Meta-Learning-Driven Autonomous Intelligence for Real-Time Adaptation in Human–AI Collaborative Systems: A Comprehensive Review

Abstract

ABSTRACT Human–Artificial Intelligence (AI) collaborative systems represent a cornerstone of modern technological innovation, enabling physical and cognitive co-adaptation across diverse domains including industrial robotics, intelligent transportation, defense operations, and digital healthcare. However, conventional machine learning models heavily rely on stationary environmental assumptions and extensive batch retraining, rendering them incapable of responding rapidly to non-stationary human behaviors, dynamic task allocations, and unexpected environment shifts. To address this fundamental limitation, meta-learning ('learning to learn') has emerged as a transformative paradigm. By leveraging prior experience across distributions of tasks, meta-learning enables autonomous agents to adapt their predictive and control policies in real time using minimal interaction data (few-shot learning). This paper presents a comprehensive review of meta-learning driven autonomous intelligence specifically tailored for real-time adaptation in human–AI collaborative frameworks. We synthesize the state-of-the-art developments in metric-based, model-based, and optimization-based meta learning, mapping their mathematical foundations to the complex dynamics of human-in-the-loop interaction. Furthermore, we categorize current real-time adaptation frameworks, highlight critical research gaps—such as safety under distribution shifts, non-stationary human behavioral modeling, and latency bounded inference—and delineate prospective future directions. This survey establishes a unified taxonomic bridge between meta-learning algorithm design and human-centered collaborative engineering. 

KEYWORDS: Meta-learning, Human–AI Collaboration, Real-time Adaptation, Few-shot Reinforcement Learning, Model-Agnostic Meta Learning (MAML), Human-in-the-Loop.