Can Industrial Robot Adoption Improve the Green Total Factor Productivity in Chinese Cities?

Abstract

Industrial robot adoption significantly affects economic growth and environmental protection, serving as a critical driver of green development. This paper empirically investigates the effects of industrial robot adoption on green total factor productivity from the perspectives of knowledge flow and spatial spillover using the Chinese cities panel dataset. The findings demonstrate that industrial robot adoption improves local green total factor productivity while generating positive spillover effects on neighboring regions, mediated by strengthened knowledge agglomeration and diffusion capacities. Central cities within urban clusters exhibit significantly stronger impacts on knowledge aggregation and diffusion capabilities than peripheral cities. Furthermore, cities with higher human capital, better transportation infrastructure, and stronger support for the AI industry show a more significant positive effect of industrial robot adoption on knowledge agglomeration and diffusion capabilities. This, in turn, facilitates the flow of knowledge between cities and improves green total factor productivity, thereby contributing to green development. This study provides city-level empirical evidence highlighting how industrial robot adoption drives green development through spatial spillovers and knowledge flow mechanisms.

1. Introduction

In the era of industrialization and globalization, unprecedented ecological challenges have emerged [1]. Resource scarcity, biodiversity loss, atmospheric pollution, and energy depletion threaten Earth’s ecology and human sustainable development [2,3]. The U.S. Energy Information Administration predicts a 50% increase in global energy consumption from 2018 to 2050. The 2023 Emissions Gap Report notes a 1.2% rise in global greenhouse gas emissions in 2021–2022, reaching 57.4 billion tons of CO₂ equivalent. The world risks surpassing the goals of the Paris Agreement if more measures are not taken, necessitating a balance of economic growth and environmental protection within planetary boundaries [4,5].

Technological innovation is central to attaining Porter’s dual environmental–economic benefits critical for sustainable development [6,7]. As transformative technologies, industrial robots signify the future of manufacturing and are crucial for fostering green production and eco-friendly economic growth [8]. The World Robotics Report 2023 documents a 7% compound annual growth rate in global industrial robot sales (2017–2022), underscoring their growing centrality in contemporary industrial systems. Functioning as Industry 4.0’s technological backbone, industrial robots have reconfigured production paradigms and catalyzed structural economic transformations [9], thereby generating innovative pathways for sustainable development. The adoption of industrial robots has gradually demonstrated significant green attributes in real-world cases. For instance, Johnson & Johnson Vision (London), one of the lighthouse factories, integrated fourth industrial revolution technologies, including adaptive process control, AI, and robotics, resulting in a 50% increase in inventory turnover, 8% improvement in customer service, and 53% reduction in carbon footprint (delivered goods).

The existing literature predominantly focuses on the economic and sociological implications of industrial robot adoption. Current scholarly debates center on labor market dynamics [10], productivity enhancements [11,12], and workforce well-being [13] in the context of widespread robotic integration. Researchers adopting critical perspectives contend that automation adoption correlates with employment contraction and wage suppression [13,14]. Conversely, proponents highlight robotic systems’ capacity to execute specialized tasks with superior precision and operational efficiency compared to human labor, potentially yielding optimized production workflows and reduced operational expenditures [15]. Regrettably, scholarly discourse on industrial robots’ use for environmental preservation, green development, and sustainable development is somewhat limited. Notably absent from this discourse is a systematic inquiry into ecological externalities. Current research exhibits methodological fragmentation, primarily analyzing isolated pollutants (e.g., particulate emissions, wastewater discharge, CO₂ equivalents) rather than holistic sustainability metrics [16,17]. Some research has found that businesses that use industrial robots to do tasks minimize their emissions of air and water pollutants [18]. Counterintuitively, some argue that automation-induced energy efficiency gains may outweigh incremental pollution from robotic operations, particularly when enabling the transition to renewable-powered smart factories [19]. This compensation hypothesis posits that labor-to-robot substitution enhances technical energy efficiency and facilitates institutional transitions toward cleaner production regimes [20].

