Location and capacity optimization of urban sanitation robot base stations using improved NSGA-II Authors: Jie Ding, Linshen Wang, Kun Pan, Nan Ji, Na Cui Published: 2025 ABSTRACT AbstractTraditional sanitation systems are increasingly challenged by contemporary waste collection demands, owing to limited service capacity, high operational costs, and uneven resource distribution. Sanitation robots, capable of high-precision waste identification and multi-robot collaboration, offer a promising alternative to improve collection efficiency. Their performance, however, critically depends on the scientific layout of base stations. This study proposes a multi-objective location and capacity optimization model based on an improved NSGA-II, incorporating multi-source data such as population density and spatiotemporal waste distribution. To strengthen the algorithm’s adaptability to complex urban environments, a self-adaptive crossover operator and a dynamic objective weighting mechanism are introduced. The TOPSIS method is then applied to evaluate and rank Pareto-optimal solutions, yielding an optimal layout and capacity plan for base stations. An empirical analysis in Lixia District, Jinan City, demonstrates that the optimized scheme reduces the number of facilities by one, improves service utility, and decreases the number of uncovered demand points to eight. Compared with existing planning approaches, the proposed model proves effective in dynamic scenarios and provides a collaborative multi-objective optimization framework for smart urban sanitation planning.Keywords: Sanitation facilities, Sanitation robot base stations, Location-allocation, Multi-objective optimizationSubject terms: Energy and society, Sustainability, Computational science, Computer science Introduction Currently, the world is facing unprecedented challenges in waste management. The United Nations Environment Programme’s Global Waste Management Outlook 2024 indicates that urban solid waste will increase from 2.1 billion tons in 2023 to 3.8 billion tons by 2050. In 2020, the direct cost of global waste management was estimated at $252 billion. When hidden costs such as pollution and climate change are considered, the cost rises to $361 billion. Without urgent waste management actions, by 2050, the global annual cost will nearly double, reaching a staggering $640.3 billion1 This imbalance has led to a quadruple crisis. Economically, sanitation investments in various cities will increase. Environmentally, open dumping of waste is a significant source of greenhouse gas emissions and a major factor contributing to global warming2. Socially, the overall prevalence of occupational injuries among sanitation workers is 32.36%3. The high risk of injury in this industry can easily lead to labor shortages and moral controversies. In terms of urban management, the variety and quantity of sanitation facilities and equipment continue to grow, with automated equipment, mainly automated garbage trucks, being widely used. However, there are challenges such as insufficient parking space, lack of location selection and scale standards, which urgently require academia to seek technological solutions to overcome these obstacles. In recent years, breakthroughs in sanitation robot technology have provided new pathways for urban waste management. Jinqiang Bai’s team developed a grassland waste recognition robot that achieves a 95% recognition accuracy through deep neural networks4. Guozhen Li’s team integrated tactile sensing with machine learning, achieving a waste classification accuracy of 94%5. Abhishek Gupta’s team optimized multi-robot path planning using collaborative algorithms, significantly reducing the energy consumption and cost of waste collection6. These technologies offer efficient and sustainable solutions for smart city waste management. The efficient operation of sanitation robots is closely tied to base stations, which act as “energy supply stations” and “data interaction hubs” for these robots, ensuring their continuous operation. The layout of base stations significantly impacts operational costs and resource utilization efficiency. Improper base station layout can result in budget overruns or resource waste. Research by Mario Castro-Gama et al. explored the optimal placement of autonomous robot charging stations within drinking water networks, employing the proposed algorithm to effectively estimate the required number of charging stations and autonomous robots in the distribution network, thereby achieving efficient resource utilization7. The study reveals that base stations significantly affect the operational efficiency of robots. Proper site selection for base stations is crucial for improving the efficiency of sanitation robot systems and reducing operational costs. In light of this, the paper introduces intelligent enhancements to existing sanitation facilities, transforming them into sanitation robot base station. The study develops a two-stage multi-objective optimization model for resolution. During the site selection phase, a model is established to maximize garbage collection efficiency, minimize economic costs, and cover the most demand points. In the capacity determination phase, a model is constructed to minimize economic costs and reduce garbage collection time. An improved Non-dominated sorting genetic algorithm II (NSGA-II) is employed to obtain the optimal solutions for the multi-objective resolution, with an empirical analysis conducted in Lixia District of Jinan City as the case study. Based on the existing facility layout plan, a spatial optimization layout for base stations was proposed, determining the optimal number of sanitation robots to be deployed at each base station. This ultimately results in a sanitation service system with greater intelligence and improved operational efficiency. This study builds on the literature related to facility location and multi-objective optimization algorithms (MOOA). Section 2 will systematically review the evolution of facility location theory, the development of multi-objective optimization methods, and their application in sanitation facilities planning. This review will provide the theoretical basis and methodological support for the model construction and algorithm improvement presented in this paper. Section 3 outlines and defines the research problem. Section 4 details the model parameters and develops the model necessary for the study. Section 5 describes the solution process of the NSGA-II. In Sect. 6, a case analysis is conducted using Lixia District, Jinan City as an example. The site selection scheme proposed in this paper is compared with the existing planning scheme to verify the applicability of the proposed model and algorithm. Finally, Sect. 7 summarizes the conclusions of this paper and offers suggestions for future research. Related work Over more than half a century, facility location theory has evolved to provide a methodological foundation for selecting location for sanitation robot base stations. Early research focused on single-objective static optimization, such as the Weber problem, which aims to minimize transportation costs8, the P-Median model, which strives to shorten service distances9, and coverage models that seek a balance between the number of facilities and coverage area10. Although these classic models laid the theoretical groundwork, they struggle to meet the complex demands of modern cities. With the rise of sustainable development concepts, multi-objective models have gradually become mainstream. The goal of multiple objective problems (MOPs) is to find a set of solutions approximating the Pareto Front (PF) with good convergence and diversity11. For example, the study by Ling Shen et al. explored the optimal locations for centralized and decentralized construction waste treatment facilities to achieve low-cost, low-transportation carbon emissions, and low social negative impact utilization12. Eun-Kyeong Kim et al. developed an advanced multi-objective location optimization model, optimizing three objectives: maximizing the utilization rate of candidate sites for establishing urban parks, maximizing the number of vulnerable groups covered by urban parks, and minimizing the total construction cost of urban parks13. Similar to the above studies, location-capacity problems usually also involve multiple conflicting objectives, meaning that improvement in one objective may lead to the deterioration of another objective14. For instance, increasing economic costs can allow for the deployment of more sanitation robots, thereby reducing waste collection time. However, since it is unlikely to find a single solution that optimizes both cost and time in reality, the Pareto optimal set is, in a sense, the best trade-off solution15. It contains various solutions that balance all objectives, where an improvement in one objective cannot be achieved without deteriorating at least one other objective16. Multiple objective evolutionary algorithms (MOEAs) are currently popular for multi objective optimization17,18, utilizing population evolution to search the PF. Mainstream algorithms can be categorized into three types19. (1) Pareto-dominance-based algorithms. These algorithms directly approach