Viability‑Constrained Reinforcement Learning via Homeostatic State Augmentation

This work introduces a reinforcement learning framework in which the agent’s state is augmented with an internal viability variable representing structural and computational homeostasis. The resulting viability‑augmented MDP incorporates a logarithmic barrier in the reward, allowing safety to emerge directly from the agent’s internal dynamics rather than from external constraints. We prove that the associated Bellman operator remains a contraction under standard boundedness assumptions, ensuring the existence and uniqueness of the optimal value function. Experiments on stochastic control benchmarks show improved survival, reduced collapse frequency, and greater stability under high‑variance conditions.