Voice-Controlled Drone Navigation : Bridging the Gap Between Natural Language and Drone Navigation

Sammanfattning

The growing field of drone usage and its technology, in both civilian and commercial sectors, is currently slowed down by a steep learning curve and complex manual controllers. To address this obstacle, this project investigates voice-controlled navi- gation designed to bridge the gap between natural human language and navigation of Unmanned Aerial Vehicles. The system was developed and evaluated in a simulated environment using ArduPi- lot SITL and Gazebo to simulate flight dynamics and sensor-based collision avoid- ance, all of this was possible through a modular Docker-containerized architecture. The voice-control architecture uses a pipeline which captures audio via Whisper.cpp, filters input through a rule-based wake and sleep word system and utilizes fastText, a lightweight machine learning classifier for intent recognition and translation into MAVLink commands. Evaluation of the system’s ability to parse text into MAVLink commands proved to be highly efficient with an average real-time processing latency of 2.2ms. End-to- end pipeline accuracy proved to be very sensitive to acoustic interference, degrading from 82.5% in a quiet environment to 62.5% in high-noise conditions. Furthermore, the project highlights the weaknesses and inherent reliability and safety trade-offs present in audio-based control of a drone in an uncontrolled environment. Ulti- mately, this project provides a successful proof-of-concept for an accessible and simple interface for voice-controlling a drone.