Computers & Industrial Engineering 200 (2025) 110767

Contents lists available at ScienceDirect

Computers & Industrial Engineering

journal homepage: www.elsevier.com/locate/caie

Joint optimization of storage assignment and order batching in robotic

mobile fulfillment system with dynamic storage depth and surplus items
Zhi Liu a,b , Jiansha Lu a,*, Chenhao Ren c , Jun Chen a , Zhilong Xu a, Guoli Zhao a,d
a
College of Mechanical Engineering, Zhejiang University of Technology, Hangzhou 310023, China
b
School of Physics and Electronic Engineering, Yancheng Teachers University, Yancheng 224007, China
c
School of Economics and Management, Zhejiang University of Science and Technology, Hangzhou 310023, China
d
Bosch Power Tools (China) Co., Ltd., Hangzhou 310052, China

A R T I C L E I N F O A B S T R A C T

Keywords: This paper studies the joint optimization of storage location assignment and order batching in robotic mobile
Storage location assignment fulfillment systems (RMFS), considering dynamic storage depth and surplus items. Firstly, a joint optimization
Order batching model of storage location assignment and order batching is established. The model is divided into two stages. The
Robotic mobile fulfillment system
first stage is the item location assignment optimization model, which is used to describe the types and quantities
Dynamic system
of items placed in each slot on each pod. The second stage is the joint optimization model of pod location
assignment and order batching, which is used to describe the coordinates of each pod, the number of order
batches, and the order combinations contained in each batch. Considering the vast solution space and the
numerous constraints of the constructed model, a two-stage greedy variable neighborhood simulated annealing
algorithm (TGVNSA) is introduced to address these challenges. Finally, numerical experiments prove that the
algorithm can effectively solve the established model. TGVNSA is evaluated against two conventional methods:
variable neighborhood search and adaptive genetic algorithms, focusing on metrics such as pod retrieval times,
comprehensive picking costs, CPU time, and CQ value (comprehensive quality in the item location assignment).
The findings demonstrate that TGVNSA boasts superior comprehensive performance. Compared with other
commonly used strategies in the field such as random, classification, and optimal relevance, the method pro­
posed in this paper demonstrates superior optimization performance, particularly when considering dynamic
storage depth and surplus items. Moreover, this paper also proves that, under the same combination of strategies,
the joint optimization method proposed in this paper reduces the comprehensive picking costs by 11.46%
compared to the separate optimization approach.

1. Introduction which helps optimize warehouse layout, inventory management, and

operational processes, helping enterprises adapt to constantly changing
The rapid development of the mobile robot industry has brought market demands (Barros and Nascimento 2021). Given its unique ad­
significant changes in the field of warehousing and logistics. The storage vantages, RMFS is gradually applied in various types of distribution
and retrieval systems have evolved from the traditional picker-to-parts centers, e-commerce warehouses, material supermarkets, and other
model to the parts-to-picker model. In 2012, Amazon acquired Kiva systems that integrate warehousing and sorting all over the world.
and applied its robots to its distribution centers. Since then, robotic The operation of RMFS mainly includes the following several pro­
mobile fulfillment systems (RMFS) have been officially applied in large- cesses: First, the items are put into storage. The items are placed on the
scale systems (Boysen et al. 2019). RMFS has greatly improved the ef­ pods (shelves) according to the established strategy. The robots carry the
ficiency of the labor-intensive sorting work, reduced the error rate, pods to the designated storage area and place them according to a
reduced costs, improved safety, and increased space utilization. At the certain storage strategy. After the customer places an order, the order
same time, RMFS also has high flexibility and scalability. Mobile robots processing operation is carried out. The warehouse management system
equipped with sensors and cameras can collect a large amount of data, assigns the picking tasks in the order to the robots. The robots move to

* Corresponding author.
E-mail address: ljs@zjut.edu.cn (J. Lu).

https://doi.org/10.1016/j.cie.2024.110767
Received 4 July 2024; Received in revised form 28 October 2024; Accepted 26 November 2024
Available online 4 December 2024
0360-8352/© 2024 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

the corresponding pod locations according to the task sequence, and preparations for task allocation. Therefore, the joint optimization of
bring the pods to the picking area, then the picking operation is carried these two stages can provide better initial values for subsequent work.
out by manual workers or picking robots. Finally, packaging and veri­ The flexibility and scalability of the RMFS enable it to meet complex
fication, outbound, and distribution are conducted (D’Andrea et al. and varying demands and environments. However, there is relatively
2008). During the entire process, the system will continuously collect little research on RMFS that takes random factors and dynamic situa­
data, including picking time, robot operating status, order accuracy, tions into account, yet it is particularly crucial at this stage. Therefore,
etc., for future analysis and system optimization. RMFS has unparalleled this paper focuses on real-world scenarios and undertakes joint opti­
advantages in picking efficiency compared to other systems, making it mization of storage assignment and order batching for the RMFS,
especially suitable for e-commerce warehouses or manufacturing ma­ considering the dynamic changes of certain parameters. The dynamic
terial supermarkets (Yang et al. 2021). parameters considered in this paper include storage depth, the surplus
Research related to RMFS can be divided into system analysis, design items in each pod, and the number of orders contained in each batch.
optimization, and operations planning and control. The operation The main content of this paper is to address the problem of deter­
planning can also be subdivided into four aspects: storage assignment, mining which items should be placed on which pods and where the pods
order batching, task allocation, and path planning (Azadeh et al. 2019). should be positioned in the storage area. This problem is divided into
The schematic diagram of the RMFS operation process is shown in two stages. The first stage involves the item location assignment, placing
Fig. 1. Conventional configurations of RMFS have been the subject of various items in different slots (storage positions in each pod), to
considerable research. In recent years, with rising land use costs, minimize the number of pod retrieval times accessed during order
intensive storage has become a topic of significant practical importance. fulfillment. The second stage focuses on the pod location assignment,
While there has been some research on dense RMFS (Yang et al. 2022), incorporating order batching for joint optimization, aiming to minimize
studies specifically addressing the storage location assignment problem the comprehensive picking costs during the order fulfillment process.
remain relatively scarce. The schematic diagram of the two stages is shown in Fig. 2.
Storage assignment is based on the characteristics and storage re­ The storage depth is allowed to change dynamically in this paper,
quirements, exploring the correlations between items, placing them in with different areas permitted to utilize varying depths. Meanwhile,
the corresponding pods in the optimal combination, and assigning the during the order fulfillment process, the surplus items on each pod are
pods to reasonable positions. The purpose is to improve the utilization recorded in real-time, allowing for a more accurate description of the
rate of warehouse space, optimize the picking efficiency, enhance the number of retrieval times for each pod. Subsequently, an integer pro­
order processing capacity, and quickly adapt to changes in demand, etc. gramming model is established to describe this problem, and a fast and
Storage assignment is crucial for ensuring the efficiency and safety of effective heuristic algorithm is developed to solve it based on the
RMFS and adapting to the constantly changing business requirements characteristics of the problem.
(Roy et al. 2019). In actual operation, there are tens of thousands of The dynamic storage depth means that the storage depth can be
types of items. Each type of item may be selected simultaneously with flexibly selected according to actual needs. If the quantity of items to be
many other items, and the quantitative relationships are intricate and stored is moderate and there is a high demand for efficient inbound and
complex. Faced with the large-scale item combinations and quantity outbound operations, a single-depth storage configuration can be used.
relationships reflected by massive orders, storage assignment aims to This configuration ensures that an adequate number of robots can
determine the types and quantities of items stored on each pod to meet operate simultaneously without causing conflicts or deadlocks, as
the diverse order-picking demands as much as possible, thereby mini­ illustrated in Fig. 3(a). If there are too many items to be stored, and some
mizing the number of pod movements required for order picking, so of them have a low turnover rate, then, certain areas can be set up as
storage assignment is extremely difficult. double-depth or multi-depth storage to increase the maximum capacity
Order batching involves combining multiple orders for picking at the of racks, allowing it to have enough space to store items that meet the
same time, which reduces repetitive tasks for robots or manual labor. order requirements, as shown in Fig. 3(b).
The goal is to reduce the number of times pods need to be moved during The dynamic surplus items in each pod mean that the consumption of
order picking, as well as to minimize the walking distance or time for items during the picking process will be taken into account. The number
robots. The strategies used in order batching typically consider princi­ of items that a pod can store is not infinite. As the picking operation
ples such as the attributes of the items, their priority, and their similarity proceeds, the storage capacity of each pod will decrease to varying de­
(Boysen et al. 2017). grees. Obviously, whether to consider the dynamic surplus items has a
There has been considerable research on the optimization of indi­ great impact on the optimization of all stages of RMFS. So, in the process
vidual stages in RMFS. However, joint optimization of two or more of optimizing storage assignment and order batching, it is also necessary
stages often leads to better results (Scholz et al. 2017; K. Zhang and Gao to consider dynamic surplus items in each pod.
2023). Storage assignment and order batching are both preliminary Order batching can significantly reduce the number of pod retrievals

Fig. 1. Schematic diagram of RMFS operation process.

2
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

Fig. 2. Item location assignment and pod location assignment.

configurations, effectively addressing the massive order demand

during peak periods.

The remaining chapters are arranged as follows: Section 2 is a

literature review; Section 3 describes and models the joint optimization
of storage assignment and order batching considering dynamic param­
eters; Section 4 designs an algorithm to solve the model; Section 5 an­
alyzes the calculation results; and Section 6 summarizes.

