End-to-End LLM Flight Planning with RAG-based Memory and Multi-modal Coach Agent

Jul 16, 2026

Upload by Skyler

Bridging the gap between human pilot intent and autonomous flight operation is critical for deploying electric vertical takeoff and landing (eVTOL) aircraft in complex airspace. FRAMe presents an end-to-end Large Language Model flight planning tool that uses retrieval augmented generation (RAG)-based memory together with a multi-modal coach agent. The system produces flight plans that satisfy mission constraints while aligning with operator preferences. Experiments in real-world-inspired scenarios show high validity across planners and improved preference-relevant metrics, enabling natural-language to safe, efficient, flexible routes.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages18

Reading Time45 min

Introduction

Motivation and limitations of classical flight planning

Prior work on LLM planning, UAV control, and preference alignment

Retrieval-augmented planning and memory-based approaches

FRAMe overview and contributions

End-to-End LLM Flight Planning with RAG-based Memory and Multi-modal Coach Agent Amin Tabrizian * 1 Arsyi Aziz * 1 Aarifah Ullah 1 Mahyar Ghazanfari 1 Pouria Razzaghi 2 Peng Wei 1 arXiv :2607 .06964v 1 [ cs .RO] 8 Jul 2026 Abstract Bridging the gap between human pilot intent and autonomous flight operation is critical for realworld electric vertical takeoff and landing (eVTOL) aircraft deployment. Flight planning traditionally relies on classic algorithms that struggle to incorporate flexible human preferences. We present FRAMe, an End-to-End Large Language Model (LLM) Flight Planning tool with RAG-based Memory and Multi-modal Coach Agent. Our system integrates a planner LLM with a multi-modal coach agent and retrieval augmented generation (RAG)-based memory to generate flight plans that satisfy mission constraints while aligning with human flight operator preferences. We demonstrate the system ina range of real-world-inspired scenarios of varying difficulty levels. Across four LLMs, the full FRAMe system (RAG and coach) yields the highest validity for every planner (up to 93.8% aggregate, 99% on Easy scenarios for the strongest planner) and shifts preference-relevant metrics in the operator-favored direction where the metric has headroom. FRAMe signifies how advanced LLMs can be deployed for human-centric mission planning, translating natural language instructions into safe, efficient, and flexible flight routes. The code is available at: [github.com/amin](github.com/amin)tabrizian/FlightPlanningLLMs 1. Introduction Advanced Air Mobility (AAM) operations are projected to grow significantly in the near future, with unmanned aircraft systems (UAS) and eVTOL vehicles performing large *Equal contribution 1 George Washington University, Washington, DC, USA 20052 2Metis Solutions Technology Inc, NASA Ames Research Center, Moffett Field, CA, USA 94035 . Correspondence to: Amin Tabrizian <amin [tabrizian@gwu.edu](tabrizian@gwu.edu) >. Proceedings of the 43 rd International Conference on Machine Learning, Seoul, South Korea. PMLR 306, 2026 . Copyright 2026 by the author(s) . volumes of flights for both cargo delivery and passenger transportation. Maintaining safe and efficient flight planning within such dense and complex airspace presents a major challenge. Flight restrictions, coupled with the often subjective and context-dependent nature of mission-specific requirements, make it difficult to define rigid, mathematical objectives that fully capture operational intent. Classical path planning algorithms, such as A∗ and rapidlyexploring random trees (RRT, RRT∗ ) (Kuffner & LaValle, 2000 ; LaValle, 2006), can compute optimal, obstacle-free trajectories; however, they depend on explicitly defined mathematical objectives and constraints, which limits their ability to capture nuanced human pilot preferences, such as trading off flight duration against waypoint complexity. Asthe scale and complexity of eVTOL operations continue to grow, there is an increasing need for automated planning systems capable of interpreting such nuanced mission intent. One promising technology to bridge this gap is to use LLMs for planning. For instance, in robotics and navigation, the SayCan framework (ichter et al., 2023) demonstrated the capability of LLMs to guide robots using natural language to do feasible and contextually appropriate actions. In planning, hybrid approaches have leveraged LLMs to generate subgoals for classical planners (Meng et al., 2024 ; Liu et al., 2023 ; Dagan et al., 2024), while other methods integrate solver heuristics to better guide LLM generated plans (Wu & Mitra, 2024 ; Hirsch et al., 2024) . Researchers have also explored combining LLMs with reinforcement learning (RL) to enhance reasoning capabilities. For instance, S2RCQL (Deng et al., 2025) augments prompts using information derived through Q-learning, and another approach employs LLMs to generate semantic hints for contextual RL in motion planning

FAQs

What problem does FRAMe address in eVTOL flight planning?

FRAMe targets the difficulty of translating subjective, context-dependent human pilot intent into autonomous flight plans while meeting mission constraints in dense airspace.

How does FRAMe combine LLM planning with RAG-based memory?

FRAMe uses retrieval augmented generation (RAG)-based memory so the planner LLM can condition route generation on preference-conditioned past flight plans.

What role does the multi-modal coach agent play?

The multi-modal coach agent helps align generated plans with human operator preferences, improving preference-relevant metrics while maintaining plan validity across scenarios.

Jul 16, 2026

Upload by Skyler