Green Total Factor Productivity (GTFP) has emerged as a critical metric integrating environmental stewardship with production efficiency in sustainability science [21,22,23]. Nevertheless, there is a dearth of research on the relationship between industrial robot adoption and GTFP, as well as the implications of information flow on geographical spillover, particularly in knowledge cities. Firstly, the geographical dimension remains underexplored in assessing automation’s dual impacts on ecological–economic system couplings. Regional economic theory posits that disruptive technologies (e.g., industrial revolution-driven urban expansion, IT-enabled polycentric development) inherently reshape urban spatial configurations through historical paradigm shifts [24]. Building on this technological determinism framework, the ongoing robotics revolution represents a next-wave spatial reorganizer within this technological determinism framework. As a disruptive innovation, industrial robots undergo accelerated diffusion across manufacturing value chains, creating complex knowledge spillover dynamics that transcend organizational and geographical boundaries [25]. So, more research is necessary to fully understand how industrial robots, a new technological revolution, affect the geographic dimensions of the economy, society, and environment. Secondly, current analyses inadequately address the network externalities inherent in knowledge flows. Robot adoption simultaneously transforms innovation ecosystems through (i) technology embodiment effects and (ii) knowledge recombination processes—dual drivers of regional development in increasingly interconnected techno-economic systems [26,27]. As these robots reshape production and living spaces, the exchange of technological talent, environmental information sharing, and green technology collaboration across industries and cities will notably affect economic growth, pollution reduction, and environmental protection [28], ultimately impacting GTFP.

Therefore, this paper takes Chinese cities as the research object to analyze the integrated impacts of transformative technologies on economic systems and environmental systems. To address these gaps, this paper first establishes a spatial econometric model to verify the existence of spatial spillover effects. Then, it employs a spatial social network model to explore the mechanisms of spatial linkages, aiming to capture the direct productivity enhancements and cross-regional knowledge externalities brought about by robot adoption. Then, it seeks to elucidate a series of pivotal inquiries: does the adoption of industrial robots precipitate a spatial spillover effect on GTFP? What intricate mechanisms facilitate this spatial spillover phenomenon? From the perspective of knowledge flow, how does the adoption of industrial robots impact the aggregation and diffusion of knowledge? The objective is to understand better how industrial robot adoption affects GTFP and how this impact diffuses and propagates spatially. This study provides empirical evidence for exploring urban and regional green development paths and building sustainable human settlements under emerging technological changes.

This paper makes three primary contributions. (1) This study takes a holistic view of the spatial correlation between industrial robot adoption and GTFP, systematically analyzing the complex relationship. It discusses the symbiotic relationship between industrial robot adoption and GTFP, viewed through the relatively uncharted terrain of spatial spillover. This exploration substantially augments our comprehension of industrial robot’s binary impact on economic development and environmental protection. (2) By deeply integrating knowledge spillover theory and proximity theory, it delves into the industrial robot adoption from the perspective of knowledge flow, examining its impact on the spatial pattern of innovation. Industrial robots are a comprehensive manifestation of advanced technologies that will significantly affect the flow and distribution of talents, technologies, and industries. As crucial hubs of knowledge flow, cities experience transformative changes in their innovation patterns due to the adoption of industrial robots. Moreover, technological innovation stands as a stalwart pillar in improving GTFP. Thus, this study shows how industrial robots can impact local and neighboring GTFP through knowledge flow. (3) By exploring heterogeneity in mechanisms, the study examines the differential effects of industrial robot adoption based on urban location characteristics and resource endowments. Given the pivotal role of urban agglomerations in China’s economic development, this research provides insights into adjusting agglomeration development models, strengthening urban competitive advantages, and improving GTFP through optimized knowledge flows.

Following is the research framework of this study, as shown in Figure 1.

2. Theoretical Framework and Research Hypotheses

Tobler’s First Law of Geography establishes the fundamental principle of spatial autocorrelation, positing that geographic entities exhibit stronger interdependencies with proximate counterparts than distant ones. Given regional integration, the use of industrial robots has an influence on GTFP that spreads to neighboring market entities and regions via value chains, industry chains, and supply networks, resulting in regional spillover effects of economic vibrancy [29]. Installing advanced technology in a location has a significant positive economic impact locally and catalyzes an economic spillover effect in adjacent urban areas [30]. This effect may manifest as neighboring city enterprises emulating advanced technologies to enhance their production efficiency, attracting innovative resources such as knowledge, capital, and talent, agglomerating emerging industries, and upgrading traditional industries, thereby improving GTFP [31].