2. Literature review

This section summarizes relevant research and is divided into the

following aspects: an overview of storage location assignment problems,
an overview of order batching problems, and an overview of joint
optimization problems. A summary of relevant studies about RMFS can
Fig. 3. Schematic diagram of RMFS with different storage depth. be found in Table 1.

and improve picking efficiency. However, handling the picking process 2.1. Storage assignment
for multiple orders simultaneously will inevitably increase the workload
and pressure on the pickers. So, the number of orders in a batch is not the Storage assignment can be divided into item location assignment,
more the better. The number of orders in each batch is determined based pod location assignment, and pod location reassignment. Storage
on a combined value of order similarity and manual picking costs. For assignment is a typical generalized assignment problem. A scientific
example, placing orders with low similarity into the same batch results storage assignment method can shorten walking distances, reduce
in only a little reduction in the number of pod retrievals, while signifi­ search times, decrease workload, and improve warehouse picking effi­
cantly increasing manual picking costs. This, in turn, causes a decline in ciency. Hausman et al. (1976) first proposed three basic assignment
the overall efficiency of the system. strategies: dedicated storage, random storage, and class-based storage.
The contributions of this paper are as follows: Subsequent literature has studied this issue from various aspects such as
throughput, COI (Cube-per-Order Index), turnover rate, and the corre­
(i). The storage location assignment of RMFS is subdivided into item lation between demand and structure.
location assignment and pod location assignment and is jointly Regarding the item location assignment problem, Onal et al. (2017)
optimized with order batching. Then, a two-stage integer pro­ proposed a chaotic storage assignment strategy and developed a queuing
gramming model for storage location assignment and order model for the item assignment. The results showed that, with the in­
batching (SLAOB) is constructed with the objectives of mini­ crease of the chaotic level, the order picking time could be reduced by up
mizing number of pod retrieval times, order relevance, and to 16 %. Boysen et al. (2017) along with Weidinger and Boysen (2018)
comprehensive picking costs. proposed a shared storage strategy that can significantly improve the
(ii). According to the characteristics of the model, a two-stage greedy picking efficiency of RMFS. Lee et al. (2020) introduced a bi-objective
variable neighborhood simulated annealing algorithm (TGVNSA) optimization model to maximize the relevance of SKUs within the
was constructed to solve the problem. same cluster to enhance efficiency, while also minimizing the maximum
(iii). Use order data to guide the completion of RMFS storage location value of SKU picking frequency across different clusters to balance
assignment, considering the dynamic storage depth and the dy­ traffic flow. Kim et al. (2020) analyzed the item assignment problem in
namic consumption of items. RMFS and proposed a re-optimization heuristic method to address
(iv). Allow the existence of multiple depth storage areas, enabling changes in item similarity values. J. Zhang et al. (2023) considered the
RMFS to accommodate more pods than traditional human influence in the item allocation of RMFS, developed a bi-
objective mixed integer programming model aimed at maximizing the

3
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

Table 1 et al. (2017) on storage location assignment for AS/RS provides theo­
Summary of literature related to RMFS. retical support for storage location assignment in RMFS. Nigam et al.
Paper Problem Method Objective Feature (2014) used a queueing model to study the issue of random storage and
nearest storage strategies in RMFS. The results showed that the random
Boysen et al. ILA & IP & DP Number of rack Shared storage
(2017) OB changes & visits storage had higher space utilization, while the nearest storage had
Lamballais PLA QN Throughput Storage zones higher throughput. Lamballais et al. (2017) constructed a queuing
et al. network model to evaluate the performance of the RMFS. They assessed
(2017) system performance using indicators such as maximum throughput,
Onal et al. ILA QN Makespan Explosive
(2017) storage
average order picking time, and robot utilization rate. The results
Weidinger ILA MIP Weighted sum of Scattered indicate that employing a zoning storage strategy based on pod turnover
and distances storage rate can increase the system’s maximum throughput by about 50 %. A
Boysen limitation mentioned is that the study assumes each pod stores only one
(2018)
type of item. Weidinger et al. (2018) conducted a study on the dynamic
Weidinger PLA MIP Total travel Interval
et al. distance scheduling pod location assignment in RMFS, constructed a mixed-integer pro­
(2018) gramming model, and solved the model using an adaptive planning
Xiang et al. ILA & MIP Number of pod Joint method. This reduced the total picking travel distance by 3.49 % and
(2018) OB movements optimization reduced the number of robot formations by 2.17 %. However, the
Yuan et al. PLA Fluid model Total travel Velocity-based
(2019) distance storage
theoretical model reflecting the dynamic relationship between pod
Lee et al. ILA MIP Correlation Traffic balanced storage locations was not been constructed. Yuan et al. (2019) further
(2020) storage explored random storage, velocity-based, and class-based shelving
Kim et al. ILA MIP Correlation Changed assignment strategies, constructing a queueing model to evaluate the
(2020) similarity values
performance of different storage strategies. They discovered that the
X. W. Li et al. ILA & MIP Correlation & Turnover-rate-
(2020) PLA distance based class-based strategy was more robust and adaptable to different ware­
decentralized house configurations and layouts in RMFS. X. W. Li et al. (2020) first
storage employed time association analysis and clustering to identify highly
Keung et al. PLA Data-driven Total travelling Consider multi- related products and store them on the same pod. Then they proposed a
(2021) cost deep costs
new turnover-based decentralized storage strategy (TRBDSP) to avoid
Keung et al. PLA IP Total cost Industrial
(2022) internet of AGV congestion. Additionally, they introduced an order picking per­
things formance evaluation method that considers AGV energy consumption.
Wang et al. PLA IP Throughout Fishbone RMFS Keung et al. (2021) proposed a data-driven approach for regional clus­
(2022)
tering and storage location allocation to improve the operational effi­
Cezik et al. PLA Continuous Total travel Velocity-based
(2022) variable distance stowage ciency of RMFS. Keung et al. (2022) discussed the use of the Industrial
Mirzaei et al. ILA & MIP Expected travel Symmetry Internet of Things (IIoT) in RMFS to improve operational efficiency and
(2022) PLA time breaking developed different storage location assignment rules and strategies to
constraints minimize operating costs. It is worth noting that Keung’s two papers
J. Zhang ILA MIP Correlation & Human energy
discuss the blocking costs of pod movement in multi-deep storage areas.
et al. energy consumption
(2023) consumption Wang et al. (2022) proposed a storage location assignment optimization
Ma et al. ILA MIP Correlation Scattered model for fishbone RMFS, aiming to maximize operational efficiency
(2023) Storage and balance aisle workloads. Lanza et al. (2022) proposed a mixed-
Liu and Poh PLA MIP Quantity of Large items
integer linear programming (MILP) model to allocate storage locations
(2023) specific items
Zarinchang ILA Nonlinear Rack stability & Worker safety
for product types and determine the order of allocation. The model
et al. MIP picking combines effective inequalities, relaxations, and two mathematical
(2024) efficiency, etc. heuristic methods to address this problem. The sorting approach in this
Ding et al. PLA IP Total travel Velocity-based paper can provide guidance for multi-deep storage solutions. Liu and
(2024) distance
Poh (2023) proposed decentralized storage with SKU-specific bulky lo­
Cheng et al. OB SMDP & Number of pod Deep
(2024) DRL movements reinforcement cations to improve picking efficiency in response to quantity fluctua­
learning tions in customer order patterns. H. R. Li et al. (2024) established a
This paper ILA & IP Comprehensive Dynamic mathematical model for the storage location assignment problem in
PLA & picking costs storage depth &
large-scale automated warehouses, considering three objectives: effi­
OB surplus items
ciency, pod stability, and load balancing. They proposed an improved
Note: ILA: Item Location Assignment; PLA: Pod Location Assignment; OB: Order vortex search algorithm based on attraction operation in the flow field,
Batching; QN: Queueing Network; IP: Integer Programming; MIP: Mixed Integer dimension-by-dimension dynamic radius, and leadership decision-
Programming; DP: Dynamic Programming; SMDP: Semi-Markov Decision Pro­ making mechanism (FDVSA). Cezik et al. (2022); Ding et al. (2024)
cess; DRL: deep reinforcement learning.
proposed a velocity-based pod storage location assignment (VRSLA)
method for unidirectional RMFS. This method optimizes the total dis­
similarity of items in the pods and minimizing the energy consumption tance of pod movement by calculating the defined pod velocity and
of pickers, then they proposed an improved knee-driven evolutionary storage location velocity.
algorithm (IKnEA) to solve the model. Ma et al. (2023) proposed a
decentralized related storage strategy based on item classification and 2.2. Order batching
correlation to improve order-picking efficiency in RMFS. Zarinchang
et al. (2024) proposed an artificial intelligence algorithm to address the The order batching problem was first proposed by Ackerman.
storage location allocation problem in warehouses. This algorithm Compared with storage assignment, order batching optimization can
combines basin-hopping and simulated annealing, demonstrating its improve picking efficiency more significantly. The order batching
accuracy comparable to commercial solvers in real case studies. problem is an NP-hard problem. Depending on whether the order in­
Regarding the issue of pod location assignment, The researches by formation is known, order batching can be divided into online order
Yang, Miao, Xue, and Qin (2015); Yang, Miao, Xue, and Ye (2015); Yang batching and offline order batching. Generally, the offline order

