An AI framework gives virtual agents internal states to adapt

An AI agent in a three-dimensional virtual world sees, smells, senses temperature and detects collisions while four internal variables move in the background: satiation, hydration, body temperature and damage. Researchers led by Woo Choong-Wan at the Institute for Basic Science and Sungkyunkwan University built this survival testbed to study a new approach called interoceptive AI, which gives those internal states a role in learning and decision-making.

The idea starts with a familiar gap between biology and conventional AI. Living organisms adjust priorities according to what is happening inside their bodies; an external cue can mean something different to a hungry animal than to a satiated one. Most AI agents instead pursue objectives defined from outside. Robots may already detect a low battery, an overheating motor or a damaged component, but usually respond through predefined rules. Interoceptive AI treats internal conditions as continuing context, not only as faults or reward inputs.

The researchers mathematically formalized how internal and external states can remain separate while influencing one another. In EVAAA—Essential Variables in Autonomous and Adaptive Agents—agents must keep their four internal variables within viable ranges while navigating resources, obstacles, predators, changing temperatures and day-night cycles. Separate test environments introduce conditions absent from training, making adaptation to unfamiliar situations part of the evaluation.

The framework also sets up decisions that resemble animal-behavior experiments. An agent may have to choose between competing resources according to its current needs, or weigh the benefit of obtaining one against the risk of physical damage. The researchers suggest that internal states could also tune learning rates, sensitivity to information and the balance between exploring new options and exploiting known ones, while helping agents learn without overwriting earlier knowledge.

Concretely, the immediate change is for AI and robotics researchers: they get a virtual benchmark for testing whether an agent can carry a stable internal reference from one environment to another. The work is published in Nature Machine Intelligence and presents a proposed framework and a virtual benchmark.