Bayesian belief-to-action planning for truck–drone emergency routing

Abstract

Post-disaster truck–drone delivery is planned while road access is only partly known and may recover during the response. Existing disruption-aware routing models usually treat accessibility as fixed before planning, which limits their ability to use field confirmations and command-center updates during execution. This paper proposes a Bayesian belief-to-action policy for truck–drone emergency routing under progressive access recovery. Sparse access observations update a discrete-time recovery belief, which is then used to generate short-horizon recovery scenarios. These scenarios are solved by an adaptive large neighborhood search, and a consensus repair rule commits only the current-stage truck–drone action that is feasible under the observed access state. The policy therefore converts access reports into executable decisions rather than treating them only as status records. The primary experiment uses a locked set of 30 instances, each with 30 demand nodes, and a generic constant-hazard prior that does not contain the generator’s recovery profile. The belief-guided policy remains feasible in all 30 instances, whereas the myopic policy fails in two; on the 28 feasible paired instances it improves the objective in 27 cases (Wilcoxon $$p<10^{-7}$$ ), mainly by lowering weighted service delay. The policy’s advantage is best interpreted at the system level: the frozen-hazard ablation shows no significant incremental effect of hazard-count updating, while status revelation, posterior scenario lookahead, and rolling commitment operate together in the full policy. Fixed spatial pooling offers no significant gain, adaptive-shrinkage pooling is significantly worse, and a generator-profile prior does not significantly improve on the generic independent-component belief in this sample. Cluster coupling is therefore retained as an optional, signal-dependent extension rather than assumed in the operational default.

The authors' abstract, as published at the source. Discover Artificial Intelligence, 2026 · DOI ↗