the PF by employing fast non-dominated sorting to filter the solution set. NSGA-II20,21 is a classic example of this category. It balances solution diversity and convergence through fast non-dominated sorting, crowding comparison, and elitist strategies, and is widely applied in multi-objective optimization problems. Strength pareto evolutionary algorithm II (SPEA2)22 introduces an environmental selection mechanism and refined archive management, significantly enhancing the uniformity of the solution set distribution, making it particularly suitable for optimization problems in high-dimensional objective spaces. (2) Performance indicator-based algorithms. These algorithms are guided by indicators such as hypervolume to direct the search. Indicator-based evolutionary algorithm (IBEA)23 assesses the quality of the solution set using custom indicators, allowing flexible adaptation to various optimization needs, though its computational complexity rises significantly with the number of objectives. Fast hypervolume indicator-based MOEA (FV-EMOA)24 integrates hypervolume indicators into the selection mechanism, using gradient information to guide the search, improving convergence speed while maintaining diversity. (3) Decomposition-based algorithms. These algorithms decompose multi-objective problems into multiple single-objective subproblems, approaching the PF through collaborative optimization. MOEA based on decomposition (MOEA/D)25 uses weight vectors to partition the objective space, combined with neighborhood update strategies, excelling under complex constraints. MOEA/D with dynamical resource allocation (MOEAD-DRA)26 introduces a dynamic resource allocation mechanism that adaptively adjusts the allocation of computational resources for subproblems, significantly enhancing the algorithm’s efficiency and robustness. This research provides methodological support for multi-objective facility location; however, its applicability in the context of sanitation robot base station layout still needs to be tested. Currently, there is limited research on location and capacity determination of sanitation robot base station, and numerous challenges are encountered. Firstly, the amount and distribution of urban waste are influenced by factors such as seasons, holidays, and city activities, resulting in significant spatiotemporal variations (Firman Firmansyah; Bruno Ribas Alzamora)27,28. Traditional location models struggle to respond in real-time to these dynamic changes, leading to base station layouts that fail to meet actual demands, causing resource waste or insufficient service. Secondly, the selection of base station locations must balance multiple objectives, including construction cost, operational cost, service coverage, and the convenience of charging and maintenance (Dongqing Jia; Guokai Li; Vinay Yadav; Constantinos Chatzouridis)29–32. These objectives are interrelated and constrain each other. Pursuing extensive coverage may increase costs, while overly compressing costs may reduce service quality. Existing MOOA have weak capabilities when handling complex constraints and goal conflicts (Ines Mehouachi; Farhad Habibi)33,34, making it difficult to find a global optimal solution. Finally, the location of base station faces coordination issues with existing urban infrastructure (Zhonglin Sheng)35. Urban space is limited and complexly laid out, requiring base station planning to consider compatibility with surrounding facilities to avoid disrupting urban operations. Notably, emerging research in the fields of drones and unmanned vehicles shows unique value: Balaji Kirubakaran and others have proposed an innovative perspective of integrating drone base station into urban infrastructure, providing new insights into the layout of drone base station and offering technical inspiration for this study36. Daria Matkovic and others integrated a location-based charging station saliency prediction model with the lion optimization algorithm to optimize the placement of additional charging stations within the established charging infrastructure, thereby improving the utilization efficiency of the charging station37. This method can be applied to the location of sanitation robot base stations to balance the conflict between service coverage and construction investment. To address the MOPs of location and capacity determination for sanitation robot base stations, this study proposes a three-tier “data-model-validation” system. At the data layer, a multi-source heterogeneous database is constructed by integrating data such as population density, satellite remote sensing images, and the number of tourists in scenic spots during holidays. The improvements at the model layer are demonstrated in two aspects: firstly, an adaptive crossover operator is designed to maintain population diversity in the later stages of NSGA-II evolution; secondly, a dynamic adjustment mechanism for objective weights is introduced. This mechanism computes different time weight vectors for weekdays, weekends, and holidays, dynamically adjusting the weights of service utility objectives to better align the optimization process with actual demand changes. At the validation layer, the improved NSGA-II is applied to the location and capacity determination study of sanitation robot base stations in Lixia District, Jinan City, assessing the distribution of existing facilities and determining the optimal location and capacity scheme. The breakthrough of this study lies not only in algorithmic improvements but also in the transformation of the research paradigm. Unlike traditional studies on location and capacity determination that rely on static empirical models, this research incorporates the spatiotemporal fluctuation characteristics of waste volume into the optimization model by constructing a “time-demand” dynamic coupling mechanism, achieving a shift from a “predict-respond” mode to a “perceive-adapt” mode. These advancements provide new theoretical tools for smart city infrastructure planning, and the methodological framework can be extended to areas such as emergency station deployment and shared charging station layout, demonstrating significant potential for multi-domain application. Problem description As urban waste continues to increase, traditional labor-intensive sanitation models are increasingly exposed to issues such as insufficient service resilience, rising labor costs, and delayed dynamic responses. These challenges are particularly pronounced during holidays when population surges lead to significant fluctuations in waste generation. Such conditions pose severe challenges to the traditional sanitation model, which relies on fixed scheduling and static resource allocation. In this context, sanitation robots, as the core components of intelligent sanitation systems, are redefining the paradigm of urban sanitation management with their advantages of round-the-clock service and intelligent scheduling. The deployment efficiency of sanitation robot systems is highly dependent on the strategic layout of the base stations. Base stations not only provide energy supply and data transmission support for the robots but also directly influence the feasibility of their service radius, operational scheduling, and route planning. Therefore, determining how to scientifically select locations and configure capacities for sanitation robot base stations in complex urban environments has become a core issue urgently needing resolution in the current field of smart sanitation. The location and capacity determination for sanitation robot base stations involve optimizing multiple objectives. Firstly, the layout of existing sanitation facilities needs to be optimized with the goals of maximizing service utility, minimizing economic costs, and covering the most demand points, based on the dynamic changes in waste generation over time. Secondly, when the existing facilities cannot cover additional demand points, new base stations should be added appropriately, with economic cost as a constraint. Finally, the number of sanitation robots should be determined by minimizing waste collection time and economic costs. This issue is a typical multi-objective location and capacity problem with spatiotemporal dynamic characteristics, which traditional static models struggle to address due to the dynamic service demands caused by temporal heterogeneity. Therefore, the study established a dynamic response, multi objective oriented optimization model for the location and capacity determination of sanitation robot base stations to address this problem. The location and capacity determination problem discussed in this paper primarily involves two types of sites: base stations that provide charging and maintenance for sanitation robots, and demand points that receive waste collection services. The base stations include existing, candidate, and newly added stations. The demand points are served by sanitation robots. Given the spatial heterogeneity of garbage distribution, such as the need for high-load continuous operation in large open areas and the requirement for flexible mobility and precise interactive services in backstreets and commercial streets—the study adopts a heterogeneous robot collaborative operation