Delivery loop optimizer (Multi-parcel delivery optimizer for micro-mobility assistance in the last meters)

AI Quick Summary

Due to the increasing volume of parcels delivered in Switzerland, the need to execute deliveries efficiently is increasingly important. Urban B2C last-mile delivery in dense Swiss cities is the costliest step in the supply chain and is difficult to standardise. Execution efficiency depends not only on last-meter topology, ease of access, and the availability of parking spaces, but also on the driver’s knowledge of the area, ability to make real-time decisions, and physical effort. The proposed solution is a tool that helps delivery drivers execute the last meters more efficiently by recommending suitable parking areas, suggesting which parcels should be grouped for each micro-mobility-assisted delivery loop, and continuously improving its recommendations based on driver decisions and real operational experience. The solution aims to improve last-meter productivity, reduce physical strain, support drivers’ decision-making, and enable a scalable urban B2C delivery system that can keep up with a rising demand. The challenge Urban B2C last-mile delivery in Switzerland is becoming harder to execute efficiently due to the fast growing demand in dense highly urbanised areas. In B2C urban logistics, the most expensive step is the highly variable and inefficient last meters of delivery execution. Efficiency is shaped by factors such as: street conditions, contested parking areas, restricted curb access, mixed use areas with restaurants and high pedestrian activity and inconsistent drop-off possibilities (building design, access barriers, etc.).  Additional factors influencing the last meter delivery efficiency are the driver’s physical effort and their knowledge of the area. Part of their job also entails making real-time decisions on where to park and the sequence in which the last meter parcels are delivered. Drivers often spend as high as 80% of their time outside of the vehicle walking to the address, handling building access, returning repeatedly to the vehicle to fetch more parcels or return parcels when the recipient is not at home, carrying bulky or irregular parcels and exception handling. All this calls for solutions that can make the last meter delivery execution faster and less physically demanding. The solution The project will develop and evaluate a two-stage decision-support concept for dense urban B2C delivery. Stage 1: In this stage, we will study current delivery operations through available data from Planzer and interviews with drivers in selected urban contexts in order to understand how drivers currently choose parking locations, organise nearby stops, carry parcels over the last meters, and handle practical constraints such as access difficulties or repeated returns to the vehicle. This will allow us to identify the main sources of inefficiency at stop level and to define representative delivery situations for further analysis. Based on this understanding, we will build a first optimization framework that treats the truck as a local delivery base and evaluates how several nearby deliveries can be served from one parking event. The framework will group parcel destinations around candidate parking areas and compare different local delivery strategies depending on factors such as walking distance, parcel volume, and the use of micro-mobility aids. It will also estimate the expected service time for different options, making it possible to compare current practice with alternative delivery concepts in a structured way. This first phase will allow us to test key design choices, including: how parcels should be grouped around a parking area; how far a driver can efficiently serve addresses from one parking event; which types of carrying aids are most promising in which situations; how the use of micro-mobility aids affects the parking event; and under which route and parcel conditions clustered delivery becomes advantageous. This stage will use a mixture of existing technologies such as VRP, OSRM, or Valhalla (for ETA, path optimization), and machine learning models (e.g. for clustering). Stage 2: In this subsequent stage, we will refine this initial framework using real delivery data (timestamped GPS data, videos and operational feedback). The goal is to move from a simple planning logic toward a more realistic assistant that accounts for practical factors that are difficult to model directly, such as parking difficulty, local access conditions, neighbourhood layout, traffic conditions, driver habits, and other practical constraints. Rather than relying only on static optimisation, the system will progressively integrate observed driver decisions and real itineraries in order to improve the quality and realism of its suggestions.  In practice, instead of stacking the models independently (clustering then path optimization then offloading optimization), we develop a machine learning model that account for all the variables in the first place (e.g. the clustering account for the path optimization, traffic, parking difficulties, etc.) and can make update suggestion on-the-fly in case of situation change (driver parks somewhere else, take a different route, etc.). This system will first learn to imitate the Stage 1 suggestion as a baseline, and then improve upon it based on additional data and more.  The long-term technical ambition is therefore not only to optimise delivery clusters in theory, but to inform the clustering with real constraints such as traffic, parcel volume, and additional knowle dge from the driver. The end goal is to develop a recommendation approach that remains useful under real operating conditions. Concretely, the project will lay the foundations for a system that can recommend promising parking areas, suggest a local sequence for nearby deliveries using micro-mobility aids, and adapt its guidance based on the context of the route and the realities of field execution. Barriers to exploration Technological:  Recommendation quality may be limited at first since the system does not sufficiently capture real-world conditions such as street layout, no-parking zones, temporary access restrictions, or other local operational constraints. This limitation would be reduced over time by learning from real driver choices, actual itineraries, and recurring delivery patterns. There may also be practical and technical constraints in integrating micro-mobility aids in the main delivery vehicle, including storage, accessibility, and ease of deployment during operations. If parcel volume, size, or shape are not adequately considered, the system may generate impractical groupings, for example by combining bulky or irregular parcels in ways that are difficult to handle efficiently in the field. The second phase optimization should mitigate this problem. Last meter topology in some neighbourhoods may impede a convenient use of micro-mobility aids   Cultural:  Drivers may be reluctant to adopt the tool if they perceive it as a monitoring system rather than as a practical form of operational support. Adoption may also be limited if the system produces recommendations that appear suboptimal or inconsistent with drivers’ experience and local knowledge. If the loading, unloading, or use of the micro-mobility aid makes the delivery process more cumbersome or time-consuming, drivers may choose not to use it despite its intended benefits. Expected Outcomes The project will deliver: a structured understanding of stop-level inefficiencies in dense Swiss urban delivery; quantified evidence on when clustered delivery improves productivity and by how much (an initial estimation is a savings of up to one hour per truck each day); practical insight into which carrying aids or micro-mobility solutions create real operational value; a concept for a parking-area and clustering recommendation tool; implementation recommendations for dense urban B2C delivery in Switzerland; a roadmap for a larger innovation project with operational deployment.