4
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

batching problem is studied by heuristic algorithms, data mining, clus­ batching, sequencing, and path planning algorithms. Boysen et al.
tering, and simulation, while the online order batching problem is (2017) expanded the issue into three sub-problems: order batching,
studied by time windows, heuristics, and other methods. Improvements order sequencing, and the sorting of pods at the picking station. The
to the research methods for offline order batching can also be used to results indicate that optimized order scheduling and pod sorting can
study online order batching problems. Existing literature uses indicators reduce the operating time of robots by at least 21 %. J. Zhang et al.
such as maximizing batch order correlation and minimizing the number (2017) studied the issue of synergistic optimization of order batching
of pod moves as optimization objectives, and mainly uses queuing the­ and sequencing under the premise of considering order delivery dead­
ory, (mixed) integer programming, clustering, and other methods to lines. They constructed an integer programming model for synergistic
study order batching problems in RMFS. optimization, achieving an order fulfillment rate of 80.365 %. Tabrizi
Henn (2012) considered a dynamic online order batching problem, et al. (2023) proposed a three-stage model for optimizing online order
grouping customer orders into batches with the goal of minimizing the picking and delivery in e-commerce. The first stage involves clustering
makespan, and adapting solutions from the offline batching problem to products and allocating storage locations to minimize transportation
the online environment. Menéndez et al. (2017) proposed a multi-start and energy costs. The second stage provides a joint model for online
variable neighborhood search method to solve the order batching order batching and picking path planning, considering the impact of
problem in warehouse operations based on different strategies. Yang learning effects on picking and search times to minimize completion
et al. (2020) developed an order batching optimization model for three time. The third stage applies the optimized decisions from the first two
typical storage systems and created an algorithm package that includes stages to the actual online order picking and delivery process, resulting
location distance algorithms, location selection algorithms, path plan­ in reductions of 35.65 % in total costs and 12.48 % in completion time.
ning algorithms, and order batching algorithms to address these models. Momeni and Jain (2024) examined the dedicated storage policy in
Gil-Borrás et al. (2020) proposed a method that combines Greedy Ran­ relation to the production planning problem and proposed a mixed-
domized Adaptive Search Procedure (GRASP) and Variable Neighbor­ integer programming model aimed at minimizing the total costs of
hood Descent (VND) to optimize online order batching problems. The production and warehousing. They initially employed lexicographical
next year, Gil-Borrás et al. (2021) conducted another research on online constraints to ensure that storage locations were allocated adjacently to
order batching problems with multiple pickers, aiming to minimize the items.
makespan, balance the picking time, and evenly distribute the workload It is particularly important to highlight three papers, one of which is
among pickers. They proposed a mixed shaking variable neighborhood by Xiang et al. (2018), who focused on storage assignment and order
descent metaheuristic method to tackle this issue. Furthermore, the batching problems. They developed a mixed-integer linear program­
study also evaluated the balance of workload among pickers and the ming model with the goal of increasing item similarity during the stor­
picking time. Kong et al. (2023) introduced an order batching method age allocation phase and reducing the frequency of pod movements in
with forecasting and delay capabilities, considering time intervals and the order batching phase. They designed a Variable Neighborhood
buyer completion rates to generate batches, aiming to minimize the total Search algorithm to solve the order batching problem, and the results
processing time of auction orders and the system response time. Gil- indicated that effective storage assignment and order batching could
Borrás et al. (2024) proposed two new time window strategies and significantly reduce pod transportation. However, the paper did not
demonstrated how these time window methods affect different objective explore the inherent correlation between storage assignment and order
functions of the online order batching problem when the number of batching, treating them as separate phases and neglecting pod location
orders and picking stations varies. Gabellini et al. (2024) used a assignment. Additionally, it assumed that the quantity of items on each
nonlinear machine learning-based predictive model to forecast the pod was sufficient for orders and did not account for item depletion on
picking time for batch orders, then employed this prediction to guide a the pods. The second paper is by Yang et al. (2022), who developed a
genetic algorithm in finding the optimal allocation of future planned Semi-Open Queueing Network (SOQN) model to describe multi-deep
batch orders to pickers, while considering each picker’s individual compact RMFS. They used Approximate Mean Value Analysis (AMVA)
learning and fatigue characteristics. Most of the above studies focus on to solve the model and obtain performance metrics such as system
manual picking, providing significant guidance for the order batching throughput, robot utilization, and queue length. This is the first study to
problem in RMFS. These studies indicate that using data mining model and evaluate multi-deep compact RMFS, providing guidance for
methods such as clustering and association rules for order batching warehouse planners and managers in the design and operation of such
research in RMFS would be an effective approach. There are few studies systems. Thirdly, Mirzaei et al. (2022) addressed the storage location
that directly address the order batching problem in RMFS, with the most optimization problem in RMFS by establishing a mixed-integer pro­
recent one being as follows. Cheng et al. (2024) introduced a semi- gramming model guided by order information. They adopted a decen­
Markov decision process framework to simulate the Mobile Robot tralized storage allocation strategy that comprehensively considers
Order Batching Problem (MROBP), enabling continuous decision- product correlation and turnover frequency for both item assignment
making in dynamic warehouse environments. The proposed method and pod location assignment. By introducing symmetry breaking con­
leverages the adaptive learning capabilities of deep reinforcement straints and deriving expressions for expected travel times, they
learning to enhance order batching decisions, outperforming traditional simplified the complexity of the optimization model. However, the pod
heuristic and metaheuristic approaches. location assignment in their work is allocated only to a specific zone,
with pods placed randomly within that zone, without considering the
2.3. Joint optimization differences between each individual storage location.
The aforementioned literature has made significant contributions to
Some scholars have conducted joint optimization of the storage research on storage location assignment and order batching. However,
location assignment problem or the order batching problem in there are few studies that consider the order batching and storage
conjunction with other issues. Lin et al. (2016) developed a model under location assignment for RMFS together. Exploring the intrinsic connec­
certain practical constraints to simultaneously determine the optimal tions between storage location assignment and order batching, and
order batching allocation and the shortest Manhattan path for pickers, conducting joint optimization, will provide good initial values for sub­
proposing an improved particle swarm optimization algorithm to solve sequent task allocation and path planning. This approach is also essen­
it. Scholz et al. (2017) proposed a model that simultaneously addresses tial for enhancing overall efficiency, reducing costs, and improving
the joint order batching, assignment, sequencing, and routing problem order fulfillment capabilities in RMFS. At the same time, most literature
to minimize the total delay time of customer orders. They introduced a on RMFS does not consider the possibility of multi-deep storage and
variable neighborhood descent algorithm that integrates various typically assumes that the quantity of items on the pods adequately

5
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

meets order demands, without dynamically accounting for changes in (iv). The number of orders in a batch cannot exceed the maximum
the remaining items on the pods. If dynamic storage depth and surplus allowed value.
items are taken into account in RMFS research, it would enhance the (v). The contents of the orders are known.
system’s adaptability, allowing for rapid adjustments in operational (vi). The robot travel distance is calculated by Manhattan distance,
strategies in response to fluctuations in orders and diverse demands, and multi-depth cost is added when retrieving pods in the middle
thereby reducing the need for manual intervention and management of multi-depth areas.
complexity. (vii). The picking process for a wave order does not include mid-way
replenishment operations.
3. Problem description and model formulation
The meanings of the symbols used in this paper are shown in Table 2,
This section describes the joint optimization of RMFS storage and the definitions and values of the decision variables are shown in
assignment and order batching considering dynamic storage depth and Table 3.
surplus items and establishes a mathematical model.
3.2. Optimization model of item location assignment
3.1. Problem description
The item location assignment problem can be described as placing
The storage assignment in this paper is divided into two stages: item items on at most m pods, each pod has Q slots, and the upper limit of the
location assignment and pod location assignment. The relationship be­ storage quantity of item i in a slot is pi. Need to complete j orders, and
tween the two stages is illustrated in Fig. 2. Stage 1 involves optimizing each order includes multiple item requirements. The requirement is to
item location assignment, which includes placing various types of items place i type of items into pods so that the number of pod retrieval times
on the pods according to specific principles. Each pod includes multiple to complete the picking task of j orders is minimized. Consider the
slots, and each slot has an upper limit on the number of items that can be consumption of items in each slot during the order fulfillment process, as
stored. As order retrieval proceeds, the number of surplus items on the this will affect the number of each pod retrieval times. The result is the
pods decreases accordingly. It is necessary to determine the number of type and quantity of items stored in each slot on each pod. This problem
pods used and the items and their quantities in each slot, so as to is essentially a bin-packing problem, and the mathematical model is as
minimize the total number of pod retrieval times needed to fulfill order- follows:
picking tasks. Once the actual number of pods in use is determined, the ∑∑
system configuration, i.e. the storage depth, can be determined. min xmj (1)
Stage 2 involves the joint optimization of pod location assignment m∈M j∈J

and order batching, where each pod is positioned at an appropriate ∑

location within the system, and orders are merged for picking based on a ymji ≥ mji , ∀j ∈ J, i ∈ I (2)
specific principle, with the goal of minimizing comprehensive picking m∈M

costs. Comprehensive picking costs include pod retrieval cost, manual ∑

ymji ≤ zmni , ∀m ∈ M, i ∈ I, n ∈ N (3)
picking cost, and multi-depth cost. The pod retrieval cost is determined
j∈J
by the location of pods and the number of pod retrieval times. The
number of pod retrieval times and the results of order batching are mji ⋅xmj ≥ ymji , ∀i ∈ I, j ∈ J, m ∈ M (4)
closely interconnected. Furthermore, whether to perform order batching
and the choice of different order batching strategies significantly impact ∑∑
zmni ≤ Ci , ∀n ∈ N (5)
the actual number of pod retrievals. The consumption of items in each i∈I m∈M
slot during the order completion process is also considered. The manual
picking cost aims to limit the number of orders in a batch. The more qmi pi ≥ zmni , ∀m ∈ M, i ∈ I, n ∈ N (6)
orders in a batch, the greater the number of orders workers must pick ∑
simultaneously, which increases workload, task difficulty, and error qmi ≤ Q, ∀m ∈ M (7)
rates, resulting in higher picking costs. So, there is no need to put fewer i∈I

related orders in the same batch. Multi-depth cost is used to describe the
additional cost of reverse stocking when retrieving pods in the middle of
multi-depth areas.
Table 2
The order batching determines the number of pod retrieval times, Definition of symbols.
and the number of pod retrieval times and the location of each pod
Symbol Definition
jointly determine the pod retrieval cost. Moreover, the different condi­
tions of storage location assignment can also influence the optimal so­ M Set of available pods, M={1, 2,…, m}
I Set of items, I={1, 2,…, i}
lution for order batching. Therefore, there is an intrinsic coupling
J Set of orders, J={1, 2,…, j}
relationship between order batching and storage optimization, necessi­ mji The demand for item i in order j
tating joint optimization. Ci The total storage volume of items i
Based on the above description and observations in actual produc­ pi The maximum number of items i that can be stored in a slot
tion, the following assumptions are made about the system discussed in Q The number of slots on a pod
Dxy Manhattan distance between (x,y) and the picking station
this article: Z Number of zones in the storage area.
M1 Set of pods actually in use, M1={1, 2,…, m1}
(i). One type of item can be stored in more than one pod, and X The horizontal coordinate of the storage area, X={1, 2,…, x}
different items can be stored in one pod, but each slot can store Y The vertical coordinate of the storage area, Y={1, 2,…, y}
K Set of order batch, K={1, 2,…, k}
only one type of item.
P Maximum order quantity per batch
(ii). There is an upper limit on the number of each item stored in a N Set of natural numbers
slot. As the retrieval progresses, the surplus items decrease. costlabor Manual picking cost
(iii). A pod only visits one picking point in a retrieval mission, and it costdepth Multi-depth cost
returns to the original position when the picking is completed, coefla Manual picking cost coefficient
coefmd Multi-depth cost coefficient
and a pod can be retrieved multiple times.

6
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

Table 3 ∑
1≤ ujk ≤ P, ∀k ∈ K (14)
The definition and value range of decision variables.
j∈J
Decision Definition and value
variables
∑ ∑∑
zmni ≥ ujk ymji , ∀i ∈ I, k ∈ K, n ∈ N (15)
xmj Binary variables: 1, if pod m provides picking service for order j, m∈M1 j∈J m∈M1
otherwise, 0
∑ ∑ ∑
ymji Non-negative integer variables: The quantity of items i picked in xmj zmni − ymji = xmjʹ zm(n+1)i , ∀m ∈ M; j, jʹ ∈ J; n ∈ N (16)
the pod m for order j i∈I i∈I i∈I
qmi Non-negative integer variables: The number of slots occupied by
item i in the pod m Z [ ( )]
zmni Non-negative integer variables: The current quantity of item i on
∑ X X Y− Z− 1
dz ⋅⌊ ⌋+ X− ⌊ ⌋⋅dz − 1 ⋅ ≤ card(M1 ) (17)
pod m after its n-th retrieval z=1
d z + 1 dz + 1 Z
pmxy Binary variables: 1, if pod m is at coordinate (x,y), otherwise, 0
ujk Binary variables: 1, if order j is in batch k, otherwise, 0 [ ( )]
P
∑ ∑ ∑
vxyz Binary variables: 1, if point (x, y) is within the z zone, otherwise, 0
costlabor = coefla ⋅ ujk = la (18)
dz Non-negative integer variables: Storage depth of the z zone.
la=1 k∈K j∈J