system. Medium-sized cleaning robots, equipped with large-capacity cargo boxes, handle garbage collection and transportation on main roads, while small-sized cleaning robots, with compact bodies and intelligent sensing modules, perform detailed operations in narrow spaces. The medium-sized cleaning robot utilizes the research findings of Qiaokang Liang et al.38. The electric unmanned sanitation vehicle (EUSV) integrates multiple sensor technologies and can operate continuously at high intensity. The small-sized cleaning sanitation robot employs the SOLIS S200 multifunctional cleaning robot launched by Candela39, which can autonomously plan cleaning paths to perform precise outdoor operations and comprehensive cleaning. The appearance and parameters of the robots are shown in Fig. 1; Table 1. Appearance of sanitation robots. Table 1Parameters of sanitation robots.ParametersEUSVSOLIS S200Length (mm)39802390Width (mm)13001360Height (mm)22801725Equipment mass (kg)1830750Water tank volume (L)1000100Garbage container volume (L)420150Cleaning operation width (m)2.21.5–1.9Cleaning operation speed (km/h)77Continuous operation time (h)88 Parameters of sanitation robots. Establishment of location and capacity determination model Model assumptions and parameter definitions The assumptions of the model are as follows: (1) Demand and service are assumed to occur at discrete points. (2) A base station can serve multiple demand points. (3) Each demand point can be served by only one base station. (4) The distance between demand points and base stations is measured using Euclidean distance. The parameters and definitions of the model are presented in Table 2. Table 2Model parameters and definitions.ParametersDefinitions Requirement set Base station location set Existing sanitation facilities set New sanitation robot base station set Set of sanitation robot types. Corresponding to medium-sized cleaning robots and small-sized cleaning robots Time period set. Corresponding to Weekdays, Weekends, and Holidays Weights for different time periods: Corresponding to Weekdays, Weekends, and Holidays R The base station’s service radius is set to 2 km Capacity of base station j The Euclidean distance between demand point i and candidate base station j The amount of waste generated at demand point i during time period t The total waste collected from all demand points covered by base station j during time period t The total waste generated at all demand points during the time period t N t jk The number of type k sanitation robots deployed at base station j during time period t The processing capacity of type k sanitation robots per time period. Medium-sized cleaning robots have a capacity of 420 L, while small-sized cleaning robots have a capacity of 150 L. The construction cost of a single base station. The value is 250 million yuan The cost of type k sanitation robots. Medium-sized cleaning robots cost 170,000 yuan, while small-sized cleaning robots cost 115,000 yuan. The average distance from base station j to each demand point within the service area Maximum operating time for sanitation robot. The value is 8 h. The operating speed of type k sanitation robots. The medium-sized cleaning robots and the small-sized cleaning robots both operate at a speed of 7 km/h. The maximum cleaning capacity of type k sanitation robots at base station j. Unit: L. Waste processing efficiency of type k sanitation robots at base station j. Unit: L/h. Base station longitude Base station latitude A binary variable. It takes a value of 1 if a base station is constructed at location j; otherwise, it is 0. A binary variable. It indicates whether demand point i is covered by base station j during time period t. It takes a value of 1 if coverage is provided; otherwise, it is 0. Model parameters and definitions. Model construction The service radius of the base station is based on the Technical code for transfer station of municipal solid waste CJJ/T47-2016. Given the higher intelligence level of sanitation robot base stations, their service radius is appropriately increased from that of existing sanitation facilities of the same level, reaching 2 km. Population density is used to calculate the amount of waste generated at different times, which in turn determines the target weight for each time period. The calculation formula is: The weights for each time period are listed in Table 3: Table 3Weights of each time period.Time periodWeightsWeekdays0.3005Weekends0.3094Holidays0.3901 Weights of each time period. A model for locating sanitation robot base stations is developed with the goals of maximizing garbage collection efficiency, minimizing economic costs, and covering the maximum number of demand points, while using the service radius as a constraint. Objective function (2) seeks to maximize the utility of waste disposal at the base station, objective function (3) aims to minimize economic costs, and objective function (4) strives to maximize the number of covered demand points. Equation (5) indicates that the coverage relationship is valid only when the demand point is within the service radius of the base station. Equation (6) specifies that each demand point can be covered by only one base station during each time period. Equation (7) states that a base station can provide services only when it is selected. Equation (8) mandates the activation of existing sanitation facilities. Equations (9) and (10) impose spatial constraints on the location of new base stations, while Eq. (11) imposes constraints on the number of new base stations. After determining the optimal locations for the sanitation robot base stations, the capacity of the medium-sized cleaning robots and the small-sized cleaning robots at each base station must be calculated. Given the differences in waste spatial distribution and the strong waste collection capabilities of the medium-sized cleaning robots, these robots are tasked with collecting 70% of the waste within their service area, while the remaining 30% is handled by the small-sized cleaning robots. The two types of sanitation robots at each base station have an ideal operating time of 8 h, which is their maximum operational duration. Consequently, the maximum clearing capacity for each type of sanitation robot within this time frame is: The waste disposal efficiency for the two types of sanitation robots at each base station is: The solution is carried out with the goal of minimizing the economic cost and the time for garbage clearance. Equation (14) aims to minimize the economic cost of the two types of sanitation robots at the base station. Equation (15) focuses on minimizing the garbage collection time at each base station. Equation (16) ensures that the number of the two types of sanitation robots at the base station is a positive integer, rounded up. Algorithm solutions Improved NSGA-II The location-capacity model proposed in this research is a multi-objective problem characterized by spatiotemporal dynamics. The NSGA-II, with its strengths in fast non-dominated sorting, elitist preservation strategy, and crowding comparison operator, effectively balances convergence and diversity, making it well-suited for multi-objective optimization problems. To further enhance the performance of the NSGA-II and ensure its capability to handle the spatiotemporal dynamic changes in urban waste distribution, the research has implemented the following improvements: (1) An adaptive crossover operator is designed to maintain population diversity in the later stages of evolution, preventing premature convergence; (2) A dynamic adjustment mechanism for objective weights is introduced by setting weight factors to balance sanitation services at different times. The algorithm flow is shown in Fig. 2. Algorithm flow of improved NSGA-II: (1) Input relevant parameters. (2) Initialize the population. (3) Perform fast non-dominated sorting and crowding distance calculation for the individuals in the population. (4) Select the parent population through tournament selection, and generate the offspring population using adaptive crossover and mutation. (5) Merge the parent and offspring populations into a new population of size 2N. Subsequently, select the top N individuals based on their non-dominated ranks and crowding distances to form the elite population. (6) Check if the generation count has reached the predefined value. If not, repeat the above steps. If it has, the evolution terminates, and the individuals in the highest non-dominated rank are output. The solution process is as follows: The population of sanitation robot base station location schemes is initialized. The population size is set to N, and the initial population Pt is generated randomly. Existing sanitation facilities are mandatorily activated. Each candidate base station has a certain probability of being selected as a sanitation robot base station. The population of sanitation robot base station location schemes is initialized. The population size is set to N, and the initial population Pt is generated randomly. Existing sanitation facilities are mandatorily activated. Each candidate base station has a certain probability of being selected as a sanitation robot base station. The dynamic objective function is