Living organisms constantly adjust their behavior according to what is happening inside their bodies. Hunger can shift attention toward food, while cold can make finding warmth more urgent. These internal signals remain meaningful even when the surrounding environment changes completely.

Artificial intelligence generally works differently. Most AI agents pursue objectives defined from the outside, and adapting those objectives or behaviors to changing conditions remains a major challenge.

A research team led by Associate Director WOO Choong-Wan of the Center for Neuroscience Imaging Research within the Institute for Basic Science (IBS) and Sungkyunkwan University, Professor HONG Seok Jun of Sungkyunkwan University, and Professor Karl J. FRISTON of University College London has proposed a new framework inspired by the way living organisms monitor and regulate their internal condition.

Called “interoceptive AI,” the framework gives artificial agents explicitly defined internal states and allows those states to serve as context for learning and decision-making. The researchers mathematically formalized how internal and external states can be separated while still interacting, and how maintaining internal stability can contribute to an agent’s rewards and behavior. Rather than treating an internal condition simply as something to monitor, the framework makes it part of the information an agent uses to interpret its surroundings, evaluate outcomes, and decide what to do next.

Current AI and robotic systems already monitor internal conditions in limited ways. A robot, for example, can detect a low battery, overheating motor, or damaged component and respond according to predefined rules. Reinforcement-learning agents can also learn complex behaviors from reward signals supplied by human designers.

The interoceptive AI framework takes a different approach. Instead of treating internal conditions only as faults to detect or variables used to calculate rewards, it treats them as a continuing source of context that can influence what an agent learns, prioritizes, and does.

This distinction becomes especially important in changing environments. External cues may disappear, become unreliable, or take on different meanings, while an agent’s internal condition remains continuously available. The researchers therefore describe internal states as a “universal and valuable context”: universal because they remain available across different situations, and valuable because they are directly tied to the agent’s internal needs. A resource that is unimportant when those needs are already satisfied, for example, may become highly valuable when an essential variable approaches a critical limit.

To study this idea, the researchers developed EVAAA (Essential Variables in Autonomous and Adaptive Agents), a three-dimensional virtual survival benchmark. Artificial agents operate in an environment containing resources, obstacles, predators, changing temperatures, and day-night cycles while regulating four internal variables—satiation, hydration, body temperature, and damage—within viable ranges.

The benchmark also includes tasks inspired by animal-behavior experiments. In one, agents must choose between competing resources according to their current internal needs. In another, they must weigh the benefit of obtaining a resource against the risk of physical damage. Because the same internal-state framework applies across different environments and tasks, researchers can test whether agents can use their internal condition as a stable reference when external circumstances change.

The researchers argue that this principle could also help address two persistent challenges in artificial intelligence: deciding when to explore new possibilities and when to rely on existing knowledge, and learning new information without overwriting what has already been learned. In living organisms, both processes are strongly influenced by internal conditions. A hungry animal, for example, may exploit a known food source, whereas a satiated animal has greater freedom to explore unfamiliar territory.

In a sense, this approach incorporates ideas from neuromodulation, the biological mechanisms through which internal states can alter perception, learning, memory, and decision-making. In artificial agents, similar modulatory mechanisms could allow internal states to adjust factors such as learning rates, sensitivity to information, or the balance between exploration and exploitation.

LEE Sungwoo, a student researcher and the first author, said, “The key to this study is that we treated internal state not simply as a signal to monitor, but as a context for judgment.” He added, “We plan to study how bodily damage changes an agent’s attention and planning.”

The approach could be particularly useful for “physical AI,” in which artificial intelligence operates through robots in the real world. Robots already use sensors to monitor conditions such as battery levels, motor temperature, mechanical wear, and damage. Interoceptive AI would go a step further by allowing these internal conditions to influence learning and decision-making rather than simply triggering predefined responses.

For example, a robot that has damaged one of its components could learn to place less stress on that part, avoid routes associated with previous collisions, or alter what information it prioritizes while navigating. In this way, the robot’s changing physical condition could become part of the context it uses to determine its next action.

The researchers emphasize that the framework is a simplified computational abstraction of biological interoception rather than an attempt to reproduce the full complexity of a living body or brain. EVAAA likewise provides a controlled test environment for investigating how artificial agents use internal-state information under changing conditions.

The framework also raises questions about how increasingly autonomous agents should balance internal requirements with goals set by humans. Internal regulation does not necessarily have to compete with an assigned mission. A planetary rover, for example, may temporarily suspend an operation when its battery or temperature reaches an unsafe level because maintaining its operational condition is necessary for completing the larger mission.

The researchers also caution against interpreting internal-state regulation as evidence of consciousness or subjective experience. Existing robots already monitor battery levels, temperature, and mechanical strain without implying that they experience those conditions. In the proposed framework, interoception is instead a computational mechanism for regulating learning and behavior.

Associate Director WOO Choong-Wan said, “For living organisms, the internal environment and the ability to regulate it are central to the intelligent behavior they display.” He added, “This study translates principles underlying intelligence in living systems into a form that AI can use, providing a starting point for research on autonomy and adaptivity.”

The study was published online in Nature Machine Intelligence on August 26. EVAAA, a three-dimensional virtual survival benchmark that provides one implementation of the interoceptive AI framework, was previously presented at NeurIPS 2025 and published as a conference paper.