Multi-Attribute Vehicle Routing Problem

• MAVRP is defined as a routing framework where decisions combine multiple operational attributes including time windows, capacity, distance limits, and more.
• It leverages diverse solution paradigms such as aggregated cost formulations, decomposition, and neural unified solvers to address interdependent constraints.
• Applications span urban logistics, agriculture spraying, and modular platooning, demonstrating its practical relevance in complex, real-world routing scenarios.

Searching arXiv for recent MAVRP-related papers to ground the article in current literature. arxiv_search query="Multi-Attribute Vehicle Routing Problem MAVRP neural routing explainability RouteFinder FiLMMeD MVMoE" max_results=10

arxiv_search query="(Zhang et al., 2020, Berto et al., 2024, Zhou et al., 2024, Corrêa et al., 30 Apr 2026, Afifi, 5 Jul 2026) Multi-Attribute Vehicle Routing Problem" max_results=10

Multi-Attribute Vehicle Routing Problem (MAVRP) denotes a class of vehicle routing problems in which routing decisions depend jointly on several customer, vehicle, temporal, network, and cost attributes rather than only on geometric distance. In the literature summarized here, the term is sometimes explicit and sometimes implicit. “Multi-Vehicle Routing Problem with Soft Time Windows” is described as a special-case MAVRP centered on temporal and capacity attributes (Zhang et al., 2020), whereas later neural work studies MAVRP directly as a unified routing setting in which capacity, time-window, distance-limit, backhaul, and open-route constraints may be activated simultaneously (Afifi, 5 Jul 2026). Foundation-model and multi-task solvers adopt the same compositional view by treating named VRP variants as subsets of a generalized attribute set rather than as unrelated problems (Berto et al., 2024, Zhou et al., 2024).

1. Conceptual scope and relation to adjacent VRP classes

MAVRP is best understood as a structural generalization of the VRP in which multiple attribute layers coexist inside one routing decision process. These attributes may be customer-side, such as linehaul demand, backhaul demand, service time, or time windows; vehicle-side, such as capacity, fleet type, or starting depot; route-side, such as open-route behavior or duration limits; or interaction-side, such as synchronization and replenishment. The unified MAVRP setup studied in recent explainability work is a directed graph $G=(V,E)$ with one depot and customers, where “any subset of constraint families may be active,” specifically capacity, time windows, distance limit, backhaul, and open-route constraints (Afifi, 5 Jul 2026).

This broad interpretation distinguishes MAVRP from several neighboring notions. First, MAVRP is not identical to a single named “rich VRP” variant. The surveyed literature includes special-case MAVRPs focused on soft time windows (Zhang et al., 2020), service levels (Bulhões et al., 2017), fuel constraints (Sundar et al., 2015), stochastic fleet-sizing and depot positioning (Corredor-Montenegro et al., 2021), modular platooning (Zhou et al., 2024), and synchronization with replenishment support vehicles (Alkaabneh, 2023). Second, MAVRP is not necessarily a multi-objective problem. Many papers optimize a single scalar cost under multiple attribute-based constraints; true Pareto-style formulations are present, but they are a subset rather than the default (Meira et al., 2017).

A recurrent modeling theme is that classical VRP variants become points in an attribute space. RouteFinder explicitly groups attributes into node attributes, global attributes, and edge attributes, then defines 24 variants by combining $O$, $B$, $L$, $TW$, and $M$ on top of base capacity (Berto et al., 2024). MVMoE uses the same compositional idea over the attribute set $\{C,O,B,L,TW\}$, training on some combinations and testing zero-shot on unseen ones (Zhou et al., 2024). FiLMMeD extends this logic to the multi-depot setting, using a Boolean task descriptor $z\in\mathbb{R}^5$ for $B,L,O,TW,I$ and evaluating 24 MDVRP variants, including 8 inter-depot formulations (Corrêa et al., 30 Apr 2026). This suggests that, in current neural optimization practice, MAVRP is increasingly treated as a problem family generated by attribute composition.

2. Modeling primitives, objectives, and constraint semantics

The surveyed literature does not present a single canonical MAVRP mixed-integer formulation. Instead, it exhibits several recurrent modeling patterns. One is the aggregated single-objective formulation: the soft-time-window multi-vehicle problem of (Zhang et al., 2020) defines an instance

$s=(v,d,e,l,\alpha,\beta),$

with multiple identical vehicles, demands, capacities, one depot, Euclidean travel, and soft time windows. Its route cost is

$O$0

where the penalty term is

$O$1

In MAVRP terms, this is multi-attribute because spatial, load, and temporal attributes are coordinated inside one routing policy, but it remains single-objective because these terms are aggregated into one scalar criterion (Zhang et al., 2020).