Industrial robot adoption represents a systematic reorganization and optimal allocation of production factors, profoundly reshaping the spatial distribution of innovation activities and influencing GTFP. Specifically, the substitution of human labor by industrial robots alters demand for both low-skilled and high-skilled labor in local industries while simultaneously creating job opportunities in management, information technology, data analytics, and research. Changes and impacts on labor market demands will further influence the spatial distribution of labor, thereby affecting the layout of regional living spaces, supporting facilities construction, and transportation organization [30]. Furthermore, as a carrier of labor, capital, and innovation, the industrial layout will be significantly impacted by technological advancements, industrial associations, efficiency improvements, and production method transformations brought about by industrial robot adoption. This will reshape regional industrial and innovation spatial patterns. Through talent mobility, technology dissemination, and industry associations, the increased regional GTFP from the broader deployment of industrial robots will also improve the overall regional GTFP and accomplish coordinated regional economic growth [32]. Zhang et al. [19] highlight that industrial robot adoption’s technological advancements and GTFP enhancements frequently influence surrounding regions. Consequently, this study posits the following research hypothesis for empirical validation.

Hypothesis 1.

The adoption of industrial robots has a spatial spillover effect that improves the GTFP level of economically neighboring cities and a significant increase in the local GTFP level.

Proximity theory postulates that geographic, industrial, and technological adjacency accelerates knowledge, technology, and experience diffusion [33,34]. Industrial robot integration marks a significant change in technological innovation and production methodologies. The endogenous growth framework emphasizes systemic knowledge accumulation as the prime mover of sustainable economic evolution [35]. In this context, industrial robots, especially the new generation robots as a product of the integration of robotics, big data, artificial intelligence, and cloud computing, are more conducive to facilitating the sharing and exchange of human knowledge and information across time and space [36]. Consequently, the knowledge flow induced by industrial robots can change the innovation landscape, thereby impacting regional GTFP [37].

Industrial robot adoption has a complex dynamic process of agglomeration–innovation–diffusion. On the one hand, during agglomeration, interactions among knowledge workers, information, entrepreneurial capital, and R&D institutions cultivate innovation ecosystems that are conducive to knowledge exchange, cooperation, and innovation [38]. During industrial robot adoption, the aggregation of innovative elements further triggers a self-reinforcing mechanism, attracting the flow of labor, capital, and information technology to the local area. The intelligent production organization model promotes the shift of manufacturing from human–machine separation to integrating industrial robots, labor, information systems, and services. This interconnected system also encourages the spatial aggregation of new innovative elements, demonstrating the knowledge agglomeration capacity [39].

On the other hand, the innovation resulting from this agglomeration—including patents, new products, and the mobilization of talent and information—spreads outward through a combination of market and non-market channels [40], reflecting the ability of knowledge diffusion. Significantly, knowledge diffusion not only integrates economic benefits with the intrinsic value of knowledge, but also initiates a process of reverse value flow [41]. This process, in turn, enhances conditions for further knowledge agglomeration, fostering a virtuous cycle between knowledge concentration and dissemination. The continuous advancement of industrial robot adoption will persistently promote the agglomeration of related high-tech enterprises, forming a larger industrial cluster and driving the constant development of the entire industry group [42]. The organizational proximity and geographic proximity of enterprises or institutions within the cluster promote mutual trust and cooperation, thereby driving collective learning and knowledge sharing among them [43]. Such interactions facilitate the dissemination of various new pieces of knowledge, new ideas, and new technologies, creating a knowledge spillover effect and significantly promoting innovation and innovation diffusion within the cluster [44].

According to the Buzz–Pipeline Theory, knowledge flow mainly occurs through two modes: buzz (informal, face-to-face communication) and pipeline (formal, institutionalized communication channels) [45]. The complementarity formed by these two modes of knowledge flow at the geographical and organizational levels is key to the effective adoption of industrial robots and their impact on GTFP. Industrial robot adoption transcends mere alterations in production methods [46] but is also a catalyst for the flow and dispersion of knowledge. This theory highlights the circulation and dissemination of information in networks, with the proliferation of industrial robots acting as a buzz of information and technology in the manufacturing sector, rapidly spreading through the network’s pipelines to different nodes. This efficient flow of knowledge allows producers to quickly access the latest technological advancements globally, thereby enhancing productivity. Employees and enterprises may benefit from intelligent production by gaining experience and skills when sophisticated industrial robot technology is introduced in a certain area [47], including increased productivity, cost reduction, and cleaner production [48]. Consequently, this study posits the following research hypothesis for empirical validation.

Hypothesis 2.

The adoption of industrial robots enhances the knowledge agglomeration capability and knowledge diffusion capability of cities, thereby strengthening knowledge flow between cities and improving GTFP.