calculated. A fast non-dominated sorting is performed on the population Pt, and the crowding degree of the location schemes is calculated. The objective function is converted into a fitness function to assess the quality of the current scheme. The base station location problem is first solved, and the weight factor is dynamically set according to the time period corresponding to the current generation count. The fitness function is shown in Eqs. (17), (18), and (19). The dynamic objective function is calculated. A fast non-dominated sorting is performed on the population Pt, and the crowding degree of the location schemes is calculated. The objective function is converted into a fitness function to assess the quality of the current scheme. The base station location problem is first solved, and the weight factor is dynamically set according to the time period corresponding to the current generation count. The fitness function is shown in Eqs. (17), (18), and (19). Where F1 represents the garbage collection efficiency, which quantifies the amount of garbage collected per unit time. denotes the weight assigned to each time period. is a binary variable indicating whether demand point i is covered by base station j during time period t, taking a value of 1 if coverage is provided and 0 otherwise. Dit signifies the garbage output generated by demand point i during time period t. The objective of this formula is to maximize garbage collection efficiency, thereby guiding the algorithm to select base station layouts that ensure efficient garbage collection. Where F2 represents the economic cost associated with base station construction. xj is a binary variable, set to 1 if a base station is constructed at location j and 0 otherwise. Cbase denotes the construction cost of a single base station. The objective of this formula is to minimize economic cost, thereby ensuring the economic feasibility of the proposed solution. Where F3 represents the number of demand points covered. The objective of this formulation is to maximize the number of covered demand points, thus ensuring service to as many waste generation points as possible. The number of sanitation robots at each base station is then determined. The objective function is converted into a fitness function, as shown in Eqs. (19) and (20): Where F4 represents the economic cost of robot configuration. Crobot, k represents the cost of robot type k. represents the number of type k robots deployed by base station j during time period t. The objective of this formulation is utilized to assess the economic efficiency of various capacity configuration schemes, with the ultimate objective of minimizing costs. Where F5 represents the total time required for completing the garbage collection task. Djt denotes the total amount of garbage at all demand points covered by base station j during time period t. Vjk denotes the garbage disposal efficiency of the k-type robot in base station j. The objective of this formulation is designed to optimize garbage collection efficiency and minimize the total collection time. The algorithm employs a tournament for solution selection. It performs non-dominated sorting based on the fitness values of all solutions in the population and calculates the crowding degree for each solution. Two solutions are randomly compared, and the one with the larger crowding degree is selected to form a parent population of size N. (3)Adaptive Crossover and Mutation. To dynamically balance global search and local optimization capabilities, an adaptive crossover operator based on evolutionary stage feedback was designed. Its core is maintaining population diversity through a dual adjustment mechanism involving crossover probability and the number of crossover points. The crossover probability Pc is defined as a function that decays non-linearly with evolutionary generations: Adaptive Crossover and Mutation. To dynamically balance global search and local optimization capabilities, an adaptive crossover operator based on evolutionary stage feedback was designed. Its core is maintaining population diversity through a dual adjustment mechanism involving crossover probability and the number of crossover points. The crossover probability Pc is defined as a function that decays non-linearly with evolutionary generations: Where t represents the current generation, T is the total number of generations, and α is the decay rate parameter. and denote the preset maximum and minimum values of the crossover probability, respectively. This function leverages the characteristics of exponential decay to maintain a high crossover probability in the early stages of evolution, thereby fully exploring the solution space and promoting gene recombination. As evolution progresses, the crossover probability gradually decreases to , reducing the risk of disrupting high-quality individuals. While still retaining a minimum level of crossover probability to prevent the complete loss of population diversity. In controlling the number of crossover points, the algorithm further incorporates a dynamic adjustment strategy inversely related to evolutionary algebra: Where kmax and kmin represent the predetermined maximum and minimum number of crossover points, respectively. In the early stages of evolution, a larger number of crossover points is used to enhance the intensity of random gene fragment exchange, thereby improving global exploration capability. In the later stages of evolution, the number of crossover points is reduced to kmin to decrease the intensity of gene recombination disturbance, preserving the integrity of advantageous gene blocks and thus supporting localized refinement searches. The specific crossover process is as follows: (a) Parent selection. Randomly select two parent chromosomes from the population, calculate Pc(t) based on the current generation t, and decide with that probability whether to perform a crossover. (b) Generation of crossover points. If a crossover is needed, determine the number of crossover points based on k(t), and randomly select k(t) unique positions on the chromosome encoding sequence for segmentation. (c) Exchange of gene segments. Split the parent chromosomes into k(t) + 1 continuous segments according to the crossover points, and alternately exchange adjacent segments to generate two offspring individuals. (d) Diversity preservation: When Pc(t) decreases to , enforce at least one crossover operation to ensure the continuous introduction of new gene combinations into the population. The crossover process is shown in Fig. 3. Crossover process. The algorithm dynamically adjusts its crossover strategy and crossover probability over evolutionary generations. In the initial stage (Pc = 0.8), three-point crossover is employed to enhance global exploration, promote population diversity, and accelerate escaping local optima. In the middle stage (Pc = 0.6), two-point crossover is employed to balance exploration and exploitation, introducing new genes while beginning to maintain superior patterns. In the final stage (Pc = 0.4), single-point crossover is employed to prioritize the protection of already discovered superior genetic structures, supplemented by a lower crossover probability, which facilitates a smooth transition of the search from global exploration to local fine-grained exploitation. In the mutation operation, offspring individuals undergo mutation with a fixed probability Pm. A base station gene position (built/not built) in the chromosome is randomly selected, and its state is flipped (0→1 or 1→0). If the base station density in a certain area is excessively high or low, the gene corresponding to that area is given priority for mutation to maximize coverage potential. The mutation process is illustrated in Fig. 4. Mutation process. A fixed mutation probability is employed in this algorithm. Under normal circumstances, gene loci within chromosomes are randomly selected for mutation. However, priority is given to the mutation of chromosome gene loci in regions with excessively high or low base station density. (4)Population merging and elite preservation. Combine the parent and offspring populations into a temporary population Rt, perform non-dominated sorting on Rt, categorize into non-dominated levels, and determine the ranking of individuals. Within each non-dominated level, calculate the distribution density of individuals in the objective space. Select the top N individuals from Rt based on their non-dominated rank and crowding distance to form a new population Pt+1, ensuring the preservation of elite solutions. The specific process is shown in Fig. 5. Population merging and elite preservation. Combine the parent and offspring populations into a temporary population Rt, perform non-dominated sorting on Rt, categorize into non-dominated levels, and determine the ranking of individuals. Within each non-dominated level, calculate the distribution density of individuals in the objective space. Select the top N individuals from Rt based on their non-dominated rank and crowding distance to form a new population Pt+1, ensuring the preservation of elite solutions. The specific process is shown in Fig. 5. (5)Repeat the evolutionary process. If t has not reached the specified maximum