∑∑
[
∑∑ ∑
] costdeep = coefmd ⋅Dxy ⋅dz ⋅vxyz (19)
(Q − qmi ) ⋅xmj ≤ Q (8) x∈X y∈Y

j∈J m∈M i∈I

The objective function (10) is to minimize the comprehensive pick­
∑ ∑ ∑ ing costs when finishing the order fulfillment process. Constraint (11)
xmj zmni − ymji = xmjʹ zm(n+1)i , ∀m ∈ M; j, jʹ ∈ J; n ∈ N (9)
i∈I i∈I i∈I
means that a pod can only occupy one position. Constraint (12) indicates
that at most one pod can be placed at each position in the storage range.
The objective function (1) is to minimize the number of pod retrieval Constraint (13) means that each order is assigned to a batch. Constraint
times during the order fulfillment process. Constraint (2) indicates that (14) means that each batch contains at least one order and at most P
the quantity of items on the pod can satisfy the needs of all orders. orders. Constraint (15) indicates that the required items in each batch do
Constraint (3) indicates that the quantity of items provided by each pod not exceed the total storage capacity in the pod. Constraint (16) in­
does not exceed its own storage capacity. Constraint (4) means that each dicates that the quantity of items stored will decrease after the pod is
order is served by at least one pod for picking. Constraint (5) means that retrieved once, with the reduction amount being related to the order
the quantity of items stored does not exceed the total number. Constraint demand. Constraint (17) indicates that the setting of depth for each zone
(6) indicates that the actual storage quantity of an item on each pod does must meet the requirements for pod placement. Constraint (18) repre­
not exceed the pod’s maximum capacity for that item. Constraint (7) sents the manual picking cost. Constraint (19) represents the multiple
indicates that the maximum number of slots used by each pod is Q. depth cost.
Constraint (8) ensures that the number of empty slots on all pods that
already stored items is less than Q so as to prevent excessive empty slots 3.4. Problem analysis
from wasting storage space. Constraint (9) indicates that the quantity of
items stored will decrease after the pod is retrieved once, with the Existing literature has established that both the storage location
reduction amount being related to the order demand. assignment problem and order batching are NP-hard (Ding et al. 2024;
Yang et al. 2020). The joint optimization of these two fundamental
3.3. Joint optimization model of pod location assignment and order problems, as discussed in this paper, also exhibits NP-hard computa­
batching tional complexity. Furthermore, the decision variables in this study
include both binary (0–1) and integer variables, which adds to the dif­
Joint optimization of pod location assignment and order batching ficulty of solving the proposed model. According to Mirzaei et al. (2022),
can be described as: In an area of size X × Y, place non-empty pods at using the GUROBI solver for pod location assignment tasks involving
different locations in the system, and the Manhattan distance from each 100 orders and 50 item types requires approximately one day of
location to the picking station is known. Divide all orders into k batches computation. In practice, the actual number of orders and items is
according to their similarity, and pick the orders in a batch together. A significantly larger, making it challenging for exact algorithms to find
batch contains at least one order and at most P orders. It requires a solutions within a reasonable time frame.
reasonable allocation of pod locations and order batching to minimize Moreover, this paper not only addresses pod location assignment but
the comprehensive picking cost. Comprehensive picking costs include also incorporates item placement and order batching. Each order re­
pod retrieval cost, manual picking cost, and multi-depth cost. The quires multiple items, establishing a connection between order demand
optimization result is the coordinates of each pod, the number of and the surplus items on the pods through the variable zmni. Addition­
batches, and the order combinations contained in each batch. This ally, different storage depths can be set in various zones, further
problem is essentially a generalized assignment problem, and the enlarging the solution space compared to addressing each problem in
mathematical model is established as follows: isolation. As the number of pods, storage locations, item types, and order
( ) quantities increases, the solution space grows exponentially. As a result,
∑ ∑∑∑∑
min Dxy pmxy xmj ujk + costlabor + costdepth (10) it is difficult to achieve solutions with exact algorithms within a limited
m∈M1 j∈J k∈K x∈X y∈Y time, highlighting the need for a fast and effective heuristic algorithm to
∑∑ address the problem.
pmxy = 1, ∀m ∈ M1 (11)
x∈X y∈Y
4. Algorithm design
∑
pmxy ≤ 1, ∀x ∈ X, y ∈ Y (12) The joint optimization of storage assignment and order batching in
m∈M
RMFS with dynamic storage depth and surplus items discussed in this
∑ paper is a multi-stage combinatorial optimization problem. There is an
ujk = 1, ∀j ∈ J (13)
k∈K
obvious coupling relationship between each stage of this problem.
However, the solution space of this problem is vast and there are so

7
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

many constraints. Obviously, it is difficult to find the optimal value approach exhibits adaptability and the ability to escape local optima
using an accurate algorithm. Since there is a clear optimization direction (Hansen and Mladenovic 2001). Due to the vast solution space of the
at each stage of this problem, it is suitable to use heuristic algorithms to model presented in this paper, it is easy to get trapped in local optima;
solve it, especially neighborhood search algorithms. The mathematical therefore, a global optimization algorithm is essential for effectively
model constructed in Section 3 may not achieve satisfactory optimiza­ finding the optimal solution. Combining simulated annealing with var­
tion results using a single algorithm; therefore, it is necessary to design iable neighborhood search is a reasonable choice, and incorporating
an algorithm that converges quickly, has high precision, and possesses a greedy search into various neighborhood structures can both accelerate
certain degree of adaptability. the search process and ensure solution quality. Greedy search selects the
locally optimal choice at each step based on the current state, without
considering the subsequent global implications. While it can often
4.1. The framework of TGVNSA quickly find a feasible solution, it does not always guarantee a global
optimal solution. In summary, the algorithm design in this paper is
The simulated annealing algorithm mimics the annealing process in based on the simulated annealing framework, incorporating variable
physics by probabilistically accepting poorer solutions to avoid getting neighborhood search and greedy search to develop a hybrid algorithm
trapped in local optima, making it a global optimization algorithm that achieves rapid convergence while guaranteeing global optimality.
(Kirkpatrick et al. 1983). Meanwhile, variable neighborhood search This algorithm is named the Two-stage Greedy Variable Neighborhood
dynamically alters the neighborhood structure during the search pro­ Simulated Annealing Algorithm (TGVNSA).
cess, exploring the solution space through various structures and The flowchart of TGVNSA solving the joint optimization of storage
selecting an appropriate one based on the current solution. This

Fig. 4. Flow chart of TGVNSA.

8
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

assignment and order batching in RFMS with dynamic storage depth and 4.2. Encoding
surplus items is shown in Fig. 4.
The first stage of the algorithm completes the optimization of item This paper adopts real number coding and uses a two-dimensional
location assignment. The first step is to initialize the types and quantities array to describe the current state of the pod, which contains informa­
of items in each pod. The initialization algorithm is shown in Section tion on the type of items, the number of items, and the pod location. An
4.3. After obtaining the initial cost, the first stage of the greedy variable example of pod encoding is shown in Fig. 5. In this example, each row in
neighborhood simulated annealing process is performed. At each tem­ the array represents information about a pod, it is assumed that each pod
perature, a Markov chain of a certain length is constructed to obtain the has 6 slots, so the first 6 numbers in each row represent the type of item,
optimal solution at that temperature. The process is as follows: first, the middle 6 numbers represent the number of items, and the last two
determine whether the current solution meets the order requirements. If numbers represent the coordinates of the pod in the system.
it does not, the greedy algorithm is used to systematically increase the Similarly, a two-dimensional array is used to encode the order
missing items. If the order requirement is met, the type or quantity of batches. An example of the encoding method is shown in Fig. 6. In this
items in the pod is changed through a random neighborhood. Then, example, each row represents a combination of orders included in a
compare the neighborhood solution with the current solution. If the batch, and the number represents the order number.
neighborhood solution is better, accept the neighborhood solution. If the
current solution is better, accept the neighborhood solution according to 4.3. Initial solution generation
the Metropolis criterion. Until the length of the Markov chain reaches a
specified value, it is considered that the optimal solution at this tem­ The TGVNSA proposed in this paper does not depend on the quality
perature has been found. Subsequently, the temperature is updated, and of the initial solution, and the initial solution at each stage is generally
the next Markov chain will be created to seek the optimal value at the based on the random principle. The initial solution for item location
next temperature. The temperature continuously decreases as the algo­ assignment is constructed through Algorithm 1. Each item is placed in a
rithm iterates a set number of times until the termination condition is slot on a random pod. One type of item is allowed to occupy multiple
met. The flowchart for the first stage is shown on the left side of Fig. 4. pods and multiple slots. The number of items in each slot does not
After the item location assignment optimization, the item combina­ exceed the maximum limit. The initial solution constructed in this way
tion in each pod is obtained to initialize the pod locations and order has sufficient randomness and also indicates that some items are origi­
batches. The initialization algorithm is shown in Section 4.3. Then the nally stored in the pods, which is more in line with the actual situation.
second stage of greedy variable neighborhood simulated annealing Algorithm 1 Constructing the initial solution for item location assignment
process is started to perform joint optimization of pod location assign­ Input: M, I, pi, Q
ment and order batching. The process of constructing the Markov chain Output: types(types of items on each pod), quantity(quantity of items on each pod)
at a certain temperature in this stage is as follows. Given the current pod (continued on next page)
location solution, order batching solution, pod location costs, and total
costs, it is necessary to find the neighborhood solutions for both the pod
location and order batching. Considering that the neighborhood of the
pod location is much larger than the neighborhood of the order batch, in
order to converge as quickly as possible, a variable neighborhood search
using the Metropolis criterion is performed in each iteration of the
Markov chain, with the aim of obtaining the optimal neighborhood so­
lution for the pod location at current temperature. Then, neighborhood
searches are conducted on the order batching. Similarly, accepting the
solutions based on the comprehensive costs and the Metropolis criterion.
Until the length of the Markov chain reaches the specified value, it is
considered that the optimal pod location solution and order batching
solution for the current temperature have been found. Then the tem­
perature is updated and iterated a certain number of times until the
termination condition is reached. The flowchart for the second stage is
shown on the right side of Fig. 4.

Fig. 6. Encoding method of batches.

Fig. 5. Encoding method of pods.