Project Idea Metadata

Project Idea Name: Delivery loop optimizer (Multi-parcel delivery optimizer for micro-mobility assistance in the last meters)
Date: 4/2/2026 12:53:51 PM
Administrators:

Project Idea Description

The challenge

Urban B2C last-mile delivery in Switzerland is becoming harder to execute efficiently due to the fast growing demand in dense highly urbanised areas. In B2C urban logistics, the most expensive step is the highly variable and inefficient last meters of delivery execution. Efficiency is shaped by factors such as: street conditions, contested parking areas, restricted curb access, mixed use areas with restaurants and high pedestrian activity and inconsistent drop-off possibilities (building design, access barriers, etc.).

Additional factors influencing the last meter delivery efficiency are the driver’s physical effort and their knowledge of the area. Part of their job also entails making real-time decisions on where to park and the sequence in which the last meter parcels are delivered. Drivers often spend as high as 80% of their time outside of the vehicle walking to the address, handling building access, returning repeatedly to the vehicle to fetch more parcels or return parcels when the recipient is not at home, carrying bulky or irregular parcels and exception handling. All this calls for solutions that can make the last meter delivery execution faster and less physically demanding.

The solution

The project will develop and evaluate a two-stage decision-support concept for dense urban B2C delivery.

Stage 1:

In this stage, we will study current delivery operations through available data from Planzer and interviews with drivers in selected urban contexts in order to understand how drivers currently choose parking locations, organise nearby stops, carry parcels over the last meters, and handle practical constraints such as access difficulties or repeated returns to the vehicle. This will allow us to identify the main sources of inefficiency at stop level and to define representative delivery situations for further analysis.

Based on this understanding, we will build a first optimization framework that treats the truck as a local delivery base and evaluates how several nearby deliveries can be served from one parking event. The framework will group parcel destinations around candidate parking areas and compare different local delivery strategies depending on factors such as walking distance, parcel volume, and the use of micro-mobility aids. It will also estimate the expected service time for different options, making it possible to compare current practice with alternative delivery concepts in a structured way.

