Depletion-Gated Reorientation: A Minimal Mechanism for Efficient Spatial Foraging

We study the problem of efficient foraging in patch-depleting spatial environments, where an agent must balance local exploitation with timely abandonment of depleted patches. We propose the Depletion-Gated Pivot (DGP), a minimal GRU-based architecture combining exploration (IEC: full memory updates, high entropy) with a depletion-gated reorientation event: when local viability falls below a threshold, the agent executes a brief high-entropy action selection step, with optional partial memory reset, that disrupts exploitation habits and promotes spatial relocation. We evaluate DGP against ablations and a capacity-matched PPO-GRU baseline across three spatial environments with varied depletion dynamics. DGP consistently outperforms IEC-only exploration and the PPO-GRU baseline in this environment family, while a richer staged variant (DGP + EIL drift phase) is substantially worse than DGP in all environments, confirming that the current EIL-style drift phase is harmful overhead in this task family. DGP fails in a non-spatial bandit setting where pivot is not coupled to physical relocation, establishing clear scope. To identify the key mechanism, we compare DGP against Viability-Conditioned Entropy PPO (VCE-PPO), which uses the same hard viability-triggered gate but replaces the structured pivot action with entropy modulation alone. VCE-PPO reduces overcommitment relative to IEC-only, but fails to improve AUC over IEC-only, whereas DGP reduces overcommitment further and improves AUC. This dissociation shows that the structured spatial pivot action, not entropy modulation alone, is the key performance-improving mechanism in this environment family. This preprint corresponds to a manuscript submitted for peer review.