3. Model Design and Variable Selection

3.1. Model Setting

(1)

Global Spatial Autocorrelation Test

The $Geary’s C$ index and the global $Moran’s I$ index are commonly used for spatial correlation analysis [49]. $\mathrm{Geary}’\mathrm{s} C$ index is a statistical measure used in spatial data analysis, primarily to measure the autocorrelation of a variable in space. The formula is as follows:

$$\mathrm{Geary}’\mathrm{s} C={\displaystyle \frac{\left(n-1\right){\sum }_{i=1}^n{\sum }_{j=1}^n{w}_{ij}{\left(Y_i-Y_j\right)}^2}{2\left({\sum }_{i=1}^n{\sum }_{j=1}^n{w}_{ij}\right){\sum }_{i=1}^n{(Y_i-\overline{Y})}^2}}$$

(1)

where $C$ is Geary’s index and n is the sample size, $\overline{Y}$ is the average GTFP of cities, $Y_i$ is the GTFP of city $i$, and $w_{ij}$ is the spatial weighting matrix. $Geary’s C$ fluctuates between 0 and 2, with values below 1 indicating a positive spatial correlation, values above 1 suggesting a negative spatial correlation, and a value of 1 indicating a random distribution.

Contrary to $Geary’s C$, which concentrates on neighboring data point comparisons, $Moran’s I$ is designed to assess the broader relational dynamics between neighboring data points [50]. The formula is as follows:

$$Mora{n}^{\prime }s I={\displaystyle \frac{{\sum }_{i=1}^n{\sum }_{j=1}^n{W}_{ij}\left(Y_i-\overline{Y}\right)\left(Y_j-\overline{Y}\right)}{S^2{\sum }_{i=1}^n{\sum }_{j=1}^n{W}_{ij}}}$$

(2)

where $Y_j$ is the GTFP of city $j$, $S^2$ represents the sample variance, and $W_{ij}$ denotes the spatial weight matrix. $Moran’s I$ has a value between −1 and 1, with values near 1 denoting a significant spatial correlation, values near −1 denoting a great spatial disparity, and values close to zero indicating a random spatial distribution.

(2)

Local Spatial Autocorrelation Test

Local spatial autocorrelation explores the spatial correlation and heterogeneity of a sample city’s GTFP with its neighboring cities from a local spatial perspective. The local Moran’s index is calculated as follows:

$$I_i=Z_i{\displaystyle \sum }_{i=1}^n{W}_{ij}{Z}_j$$

(3)

where $Z_i$ and $Z_j$ represent the normalized values of green total factor productivity observations for cities $i$ and $j$, respectively. Four quadrants—high–high (HH), low–high (LH), low–low (LL), and high–low (HL)—represent distinct local spatial correlation patterns based on the Moran scatter plot analysis of local spatial autocorrelation. Positive spatial correlation is shown by HH and LL, whereas negative spatial correlation is indicated by LH and HL.

(3)

Spatial Econometric Models

Given the extensive and close economic links between regions in China, this paper posits that industrial robot adoption in one region may impact the GTFP of surrounding regions. To explore this, we set the following more general Spatial Durbin Model (SDM):

$$\begin{array}{ll}GTF{P}_{it}={\beta }_0+\rho & {\displaystyle {\displaystyle \sum }_{j=1,j\ne i}^n}{W}_{ij}GTF{P}_{it}+{\beta }_{1}Robot\_exposur{e}_{it}+\phi {\displaystyle {\displaystyle \sum }_{j=1,j\ne i}^n}{W}_{ij}Robot\_exposur{e}_{it}+\gamma contro{l}_{it}\\ & +\vartheta {\displaystyle {\displaystyle \sum }_{j=1,j\ne i}^n}{W}_{ij}contro{l}_{it}+{\delta }_i+{\mu }_t+{\epsilon }_{it}\end{array}$$

(4)

where $GTF{P}_{it}$ denotes the green total factor productivity of city $i$ during a period $t$, $Robot\_exposur{e}_{it}$ denotes the robot penetration of city $i$ in the period $t$, $control$ denotes the set of control variables that affect urban green total factor productivity and change with $i$ and $t$. The term ${\delta }_i$ represents city fixed effects, which control for individual effects that affect urban green total factor productivity but do not vary over time, ${\mu }_t$ is time fixed effects, which control for time effects that affect urban green total factor productivity but do not vary over time, ${\epsilon }_{it}$ is a random disturbance factor, $\rho$ is the spatial autoregressive coefficient, indicating that there is also an effect of neighboring cities’ GTFP on local GTFP, and $\phi$ is the extent to which the penetration of industrial robots in other cities affects local GTFP.