A second pattern is the constrained selective-routing formulation. In the Vehicle Routing Problem with Service Level Constraints, customers belong to disjoint groups $O$2, and service obligations are imposed by

$O$3

The objective minimizes routing cost plus lost profits from unserved customers, so service heterogeneity enters as group-indexed constraints rather than as a separate objective (Bulhões et al., 2017). This is a clear example of MAVRP as attribute-indexed feasibility.

A third pattern is strategic-operational coupling under uncertainty. The Vehicle Fleet Sizing, Positioning and Routing Problem with Stochastic Customers uses a two-stage stochastic program,

$O$4

where first-stage decisions open depots and allocate vehicles, and the second stage solves a multi-depot routing problem after customer activity and demand are realized (Corredor-Montenegro et al., 2021). Here, the “attributes” include facility opening, depot-specific fleet sizing, stochastic customer activation, and time-window and capacity feasibility inside the recourse problem.

A fourth pattern is genuine multi-objective MAVRP. PostVRP formulates

$O$5

where $O$6 is total length, $O$7 is number of vehicles, and $O$8 is route-length dispersion, subject to a maximum route-length constraint (Meira et al., 2017). This directly contradicts the common misconception that all MAVRPs are scalarized. Some are; some are not.

Taken together, these formulations show that “multi-attribute” refers to heterogeneity in operational dimensions, not to a single objective architecture. Attributes may appear as objective terms, hard constraints, dynamic state variables, route-state recursions, or task descriptors. This suggests that the most stable conceptual distinction is between attribute multiplicity and objective multiplicity.

3. Major attribute families

The literature covers several recurring attribute families. They differ in whether they alter feasibility, route-state dynamics, or arc costs, but in all cases they change the routing problem’s combinatorial structure.

Attribute family	Representative structure	Representative paper
Temporal	Soft or hard time windows, service times, time-dependent travel, synchronization	(Zhang et al., 2020)
Resource and energy	Capacity, fuel, battery-distance limits, replenishment, backhaul load evolution	(Sundar et al., 2015)
Fleet and network structure	Multiple depots, heterogeneous technologies, open routes, inter-depot movement	(Corrêa et al., 30 Apr 2026)
Service-mode choice	Door-to-door vs SDL, customer acceptance of AEV or SDL service	(Zhao et al., 7 Mar 2025)
Interaction and synchronization	Refill meetings, support-vehicle coordination, multiple synchronization constraints	(Alkaabneh, 2023)
Vehicle configuration	Docking, splitting, platoon size, modular shared travel costs	(Zhou et al., 2024)
Strategic uncertainty	Depot positioning, fleet sizing, stochastic customers and demands	(Corredor-Montenegro et al., 2021)

Temporal attributes range from soft windows, where earliness and lateness are penalized but not forbidden (Zhang et al., 2020), to strict windows in hybrid AEV/shared-location routing with time-dependent travel (Zhao et al., 7 Mar 2025), to endogenous synchronization times in precision-agriculture replenishment routing, where refill start is constrained by both sprayer completion and tanker arrival (Alkaabneh, 2023). This indicates that “time” in MAVRP is not one attribute but a family of attribute types: exogenous windows, endogenous waiting, travel-time dependence, and synchronization.

Resource attributes likewise vary in semantics. Fuel-constrained multiple-depot routing treats fuel as a route-state variable reset by depot refueling, with reachability quantities

$O$9

used to strengthen exact formulations (Sundar et al., 2015). Pickup-and-delivery with time windows embeds service-state, precedence, and capacity into a multi-dimensional state-space-time network (Mahmoudi et al., 2016). In synchronized spraying, sprayer and tanker inventory states coevolve through refill quantities and support-vehicle trips (Alkaabneh, 2023). A plausible implication is that MAVRP becomes hardest when resources are both cumulative and transferable.

