Differentiated Pickup Point Offering for Emission Reduction in Last-Mile Delivery

Albina Galiullina, Wouter van Heeswijk, Tom van Woensel

TL;DR

This paper introduces a Differentiated Pickup Point Offering (DPO) policy that uses Deep Reinforcement Learning to minimize the total carbon footprint of last-mile delivery. By strategically recommending a single, specific pickup point to each customer rather than offering unrestricted choice, the system balances delivery truck efficiency with the risk of "carbon leakage" from customers driving to collect their packages.

What Problem Does This Paper Address?

Pickup points (like lockers or local shops) are often touted as a "green" alternative to home delivery because they allow trucks to consolidate many orders into a single stop. However, there is a catch: if a pickup point is too far away, customers are more likely to drive their cars to get there. This "carbon leakage" from individual car trips can actually result in higher total emissions than if the truck had just driven to everyone’s front door.

Current industry practice usually gives customers an unrestricted choice of all available pickup points. The authors argue this is sub-optimal because it leads to fragmented delivery routes. The challenge is to design a dynamic system that recommends pickup points that are close enough to encourage walking/cycling but centralized enough to create efficient truck routes—all while accounting for the fact that orders arrive one by one throughout the day.

Methods

The authors model the problem as a finite-horizon Markov Decision Process (MDP) and solve it using Deep Reinforcement Learning (DRL).

  • State Representation: To handle the varying number of customers and their spatial relationships, the system state is represented as a graph.
  • Graph Attention Network (GAT): The "brain" of the agent uses a GAT encoder. This allows the model to dynamically weight the importance of neighboring nodes (like nearby customers or already-visited pickup points) to produce embeddings that capture the spatial "consolidation potential."
  • Proximal Policy Optimization (PPO): The agent is trained using PPO, an on-policy RL algorithm, to learn which pickup point to offer to an arriving customer to minimize the expected total emissions (truck + customer travel).
  • Two-Phase Process:
    1. Capture Phase: Orders arrive sequentially; the agent makes a recommendation; the customer chooses (stochastically) between the recommendation and home delivery.
  • Fulfillment Phase: Once all orders are in, a Traveling Salesman Problem (TSP) is solved to find the optimal truck route.

Datasets

The study uses synthetic instances generated to simulate urban service regions.

  • Spatial Parameters: Instances vary the service region radius ($L \in \{0.5, 8\}$ km) and the number of pickup points ($|M| \in \{4, 30\}$).
  • Customer Behavior: They use a multinomial logit model based on real-world survey data (Niemeijer and Buijs, 2023) to simulate how distance influences a customer's choice of delivery mode and transportation (car, walk, cycle).
  • Benchmark: The authors compare DPO against "Home-only," "Unrestricted Choice," "Nearest Pickup Point," and a "Perfect Information" lower bound.

Key Insights and Ideas

  • The "Unrestricted Choice" Trap: Counter-intuitively, giving customers more choices often increases emissions. When customers can pick any point, they choose the one most convenient for them, which scatters the truck's stops across the city.
  • Anticipatory Consolidation: The DPO policy learns to be "anticipatory." It might recommend a pickup point that is slightly further for the current customer if it knows that point is likely to become a hub for future orders arriving later in the day.
  • Dynamic Adaptability: As the "Capture Phase" progresses, the agent becomes more rigid. In the latter half of the day, over 80% of recommendations are for pickup points already "locked in" by previous customers, effectively forcing consolidation.
  • Urban vs. Rural: DPO is most effective in dense urban settings. In large, sparse regions, the "Home-only" delivery often wins because the distances to pickup points are so great that car trips become inevitable.

Key Results

  • Emission Reductions: The DPO policy reduced total carbon emissions by up to 9% compared to home-only delivery.
  • Superiority over Heuristics: DPO outperformed the "Nearest" and "Unrestricted" policies by an average of 2%. While 2% sounds small, in the logistics industry, this represents a significant margin when scaled across thousands of daily deliveries.
  • Route Efficiency: In dense areas ($L < 1$ km), DPO consolidated orders so effectively that the truck visited only 1–2 pickup points on average, compared to the "Nearest" policy which required visiting many more distinct locations.

Limitations and Future Work

  • Capacity Constraints: The current model assumes pickup points (lockers) have infinite capacity. In reality, lockers fill up, which would add a significant constraint to the recommendation engine.
  • Time Windows: The study focuses on next-day delivery without specific time slots. Integrating temporal demand management (e.g., offering discounts for specific delivery windows) is a suggested next step.
  • Customer Satisfaction: The paper assumes customers accept recommendations based on a probability model. Future work could explore the behavioral impact of "restricting" choice and whether it affects long-term customer retention.

Summary updated

Abstract

Pickup points are widely recognized as a sustainable alternative to home delivery, as consolidating orders at pickup locations can shorten delivery routes and improve first-attempt success rates. However, these benefits may be negated when customers drive to pick up their orders. This study proposes a Differentiated Pickup Point Offering (DPO) policy that aims to jointly reduce emissions from delivery truck routes and customer travel. Under DPO, each arriving customer is offered a single recommended pickup point, rather than an unrestricted choice among all locations, while retaining the option of home delivery. We study this problem in a dynamic and stochastic setting, where the pickup point offered to each customer depends on previously realized customer locations and delivery choices. To design effective DPO policies, we adopt a reinforcement learning-based approach that accounts for spatial relationships between customers and pickup points and their implications for future route consolidation. Computational experiments show that differentiated pickup point offerings can substantially reduce total carbon emissions. The proposed policies reduce total emissions by up to 9% relative to home-only delivery and by 2% on average compared with alternative policies, including unrestricted pickup point choice and nearest pickup point assignment. Differentiated offerings are particularly effective in dense urban settings with many pickup points and short inter-location distances. Moreover, explicitly accounting for the dynamic nature of customer arrivals and choices is especially important when customers are less inclined to choose pickup point delivery over home delivery.
BibTeX
@misc{galiullina2026differentiated,
  author = {Albina Galiullina and Wouter van Heeswijk and Tom van Woensel},
  title = {{Differentiated Pickup Point Offering for Emission Reduction in Last-Mile Delivery}},
  year = {2026},
  eprint = {2601.14196},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.14196},
}
Figure