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.
