Diseño e implementación de arquitecturas basadas en agentes LLM para la planificación de trayectorias de drones en lenguaje natural
Year: 2026
Type: dissertation
Abstract: The integration of Large Language Models (LLMs) into robotic platforms has marked a paradigma shift in Imman-robot interaction, moving from rigid, command-based pro- gramming toward intuitive, semantic-driven mission planning. However, the translation of bighlevel human intent into precise, deterministic flight trajectories remains a criti- cal challenge due to the stochastic nature of probabilistic language generation and the complexity of spatial reasoning in three-dimensional environments. m the present study, the capabilities of LLMs to function as autonomous high-level planners for Unmanned Aerial Systems (UAS) are analyzed. We propose a framework that bridges the gap between natural language and the robotic control layer, specifically leveraging the ROS? middleware and the Aerostack? software stack. This is achieved through a comparative analysis of three distinet architectural paradigms: direct zero- shot generation, tool-angmented agentic workflows for deterministic code synthesis, and decoupled hierarchical planning. Ey implementing a Model Context Protocol (MCP) server for lfeeyele automation and a multimodal “LLM-as-a-Judge” feedback loop for visual validation, we test the systenvs reliability in mission synthesis. We find that the integration of Python-based tool calling significantly enhances trajectory precision, as the system adopts a methodology alin to human engineering—synthesizing mathematical logic rather than predicting coordinate sequences. Performance is assessed through a dual-metzic evaluation approach: a tradi- tional mumerical analysis and a human-expert-reviewed rubric based on seven abstract qualitative dimensions, capturing facets of semantic intent and spatial topology often missed by purely quantitative measurements. These findings demonstrate that when equipped with deterministic tools and visual feedback mechanisms, multimodal LLMs exhibit robust capabilities for autonomons tra- jectory generation. This research confirms the viability of embodied Al agents in complex robotie ecosystems, opening the door to UAS control through high-level cognitive inter- faces.
Author: Fernández Barrera, Pablo
Language: Spanish
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7215290281