Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Contextualized Autonomous Drone Navigation using LLMs Deployed in Edge-Cloud Computing

Hongqian Chen, Yun Tang, Antonios Tsourdos, Weisi Guo

TL;DR

The paper tackles dynamic autonomous drone navigation by enabling LLMs to incorporate human-described semantic context into real-time adjustments of navigation maps and sub-task instructions. It introduces a two-level framework where path planning operates on a contextually updated map via Map Information Parsing and LLM-adjusted Potential Fields, while sub-task instructions are generated by a Moonshot LLM and translated into flight actions for a sequence of waypoints using $A^*$ path planning on a $20\times 20$ grid. Through experiments comparing multiple LLMs, the study demonstrates a trade-off: small, edge-friendly models (e.g., Llama3:8b, Moonshot v1-8k) enable low-latency edge deployment, whereas large models (e.g., GPT-4, GPT-3.5T, Llama3:70b) deliver higher accuracy in dynamic contexts when hosted in the cloud. The work outlines a practical architecture for native LLM navigation within ORAN-enabled 6G networks, highlighting both the latency and sovereignty advantages of edge deployment and the richer contextual reasoning available in cloud-based LLMs. This framework provides a pathway toward 6G-native AI navigation, enabling real-time semantic adaptation and robust multi-task execution in future urban mobility scenarios.

Abstract

Autonomous navigation is usually trained offline in diverse scenarios and fine-tuned online subject to real-world experiences. However, the real world is dynamic and changeable, and many environmental encounters/effects are not accounted for in real-time due to difficulties in describing them within offline training data or hard to describe even in online scenarios. However, we know that the human operator can describe these dynamic environmental encounters through natural language, adding semantic context. The research is to deploy Large Language Models (LLMs) to perform real-time contextual code adjustment to autonomous navigation. The challenge not evaluated in literature is what LLMs are appropriate and where should these computationally heavy algorithms sit in the computation-communication edge-cloud computing architectures. In this paper, we evaluate how different LLMs can adjust both the navigation map parameters dynamically (e.g., contour map shaping) and also derive navigation task instruction sets. We then evaluate which LLMs are most suitable and where they should sit in future edge-cloud of 6G telecommunication architectures.

Contextualized Autonomous Drone Navigation using LLMs Deployed in Edge-Cloud Computing

TL;DR

The paper tackles dynamic autonomous drone navigation by enabling LLMs to incorporate human-described semantic context into real-time adjustments of navigation maps and sub-task instructions. It introduces a two-level framework where path planning operates on a contextually updated map via Map Information Parsing and LLM-adjusted Potential Fields, while sub-task instructions are generated by a Moonshot LLM and translated into flight actions for a sequence of waypoints using path planning on a grid. Through experiments comparing multiple LLMs, the study demonstrates a trade-off: small, edge-friendly models (e.g., Llama3:8b, Moonshot v1-8k) enable low-latency edge deployment, whereas large models (e.g., GPT-4, GPT-3.5T, Llama3:70b) deliver higher accuracy in dynamic contexts when hosted in the cloud. The work outlines a practical architecture for native LLM navigation within ORAN-enabled 6G networks, highlighting both the latency and sovereignty advantages of edge deployment and the richer contextual reasoning available in cloud-based LLMs. This framework provides a pathway toward 6G-native AI navigation, enabling real-time semantic adaptation and robust multi-task execution in future urban mobility scenarios.

Abstract

Autonomous navigation is usually trained offline in diverse scenarios and fine-tuned online subject to real-world experiences. However, the real world is dynamic and changeable, and many environmental encounters/effects are not accounted for in real-time due to difficulties in describing them within offline training data or hard to describe even in online scenarios. However, we know that the human operator can describe these dynamic environmental encounters through natural language, adding semantic context. The research is to deploy Large Language Models (LLMs) to perform real-time contextual code adjustment to autonomous navigation. The challenge not evaluated in literature is what LLMs are appropriate and where should these computationally heavy algorithms sit in the computation-communication edge-cloud computing architectures. In this paper, we evaluate how different LLMs can adjust both the navigation map parameters dynamically (e.g., contour map shaping) and also derive navigation task instruction sets. We then evaluate which LLMs are most suitable and where they should sit in future edge-cloud of 6G telecommunication architectures.

Paper Structure

This paper contains 14 sections, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Review of LLM-informed Navigation
  3. Gap and Novelty
  4. System Setup and Methodology
  5. LLM-informed Navigation: Path Planning Level
  6. Map Information Parsing
  7. LLM-adjusted Potential Fields
  8. LLM-informed Navigation: Sub-Task Instruction Level
  9. Experimental Setup
  10. Evaluation Criteria
  11. Comparison of Different LLMs in Path Planning
  12. Next Steps: Future 6G Architecture for Native LLM-Navigation
  13. Conclusions and Future Work
  14. Appendix A - Experimental Setup

Figures (3)

  • Figure 1: System Diagram of LLM-navigation: (A) 3D city map, (B) Mission definition, (C) LLM generated instruction set for sub-tasks, (D) Navigation via way points and potential fields, and (E) Contextualized potential field adjustment using LLMs.
  • Figure 2: A) Map parsing with descriptive inputs, B) Potential field adjustment with contextual inputs.
  • Figure 3: Native-AI implemented inside ORAN architecture to achieve near real-time path planning using LLMs and real-time tactical control of sub-tasks using smaller language models.