number of generations, continue repeating the above steps. Repeat the evolutionary process. If t has not reached the specified maximum number of generations, continue repeating the above steps. (6)Once the specified number of cycles is reached, the evolution ends, and all individuals in Pt at the highest non-dominated level are output, forming the optimal solution set for the sanitation robot base station location and capacity determined by the algorithm. Once the specified number of cycles is reached, the evolution ends, and all individuals in Pt at the highest non-dominated level are output, forming the optimal solution set for the sanitation robot base station location and capacity determined by the algorithm. Execution steps of the elitism preservation strategy. The process comprises two key stages: (1) merging the parent and offspring populations, and performing fast non-dominated sorting and crowding distance calculation on the merged population; (2) iteratively selecting individuals with higher non-dominated ranks, and then employing crowding distance to maintain population diversity while removing individuals with lower crowding distances. This strategy effectively preserves high-quality solutions while promoting diversity and convergence towards the Pareto optimal front. Parameter study In the design and application of MOEAs, parameter settings have a significant impact on the algorithm’s convergence speed, solution set quality, and computational efficiency. To further optimize the performance of improved NSGA-II, this study conducted a systematic analysis of two key parameters: population size and the number of generations. Through a comparative analysis of the algorithm’s performance under various parameter settings, the most suitable configuration for this problem was identified, balancing both convergence and diversity. Population size directly affects the algorithm’s global search capability and computational complexity. A larger population helps maintain diversity and avoid premature convergence but significantly increases the computational burden. Conversely, a smaller population, while improving efficiency, may reduce the comprehensiveness and robustness of the solution set. To scientifically determine the optimal population size, 50 independent experiments were conducted for each of four configurations (population sizes of 25, 50, 75, and 100) under a fixed number of generations. The statistical results based on the Hypervolume (HV) indicator are presented in Table 4. A higher HV value signifies superior comprehensive performance of the solution set, encompassing both convergence and diversity. Table 4HV performance under different population sizes.Population sizemeanHV±stdHVLiftFromPrev (%)259.63896± 0.00205-509.65375± 0.000690.1535759.65840± 0.000500.04821009.66053± 0.000290.0220 HV performance under different population sizes. The results indicate that as the population size increases, the mean HV steadily rises, signifying a continuous enhancement in solution set quality. Concurrently, the standard deviation of HV (stdHV) progressively diminishes, suggesting an improvement in algorithm stability. However, an examination of the lift from previous (LiftFromPrev) reveals that when the population size was increased from 50 to 75, the HV improvement was only 0.0482%. Furthermore, upon increasing the population from 75 to 100, this improvement further declined to 0.0220%, clearly demonstrating significantly diminishing returns. This suggests that beyond a population size of 50, the performance gains from further expansion become very limited, while computational costs escalate significantly. Therefore, considering both solution set quality and computational efficiency, setting the population size to 50 emerges as a highly suitable choice. This configuration not only ensures high HV values and stability but also prevents inefficient computations associated with an excessively large population, thereby achieving an effective balance between convergence and diversity. The number of generations is another key parameter affecting algorithm convergence. A sufficient number of generations is a necessary condition for ensuring the algorithm thoroughly explores the solution space and approximates the true PF. To determine an appropriate number of generations, this study tested the performance of the improved NSGA-II on the ZDT1 and ZDT2 functions under 100, 200, and 300 generations, with a fixed population size of 50. Figures 6 and 7, and Fig. 8 illustrate a comparison between the PFs obtained by the proposed algorithm on the ZDT1 and ZDT2 functions and the theoretical PF for each of the three generation counts. When 100 generations were employed, a significant discrepancy was observed in the distribution of solution sets on both functions, indicating that the algorithm had not fully converged. As the number of generations increased to 200 and 300, the solution set closely approximated the theoretical PF, exhibiting excellent convergence performance. Comparison of PFs obtained by the improved NSGA-II and the true PF with 100 generations. (a) ZDT1; (b) ZDT2. With 100 generations, the solution set significantly deviates from the true PF, indicating insufficient convergence. Comparison of PFs obtained by the improved NSGA-II and the true PF with 200 generations. (a) ZDT1; (b) ZDT2. After 200 generations, the solution set tightly aligns with the true PF, demonstrating a marked enhancement in both convergence and diversity. Comparison of PFs obtained by the improved NSGA-II and the true PF with 300 generations. (a) ZDT1; (b) ZDT2. It was observed that at 300 generations, the results nearly overlapped with those at 200 generations, suggesting that the algorithm stabilizes after approximately 200 generations. To further assess the influence of different number of generations on the solution set’s distribution characteristics, population diversity was statistically analyzed across different number of generations. A larger indicator value suggests a broader distribution range and greater diversity within the population. The results are presented in Table 5. Table 5Population diversity of the proposed algorithm across different generations.GenerationPopulation Diversity on ZDT1Population Diversity on ZDT21000.53760.44942000.50740.51563000.50000.4853 Population diversity of the proposed algorithm across different generations. The study revealed that the impact of the number of generations on population diversity varied depending on the characteristics of the function. On ZDT1, diversity was observed to gradually decrease with an increasing number of generations, suggesting a gradual transition of the algorithm from global exploration to refined exploitation. Conversely, on ZDT2, diversity showed an initial increase followed by a decrease, achieving its optimal distribution state at 200 generations. A comprehensive analysis of both convergence and diversity performance led to the selection of 200 as the optimal number of generations for the algorithm. With this parameter, the solution set was able to sufficiently approach the theoretical PF on both ZDT1 and ZDT2, thereby satisfying the convergence requirements. Simultaneously, good diversity was maintained on ZDT1, and optimal diversity was achieved on ZDT2. While a slight improvement in convergence on ZDT1 was observed with 300 generations, the effect was not significant, and the computational cost increased. Furthermore, the diversity on ZDT2 even decreased. Therefore, 200 generations were determined to achieve the optimal balance among computational efficiency, convergence, and diversity. Consequently, the algorithm parameters were set as follows: a population size of 50 and 200 generations. Performance testing of algorithms To verify the improved NSGA-II’s superiority over the traditional NSGA-II, the study compares the solution quality and optimal gap of both algorithms, thereby validating the effectiveness of the proposed approach. The ZDT1 and ZDT2 functions were employed as benchmark bi-objective optimization problems. Figure 9 presents a comparison of solutions from the traditional NSGA-II after 200 generations with the theoretical PF. As shown, the traditional NSGA-II, when subjected to identical conditions, is observed to be less effective in approximating the theoretical PF compared to the proposed algorithm (Fig. 7), which suggests a marginally lower performance. The PFs obtained by NSGA-II on ZDT1 and ZDT2 are compared with the true PF. To comprehensively evaluate the proposed algorithm’s performance, a further comparison was conducted with the SPEA2 using the ZDT1 and ZDT2 functions. As a classic MOEA, SPEA2’s solution set convergence and diversity are widely employed in algorithm comparisons. Figure 10 presents a comparison of the PF obtained by SPEA2 after 200 generations with the true PF. The results indicate that SPEA2 approximates the true PF on both ZDT1 and ZDT2. However, it exhibits certain deviations in solution distribution uniformity and convergence when compared to the true PF. In contrast, the improved NSGA-II proposed in this paper demonstrates superior convergence and diversity in both sets of functions (Fig. 7), with its solution set almost completely covering