Machine Learning and the Internet of Things for Optimisation of the Last Mile Delivery

Project Information

SmartDelivery

Grant agreement ID: 101110022

DOI

10.3030/101110022

Project terminated on 18 January 2026

EC signature date 27 April 2023

Start date 1 September 2024

End date 28 February 2027

Funded under

• Marie Skłodowska-Curie Actions (MSCA)

Total cost

No data

EU contribution

€ 206 641,20

Coordinated by

UNIVERSIDAD DE SALAMANCA
Spain

Periodic Reporting for period 1 - SmartDelivery (Machine Learning and the Internet of Things for Optimisation of the Last Mile Delivery)

Reporting period: 2024-09-01 to 2027-02-28

Summary of the context and overall objectives of the project

Last-mile delivery (LMD) is the final stage of the logistics process, where goods move from a hub or distribution center to the customer. It is widely recognized as the most expensive and operationally challenging segment of the logistics chain, accounting for up to 41% of total logistics costs. Beyond direct operational expenses (fuel, maintenance, wages), LMD generates significant negative externalities, including urban congestion, accidents, pollution, and infrastructure wear. The rapid expansion of e-commerce, accelerated by post-pandemic shifts in consumer behavior, has intensified the urgency of addressing the inefficiencies of the LMD process. This aligns with the European Green Deal and UN Sustainable Development Goals (SDGs 3 and 11), which aim to foster sustainable mobility and improve citizens' quality of life. From an operational research perspective, route planning in LMD is typically modeled, in its baseline academic form, through the Capacitated Vehicle Routing Problem (CVRP), and scientific literature has proposed several novel methodologies to solve CVRP instances. However, these theoretical advances often fail to produce the expected results in real-world scenarios. This is largely caused by inaccurate travel data estimates (e.g. derived from online open maps) that fail to account for the unpredictable events occurring during deliveries. Furthermore, the selection of the most suitable solver for specific problem instances is frequently manual and biased. Moreover, standard solvers typically ignore the "human factor" (specifically, driver familiarity with delivery zones) which directly affects service times and customer experience.
The SmartDelivery project aims to bridge these gaps by developing a novel, layered platform for LMD optimization structured around three core technical objectives: (1) the design of a Machine Learning-based algorithm selection module to automatically choose the most suitable heuristic or metaheuristic solver from a portfolio based on instance features and performance metrics, overcoming the limitations of manual selection; (2) the implementation of a robust methodology to replace inaccurate, static travel time data with real-time estimations via custom IoT devices, ensuring high-fidelity inputs for the routing process; and (3) the development of a novel "sixth sense" concept to dynamically assign routes to drivers based on their efficiency and familiarity with delivery zones, replacing standard random assignment strategies and significantly reducing service times.
By integrating Machine Learning (ML) techniques with real-world sensing, SmartDelivery aims to significantly enhance LMD efficiency, thereby reducing operational costs for logistics companies and mitigating environmental externalities, thus contributing to the realization of smarter, greener, and more efficient European cities.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

Before its early termination, the SmartDelivery project delivered three significant contributions to advance Last Mile Delivery (LMD) optimization with ML-based algorithm selection, realistic travel-time costs estimation and driver-aware route assignment.
First, the project implemented a travel-time–based CVRP instance generation mechanism that replaces purely Euclidean costs with a traffic-aware travel-time cost matrix. To avoid a simple linear proportionality to distance, a novel parametrized traffic model was designed to incorporate congestion effects, intersection-related delays, and weather-induced latencies, producing realistic travel time simulations. The generator script also enables the systematic variation of heterogeneous geometric characteristics, including depot positioning, customer distribution patterns, demand profiles, and route sizes. It was utilized to generate a large-scale, balanced synthetic dataset comprising 10,000 CVRP instances ranging from 30 to 200 customers. Instances are exported in a CVRPLIB-like format with explicit full travel-time matrices and metadata in the instance header, enabling traceable and reproducible experiments. This work establishes the necessary data infrastructure for the subsequent training of the ML-based algorithm selection module, which was planned but not implemented within the project period.
Second, the project provided the technical specifications for the SmartPos IoT prototype, designed to overcome the limitations of static travel time estimates derived from open maps. The device is intended to be installed on each fleet vehicle, where it continuously captures and transmits real-time travel data to a central cloud infrastructure. Furthermore, the project defined a data processing framework to organize such observations into specific, continuously updated time-window buffers (e.g. morning, afternoon, evening). This system prioritizes the latest measurements by discarding the oldest entries when buffers are full, and subsequently employs a Linearly Weighted Moving Average to calculate travel costs. This approach ensures that routing algorithms utilize dynamic estimates derived from recent operational data rather than static open map values.
Third, the project formalized a novel route assignment methodology centered on the "sixth sense" parameter, mathematically defined as the ratio between the global minimum recorded service time and a specific driver's service time at a destination. The route assignment process was modeled as an Integer Programming problem designed to maximize the total sixth-sense value across the fleet. To validate this approach, a Python-based simulation tool was developed. The software ingests CVRP solutions in JSON format, generates simulated driver–destination familiarity matrices, and solves the optimization model under standard assignment constraints (single driver-to-route mapping). The prototype outputs a structured assignment plan and a heatmap visualization to analyze the distribution of driver-route familiarity.
Finally, a complementary research line addressed the challenge of "micro-navigation" at delivery sites, identified as a critical factor affecting service time. In large-scale infrastructures (e.g. airports, hospitals, multi-building campuses), a significant portion of service time is consumed identifying access points and navigating interior spaces. Consequently, the project conducted a focused study on indoor positioning technologies and proposed a hybrid, easy-to-deploy indoor localization architecture designed to guide drivers in these complex environments. While distinct from the core SmartDelivery platform, this component is conceptually consistent with the project's goal of minimizing service time, offering a further layer of optimization of LMD in complex delivery scenarios.
Results were published in high-impact peer-reviewed journals and presented at international conferences and workshops. Furthermore, to ensure transparency and support reproducibility, all developed scripts, the generated dataset, and the related technical documentation have been made freely available on GitHub. The dataset is also permanently archived and accessible via the Zenodo repository.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

SmartDelivery advanced the state of the art by defining a comprehensive, data-driven, and driver-centric architecture that integrates three traditionally isolated components: an ML-based Algorithm Selection module to identify the most suitable solver per instance, mitigating manual selection bias; a data processing methodology that replaces static map-based estimates with real-time travel times collected via IoT devices, ensuring routing decisions reflect actual operational conditions; and a novel “sixth sense” metric that quantifies driver familiarity. This shifts route assignment from random allocation to an optimization model integrating human expertise, thereby significantly reducing service times and operational inefficiencies.
The project also developed and released an open-source CVRP instance generator that models routing costs as travel times rather than standard Euclidean distances between destinations, bringing synthetic instances closer to operational settings. The generator implements a parameterised travel-time model that accounts for dynamic factors such as traffic congestion, intersection delays, and weather conditions, while preserving key geometric properties essential for representing realistic routing scenarios.
It outputs VRPLIB-format CVRP instances with travel-time costs encoded as FULL_MATRIX, together with per-instance summary statistics to support future machine-learning pipelines. The generator was used to produce a balanced, FAIR-compliant dataset of 10,000 instances, which is freely accessible via GitHub and Zenodo to foster reproducible research within the community.