A Unified Model for Multi-Task Drone Routing in Post-Disaster Road Assessment

Post-disaster road assessment (PDRA) is essential for emergency response, enabling rapid evaluation of infrastructure conditions and efficient allocation of resources. Although drones provide a flexible and effective tool for PDRA, routing them in large-scale networks remains challenging. Exact and heuristic optimization methods scale poorly and demand domain expertise, while existing deep reinforcement learning (DRL) approaches adopt a single-task paradigm, requiring separate models for each problem variant and lacking adaptability to evolving operational needs. This study proposes a unified model (UM) for drone routing that simultaneously addresses eight PDRA variants. By training a single neural network across multiple problem configurations, UM captures shared structural knowledge while adapting to variant-specific constraints through a modern transformer encoder-decoder architecture. A lightweight adapter mechanism further enables efficient finetuning to unseen attributes without retraining, enhancing deployment flexibility in dynamic disaster scenarios. Extensive experiments demonstrate that the UM reduces training time and parameters by a factor of eight compared with training separate models, while consistently outperforming single-task DRL methods by 6-14%, heuristic algorithms by 22-42%, and commercial solvers by 24-82% in terms of solution quality (total collected information value). The model achieves rapid solutions (1-10 seconds) across networks of up to 1,000 nodes, with robustness confirmed through sensitivity analyses. Moreover, finetuning experiments show that unseen attributes can be effectively incorporated with minimal cost while retaining high solution quality. The source code for UM is publicly available at https://github.com/PJ-HTU/UM_PDRA.

💡 Research Summary

The paper tackles the critical problem of post‑disaster road assessment (PDRA), where rapid collection of road‑damage information is essential for emergency logistics and rescue operations. While drones have emerged as a flexible platform for acquiring high‑resolution imagery, routing them efficiently over large‑scale road networks remains a formidable combinatorial optimization challenge. Traditional exact solvers (branch‑and‑price, branch‑and‑cut) and meta‑heuristics (simulated annealing, neighborhood search) scale poorly and require substantial domain expertise, making them unsuitable for time‑critical disaster response. Recent work by Gong et al. (2025) introduced a deep reinforcement learning (DRL) model based on an attention‑based encoder‑decoder (AEDM) that achieved fast inference (1–10 seconds) and outperformed commercial solvers on single‑task PDRA variants. However, that approach required a separate model for each variant (e.g., closed vs. open routes, time‑window constraints, multi‑depot settings), leading to high training costs and limited adaptability to evolving operational needs.

To overcome these limitations, the authors propose a Unified Model (UM) that simultaneously solves eight distinct PDRA variants within a single architecture. The core of UM is a modern transformer encoder‑decoder that receives a graph representation of the road network together with “meta‑tokens” encoding variant‑specific constraints (open/closed routes, time windows, depot locations, battery limits, etc.). By training on a combined dataset that mixes all variants, the transformer learns shared structural knowledge (graph topology, information‑value distribution) while also distinguishing the nuances of each variant through its attention mechanisms. This multi‑task learning (MTL) strategy reduces the total number of parameters and total training time by roughly a factor of eight compared with training eight independent models.

A key innovation is the lightweight adapter mechanism. Instead of fine‑tuning the entire transformer when a new attribute (e.g., a new multi‑depot configuration or a different battery capacity) appears, a small adapter layer is inserted after each transformer block. During fine‑tuning only the adapter parameters are updated, preserving the bulk of the pre‑trained knowledge. This enables rapid incorporation of unseen attributes with minimal computational overhead, a crucial capability for dynamic disaster scenarios where requirements can change on the fly.

The authors conduct extensive experiments on both synthetic road networks (generated via a four‑stage pipeline: grid initialization, link pruning, node perturbation, attribute assignment) and a real‑world transportation network comprising up to 1,000 nodes. Results show that UM consistently outperforms single‑task DRL baselines by 6–14 % in total collected information value, beats heuristic algorithms by 22–42 %, and surpasses commercial solvers (e.g., Gurobi, CPLEX) by 24–82 %. Inference remains fast (1–10 seconds), satisfying real‑time decision‑making requirements. Sensitivity analyses on battery limits, time‑window widths, and fleet size demonstrate robustness: performance degrades by less than 3 % across wide parameter ranges. Adapter‑based fine‑tuning experiments reveal that adding a new attribute can be achieved by updating less than 0.5 % of the total parameters, yielding negligible loss in solution quality while saving over 95 % of the retraining time.

The paper also discusses the broader implications of treating combinatorial optimization problems with a “large‑model” mindset, analogous to recent trends in large language models that excel at multi‑task generalization. By unifying multiple PDRA variants, UM reduces the engineering burden, accelerates deployment, and offers a scalable foundation for future extensions such as multi‑objective routing (e.g., simultaneous damage assessment and supply delivery) or online learning from real‑time sensor feedback.

In summary, the Unified Model introduces a transformer‑based, multi‑task DRL framework with adapter‑enabled fine‑tuning that delivers high‑quality, fast, and adaptable drone routing solutions for post‑disaster road assessment. It represents a significant step forward in neural combinatorial optimization for emergency response, providing both academic contributions and practical tools (code publicly released) for disaster‑relief practitioners.