The fundamental motivation for research in human-machine collaborative augmented intelligent control stems from the inherent and complementary limitations of human intelligence and artificial intelligence (AI). Human intelligence excels in creative thinking, contextual understanding, ethical judgment, and handling uncertainty, but is constrained by information processing speed, memory capacity, and subjective cognitive biases. AI, on the other hand, is adept at high-speed computation, massive data processing, and stable execution but lacks common-sense reasoning, interpretability, and adaptability to open environments. Therefore, relying solely on humans or machines is insufficient for efficiently and reliably tackling complex tasks.

Developing hybrid-augmented intelligence, where the strengths of both are fused through deep human-machine collaboration to achieve a synergistic "1+1>2" effect, has become a core direction for the new generation of AI. The key to realizing this vision lies in human-in-the-loop (HiTL) control systems, whose central scientific and engineering challenge is the design of collaborative control algorithms. This research holds significant theoretical value for unifying human-machine interaction theories and breaking through existing control paradigms, and it represents a pressing practical need to drive technological transformation in fields such as intelligent healthcare, advanced industry, and future military applications, enabling safe, efficient, and trustworthy human-machine coexistence.

This article systematically reviewed the current research status in this field from aspects such as human behavior modeling and machine collaborative control.

• Human control behavior modeling: Addressing the "black-box" nature of human behavior, modeling methods are systematically categorized into four types. 1) Control theory-based methods (e.g., quasi-linear models, optimal control models, model predictive control): These models offer strong interpretability and a solid theoretical foundation, suitable for tasks with relatively fixed patterns and clear mechanisms, but their ability to characterize strongly nonlinear or non-rational behavior is limited. 2) AI technology-based methods (e.g., fuzzy control, neural networks): Particularly represented by deep learning, these methods possess powerful nonlinear fitting and feature learning capabilities, handling high-dimensional, complex behaviors. However, they suffer from "black-box" issues, poor interpretability, and challenges in safety verification. 3) Probability theory-based methods (e.g., Markov decision processes, Gaussian mixture models): These excel at capturing the uncertainty and randomness of behavior, suitable for scenarios like inverse reinforcement learning and imitation learning, but face challenges such as the curse of dimensionality and high online computational complexity. 4) Physiological structure-based methods (e.g., Hess model): Starting from the biological mechanisms of perception-cognition-execution, these offer the highest interpretability but involve complex models with numerous parameters, often used for mechanistic analysis and high-fidelity simulation.

• Human-machine collaborative augmented intelligent control methods: The core is to achieve effective collaboration between the machine and its human partner, primarily through two architectures. 1) Switching control: Authority is transferred distinctly between full human control and full automation based on conditions like task state and operator status. The key lies in smooth, safe timing judgment and transition for switching. 2) Shared control: Humans and machines jointly apply control inputs within the same time frame, representing the mainstream paradigm for deep intelligent integration. Shared control design methods can be further divided into model-based design and model-free design. Model-based design: Utilizing the aforementioned human models, machine collaborative strategies are designed within frameworks like dynamic game theory (Stackelberg, Nash, cooperative games) and model predictive control. This approach is logically rigorous but depends on model accuracy. Model-free design: This does not rely on precise human models but directly uses human input measurements to achieve collaboration through methods like dynamic control authority allocation (e.g., based on fuzzy rules, optimized weights) and reinforcement learning (especially deep reinforcement learning). It offers high flexibility but faces challenges in stability and safety assurance.

Research on human-machine collaborative augmented intelligent control has achieved substantial results in theoretical modeling (encompassing control theory, AI, probability, and other paradigms), collaborative architectures (switching and shared), control design (model-based and model-free), and extensions to multi-agent systems, showing great application potential in several key fields. However, challenges remain, including the lack of a unified theoretical framework, insufficient model generalization and safety assurance in complex scenarios, and an evaluation system for collaboration efficiency that needs further refinement. Future research will be propelled by emerging AI technologies, presenting the following trends: 1) Model-data hybrid driving: A new paradigm of hierarchical collaborative control will emerge, combining the strong interpretability and safety constraint embedding capability of model-based methods with the powerful environmental perception, intent understanding, and end-to-end mapping capabilities of data-driven methods, particularly large model technology. 2) Embodied learning and situational understanding: Machines will evolve from passive interaction to active "embodied" learning. Through continuous interaction with the physical environment and humans, a deeper understanding of task contexts and human intentions will be achieved, enabling more adaptive and personalized collaborative behaviors.