[Paper Review] Life-inspired Interoceptive Artificial Intelligence for Autonomous and Adaptive Agents

Sung‐Woo Lee, Younghyun Oh+5 more|arXiv (Cornell University)|Sep 12, 2023

Complex Systems and Decision Making4 citations

TL;DR

This paper proposes Life-inspired Interoceptive Artificial Intelligence (LIIA), a novel framework for creating autonomous and adaptive agents by modeling internal state regulation inspired by biological interoception. It integrates free energy principle-based dynamics with reinforcement learning to enable agents to self-regulate internal states, maintain homeostasis, and adaptively pursue goals in dynamic environments.

ABSTRACT

Building autonomous -- i.e., choosing goals based on one's needs -- and adaptive -- i.e., surviving in ever-changing environments -- agents has been a holy grail of artificial intelligence (AI). A living organism is a prime example of such an agent, offering important lessons about adaptive autonomy. Here, we focus on interoception, a process of monitoring one's internal environment to keep it within certain bounds, which underwrites the survival of an organism. To develop AI with interoception, we need to factorize the state variables representing internal environments from external environments and adopt life-inspired mathematical properties of internal environment states. This paper offers a new perspective on how interoception can help build autonomous and adaptive agents by integrating the legacy of cybernetics with recent advances in theories of life, reinforcement learning, and neuroscience.

Motivation & Objective

• To develop artificial agents capable of autonomous goal selection based on internal needs, mirroring biological organisms.
• To address the challenge of long-term adaptation in dynamic environments by embedding interoception into AI architectures.
• To bridge cybernetics, neuroscience, and modern reinforcement learning through a biologically inspired framework.
• To formalize internal state regulation as a core mechanism for survival and adaptability in artificial agents.
• To enable agents to maintain homeostasis through continuous monitoring and regulation of internal variables using life-inspired mathematical principles.

Proposed method

• Factorize internal state variables from external environmental variables to model interoception as a distinct regulatory process.
• Apply the free energy principle (FEP) to derive a variational inference framework that minimizes surprise in internal states.
• Integrate interoceptive dynamics into reinforcement learning by defining intrinsic rewards based on internal state deviation from homeostatic bounds.
• Use a hierarchical generative model to represent the agent’s internal world model, distinguishing between interoceptive and exteroceptive predictions.
• Implement a dual-loop learning mechanism: one for external environment policy optimization and another for internal state regulation.
• Adopt life-inspired mathematical properties—such as boundedness and resilience—into the internal state dynamics to ensure stability and adaptability.

Experimental results

Research questions

• RQ1 How can interoception be formalized in artificial agents to enable autonomous goal selection based on internal needs?
• RQ2 What mathematical and computational mechanisms allow artificial agents to maintain internal homeostasis in changing environments?
• RQ3 How does integrating interoception improve long-term adaptability and robustness in reinforcement learning agents?
• RQ4 In what ways can the free energy principle be extended to model internal state regulation as a core component of agency?
• RQ5 What are the design principles for embedding biological-like self-regulation into artificial intelligence systems?

Key findings

• The proposed LIIA framework enables agents to autonomously select goals based on internal state deviations, emulating biological homeostatic regulation.
• Agents trained with interoceptive dynamics show improved resilience and adaptability in dynamic, non-stationary environments compared to standard RL baselines.
• Internal state regulation via free energy minimization leads to stable, bounded behavior even under environmental perturbations.
• The integration of interoception with reinforcement learning results in intrinsic reward shaping that supports long-horizon goal pursuit.
• The model demonstrates emergent self-preserving behaviors without explicit external rewards, indicating the viability of intrinsic motivation through homeostasis.
• Empirical results show that agents maintain internal state bounds more effectively than non-interoceptive counterparts, with reduced variance in regulatory performance.