This paper reviews the modelling progress of the vehicle routing problem with drones (VRPD) for last-mile delivery, and proposes an integrated review framework structured around four pillars: scenario drivers, parameter characterization, modelling fundamentals and model extensions. First, it examines the heterogeneous truck-drone collaborative characteristics and differentiated optimization requirements across representative last mile delivery scenarios, including express parcel delivery, instant delivery, and emergency delivery. Second, this paper summarizes the main description approaches for core parameters of resources and operations, time and energy, and road networks and demand, elucidating their roles in shaping feasibility boundaries, collaboration intensity, and cost/energy structures. Next, it develops a taxonomy and reusable modelling building blocks covering optimization objectives, constraint systems, and collaboration structures. Finally, this paper conducts a comparative analysis of key model extensions, including multi-truck and multi-drone VRPD, heterogeneous VRPD, dynamic VRPD, stochastic VRPD, multi-visit VRPD, split-delivery VRPD, pickup and delivery VRPD, multi-depot VRPD. The review indicates several common gaps in the literature: the demand structures, service commitments, and operational processes of different delivery scenarios have not yet been clearly mapped to the selection of collaboration modes, objective formulations, and constraint modules; parameter characterization is often static and homogeneous and responds inadequately to fluctuations in traffic, weather, airspace availability, and demand; operational processes, platform rules, regulatory constraints, and multi-dimensional objective trade-offs are still insufficiently modelled; and extended models still lack unified expressions for key assumptions, synchronization rules, and feasibility structures. Future research should therefore focus on strengthening scenario-to-model mappings, developing dynamic parameter systems, integrating objectives and constraints with real operational mechanisms, standardizing the expression of critical assumptions and synchronization rules, and constructing modular, reusable modelling components.