Fleet and service-mode attributes broaden the notion further. FiLMMeD’s MDVRP catalog includes open routes, backhauls, route-length limits, time windows, and inter-depot routes inside a single attribute-conditioned model (Corrêa et al., 30 Apr 2026). The hybrid AEV/conventional problem introduces customer-specific acceptance of AEV delivery and SDL service, paired SDL nodes, driver working-time constraints for conventional vehicles, and maximum route distance for AEVs (Zhao et al., 7 Mar 2025). Modular vehicle routing adds route-state-dependent costs through platoon size $B$0, with arc cost

$B$1

so the economic effect of traversal depends on vehicle coupling decisions (Zhou et al., 2024). This suggests that MAVRP increasingly includes attributes that change the meaning of a route itself, not merely its admissible node set.

4. Exact, decomposition, and metaheuristic solution paradigms

Exact MAVRP methods usually rely on state propagation or decomposition, because the extra attributes destroy the simplicity of classical route partitioning. In fuel-constrained multiple-depot routing, four MILP formulations are proposed, with the strengthened arc-based model $B$2 clearly outperforming lifted node-based alternatives and solving almost all 40-target instances within the one-hour limit (Sundar et al., 2015). In stochastic location-routing-fleet-sizing, a multicut L-shaped algorithm combines a master problem over depot openings and vehicle counts with MDVRP recourse solved by column generation and ESPPRC-R pricing (Corredor-Montenegro et al., 2021). In multi-vehicle pickup-and-delivery with time windows, assignment-routing constraints are embedded directly into a vehicle–space–time–service-state network and solved by forward dynamic programming on local clusters (Mahmoudi et al., 2016). These are different technical strategies, but they share one principle: represent the troublesome attribute as state, not as a weak after-the-fact side constraint.

The same principle appears in heuristic design. The adaptive hybrid genetic search and large neighborhood search method for a warehouse MAVRP treats each individual as a solution plus its own ALNS configuration, then uses an adaptive survival mechanism and a cooperative evolution mechanism to select instance-appropriate parameterizations (Liu et al., 2024). In synchronized multi-trip spraying, ALNS combines destroy operators keyed to waiting time, route structure, and refill positions with a 60-second exact intensification step that fixes sprayer routes and optimizes refill nodes and tanker routing (Alkaabneh, 2023). In modular vehicle routing, tabu search uses an ST-DAG and a Gantt-chart representation together with serial merging, parallel merging, and relocate operators to manipulate platoon structure directly (Zhou et al., 2024). These methods are attribute-aware in a literal sense: neighborhoods are defined over the special structure induced by the attribute set.

The empirical record is consistent with the modeling difficulty. Strategic stochastic MAVRP instances were solved in about 1.7 hours for the expected-value solution and about 5.8 hours for the conservative solution, while 75% of daily MDVRP recourse instances were solved in under 3 minutes (Corredor-Montenegro et al., 2021). Modular vehicle routing’s tabu search matched or improved small-instance MILP solutions while solving instances up to 67 nodes in seconds (Zhou et al., 2024). The synchronized spraying ALNS improved practical solutions by an overall average of 5.61% (Alkaabneh, 2023). This suggests that exact methods remain central for formulation strength and subproblem optimization, whereas rich-instance deployment continues to depend heavily on decomposition-aware heuristics and matheuristics.

5. Neural unified solvers and explainability

Neural MAVRP research has evolved from special-case learned solvers toward unified, attribute-conditioned architectures. The earliest paper in this set, on soft-time-window multi-vehicle routing, treats route generation as a sequential multi-agent decision process and uses an attention encoder-decoder with REINFORCE to optimize travel plus soft-window penalties (Zhang et al., 2020). Its importance for MAVRP is methodological: multiple vehicles, capacities, visited-set status, and soft temporal penalties are coordinated in one autoregressive policy, even though the problem remains a special-case MAVRP and not a general attribute framework.

Later work makes attribute composition explicit. RouteFinder defines a unified environment whose most general form supports open routes, backhauls, duration limits, time windows, and mixed backhauls, then trains a single transformer policy with global attribute embeddings, mixed batch training, and efficient adapter layers for unseen attributes (Berto et al., 2024). MVMoE addresses the same multi-variant family with mixture-of-experts layers in encoder and decoder, training on six variants and evaluating zero-shot on ten unseen combinations; its key claim is that a solver should be defined over the union of attributes and generalize compositionally across them (Zhou et al., 2024). FiLMMeD extends the unified-solver idea to 24 multi-depot variants and 16 single-depot variants, using FiLM conditioning on the active constraint vector, a targeted curriculum over attribute combinations, and preference optimization as an alternative to REINFORCE (Corrêa et al., 30 Apr 2026). Together, these papers operationalize MAVRP as a learned policy over an attribute basis.