the true Pf. Further verified the effectiveness and superiority of the proposed algorithm in solving complex MOPs. The PFs obtained by SPEA2 on ZDT1 and ZDT2 are compared with the true PF. To further evaluate the quality of the solutions obtained by the improved NSGA-II, this study employs generational distance (GD) and spacing (SP) to assess the convergence of the solutions generated by the two algorithms. GD represents the average minimum distance from the set of optimal solutions obtained by the algorithm to the theoretical PF. A smaller GD value indicates a smaller gap between the optimal solutions and the theoretical PF, signifying better convergence and smaller optimality gaps. SP refers to the average distance between solutions on the PF, reflecting the diversity and uniformity of the solution distribution. A smaller SP value generally indicates a more uniform distribution of solutions along the PF. The study conducts a comparative analysis of the convergence between the improved NSGA-II and the traditional NSGA-II under consistent parameter configurations. The specific parameters are as follows: a population size of 50, 200 generations, a crossover probability of 0.8, a mutation probability of 0.2, a maximum crossover probability of 0.8, a minimum crossover probability of 0.4, a decay coefficient of 0.2, a maximum of 3 crossover points, and a minimum of 1 crossover point. The comparison results are shown in Figs. 11, 12 and 13, and Fig. 14, which respectively present the convergence performance of the two algorithms on the ZDT1, ZDT2, ZDT3, and ZDT4 functions. As clearly illustrated in the figures, the improved NSGA-II demonstrates a significant advantage in convergence over the traditional NSGA-II. It converges more rapidly and reaches the optimal solution in a shorter amount of time. Regarding solution distribution, the improved NSGA-II converges to a smaller value and achieves a more uniform distribution of solutions. These results fully validate the effectiveness of the improvements in enhancing algorithm performance and strongly confirm that the improved NSGA-II delivers superior performance in addressing the spatial dynamic variation problem in urban waste distribution. Convergence of generational distance and spacing of ZDT1 function. This function is employed to evaluate the algorithm’s performance on a convex PF. Compared to the traditional algorithm, the improved NSGA-II achieved superior convergence proximity and distribution uniformity, demonstrating its enhanced performance in handling continuous convex problems. Convergence of generational distance and spacing of ZDT2 function. This function is employed to evaluate the algorithm’s performance on a concave PF. The improved NSGA-II exhibited accelerated convergence and an improved distribution of solutions, underscoring its effectiveness in handling concave objective spaces. Convergence of generational distance and spacing of ZDT3 function. This function is employed to evaluate the algorithm’s performance on a discrete PF. The improved NSGA-II maintained consistent convergence and distribution on discontinuous segments, outperforming NSGA-II in terms of diversity preservation. Convergence of generational distance and spacing of ZDT4 function. This function is employed to evaluate the algorithm’s performance on multimodal PFs. The improved NSGA-II demonstrated reliable global search capability, effectively escaping local optima and converging to positions closer to the true PF, all while achieving a superior distribution compared to NSGA-II. Selection of the best solution An improved NSGA-II is employed to solve the multi-objective location model, resulting in a set of Pareto optimal solutions. These solutions can be evaluated and ranked using the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) method. The steps are as follows: (1) Construct the decision matrix. Construct the decision matrix using the objective function values of each solution in the Pareto optimal solution set obtained through the improved NSGA-II. Where Xij represents the value of the j-th objective function for the i-th site selection plan of the sanitation robot base station, where i = 1,…,m, and j = 1,…,n. The objective of this formulation is to construct a matrix containing the objective function values of all Pareto-optimal solutions, thereby providing a data foundation for subsequent decision-making analysis. (2) Positive transformation of indicators. Transform the goals of minimizing economic cost and minimizing waste collection time for each plan into goals of maximization. (3) Standardize the decision matrix. To eliminate the influence of different dimensions of objective functions, the decision matrix is standardized using the vector normalization method, as shown in the following formula: Where rij represents the standardized indicator value. The objective of this formulation is designed to eliminate the dimensional discrepancies across various objective functions, thereby ensuring that all indicator values are on a comparable scale, which simplifies subsequent comparison and weighted computation. The standardized decision matrix R is expressed as: (4) Determine the weight vector. The entropy weight method is used to establish the weight vector for each objective function. This method, grounded in information theory, is an objective weighting approach that effectively reflects the significance of each indicator in a comprehensive evaluation. It involves calculating the proportion of each indicator in the normalized decision matrix, determining the information entropy of each indicator based on these proportions, and then calculating the redundancy of the information entropy. Finally, the weight of each indicator is derived from the redundancy of the information entropy. (5) Construct the weighted normalized decision matrix. Multiply the normalized decision matrix R by the weight vector to obtain the weighted normalized decision matrix V. Where v represents the weighted standardized metric value. The objective of this formulation is to incorporate the relative importance of each objective function by weighting the standardized metric values, thereby reflecting the varying contributions of each metric to the overall evaluation. (6) Determine the positive and negative ideal solutions. Identify vector V+ and V-, which consists of the optimal values and the worst values for each evaluation criterion. Where V+ represents the value of the positive ideal solution on the j-th metric. vij is the value of the j-th criterion for the i-th solution in the weighted normalized decision matrix. The objective of this formulation is to determine the optimal value of each metric in the Pareto solution set and construct an ideal optimal solution as an evaluation benchmark. Where V− represents the value of the negative ideal solution for the j-th index. The objective of this formulation is to identify the least desirable values for each index within the Pareto solution set, thereby establishing a worst-case scenario as an evaluation benchmark. (7) calculate the distances and from each option to the positive and negative ideal solutions Where represents the Euclidean distance between the i-th alternative and the positive ideal solution. The objective of this formulation is to quantify the degree of difference between each alternative and the ideal optimal solution, where a smaller distance indicates a more superior alternative. Where represents the Euclidean distance between the i-th alternative and the negative ideal solution. The objective of this formulation is to quantify the degree of discrepancy between each alternative and the worst-case solution, where a larger distance signifies a more superior alternative. (8) Calculate the relative closeness. Determine the relative closeness Ci of each solution to the ideal solution. Where Ci represents the relative closeness of the i-th scheme, ranging from 0 to 1. The objective of this formulation is to comprehensively considers the distances between a scheme and both positive and negative ideal solutions to calculate a comprehensive evaluation index. A larger value indicates the scheme is closer to the ideal solution. (9) Rank and select the optimal solution. Order all solutions by their relative closeness Ci and choose the one with the highest relative closeness as the optimal solution. Case study Overview of the study area Lixia District in Jinan City is located in the core area of Jinan, covering a total area of 100.89 square kilometers, with a permanent population of 819,100. It is a typical high-density urbanized area. The district is home to scenic spots such as Baotu Spring, Daming Lake, and Qianfo Mountain. During holidays, the tourist flow is dense, resulting in a distinct heterogeneity in waste generation. This characteristic closely aligns with the previously mentioned research framework, providing a typical scenario for verifying the collaborative model of medium-sized cleaning robots and small interactive robots under the dynamic partition operation mechanism. According to research data, there are currently six sanitation facilities in Lixia District, with plans to construct 13 more, as shown in Fig. 15a. The large number of tourists in scenic