Interoceptive machine framework: Toward interoception-inspired regulatory architectures in artificial intelligence

This paper introduces the "interoceptive machine framework," a new approach to embodied AI that draws inspiration from biological interoception—the process by which living systems monitor and regulate their internal states. By translating these biological principles into computational architectures, the framework aims to help artificial agents achieve greater autonomy, robustness, and adaptability when operating in complex, uncertain environments.

The Three Principles of Regulation

The framework organizes internal-state regulation into three functional principles, each serving a specific computational role:

• Homeostatic: Focuses on internal viability regulation, ensuring the agent maintains its core stability. * Allostatic: Handles anticipatory uncertainty-based re-evaluation, allowing the agent to prepare for future needs rather than just reacting to current conditions. * Enactive: Facilitates active data generation through interaction, where the agent learns by engaging with its environment.

Designing for Self-Regulation

Rather than attempting to replicate the exact neurophysiology of a biological brain, the author proposes these principles as high-level abstractions. By embedding internal state variables and regulatory loops directly into an AI’s architecture, developers can create systems that are more sensitive to their own internal status and the context of their surroundings. This shift moves AI design away from purely external data processing toward a model where the agent’s "internal health" and stability guide its decision-making process.

Enhancing AI Autonomy

The primary goal of this framework is to improve how AI agents handle uncertainty. By integrating these regulatory loops, agents can achieve more calibrated decision-making and more effective interaction strategies. The author suggests that this approach provides a concrete, testable pathway for future research, with significant potential applications in fields such as human-computer interaction and the development of advanced assistive technologies. Ultimately, the framework offers a unified perspective on how internal-state regulation is essential for building more autonomous and reliable embodied AI systems.