9
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

(continued ) variety of neighborhoods. In order to meet the order requirements

Algorithm 1 Constructing the initial solution for item location assignment during the item location assignment, two different neighborhoods are
1: types = zero(length(M), Q);
used: one based on greedy search and the other based on random
2: quantity = zero(length(M), Q); principle. Before creating a neighborhood solution, check whether the
3: for i = 1: length(M) current solution can meet the order requirements. If not, use the
4: types[i,:] = randperm(length(I), min(length(I), Q)); neighborhood based on a greedy search to increase the missing items in
5: for j = 1: Q
the pods. If the order requirements are met, a neighborhood based on the
6: n = types[i, j];
7: quantity [i, j] = randperm(pn, Q); random principle is used. The two neighborhood structures are shown in
8: end Fig. 7. Fig. 7(a) shows a greedy neighborhood. When it is detected that
9: end the number of items with sequence number 15 is insufficient during the
10: return types and quantity; order retrieval process, the number of items with sequence number 15 is
increased in the neighborhood solution. Fig. 7(b) shows a neighborhood
based on the random principle. We select a slot of a pod and change the
The initial solution for pod location assignment and order batching is
type and quantity of items stored in it. In order to increase the speed of
constructed through Algorithm 2. The pods are randomly arranged to
the algorithm, in the random neighborhood, the quantity of items is
storage locations in the system until all pods are assigned locations. The
changed to any value between the current value and the maximum
initial solution of order batching is to place only one order in each batch,
value.
which is the worst solution in order batching. TGVNSA is also applicable
In the joint optimization of pod location assignment and order
to such initial solutions.
batching, the neighborhood range of pod location is much larger than
the neighborhood range of order batching. For pod location assignment,
Algorithm 2 Constructing the initial solution for pod location assignment and order
batching convergence is too slow if a single neighborhood is used. In order to
converge as quickly as possible, a variable neighborhood search is per­
Input: M1, J, K, location (system configuration)
Output: position (location of each pod), batch (combinations of orders in each batch) formed in each iteration of the Markov chain. Since the temperature
1: position = zeros(length(M1), 2); information is already available, in the variable neighborhood search,
2: index = 1; the Metropolis criterion is also used to accept poor neighborhood solu­
3: for i = 1: size(location, 1) tions, with the goal of obtaining the optimal neighborhood solution for
4: for j = 1:size(location, 2)
5: if point (i, j) is storage location
the pod location at the current temperature. Pod location assignment
6: position(index,:) = [j, i]; uses three different neighborhood structures: 9-grid optimal neighbor­
7: index = index + 1; hood, column optimal neighborhood, and random neighborhood. Fig. 8
8: Check whether each pod has been assigned coordinates, if so, break; (a) shows the 9-grid optimal neighborhood. The first number in the
9: end
square is the serial number of a pod, and the number in the bracket is the
10: end
11: end number of retrieval times for this pod. This neighborhood is to arrange
12: batch = zeros(length(J), K); the pods in ascending order according to the number of retrieval times
13: numbers = randperm(length(J)); within the 9 grids. Fig. 8(b) shows the column optimal neighborhood,
14: batch(:, 1) = numbers.’; and this neighborhood is to arrange the pods in ascending order ac­
15: return position and batch;
cording to the number of retrieval times within a column. The random
neighborhood is to swap any two pod locations in the system (not shown
in the figure).
4.4. Neighborhood solution generation The greedy neighborhood is suitable for order batching. The opera­
tion is as follows: select order A in any batch m in the current order
The neighborhood has a direct impact on the algorithm’s exploration batching solution, and find the order B with the greatest similarity to A
ability and convergence. The TGVNSA proposed in this paper adopts a in other batches. If batch m is not full, add order B to it. As shown in

Fig. 7. Neighborhood structures of item location assignment.

10
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

Fig. 8. Neighborhood structures of pod location assignment.

Fig. 9. Neighborhood structures of order batching.

Fig. 9(a), the upper number in the square represents the serial number of
an order, and the number in the square brackets represents the similarity
with order A. If batch m is full, find order C in batch m which has the
lowest similarity to order A. If the similarity value of B is greater than C,
4.5. Energy function calculation
swap B and C, otherwise skip, as shown in Fig. 9(b).
Jaccard similarity is a fast and efficient metric used to calculate the
The energy function of the TGVNSA is calculated according to the
similarity between two sets. Since the basic Jaccard similarity may not
objective functions of each stage in Section 3. Each stage of this algo­
accurately describe the similarity between sets containing duplicate el­
rithm needs to calculate the number of pod retrieval times. It is neces­
ements, this paper proposes a Weighted Jaccard Similarity Calculation
sary to consider the consumption of the number of items in each pod
Algorithm to characterize the similarity between orders. The pseudo­
during the order retrieval process, which is a practical approach. Algo­
code is presented in Algorithm 3.
rithm 4 is the calculation process of the number of pod retrieval times
Algorithm 3 Calculate order similarity
considering dynamic surplus items.
Input: order A、order B Algorithm 4 Calculate the number of pod retrieval times considering dynamic surplus
Output: similarity (similarity between orders A and B) items
1: jaccard_similarity = 0;
2: unique_elements = union(A, B); % Find all unique elements in A and B. Input: types(types of items on each pod), quantity(quantity of items on each pod),
3: count_A = histc(A, unique_elements); % Frequency of each unique element in A orders(item requirements in the order)
4: count_B = histc(B, unique_elements); % Frequency of each unique element in B Output: cost(Costs associated with the number of pod retrieval times)
5: intersection = sum(min(count_A, count_B)); % Intersection (use minimum 1: calls = 0; % Record the number of pod retrieval times
frequency) 2: for i = 1:size(orders, 1) % Traverse all orders
6: union_count = sum(max(count_A, count_B));% Union (use maximum 3: row_orders = orders(i,:); % Extract a required item from the order
frequency) 4: for j = 1:length(types) % Traverse all pods
7: jaccard_similarity = intersection / union_count; % Calculate the Jaccard 5: id = find(ismember(types(j,:), row_order));
similarity 6: if any(id) && any(quantity(j, id)) % The pod contains this item
8: return jaccard_similarity; 7: calls = calls + 1; % The pod is called
(continued on next page)

11
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

(continued ) 5.1. Calculation results

Algorithm 4 Calculate the number of pod retrieval times considering dynamic surplus
items The most prominent advantage of considering dynamic storage
8: for n = 1:length(id) % Find the slot where the item is located in the depth in RMFS compared to traditional systems is that the configuration
pod can be flexibly changed according to order requirements in the item
9: need_product = types(j, id(n)); location assignment stage. The main manifestations are allowing the
10: count = numel(find(row_order == need_product)); % Item demand
number of pods to exceed the maximum quantity limit of traditional
11: if the number of items on the pod is sufficient for order i
12: quantity(j, id(n)) = quantity(j, id(n)) – count; % Reduce surplus systems, and adopting different storage depths for high-frequency and
items low-frequency pods to improve system efficiency. This section sets
13: row_order(indices(1:count)) = –1; % Mark this item as retrieved different parameters and order requirements for systems of different
14: else % the number of items on the pod is not sufficient for order i sizes and solves them through the TGVNSA. The calculation process is
15: row_order(indices(1: quantity (j, id(n))) = –1; % Part of items
retrieved
written in MATLAB R2018b and tested by running the program on a
16: quantity (j, id(n)) = 0; % Set the surplus items on this pod to 0 device with Intel(R) Core(TM) i7-7500U CPU 2.70 GHz processor and
17: end 8G RAM.
18: orders(i,:) = row_order; % Update orders The data in the experimental cases is derived from a material su­
19: end
permarket and semi-finished product warehouse of a manufacturing
20: end
21: end company in eastern China. Three warehouses of different sizes were
22: end selected for the experiments. The size of the small warehouse is 10 × 10
23: aa = sum(sum(new_orders > 0)); (with side lengths of the cargo space and the aisle both equal to 1); the
24: cost = calls+(10 * aa); % Add non-completion penalty medium warehouse is 20 × 20; and the large warehouse is 30 × 30. The
25: return cost;
parameters for each warehouse are shown in Table 4, where the upper
limit of types of items is set to 96 for the small warehouse, 220 for the
medium warehouse, and 380 for the large warehouse. Each warehouse
contains three configurations: Configuration I uses the traditional single-
In the joint optimization of pod location assignment and order batching,
depth storage, as shown in
the energy function is the comprehensive picking cost in the current
Fig. 10(a). Configuration II uses a combination of single-depth and
state, which is composed of pod retrieval cost, manual picking cost, and
double-depth storage, as shown in Fig. 10(b). Configuration III adopts
multi-depth cost. Combined with the cost function in Section 3.3, the
single-depth storage in the high-frequency zone (zone 3), double-depth
energy function reflecting the comprehensive picking cost is calculated
storage in the middle zone (zone 2), and multi-depth storage in the low-
by Algorithm 5. The pod retrieval cost is equal to the product of the
frequency zone (zone 1), as shown in Fig. 10(c). The picking station in
number of pod retrieval times and the distance from the pod to the
the system is located in the middle of the bottom.
picking station. The number of pod retrieval times is also calculated by
In this section, the order demand is set as follows: 20 % of the items
Algorithm 4.
have a higher demand, which can satisfy 80 % of the order content. The
Algorithm 5 Calculate the comprehensive picking cost
relevant parameter settings in the example are shown in Table 5. The
Input: types(types of items on each pod), quantity(quantity of items on each pod), number of iterations is set to 100, 200, and 500 for the small, medium,
orders(item requirements in the order), batch(combinations of orders in each batch),
and large warehouses, respectively, since larger systems require more
position(location of each pod), IO_point(picking station location)
Output: cost(comprehensive picking cost) iterations to achieve an optimal solution. The manual picking cost co­
1: batch_order_list records the items required for all orders in each batch; efficient coefla, is assigned values of 1, 1, 1.2, 1.4, and 1.8, corresponding
2: actual_calls records the number of pod retrieval times calculated by to order quantities of 1, 2, 3, 4, and 5 in a single batch, this data is
Algorithm 4; approximately obtained from the results regarding personnel energy
3: distances = abs(position – IO_point);
4: manhattan_distances = sum(distances, 2);
consumption at picking stations from the literature by J. Zhang et al.
5: result = manhattan_distances.* actual_calls; (2023). The multi-depth cost coefficient coefmd, takes values of 1, 1.2,
6: retrieval_cost = sum(result(:)); 1.5, and 2, corresponding to storage depths of 1, 2, 3, and 4, respec­
7: for i = 1:size(batch, 1) tively, this data is approximately obtained from the picking costs for
8: if the number of elements in batch(i,:) is 3
multi-depth storage as reported by Lanza et al. (2022) and Yang et al.
9: num3 = num3 + 1;
10: if the number of elements in batch(i,:) is 4 (2022).
11: num4 = num4 + 1; The results of the corresponding examples are shown in Table 6.
12: if the number of elements in batch(i,:) is 5 According to Table 6, considering the dynamic storage depth in
13: num5 = num5 + 1; RMFS shows outstanding advantages when dealing with orders with
14: end
15: manual_cost = length(position) * (coef3 * num3 + coef4 * num4 + coef5 *
different demands: it can meet the order requirements that are close to
num5); or exceed the system’s saturation capacity. In case C1, the joint opti­
16: mdepth records the distance between each coordinate and the picking mization of storage assignment and order batching is carried out in a 10
station; × 10 system using configuration I, so that it could meet the picking task
17: mdepth_d records the depth of each pod;
of completing 30 orders (requiring 1,200 items) at one time. The heat
18: mdepth_cost = coefmd * (mdepth.* mdepth_d);
19: cost = retrieval_cost + manual_cost + mdepth_cost; map of the number of pod retrieval times for case C1 is shown in Fig. 11
20: return cost;

Table 4
The parameters of the warehouse.
Parameter Value
5. Computational experiments
Types of items 96/220/380
Quantity of items in each order 40
This section will analyze the results of solving the joint optimization Demand skewness 20/80
of storage assignment and order batching in RMFS with dynamic storage Maximum capacity of items in a slot 8(90 %)/6(10 %)
depth and surplus items through TGVNSA, and verify the performance of Q 6
P 5
the algorithm.