This first phase will allow us to test key design choices, including:

how parcels should be grouped around a parking area;
how far a driver can efficiently serve addresses from one parking event;
which types of carrying aids are most promising in which situations;
how the use of micro-mobility aids affects the parking event;
and under which route and parcel conditions clustered delivery becomes advantageous.

This stage will use a mixture of existing technologies such as VRP, OSRM, or Valhalla (for ETA, path optimization), and machine learning models (e.g. for clustering).

Stage 2:

In this subsequent stage, we will refine this initial framework using real delivery data (timestamped GPS data, videos and operational feedback). The goal is to move from a simple planning logic toward a more realistic assistant that accounts for practical factors that are difficult to model directly, such as parking difficulty, local access conditions, neighbourhood layout, traffic conditions, driver habits, and other practical constraints. Rather than relying only on static optimisation, the system will progressively integrate observed driver decisions and real itineraries in order to improve the quality and realism of its suggestions.

In practice, instead of stacking the models independently (clustering then path optimization then offloading optimization), we develop a machine learning model that account for all the variables in the first place (e.g. the clustering account for the path optimization, traffic, parking difficulties, etc.) and can make update suggestion on-the-fly in case of situation change (driver parks somewhere else, take a different route, etc.). This system will first learn to imitate the Stage 1 suggestion as a baseline, and then improve upon it based on additional data and more.

The long-term technical ambition is therefore not only to optimise delivery clusters in theory, but to inform the clustering with real constraints such as traffic, parcel volume, and additional knowle

dge from the driver. The end goal is to develop a recommendation approach that remains useful under real operating conditions. Concretely, the project will lay the foundations for a system that can recommend promising parking areas, suggest a local sequence for nearby deliveries using micro-mobility aids, and adapt its guidance based on the context of the route and the realities of field execution.

Barriers to exploration

Technological:

Recommendation quality may be limited at first since the system does not sufficiently capture real-world conditions such as street layout, no-parking zones, temporary access restrictions, or other local operational constraints. This limitation would be reduced over time by learning from real driver choices, actual itineraries, and recurring delivery patterns.
There may also be practical and technical constraints in integrating micro-mobility aids in the main delivery vehicle, including storage, accessibility, and ease of deployment during operations.
If parcel volume, size, or shape are not adequately considered, the system may generate impractical groupings, for example by combining bulky or irregular parcels in ways that are difficult to handle efficiently in the field. The second phase optimization should mitigate this problem.
Last meter topology in some neighbourhoods may impede a convenient use of micro-mobility aids

Cultural:

Drivers may be reluctant to adopt the tool if they perceive it as a monitoring system rather than as a practical form of operational support.
Adoption may also be limited if the system produces recommendations that appear suboptimal or inconsistent with drivers’ experience and local knowledge.
If the loading, unloading, or use of the micro-mobility aid makes the delivery process more cumbersome or time-consuming, drivers may choose not to use it despite its intended benefits.

Expected Outcomes

The project will deliver:

a structured understanding of stop-level inefficiencies in dense Swiss urban delivery;
quantified evidence on when clustered delivery improves productivity and by how much (an initial estimation is a savings of up to one hour per truck each day);
practical insight into which carrying aids or micro-mobility solutions create real operational value;
a concept for a parking-area and clustering recommendation tool;
implementation recommendations for dense urban B2C delivery in Switzerland;
a roadmap for a larger innovation project with operational deployment.

The proposed solution is a tool that helps delivery drivers execute the last meters more efficiently by recommending suitable parking areas, suggesting which parcels should be grouped for each micro-mobility-assisted delivery loop, and continuously improving its recommendations based on driver decisions and real operational experience. The solution aims to improve last-meter productivity, reduce physical strain, support drivers’ decision-making, and enable a scalable urban B2C delivery system that can keep up with a rising demand.