DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES
M.Tech Research Thesis {Colloquium}
Speaker: Mr. Kautuk Astu
S.R. Number: 06-18-01-10-22-24-1-25480
Title: “Reliable Autonomous UAV Programming through Structured LLM-guided Middleware”
Research Supervisor: Prof. Yogesh Simmhan
Date & Time : July 07, 2026 (Tuesday), 15:00 PM
Venue : #102, CDS Seminar Hall
ABSTRACT
Recent advances in Unmanned Aerial Vehicles (UAVs), edge computing and AI have enabled aerial applications in precision agriculture, infrastructure inspection, disaster response, and urban services. However, developing such applications remains challenging due to the tight coupling of navigation, sensing, analytics, and edge-cloud resources. The existing frameworks still rely on low-level APIs that require manual integration of mission logic, scheduling, and deployment. Also, despite Large Language Models (LLMs) offering new opportunities for automating drone application development, the area remains underexplored due to lack of reliability in mission critical systems. This thesis advances the vision of Drone-as-a-Service (DaaS) through three complementary contributions: (1) AeroDaaS, extending a service-oriented programming framework and runtime, (2) AeroGen, for LLM-assisted generation of drone programs and automated execution, and (3) AeroEval, a mission validation framework for LLM-generated applications. Together, these contributions establish an end-to-end ecosystem for reliable autonomous aerial applications across heterogeneous drones. To address the challenge of programmability and runtime support for DaaS spanning edge and cloud environments, we extend prior work on AeroDaaS to support five representative mission patterns and develop a service-oriented programming framework with reusable APIs. It includes modular analytics and trajectory scheduling policies, and a portable containerized runtime for heterogeneous UAV platforms and simulators.
Our evaluation of seven representative applications and scheduling algorithms demonstrates concise application code (≤ 40 lines of code), low runtime overhead (≤20 ms/frame), and low memory footprint (under 1 GB). AeroDaaS serves as an efficient programming framework and platform for aerial DaaS applications. Building upon this, we address the gap of code automation through AeroGen. This leverages LLMs to automatically generate DaaS application code from natural-language mission descriptions. It combines the AeroDaaS APIs with structured guardrail prompting that encodes API specifications, operational constraints and flight rules to generate deployment-ready drone programs. In evaluations performed across 20 navigation tasks and 5 analytical missions in simulation and real-world deployments, AeroGen achieves 100% single-pass success while consistently satisfying mission objectives. These demonstrate that structured prompting and well-designed programming abstractions significantly improve the reliability of AI-generated drone applications and automation. Finally, to improve reliability of generated missions we propose AeroEval, a middleware that enhances the reliability of AeroGen generated code through closed-loop multi-stage validation. It performs static pre-flight analysis like syntax correctness, API compliance and intent verification, further verified by simulation-based trajectory validation. Evaluation on 20 navigation and 5 analytical missions improves navigation success from 55% to 95% and analytical mission success by over 50% compared to AeroGen across multiple iterations, while revealing illusion of correctness in generated programs. These show that automated mission validation can result in reliable AI-assisted DaaS systems.
Collectively, these contributions establish an end-to-end Drone-as-a-Service ecosystem spanning programming, AI-assisted application generation and mission validation. Together, they bridge the gap between high-level mission specifications and dependable real-world deployment of autonomous aerial applications.
ALL ARE WELCOME