12
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

Fig. 10. System configuration diagram.

III can meet the needs of 200 orders. The heat map of case C8 is shown in
Table 5
Fig. 11(f).
Parameters for the TGVNSA algorithm.
In cases C9, C10, and C11, with a larger number of pods and 380
Parameter Value types of items, the highest order quantity reached 500 (requiring 20,000
The initial temperature 1000 items), verifying that the proposed algorithm still has considerable ef­
Cooling coefficient 0.95 fects in large-scale systems. Heat maps about cases C9, C10, and C11 are
Markov chain length 50
shown in Fig. 11(g), (h) and (i).
Iterations 100/200/500
Manual picking cost coefficient coefla 1/1/1.2/1.4/1.8 The heatmap in Fig. 11 illustrates the pod retrieval times at various
Multi-depth cost coefficient coefmd 1/1.2/1.5/2 locations in the storage area that are accessed during the completion of
all orders. The colored squares in the figure represent the pods, with
colors closer to yellow indicating a higher number of retrieval times,
Table 6
while colors closer to blue indicate fewer retrievals. The bar chart on the
Results of different cases. right displays the number of accesses corresponding to each color. From
the heatmap, it is visually apparent that pods with a higher retrieval
Case Size Cfg MNP QO ANPU NPRT NB CPC
frequency are located closer to the picking station, while those with a
C1 10 × 10 I 40 30 40 170 7 3412 lower retrieval frequency are farther away. Additionally, the pods
C2 10 × 10 I 40 50
placed on the inner side of the multi-depth area have the least number of
− − − −
C3 10 × 10 II 48 50 48 352 12 6327
C4 10 × 10 III 56 50 50 339 11 6764 retrieval times. This figure demonstrates that the model and algorithm
C5 20 × 20 I 170 130 170 1725 28 50,725 proposed in this paper yield reasonable results.
C6 20 × 20 II 206 130 182 1651 27 47,836
C7 20 × 20 II 206 200 − − − −
C8 20 × 20 III 224 200 224 2751 45 91,692 5.2. Parameter sensitivity analysis
C9 30 × 30 I 390 300 390 5269 71 227,303
C10 30 × 30 II 460 400 460 6982 109 307,282
This section conducts parameter sensitivity analysis on the proposed
C11 30 × 30 III 525 500 525 8502 135 410,371
system and algorithm. The parameters considered include item demand
NOTE: Cfg: Configuration, MNP: Maximum number of pods, QO: Quantity of skewness, item types, maximum order quantity per batch, and average
order, ANPU: Actual number of pods in use, NPRT: Number of pod retrieval storage depth (system configuration).
times, NB: Number of batches, CPC: Comprehensive picking costs.
First, we analyze the impact of different item demand skewness on
system operational results. Using the basic data of case C4, calculate the
(a). In case C2, the 10 × 10 system in the configuration I cannot meet the number of pod retrieval times (NPRT) and the comprehensive picking
item location assignment requirements of 50 orders (requiring 2,000 costs (CPC) under the demand skewness of 20/80, 40/80, and 60/80,
items) because there are insufficient pods in the system. In case C3, the where 20/80 indicates that 20 % of the items satisfy 80 % of the order
system is modified to configuration II, where some areas adopted demand. To avoid chance, each data is averaged from the results of 10
double-depth storage, increasing the maximum number of pods repeated runs. The results are shown in Fig. 12(a). It can be seen that
accommodated in the system. The results showed that it could complete under the same conditions, the fewer the high-demand items, the fewer
50 orders. The heat map of case C3 is shown in Fig. 11(b). In case C4, the the number of pod retrieval times, and the lower the comprehensive
system is changed to configuration III, which involves multi-depth picking costs. Conversely, the more the item demand skewness tends to
storage in some areas, further increasing the maximum number of average, the greater the number of pod retrieval times and the higher the
pods that can be accommodated. The results showed that the frequency comprehensive picking costs.
of pod usage was close to C3, but it incurred a higher comprehensive Upon further analysis of the impact of the different number of item
picking cost. This is because the multi-depth storage areas require types on the system operational results, calculations are performed for
additional costs for retrieval. The heat map of case C4 is shown in Fig. 11 cases with 50, 100, and 150 item types, based on case C4, as shown in
(c). However, the configuration III used in the C4 can obviously support Fig. 12(b). It is evident that, under the same conditions, the greater the
more order retrieval tasks. number of item types, the more frequent the pod retrieval instances, and
Comparing C5 and C6, it can be seen that under the same conditions, the higher the comprehensive picking costs. This means that orders of
different outcomes are obtained in different configurations. For the same multiple varieties and small batches often require more pods to meet the
task of completing 130 orders (requiring 5,200 items), configuration II, demand than orders of fewer varieties and large batches. The system
which adopts double-depth storage, enables a denser placement of racks, proposed in this paper, which takes into account the dynamic storage
ultimately resulting in better performance than configuration I. The heat depth and surplus items, has a stronger ability to cope with the demand
map of case C5 and case C6 are shown in Fig. 11(d) and (e). However, for multiple varieties and small batches.
when the order demand increases to 200 (requiring 8,000 items), case The impact of the maximum order quantity per batch on the system
C7 shows that the system using configuration II cannot complete the operation results is shown in Fig. 12(c). The accumulation of manual
assignment task for so many orders. Case C8 indicates that configuration picking costs is also taken into account in the calculation process. It can

13
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

Fig. 11. The heat map of the number of pod retrieval times.

Fig. 12. Parameter sensitivity analysis.

be concluded that the larger the maximum number of orders allowed in Average storage depth (system configurations) will affect the system
each batch, the less frequently pods need to be retrieved, and the operation results. The sensitivity of this needs to be verified. In order to
comprehensive picking costs are subsequently reduced. In practice, if compare the three configurations, a 10 × 10 system is selected to
the maximum order quantity per batch is set too high, it could over­ complete the retrieval task of 40 orders. The results are shown in Fig. 12
whelm manual picking operations, increase error rates, and cause (d). It can be seen that under the same circumstances, adopting a
congestion with pods and robots piling up near the picking station. configuration with a high storage depth will reduce the number of pod
Therefore, setting a predetermined limit for the maximum order quan­ retrieval times, because it provides more pods for use. However, when
tity per batch before the system is operational is crucial. This article the storage depth is greater than 2, there is an additional multi-depth
serves as a foundation for establishing this limit. cost, which increases the comprehensive picking costs.
Different system configurations, that is, different storage depths, will It can also be seen from Fig. 12 that changes in item types and
affect the system operation results. The sensitivity of this item needs to maximum order quantity per batch have a greater impact on the number
be verified. In order to compare the three configurations, a 10 × 10 of pod retrieval times and the comprehensive picking costs, which is far
system is selected to complete the retrieval task of 40 orders. The results greater than the impact of changes in other parameters.
are shown in Fig. 12(d).

14
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

5.3. Comparison of different algorithms number of pod retrieval times, the RNSA and GVNS algorithms have
significantly lower accuracy compared to TGVNSA. IAGA and TGVNSA
In this section, the proposed TGVNSA algorithm is compared with perform similarly well, with only a 0.15 % average difference in their
other algorithms to verify its performance. In the two categories of al­ results. In comparing CQ values, RNSA and GVNS perform worse than
gorithms, namely, neighborhood search and population optimization, TGVNSA. IAGA outperforms TGVNSA in small to medium systems but
several algorithms are selected to be applied to solve the problem pro­ falls short in large-scale systems. It’s important to note that the number
posed in this article. The comparative algorithms include the Random of pod retrieval times and CQ values are merely process variables and
Neighborhood Simulated Annealing (RNSA), Greedy Variable Neigh­ not the ultimate optimization objectives.
borhood Search (GVNS), and Improving Adaptive Genetic Algorithm Upon comparing the comprehensive picking costs for each case, it is
(IAGA). evident that RNSA has the lowest accuracy, exhibiting an average dif­
The initial solution generation in the comparison algorithms follows ference of 20.73 % from TGVNSA. GVNS shows a better performance
the method outlined in Section 4.3 of this paper, while the calculation of than RNSA, with its average accuracy difference from TGVNSA being
the fitness function uses the approach described in Section 4.5. In the 6.9 %. In small and medium-sized systems, IAGA achieves nearly the
RNSA algorithm, random neighborhoods are employed for constructing same level of accuracy as TGVNSA. In scenarios of large systems, the
the Markov chain. The GVNS algorithm utilizes both random variable IAGA algorithm exhibits a substantial decline in accuracy. In the 30 × 30
neighborhoods and greedy neighborhoods, and the design of the warehouse, IAGA performs 11.62 % worse than TGVNSA, as shown by
neighborhood structures in the GVNS algorithm is the same as that in the the blue segment in the CPC column of Fig. 13 This is because, as the
TGVNSA algorithm. The IAGA dynamically adjusts the crossover and system size increases, the solution space expands significantly, making it
mutation probabilities to better balance global search and local explo­ difficult for 100 individuals in the population to cover most of the space,
ration, enhancing the convergence speed of the algorithm. The calcu­ resulting in a decline in search quality. Increasing the population would
lation methods for crossover and mutation probabilities refer to Y. Li greatly increase the computational load and time. Therefore, it is
et al. (2022). The parameter values for the three comparison algorithms evident that the TGVNSA proposed in this paper demonstrates greater
are shown in Table 7. advantages in larger-scale scenarios.
Different algorithms are used to calculate the feasible cases in Sec­ Regarding CPU time, IAGA, which utilizes the concept of population
tion 5.1, and the comparison of the following objective values is made: optimization, requires considerably more time compared to other al­
number of pod retrieval times, comprehensive picking costs, CPU time, gorithms. Due to its use of the random neighborhood, RNSA outperforms
and the comprehensive quality in the item location assignment (CQ). both GVNS and TGVNSA in terms of CPU time efficiency.
Among them, the comparison of the number of pod retrieval times, Considering all comparison items, TGVNSA exhibits excellent per­
comprehensive picking costs, and CPU time is calculated by the formula formance in both computational accuracy and CUP time, especially in
VALUE(other)/VALUE(TGVNSA) × 100 %. The larger the value, the larger-scale scenarios where its advantages are even more pronounced.
worse the comparative algorithm performs.
CQ represents the effect of the first stage of the system, which is 5.4. Comparison of different strategies
calculated by formula (20). The parameters involved in determining CQ
are the actual number of pod retrieval times AN, the variance of the The strategy adopted by the storage location assignment and order
number of pod retrieval times VN, and the number of unused slots NN. batching joint optimization model (SLAOB) proposed in this article
The higher the CQ value, the worse the performance of the comparative combines the lowest pod recall frequency and the highest order rele­
algorithm, indicating that the ability to cope with order fluctuations vance into a cost-optimal strategy. This section provides a comparison
after item location assignment is very weak. The purpose of this com­ between SLAOB and various other strategies. For item location assign­
parison is to clearly understand the key metrics during the operation of ment, the strategies considered are Random Storage (RS), Classified
each algorithm, allowing for a more reasonable evaluation of each one. Storage (CS), Optimal Clustering (OC), and Lowest Retrieval Times (LR).
( ) For pod location assignment, the strategies examined are Random
1 AN(other) VN(other) NN(other)
CQ = + + × 100% Storage (RS), Classified Storage (CS), Decentralized Storage (DS), and
3 AN(TGVNSA) VN(TGVNSA) NN(TGVNSA)
Optimal Distance (OD). Additionally, the order batching strategies
(20)
analyzed include Fixed Quantity (FQ), Time Window (TW), and Optimal
The comparison results are shown in Relevance (OR).
Fig. 13. It can be seen that TGVNSA has good performance in solution Different strategy combinations are compared with SLAOB,
quality and time compared with other algorithms. First, in terms of including CS-CS-FQ, CS-CS-OR, OC-CS-OR, LR-CS-OR, OC-OD-OR, and
LR-OD-OR. The meaning of CS-CS-FQ is that the item location assign­
ment phase adopts the Classified Storage (CS) strategy, the pod location
Table 7 assignment uses the Classified Storage (CS) strategy, and order batching
Parameters of the comparison algorithms.
adopts the Fixed Quantity (FQ) strategy, with the others following in a
Name Parameter Value similar manner. Random Storage (RS) does not meet the requirements of
RNSA The initial temperature 1000 this paper, Decentralized Storage (DS) aims to reduce congestion, and
Cooling coefficient 0.95 the Time Window method (TW) involves issues of order sequencing,
Markov chain length 50 therefore they are not discussed in this paper. When calculating the
Iterations 100/200/500
outcomes of the mentioned strategy combinations, if optimization is
required, the TGVNSA algorithm framework introduced in this paper
GVNS Neighborhood structure Same as TGVNSA should be used. This ensures that the conditions for comparison are
Greedy search depth 5
Iterations 100/200/500
consistent.
Comparative experiments are carried out in the following two cases:
Case 1: In a 10 × 10 system with configuration III, 40 orders are
IAGA Population size 100
Max/min Crossover Probability 0.9/0.4 completed, each order contains 40 item requirements, and the types of
Crossover operator 0.6 items are 40; Case 2: In a 20 × 20 system with configuration I, 40 orders
Max/min mutation probability 0.1/0.005 are completed, each order contains 40 item requirements, and the types
Mutation Operator 0.025 of items are 40.
Iterations 100/200/500
The comparison results of different strategy combinations are shown

