A DATA-DRIVEN APPROACH: YOUTUBE AS A RESOURCE FOR AI, ML AND ROBOTIC DEVELOPMENT

Abstract

YouTube has become one of the largest repositories of human knowledge, offering multimodal content across fields like education, software development, robotics, and design. Unlike traditional machine learning (ML) systems that rely on curated datasets, YouTube provides a scalable, context-rich alternative by combining speech, visuals, and on-screen text in real-world instructional settings. This study presents a modular, data-driven framework for extracting and transforming YouTube content into structured training material for intelligent systems. While the framework is broadly applicable, Python programming is used as a demonstrative case. YouTube explicitly permits the use of publicly available content for research purposes under its Terms of Service, provided that attribution and ethical use guidelines are followed. In some cases, accessing YouTube’s extended datasets (e.g., via the YouTube Data API or Research Program) requires registration as a verified academic or developer, ensuring compliance with responsible data practices. The methodology follows a mixed-methods, modular system engineering approach, combining automated video data mining with empirical system design and lightweight deployment testing. Key techniques include transcript mining, audio signal analysis, and frame-based image extraction. Videos are programmatically accessed using Selenium, and aligned audiovisual segments are extracted using FFMPEG. Natural language processing (NLP), automatic speech recognition (ASR), and community-based relevance scoring help structure and refine the data. In the case study, a chat-bot is trained using Python tutorials, with cleaned and aligned multimodal data used to fine-tune a transformer model (e.g., CodeT5 or GPT). Performance metrics show high effectiveness: 100% transcript retrieval where subtitles exist, 93% metadata alignment via NLP, 88.5% command extraction precision, and 92% sentiment/intent classification accuracy. The system processes 5-minute videos in ~2.5 minutes on a Raspberry Pi and stores each structured dataset in just 1.5MB.Despite promising results, limitations include missing transcripts in some videos, variable content quality, reliance on accurate transcription, and moderate computational demands. Nonetheless, the framework demonstrates YouTube’s potential as a cost-effective, scalable resource for training intelligent, domain-specific systems—particularly in low-resource environments.