spots like Baotu Spring and Daming Lake during holidays also leads to a very noticeable spatiotemporal variation in population density and waste output in Lixia District, as shown in Fig. 15b–d. Since scenic spots like Baotu Spring and Daming Lake are located on the west side of Lixia District, most sanitation facilities are concentrated in the western part of the study area, with fewer facilities on the east side, resulting in an uneven distribution of sanitation facilities throughout the area. Furthermore, the entire sanitation service model is static and cannot meet the needs of the study area. The spatial distribution of sanitation facilities and the population density at different periods in Lixia District. Data sources and processing The population density data used in the study is sourced from World Pop40 (https://hub.worldpop.org/) and the Lixia District Tourism Public Service Center. The location data of sanitation facilities is sourced from field research. Given the extensive land use of the study area, the dense population distribution, and the comprehensive coverage of base stations, the waste clearance demand in the study area is discretized. The study area was divided into 800 m × 800 m grids using the Fishnet tool in ArcGIS Desktop 10.8.141 (URL: https://www.esri.com), with the centroid of each grid representing a waste clearance demand point. After excluding invalid demand points, 181 uniformly distributed waste collection demand points are generated within the study area, as shown in Fig. 16. The waste generated in each divided area is allocated to these demand points, determining the waste amount at each demand point. The distribution of demand points for garbage cleaning in the study area. Location and capacity determination result analysis The experiment was conducted on a Windows 10 × 64 system with a 3.70 GHz Intel Core i9-10900 K processor. The code was implemented in Matlab R2024a42 (https://ww2.mathworks.cn). Using the previously mentioned model and the improved NSGA-II to obtain solutions, the TOPSIS method is employed to rank the solution set. The model sets the maximum crossover probability of the improved NSGA-II at 0.8, the minimum crossover probability at 0.4, the maximum number of crossover points at 3, the minimum number of crossover points at 1, the mutation probability at 0.2, a decay coefficient of 0.2, the initial population at 50, and the number of generations at 200. First, the base station location scheme is solved, ultimately resulting in a set of balanced solutions, which constitutes the Pareto optimal solution set. The PF curve is shown in Fig. 17. Pareto optimal solution set under F1, F2, F3 objectives. The information entropy and entropy weight of each evaluation criterion in the calculated PF are presented in Table 6. Table 6Information entropy and entropy weight of each evaluation index.Evaluation criterioninformation entropyEntropy weightF10.9435050.327906F20.9321760.393661F30.9520290.278433 Information entropy and entropy weight of each evaluation index. The relative closeness of each program is shown in Fig. 18. Relative closeness of each program. As shown in Fig. 18, Scheme 42 has the highest relative closeness to the ideal solution, making it the optimal location scheme. The corresponding fitness function values are F1 = 3.39711 × 10⁸, F2 = 4.125 × 10¹⁰, and F3 = 8. This plan, building upon existing base stations, selects 9 candidate base stations and 3 new base stations, maximizing the fulfillment of waste collection needs, as shown in Fig. 19. From a spatial distribution perspective, the current planning scheme is relatively concentrated on the west side of the study area, which is closely related to the spatial locations of scenic spots such as Baotu Spring and Daming Lake. However, the concentration of a large number of facilities also leads to uneven spatial distribution and resource waste. The optimized plan integrates resources, offering a more rational location scheme and a more balanced distribution of facilities, thereby serving more demand points. The best location scheme of base stations. The optimized location scheme derived from solving the model is compared with the existing planning scheme, as shown in Table 7. The results indicate that the service efficiency of the optimized location scheme is significantly enhanced compared to the existing layout. Furthermore, it requires fewer facility points while covering more demand points, thereby meeting the garbage collection needs of areas that were previously uncovered by the existing planning scheme. Table 7Comparison between model site selection results are and existing planning schemes.Solution ComparisonNumber of FacilitiesService UtilityNumber of uncovered demand pointsBest location scheme183.39711 × 1088Existing planning scheme193.19603 × 10848 Comparison between model site selection results are and existing planning schemes. Next, we will determine the capacity of each base station, specifically the number of each of the two types of sanitation robots that should be equipped at each base station. The PF curve is shown in Fig. 20. Pareto optimal solution set under F4, F5 objectives. The information entropy and entropy weight of each evaluation criterion in the calculated PF are presented in Table 8. Table 8Information entropy and entropy weight of each evaluation index.Evaluation criterionInformation entropyEntropy weightF40.9328560.535299F50.9417120.464701 Information entropy and entropy weight of each evaluation index. The relative closeness of each program is shown in Fig. 21. Relative closeness of each program. As shown in Fig. 21, Scheme 38 has the highest relative closeness to the ideal solution, making it the optimal configuration scheme for sanitation robots. The corresponding fitness function values are F4 = 4.6055 × 107 and F5 = 7.53, which best meets the requirement of minimizing garbage collection time within the service range of the base station. The number of the two types of robots that should be equipped at each base station is shown in Table 9. Table 9Number of sanitation robots equipped at each base station.Base StationBase Station StateMedium-sizedSmall-sizedBS 1Candidate1924BS 2Existing2520BS 3Candidate1617BS 7Candidate1210BS 9Existing109BS 10Candidate98BS 11Candidate2119BS 12Candidate22BS 13Candidate12BS 14Candidate11BS 15Candidate76BS 16Existing67BS 17Existing1114BS 18Existing1010BS 19Existing11Newly Added Stations 1Newly Added22Newly Added Stations 2Newly Added23Newly Added Stations 3Newly Added76 Number of sanitation robots equipped at each base station. Based on the capacity scheme of sanitation robots, this study further demonstrates the optimization effects of sanitation robots on garbage collection efficiency, cost, and environmental benefits. By comparing the traditional manual cleaning mode with the automated cleaning mode of sanitation robots, the comprehensive benefits are highlighted in the following three aspects: Improve the efficiency of garbage cleaning. Traditional manual cleaning requires 10 h to complete the garbage collection task in the area and is constrained by physical stamina and shift schedules, resulting in insufficient operational continuity. In the optimized plan, sanitation robots operate around the clock with intelligent scheduling, reducing the collection time to within 8 h and improving efficiency by over 20%. Additionally, these robots are equipped with intelligent sensing systems that can monitor waste capacity in real-time, predict high-load areas, and dynamically optimize collection routes and frequencies, minimizing instances of empty or inefficient transport. Improve the efficiency of garbage cleaning. Traditional manual cleaning requires 10 h to complete the garbage collection task in the area and is constrained by physical stamina and shift schedules, resulting in insufficient operational continuity. In the optimized plan, sanitation robots operate around the clock with intelligent scheduling, reducing the collection time to within 8 h and improving efficiency by over 20%. Additionally, these robots are equipped with intelligent sensing systems that can monitor waste capacity in real-time, predict high-load areas, and dynamically optimize collection routes and frequencies, minimizing instances of empty or inefficient transport. Lower costs. In the study area, there is a resident population of 891,000 people. Based on the allocation of 30 sanitation workers per 10,000 people, a total of 2,673 sanitation workers are required. With an average annual salary of 60,000 yuan per worker, the annual labor cost amounts to 160.38 million yuan. Additionally, there is a need for ongoing investment in the purchase of sanitation vehicles, fuel, and equipment maintenance. By implementing the sanitation robot replacement plan, the initial equipment procurement and deployment cost is 46 million yuan, with subsequent expenses only for maintenance and electricity, significantly reducing investment costs. Lower costs. In the study area, there is a resident population of 891,000 people. Based on the allocation of 30 sanitation workers per 10,000 people, a total of 2,673 sanitation workers are required. With an average annual salary of 60,000 yuan per worker, the annual labor cost