15
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

Fig. 13. Comparison of different algorithms.

in Fig. 14. Since the result of the RS strategy is much larger than other OD-OR and decreased by 15.9 % compared to OC-OD-OR. In Case 2, it
types, it is not marked in Fig. 14. Obviously, the OC and LR strategies are decreased by 10.2 % compared to LR-OD-OR and decreased by 8.1 %
better than the RS and CS strategies in the item location assignment compared to OC-OD-OR. It can be seen that under the storage configu­
stage. As shown in Fig. 14, the OC-OD-OR and LR-OD-OR strategies are ration with higher depth, SLAOB has more outstanding optimization
compared. For high storage densities like Case 1, the LR strategy proves effects.
more effective due to requiring fewer pods. Conversely, in more spacious Based on the comparison results of various strategies in Fig. 14, the
environments such as Case 2, the wide choice of pod locations makes the conclusions are summarized as follows: (i) During the item location
OC strategy more efficient. assignment phase, the OC and LR strategies can bring the most signifi­
In the pod location assignment stage, OD is better than CS and RS. cant optimization effect improvements, and the comprehensive picking
Comparing the results of the OC-OD-OR strategy combination in Case 1 costs are on average 22.71 % and 28.94 % lower than the CS strategy.
and Case 2, it can also be concluded that the OD strategy has more (ii) During the pod location assignment phase, the OD strategy can bring
significant advantages in a larger storage area. the most significant optimization effect improvements, and the
During the order batching phase, the effect of the OR strategy is comprehensive picking costs are on average 31.22 % to 47.08 % lower
superior to FQ. However, the effect of the OR strategy also depends on than the CS strategy. (iii) Joint optimization has better effects than in­
the composition of the orders. The greater the similarity of the orders, dividual optimization at each stage, it can reduce the comprehensive
the better the effect of using the OR strategy. picking costs by 11.46 %.
The SLAOB proposed in this paper is a joint optimization of the LR-
OD-OR strategy combination. As can be seen in Fig. 14, the effect of 6. Conclusions
SLAOB is better than other strategy combinations.
As can be seen in Fig. 14, the effect of SLAOB is significantly better This paper studies the joint optimization of storage location assign­
than other strategy combinations and is also better than the LR-OD-OR ment and order batching in RMFS, taking into account dynamic storage
strategy combination. The reason is that SLAOB is a joint optimization of depth and surplus items. Firstly, a joint optimization model of storage
the LR-OD-OR strategy combination, and each part of the LR-OD-OR location assignment and order batching is established. The model is
strategy combination is optimized separately. The comprehensive divided into two stages. The first stage is the item location assignment
picking costs of SLAOB in Case 1 decreased by 10.4 % compared to LR- optimization model, which is used to describe the types and quantities of

Fig. 14. Comparison of different strategy combinations.

16
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

items placed in each slot on each pod. The second stage is the joint Cheng, B. Y., Wang, L. J., Tan, Q., & Zhou, M. (2024). A deep reinforcement learning
hyper-heuristic to solve order batching problem with mobile robots. Applied
optimization model of pod location assignment and order batching,
Intelligence, 54(9–10), 6865–6887. https://doi.org/10.1007/s10489-024-05532-9
which is used to describe the coordinates of each pod, the number of D’Andrea, R., Wurman, P., & IEEE. (2008, Nov 10-11). Future challenges of coordinating
order batches, and the order combinations contained in each batch. hundreds of autonomous vehicles in distribution facilities. Paper presented at the IEEE
Given the unique nature of the problem, a two-stage greedy variable International Conference on Technologies for Practical Robot Applications, Woburn,
MA.
neighborhood simulated annealing algorithm (TGVNSA) is proposed to Ding, T. R., Zhang, Y. K., Wang, Z., & Hu, X. P. (2024). Velocity-based rack storage
solve the model. Finally, numerical experiments prove that the algo­ location assignment for the unidirectional Robotic Mobile Fulfillment system.
rithm can effectively solve the established model. TGVNSA is compared Transportation Research Part E-Logistics and Transportation Review, 186, 23. https://
doi.org/10.1016/j.tre.2024.103533
with other algorithms in terms of number of pod retrieval times, Gabellini, M., Calabrese, F., Regattieri, A., Loske, D., & Klumpp, M. (2024). A hybrid
comprehensive picking costs, CPU time, and CQ value. The results show approach integrating genetic algorithm and machine learning to solve the order
that the overall performance of TGVNSA is superior to that of other al­ picking batch assignment problem considering learning and fatigue of pickers.
Computers & Industrial Engineering, 191, 15. https://doi.org/10.1016/j.
gorithms. By comparing the strategy used in this paper with other cie.2024.110175
strategies, it has been demonstrated that, when considering dynamic Gil-Borrás, S., Pardo, E. G., Alonso-Ayuso, A., & Duarte, A. (2021). A heuristic approach
storage depth and surplus items, the optimization effect of the proposed for the online order batching problem with multiple pickers. Computers & Industrial
Engineering, 160, 20. https://doi.org/10.1016/j.cie.2021.107517
method in this paper is significantly better than other strategy combi­ Gil-Borrás, S., Pardo, E. G., Jiménez, E., & Sörensen, K. (2024). The time-window
nations. Moreover, under the same combination of strategies, the joint strategy in the online order batching problem. International Journal of Production
optimization method proposed in this paper reduces the comprehensive Research, 62(12), 4446–4469. https://doi.org/10.1080/00207543.2023.2263884
Hansen, P., & Mladenovic, N. (2001). Variable neighborhood search: Principles and
picking costs by 11.46 % compared to the separate optimization
applications. European Journal of Operational Research, 130(3), 449–467. https://doi.
approach. org/10.1016/s0377-2217(00)00100-4
This paper successfully achieves the joint optimization of storage Hausman, W. H., Schwarz, L. B., & Graves, S. C. (1976). Optimal storage assignment in
location assignment and order batching in RMFS. Experimental results automatic warehousing systems. Management Science, 22(6), 629–638. https://doi.
org/10.1287/mnsc.22.6.629
demonstrate the effectiveness of the model and algorithm. However, for Henn, S. (2012). Algorithms for on-line order batching in an order picking warehouse.
computational convenience, Manhattan distance is used to describe the Computers & Operations Research, 39(11), 2549–2563. https://doi.org/10.1016/j.
proximity between pods and the picking station, which simplifies the cor.2011.12.019
Keung, K. L., Lee, C., & Ji, P. (2022). Industrial internet of things-driven storage location
actual travel paths of AGVs. Future research on task allocation and path assignment and order picking in a resource synchronization and sharing-based
planning in RMFS could also consider dynamic depth and remaining Robotic Mobile Fulfillment system. Advanced Engineering Informatics, 52, 19. https://
item quantities. For the issue of pod retrieval in multi-depth areas, doi.org/10.1016/j.aei.2022.101540
Keung, K. L., Lee, C. K. M., & Ji, P. (2021). Data-driven order correlation pattern and
collaborative handling by multiple robots could be explored to enhance storage location assignment in robotic mobile fulfillment and process automation
system efficiency. Additionally, an attempt could be made to achieve system. Advanced Engineering Informatics, 50, 15. https://doi.org/10.1016/j.
full-process joint optimization for RMFS by simultaneously considering aei.2021.101369
Kim, H. J., Pais, C., & Shen, Z. J. M. (2020). Item assignment problem in a Robotic Mobile
storage location assignment, order batching, task allocation, and path Fulfillment system. IEEE Transactions on Automation Science and Engineering, 17(4),
planning, and designing algorithms that incorporate AI decision-making 1854–1867. https://doi.org/10.1109/tase.2020.2979897
to increase solution reliability and efficiency. Kirkpatrick, S., Gelatt, C. D., Jr., & Vecchi, M. P. (1983). Optimization by simulated
annealing. Science (New York, N.Y.), 220(4598), 671–680. https://doi.org/10.1126/
science.220.4598.671
CRediT authorship contribution statement Kong, X. T. R., Zhu, M. H., Liu, Y., Qin, K. D., & Huang, G. Q. (2023). An advanced order
batching approach for automated sequential auctions with forecasting and
Zhi Liu: Writing – original draft, Validation, Methodology, Investi­ postponement. International Journal of Production Research, 61(12), 4180–4195.
https://doi.org/10.1080/00207543.2021.2022234
gation. Jiansha Lu: Writing – review & editing, Supervision, Resources, Lamballais, T., Roy, D., & De Koster, M. B. M. (2017). Estimating performance in a
Methodology. Chenhao Ren: Writing – review & editing, Validation. Robotic Mobile Fulfillment system. European Journal of Operational Research, 256(3),
Jun Chen: Validation. Zhilong Xu: Visualization. Guoli Zhao: Writing 976–990. https://doi.org/10.1016/j.ejor.2016.06.063
Lanza, G., Passacantando, M., & Scutellà, M. G. (2022). Assigning and sequencing storage
– review & editing, Resources. locations under a two level storage policy: Optimization model and matheuristic
approaches x2729. Omega-International Journal of Management Science, 108, 17.
Acknowledgements https://doi.org/10.1016/j.omega.2021.102565
Lee, I. G., Chung, S. H., & Yoon, S. W. (2020). Two-stage storage assignment to minimize
travel time and congestion for warehouse order picking operations. Computers &
This article is supported by Zhejiang Science and Technology Plan Industrial Engineering, 139, 13. https://doi.org/10.1016/j.cie.2019.106129
Project (Grant No. 2018C01003, 2024C01208). Li, H. R., Liu, J. S., Hu, P., & Zhou, H. (2024). Solving the storage location assignment of
large-scale automated warehouse based on dynamic vortex search algorithm. Swarm
and Evolutionary Computation, 91, 27. https://doi.org/10.1016/j.
Data availability swevo.2024.101725
Li, X. W., Hua, G. W., Huang, A. Q., Sheu, J. B., Cheng, T. C. E., & Huang, F. Q. (2020).
Storage assignment policy with awareness of energy consumption in the Kiva mobile
Data will be made available on request.
fulfilment system. Transportation Research Part E-Logistics and Transportation Review,
144, 18. https://doi.org/10.1016/j.tre.2020.102158
References Li, Y., Wan, Y., Zhang, Y., & Kuang, H. (2022). Path planning for warehouse robot based
on the artificial bee colony-adaptive genetic algorithm. Chinese Journal of Scientific
Azadeh, K., De Koster, R., & Roy, D. (2019). Robotized and automated warehouse Instrument, 43(4), 282–290.
systems: Review and recent developments. Transportation Science, 53(4), 917–945. Lin, C. C., Kang, J. R., Hou, C. C., & Cheng, C. Y. (2016). Joint order batching and picker
https://doi.org/10.1287/trsc.2018.0873 Manhattan routing problem. Computers & Industrial Engineering, 95, 164–174.
Barros, I. R. D., & Nascimento, T. P. (2021). Robotic Mobile Fulfillment systems: A survey https://doi.org/10.1016/j.cie.2016.03.009
on recent developments and research opportunities. Robotics and Autonomous Liu, M. Y., & Poh, K. L. (2023). E-commerce warehousing: An efficient scattered storage
Systems, 137, 15. https://doi.org/10.1016/j.robot.2021.103729 assignment algorithm with bulky locations. Computers & Industrial Engineering, 181,
Boysen, N., Briskorn, D., & Emde, S. (2017). Parts-to-picker based order processing in a 18. https://doi.org/10.1016/j.cie.2023.109236
rack-moving mobile robots environment. European Journal of Operational Research, Ma, Z. Q., Wu, G. H., Ji, B., Wang, L., Luo, Q. Z., & Chen, X. J. (2023). A Novel scattered
262(2), 550–562. https://doi.org/10.1016/j.ejor.2017.03.053 storage policy considering commodity classification and correlation in Robotic
Boysen, N., de Koster, R., & Weidinger, F. (2019). Warehousing in the e-commerce era: A Mobile Fulfillment systems. IEEE Transactions on Automation Science and Engineering,
survey. European Journal of Operational Research, 277(2), 396–411. https://doi.org/ 20(2), 1020–1033. https://doi.org/10.1109/tase.2022.3178934
10.1016/j.ejor.2018.08.023 Menéndez, S., Pardo, E. G., Alonso-Ayuso, A., Molina, E., & Duarte, A. (2017). Variable
Cezik, T., Graves, S. C., & Liu, A. C. (2022). Velocity-based stowage policy for a neighborhood search strategies for the order batching problem. Computers &
semiautomated fulfillment system. Production and Operations Management, 16. Operations Research, 78, 500–512. https://doi.org/10.1016/j.cor.2016.01.020
https://doi.org/10.1111/poms.13745 Mirzaei, M., Zaerpour, N., & de Koster, R. B. M. (2022). How to benefit from order data:
Correlated dispersed storage assignment in robotic warehouses. International Journal

