global geodetic coordinates and the robot's local odometry using a dynamic initialization maneuver and rotation matrix transformation.
• Curvature-Adaptive Sampling: The global path is sampled non-uniformly; fine sampling (3m) is used at high-curvature intersections, while coarse sampling (20m) is used on straight segments to reduce computational load.
• Last-Mile Exploration: Upon reaching the target vicinity, the system switches to a frontier-based exploration mode.
  • Semantic Voxel Mapping: It fuses RGB-D, LiDAR, and IMU data to build an open-set semantic voxel map.
  • Open-Vocabulary Search: An image encoder performs real-time, query-conditional semantic segmentation on the voxel map.
  • Target Localization: Voxels exceeding a semantic similarity threshold are clustered (DBSCAN), and the dominant cluster's centroid is projected to the ground plane as the final navigation goal.

Key Contributions

1. GeoQA for Global Route Generation: The design of a geospatial QA module based on generative retrieval that aligns natural language commands with OSM entities, outputting precise coordinates while avoiding hallucinations through constrained decoding.
2. RAPPER: A hierarchical planner that leverages a locally deployed LLM to parse tasks, query GeoQA, and synthesize executable Behavior Trees, ensuring safe, offline task planning.
3. NæVIS: A seamless integration of long-range navigation and "last-mile" exploration, utilizing frontier-based navigation coupled with open-set semantic voxel mapping to robustly localize open-vocabulary targets in outdoor environments.

Experimental Results

The framework was evaluated in both simulation (NVIDIA Isaac Sim) and real-world deployments on an Agilex Scout Mini UGV.

• Geospatial Reasoning: In a dataset of 5,000 queries, GeoQA achieved a Recall@1 of 97.6% (Easy) and 74.8% (Hard), outperforming local baselines (BM25, DPR, Vanilla Qwen2.5) and demonstrating superior reliability over cloud-based GPT-4o in offline scenarios.
• Simulation Performance:
  • Long-Range Navigation: G-DRAGON achieved a 100% Success Rate (SR) at 300m and 95.56% SR at 800m in Task 1 (building navigation), significantly outperforming baselines NoMaD (0% SR at 800m) and OPEN (26.67% SR at 800m).
  • Last-Mile Exploration: In Task 2 (object search), G-DRAGON maintained a 62.22% SR at 800m, whereas OPEN failed due to limitations in 2D projections and lightweight encoders.
• Real-World Deployment: The system successfully completed four long-range navigation trials (200m–500m) on a university campus. Baselines failed due to heading errors, dynamic obstacle collisions, or inability to filter distractors. G-DRAGON was the only method to complete all missions.
• Ablation Study: Removing the RAPPER module (bypassing OSM-based reasoning) resulted in a failure rate of 3 out of 4 trials, confirming that topological reasoning is a prerequisite for scalable outdoor navigation.

Significance and Claims

The paper claims that G-DRAGON addresses the critical gap between high-level semantic reasoning and low-level control in large-scale outdoor environments. By leveraging locally deployed models and retrieval-augmented generation, the framework achieves fully offline operation without the latency or hallucination risks associated with cloud-based LLMs. The system demonstrates that decoupling reasoning from control allows for robust, scalable navigation that can handle complex, open-vocabulary instructions over distances up to 500m. The authors position this work as a step toward truly autonomous agents capable of operating in connectivity-denied environments while maintaining high precision in both global routing and local semantic search.