amounts to 160.38 million yuan. Additionally, there is a need for ongoing investment in the purchase of sanitation vehicles, fuel, and equipment maintenance. By implementing the sanitation robot replacement plan, the initial equipment procurement and deployment cost is 46 million yuan, with subsequent expenses only for maintenance and electricity, significantly reducing investment costs. Enhance the precision of urban management. The use of sanitation robots significantly improves the precision of urban management through data integration and intelligent analysis. With real-time data feedback from sanitation robots on waste distribution, cleaning frequency, and equipment status, management departments can build dynamic decision models to accurately predict waste generation trends and adjust resource allocation accordingly. Additionally, the intelligent scheduling system supports collaborative operations among multiple robots, allowing for swift responses to emergencies, such as increased waste volumes during holidays, thereby avoiding the delays of manual scheduling in traditional models. The introduction of robotic technology also simplifies management by enabling the monitoring of all equipment operations through a unified platform, advancing the digitalization and centralization of urban sanitation systems. Enhance the precision of urban management. The use of sanitation robots significantly improves the precision of urban management through data integration and intelligent analysis. With real-time data feedback from sanitation robots on waste distribution, cleaning frequency, and equipment status, management departments can build dynamic decision models to accurately predict waste generation trends and adjust resource allocation accordingly. Additionally, the intelligent scheduling system supports collaborative operations among multiple robots, allowing for swift responses to emergencies, such as increased waste volumes during holidays, thereby avoiding the delays of manual scheduling in traditional models. The introduction of robotic technology also simplifies management by enabling the monitoring of all equipment operations through a unified platform, advancing the digitalization and centralization of urban sanitation systems. Reduce carbon emissions. Sanitation robots use a fully electric drive mode, achieving zero carbon emissions during operation. Compared to traditional methods, the clean nature of electric energy significantly reduces the lifecycle carbon footprint, directly supporting the achievement of the “dual carbon” goals. Reduce carbon emissions. Sanitation robots use a fully electric drive mode, achieving zero carbon emissions during operation. Compared to traditional methods, the clean nature of electric energy significantly reduces the lifecycle carbon footprint, directly supporting the achievement of the “dual carbon” goals. Reduce the incidence of diseases among sanitation workers. In traditional sanitation operations, workers are exposed to environments with garbage odors and pathogenic microorganisms for extended periods, leading to a higher risk of respiratory infections, skin diseases, and accidental injuries. With sanitation robots replacing manual labor, only a small number of technicians are needed for remote monitoring, thereby reducing workers’ exposure to harmful substances. Reduce the incidence of diseases among sanitation workers. In traditional sanitation operations, workers are exposed to environments with garbage odors and pathogenic microorganisms for extended periods, leading to a higher risk of respiratory infections, skin diseases, and accidental injuries. With sanitation robots replacing manual labor, only a small number of technicians are needed for remote monitoring, thereby reducing workers’ exposure to harmful substances. Conclusions and recommendations This paper achieves scientific decision-making for the location and capacity optimization of sanitation robot base stations. This was accomplished by constructing a multi-objective optimization model that integrates spatio-temporal dynamic features, combined with an improved NSGA-II. The conclusions are as follows: The improved NSGA-II demonstrated excellent performance in the case study of Lixia District, Jinan. The optimized solution resulted in one less facility compared to the original plan, with only 8 demand points remaining uncovered, thereby exhibiting strong adaptability to spatio-temporal fluctuations in waste volume. The improved NSGA-II demonstrated excellent performance in the case study of Lixia District, Jinan. The optimized solution resulted in one less facility compared to the original plan, with only 8 demand points remaining uncovered, thereby exhibiting strong adaptability to spatio-temporal fluctuations in waste volume. The base station layout expanded eastward from densely populated scenic areas, leading to a more balanced distribution and effectively eliminating the coverage blind spots present in the original plan. The base station layout expanded eastward from densely populated scenic areas, leading to a more balanced distribution and effectively eliminating the coverage blind spots present in the original plan. The collaboration between medium-sized and small-sized sanitation robots effectively improves waste collection efficiency and reduces costs, demonstrating that heterogeneous collaboration can balance both efficiency and economy. The collaboration between medium-sized and small-sized sanitation robots effectively improves waste collection efficiency and reduces costs, demonstrating that heterogeneous collaboration can balance both efficiency and economy. Based on the findings of this study, the following recommendations are proposed: Promote the intelligent upgrade of environmental sanitation facilities by integrating IoT and edge computing devices, enabling operational status monitoring and dynamic service adjustment, and thereby enhancing responsiveness to holidays and emergencies. Promote the intelligent upgrade of environmental sanitation facilities by integrating IoT and edge computing devices, enabling operational status monitoring and dynamic service adjustment, and thereby enhancing responsiveness to holidays and emergencies. Incorporate base station deployment into smart city top-level design, establish cross-departmental coordination mechanisms, and formulate unified site selection standards. Incorporate base station deployment into smart city top-level design, establish cross-departmental coordination mechanisms, and formulate unified site selection standards. The proposed model can be extended to optimize the layout of public facilities such as emergency stations and charging piles. A multi-agent scheduling framework incorporating drones and unmanned vehicles can be established to enhance overall urban resource allocation efficiency. The proposed model can be extended to optimize the layout of public facilities such as emergency stations and charging piles. A multi-agent scheduling framework incorporating drones and unmanned vehicles can be established to enhance overall urban resource allocation efficiency. Acknowledgements We thank Candela (Shenzhen) Technology Innovation Co., Ltd. for supporting this study. The small-sized cleaning robot used in this study was provided by them. We also thank the Lixia District Tourism Public Service Center for providing data on holiday tourist volume, which greatly facilitated this study. In addition, we used ChatGPT 4.0 (https://chatgpt.com/) solely for grammar and language editing support. No content related to data interpretation, analysis, results, conclusions, or scientific contributions was generated by the tool. We take full responsibility for the final content of the manuscript. Author contributions L.S.W. proposed the research direction, provided partial funding support, and offered valuable suggestions for the manuscript. J.D. further developed the research content and was responsible for drafting the manuscript. K.P. and N.J. contributed data support, and N.C. provided partial funding support. All authors participated in reviewing the manuscript. Funding This research is supported by Grants: Research and Application of Key Technologies for Behavior Monitoring Robots in Public Spaces of Smart Communities (Shandong Province Key Research and Development Program, Innovation Capacity Improvement Project for Science and Technology-based SMEs, 2025TSGCCZZB0009), Research on the Location and Allocation Strategies for Emergency Material Storage and Transportation Facilities under a Multi-level, Cross-regional Collaborative Model (ZR2024MG033), Research project of the Department of Natural Resources of Shandong Province in 2024: “The Research on ‘Human-Robot-City’ Coexistence Scenarios and Guidelines for Robot-Friendly Urban Space” (2024-17) and Key research project of the Shandong Urban Renewal Society: “Research on the Deterministic Transformation of Urban Space in the Age of Robotics” (250102). Data availability The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Linshen Wang, Email: cea_wangls@ujn.edu.cn. Kun Pan, Email: pankun_503@163.com. References Associated Data Data Availability Statement The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.