17
Z. Liu et al. Computers & Industrial Engineering 200 (2025) 110767

of Production Research, 60(2), 549–568. https://doi.org/10.1080/ Yang, P., Miao, L. X., Xue, Z. J., & Qin, L. (2015). An integrated optimization of location
00207543.2021.1971787 assignment and storage/retrieval scheduling in multi-shuttle automated storage/
Momeni, M. A., & Jain, V. (2024). An integrated model for the production planning and retrieval systems. Journal of Intelligent Manufacturing, 26(6), 1145–1159. https://doi.
the allocation of locations to items in a warehouse. Annals of Operations Research, 35. org/10.1007/s10845-013-0846-7
https://doi.org/10.1007/s10479-024-06085-3 Yang, P., Miao, L. X., Xue, Z. J., & Ye, B. (2015). Variable neighborhood search heuristic
Nigam, S., Roy, D., de Koster, R., & Adan, I. J. (2014). Analysis of class-based storage for storage location assignment and storage/retrieval scheduling under shared
strategies for the mobile shelf-based order pick system. storage in multi-shuttle automated storage/retrieval systems. Transportation Research
Onal, S., Zhang, J. R., & Das, S. (2017). Modelling and performance evaluation of Part E-Logistics and Transportation Review, 79, 164–177. https://doi.org/10.1016/j.
explosive storage policies in internet fulfilment warehouses. International Journal of tre.2015.04.009
Production Research, 55(20), 5902–5915. https://doi.org/10.1080/ Yang, P., Peng, Y. F., Ye, B., & Miao, L. X. (2017). Integrated optimization of location
00207543.2017.1304663 assignment and sequencing in multi-shuttle automated storage and retrieval systems
Roy, D., Nigam, S., de Koster, R., Adan, I., & Resing, J. (2019). Robot-storage zone under modified 2<i>n</i>-command cycle pattern. Engineering Optimization, 49(9),
assignment strategies in mobile fulfillment systems. Transportation Research Part E- 1604–1620. https://doi.org/10.1080/0305215x.2016.1261128
Logistics and Transportation Review, 122, 119–142. https://doi.org/10.1016/j. Yang, P., Zhao, Z. J., & Guo, H. J. (2020). Order batch picking optimization under
tre.2018.11.005 different storage scenarios for e-commerce warehouses. Transportation Research Part
Scholz, A., Schubert, D., & Wäscher, G. (2017). Order picking with multiple pickers and E-Logistics and Transportation Review, 136, 18. https://doi.org/10.1016/j.
due dates - Simultaneous solution of order batching, batch assignment and tre.2020.101897
sequencing, and picker routing problems. European Journal of Operational Research, Yang, P., Zhao, Z. J., & Shen, Z. J. M. (2021). A flow picking system for order fulfillment
263(2), 461–478. https://doi.org/10.1016/j.ejor.2017.04.038 in e-commerce warehouses. IISE Transactions, 53(5), 541–551. https://doi.org/
Tabrizi, A. M., Vahdani, B., Etebari, F., & Amiri, M. (2023). A three-stage model for 10.1080/24725854.2020.1772525
clustering, storage, and joint online order batching and picker routing problems: Yuan, R., Graves, S. C., & Cezik, T. (2019). Velocity-based storage assignment in semi-
Heuristic algorithms. Computers & Industrial Engineering, 179, 27. https://doi.org/ automated storage systems. Production and Operations Management, 28(2), 354–373.
10.1016/j.cie.2023.109180 https://doi.org/10.1111/poms.12925
Wang, Y. Y., Man, R. J., Zhao, W. M., Zhang, H. L., & Zhao, H. (2022). Storage Zarinchang, A., Lee, K., Avazpour, I., Yang, J., Zhang, D., & Knopf, G. K. (2024). Adaptive
assignment optimization for fishbone Robotic Mobile Fulfillment systems. Complex & warehouse storage location assignment with considerations to order-picking
Intelligent Systems, 8(6), 4587–4602. https://doi.org/10.1007/s40747-021-00597-2 efficiency and worker safety. Journal of Industrial and Production Engineering, 41(1),
Weidinger, F., & Boysen, N. (2018). Scattered storage: How to distribute stock keeping 40–59. https://doi.org/10.1080/21681015.2023.2263009
units all around a mixed-shelves warehouse. Transportation Science, 52(6), Zhang, J., Wang, X. P., Chan, F. T. S., & Ruan, J. H. (2017). On-line order batching and
1412–1427. https://doi.org/10.1287/trsc.2017.0779 sequencing problem with multiple pickers: A hybrid rule-based algorithm. Applied
Weidinger, F., Boysen, N., & Briskorn, D. (2018). Storage assignment with rack-moving Mathematical Modelling, 45, 271–284. https://doi.org/10.1016/j.apm.2016.12.012
M bile robots in KIVA warehouses. Transportation Science, 52(6), 1479–1495. Zhang, J., Zhang, N., Tian, L. K., Zhou, Z. J., & Wang, P. (2023). Robots? picking
https://doi.org/10.1287/trsc.2018.0826 efficiency and pickers? energy expenditure: The item storage assignment policy in
Xiang, X., Liu, C. C., & Miao, L. X. (2018). Storage assignment and order batching Robotic Mobile Fulfillment system. Computers & Industrial Engineering, 176, 16.
problem in Kiva mobile fulfilment system. Engineering Optimization, 50(11), https://doi.org/10.1016/j.cie.2022.108918
1941–1962. https://doi.org/10.1080/0305215x.2017.1419346 Zhang, K., & Gao, C. H. (2023). Improved formulations of the joint order batching and
Yang, P., Jin, G., & Duan, G. F. (2022). Modelling and analysis for multi-deep compact picker routing problem. International Journal of Production Research, 61(21),
robotic mobile fulfilment system. International Journal of Production Research, 60(15), 7386–7409. https://doi.org/10.1080/00207543.2022.2149872
4727–4742. https://doi.org/10.1080/00207543.2021.1936264

18