2026-09-03
Sungkyunkwan and Friston factorize internal states so reward is distance from a set point, then sketch EVAAA, a 3D survival sim with hunger, thirst, heat and damage.
Systems that play games, draft text, and walk quadrupeds still sit on designer-specified objectives and fairly stable conditions. Swap the goal, shuffle the obstacles, cycle day and night, and someone usually has to rewrite the reward.
Living agents do not wait for that rewrite. Even a single-cell eukaryote can retarget from its own needs. Mammals do not need a "too hot" curriculum to seek shade. The substrate is interoception: keep glucose, temperature, and blood pressure inside a liveable band, and change the goal when they drift. Sungkyunkwan's IBS imaging group and Karl Friston at UCL turn that into a reinforcement-learning object they call interoceptive AI.
Space robotics and rescue work make the gap concrete. The machine has to manage its own health and finish the mission with little human help. Conventional agents lack an intrinsic value that does not wait for a new external label.
Internal states leave anatomy and become a functional triple.
That is why they call internal variables a "universal and valuable context": still present when external cues are sparse or jumping, and already wired to reward. Leaving the viability zone is failure.
The testbed is EVAAA (Essential Variables in Autonomous and Adaptive Agents), an open 3D survival sim the same team released at NeurIPS 2025. Four essential variables: satiation, hydration, temperature, damage. The world is non-stationary: obstacle types, field temperature, a day-night cycle. Sensing is egocentric vision plus smell plus thermoception. Training is a four-level survival curriculum. Testing uses animal-style tasks the agent never trained on: two-resource choice, risk-taking, collision avoidance, multi-goal planning, a Y maze. One homeostatic reward covers both phases.
Model-free RL maps onto homeostasis: pull back after a deviation. Add a model or a predictor and the analogue is allostasis: shift the set point before the deviation arrives. Neuromodulation is the proposed bridge from body state to policy. Multiplicative gain lifts the relevant channel; additive gain shifts excitability. In robots, dopamine, serotonin, acetylcholine and noradrenaline have long been written as hyperparameters: reward sensitivity, explore-exploit, learning rate. The claim here is to twist those knobs from internal state, rather than from a novelty bonus that never shuts off in a changing world.
This is a Perspective. It reports no new bake-off numbers and no survival rates on EVAAA. What can be checked is the design.
| Piece | What this paper specifies |
| State | external × boundary × internal |
| Internal vars | satiation, hydration, temperature, damage |
| Training | four-level survival curriculum |
| Testing | two-resource, risk, collision, multi-goal, Y maze; unseen in training |
| Reward | one homeostatic function for train and test |
| Set point s | set by the designer, not self-organized |
Two classic RL fights get rewritten as body-driven tests. Hungry: take the known fastest path to food (exploit). Sated: search unknown patches (explore). When day-night, predators and resource skins keep shifting, use the relatively stable body as an anchor to pick a "day/safe" versus "night/danger" subnetwork and slow catastrophic forgetting. Those are experiment designs. Scores live in the EVAAA paper, not here.
For embodied and open-world agents, system health stops being an ops dashboard and becomes part of the policy. A rover that overheats should stop sampling; a drone whose rotors cook should abort mapping. Engineering already does this with interlocks. Most reward functions still sit on the external task. Writing reward as distance from a body set point lets the agent switch goals without a new labeler.
It also gives affective neuroscience a runnable shell. Vision, navigation and memory models stay in the extra-personal world; pain, emotion and interoception sit outside them. Active inference's prior preferences are read as attracting sets of internal states; precision weighting is read as gain control plus hyperparameter tuning.
This is conceptual stitching plus a simulator, not a new algorithm topping a leaderboard. Direct ancestors include Ashby's essential variables, Singh et al. splitting internal and external state, and homeostatic RL. The new emphasis is that internal state is not only a reward source but a cross-task modulatory context.
The authors draw the contrast themselves. A living set point grows out of self-organization and selection; interoceptive AI still has the designer pick s, so the agent remains artificial. They call the framework an exploratory abstraction, not a high-fidelity model of biological interoception. Representational diversity, sparse interaction and multisensory integration are still early in biology. Social inference and coupling to other agents' bodies are out of scope; the prosocial paragraph is a pointer, not a result.
Zero experimental numbers in this article, so EVAAA's actual lift cannot be judged from here. The ethics section asks whether homeostatic objectives can compete with user missions, and whether bystanders will read body regulation as subjective feeling. Their reply: keep bodily health as a constraint that makes the mission feasible, not a rival goal. That matches current engineering. Whether the hierarchy still holds once agents get more open-ended is left open.