Explainability has also become a first-class design axis. The 2026 MAVRP study on encoder probing and decoder attribution analyzes six solver variants combining attention or graph encoders with Hard-Mask or Recourse decoders (Afifi, 5 Jul 2026). On the encoder side it uses linear probes, spontaneous-organization metrics, effective-rank measures, and discovered-direction interventions to test whether latent space organizes active constraint families. On the decoder side it uses gradient, integrated gradients, and DeepLIFT under abductive, contrastive, and counterfactual views, scoring explanations on fidelity, concentration, stability, sanity, and actionability. Two findings are especially important. First, graph inductive bias improves latent predictability and decoder sanity. Second, recourse-trained decoders—not merely softer masks—produce make-feasible counterfactuals that hard-masked policies fail to produce (Afifi, 5 Jul 2026). This suggests that, in neural MAVRP, infeasibility handling is not just a deployment detail but part of the representational problem.

These neural frameworks remain narrower than a fully open-ended MAVRP. Their attribute vocabularies are fixed in advance; feasibility is largely mask-driven; and classical OR solvers often remain stronger on absolute solution quality, especially in RouteFinder’s comparison to HGS-PyVRP and OR-Tools (Berto et al., 2024). Still, they establish a clear design pattern: a common state representation, attribute-conditioned embeddings or gating, dynamic masking, and route-level scalar reward.

6. Applications, benchmarks, and open problems

MAVRP appears across a wide range of operational settings. Urban logistics with soft time windows and capacity constraints is one explicit special case (Zhang et al., 2020). Warehouse CTU scheduling induces a multi-depot, multi-trip, mixed linehaul-backhaul problem with makespan minimization (Liu et al., 2024). Strategic planning under uncertain customers combines depot positioning, fleet sizing, and scenario-based multi-depot recourse (Corredor-Montenegro et al., 2021). Agriculture spraying yields a synchronized support-vehicle routing problem in which a tanker must meet sprayers at refill nodes (Alkaabneh, 2023). Modular logistics uses docking, splitting, and platoon-size-dependent cost savings (Zhou et al., 2024). Hybrid last-mile delivery with AEVs and SDLs adds customer acceptance of autonomous service and alternative service modes to time-dependent routing (Zhao et al., 7 Mar 2025). These settings show that MAVRP is not a niche abstraction; it is a common consequence of operational realism.

Benchmark design reflects the same diversity. RouteFinder studies 24 attribute-composed variants in a unified environment (Berto et al., 2024). FiLMMeD evaluates 24 MDVRP variants plus 16 single-depot variants (Corrêa et al., 30 Apr 2026). MVMoE formalizes seen and unseen combinations of $B$3, $B$4, $B$5, $B$6, and $B$7 for zero-shot transfer (Zhou et al., 2024). PostVRP contributes a large-scale, real-world, multi-objective benchmark on a planar road graph with up to 30,000 deliveries per day, embedding street region, type, zone, side-of-street, crossing cost, and route-length bounds (Meira et al., 2017). This suggests that one unresolved issue in MAVRP is benchmark fragmentation: papers often define strong but local notions of “attribute-rich” routing.

Several open problems recur. One is attribute extensibility. RouteFinder, MVMoE, and FiLMMeD all rely on a fixed attribute basis, even when they adapt to unseen combinations or add one new attribute through adapters (Berto et al., 2024, Zhou et al., 2024, Corrêa et al., 30 Apr 2026). Another is state completeness. In the soft-time-window attention model, the cost depends on arrival times $B$8, but the policy formulas are not fully explicit about how current clock time is represented during decoding (Zhang et al., 2020). A third is interpretability under multiple active constraints. The explainability paper explicitly notes dependence on a supervised concept dictionary and first-order counterfactual search with fixed budgets (Afifi, 5 Jul 2026). A fourth is integration breadth: no single framework in this set jointly handles arbitrary heterogeneous fleets, stochasticity, synchronization, multi-period coupling, and open-ended service modes.

The current literature therefore supports a precise but limited conclusion. MAVRP is best viewed not as one canonical model, but as a family of routing formulations in which multiple operational attributes are structurally active at once. Some instances are single-objective, some multi-objective; some are exact and decomposition-friendly, some are best attacked by specialized heuristics; some are now approached through attribute-conditioned neural policies. The enduring technical challenge is not the presence of many constraints by itself, but the fact that these constraints alter one another’s state transitions, feasible action sets, and economic trade-offs.