Given the feedback effect inherent in spatial lag terms, the point estimates from the SDM might not accurately represent the impact of industrial robot adoption on GTFP. To mitigate potential biases in direct regression, this study employs partial differentiation to decompose the impacts into three distinct components: direct, indirect, and total effects [51]. The specific calculation formulas are as follows:

$$Direct Effec{t}_k={\displaystyle \frac{1}{n}}{\displaystyle \sum }_{i=1}^n{S}_k\left(i,i\right); Indirect Effec{t}_k={\displaystyle \frac{1}{n\left(n-1\right)}}{\displaystyle \sum }_{i=1}^n{\displaystyle \sum }_{j\ne i}{S}_k\left(i,j\right)$$

(5)

where $S_k={\left(I-\rho W\right)}^{-1}\times \left({\beta }_{1k}I+{\phi }_{k}W\right)$ is the effect matrix calculated based on the estimated parameters of the SDM. The direct effect is the average value of the diagonal elements of the above matrix, reflecting the impact of industrial robot adoption in the local area on GTFP. The indirect effect is the average value of the non-diagonal elements of the matrix, reflecting the impact of industrial robot adoption in the local area on the GTFP of other regions through spatial associations. The total effect is the sum of the direct and indirect effects.

3.2. Variable Selection and Data Description

(1)

Core Explanatory Variable: Industrial Robot Penetration

The current approach to measuring robot penetration is mainly based on the research findings of Acemoglu and Restrepo (2020) [15]. They used the “Bartik instrument variable” construction method to measure robot penetration at the commuting area level in the United States from 1993 to 2007. As a result of the International Federation of Robotics (IFR) limited availability of robot stock data, which are primarily focused on industrial robots at the national industry level, numerous researchers have expanded the scope of their study on robot adoption by measuring robot penetration at the micro (enterprise), meso (industry), and macro (city, province, or country) levels, using Acemoglu and Restrepo’s method [15]. China has witnessed a remarkable growth in industrial robot adoption, emerging as a significant global market. Many scholars have used this method to depict the penetration of industrial robots at the enterprise and city levels in China. For example, Zhang et al. (2025) use the Bartik instrumental variable method to measure robot adoption in China at the city level [52]. Many scholars have confirmed the consistency of the results obtained from this method with the development trend in China [53]. Thus, this paper constructs the industrial robot penetration at the city level in China using the Bartik instrument variable method, referring to Wang and Dong [54], as shown in Equation (6):

$$Robot\_exposur{e}_{it}={\displaystyle \sum }_{j=1}^J{\displaystyle \frac{wor{k}_{ij,t=2006}}{wor{k}_{i,t=2006}}}{\displaystyle \frac{robo{t}_{jt}}{L_{j,t=2006}}}$$

(6)

where $wor{k}_{ij,t=2006}$ represents the employment count in different manufacturing industries $j$ in city $i$ in the base period of 2006, $wor{k}_{i,t=2006}$ represents the total number of employment posts in the manufacturing industry in city $i$ in the base period of 2006, $robo{t}_{jt}$ quantifies the national count of industrial robots in industry $j$ for period $t$, and $L_{j,t=2006}$ denotes the total employment across the nation in different manufacturing sectors for the base year. The selection of 2006 as the foundational period for employment data is informed by the rationale that historical data precludes contemporaneous correlations with shifts in employment structures at both national and industry levels, thereby safeguarding the homogeneity in constructing share weights.

(2)

Explained Variable: Green Total Factor Productivity (GTFP)

Using a non-radial, non-angular directional distance function as a framework, this paper constructs a green total factor productivity DEA model based on the global technological environment and in the form of a Luenberger index to measure GTFP at the city level, referring to Liu et al. (2020) [55]. The generalized non-radial directional distance function under the global technological frontier is defined as follows:

$$\overrightarrow{D^o}\left(x,y,b;g\right)=sup\left\{w^T\beta :\left(x,y,b\right)+g\times diag\left(\beta \right)\in P^o\left(x\right)\right\}$$

(7)

where $w={\left(w_n^x,w_m^{\gamma },w_i^b\right)}^T$ denotes the weight vector associated with input factors, desirable outputs, and undesirable outputs, $g=\left(-g_x,g_y,-g_b\right)$ is the direction vector, which indicates the expected direction of efficiency improvement, that is, the reduction of input factors, the increase in desired outputs, and the decrease in undesired outputs, and $\beta ={\left({\beta }_{nx,}{\beta }_{my},{\beta }_{ib}\right)}^T\ge 0$ represents the directional distance function value of each variable, also known as the proportion factor, indicating the possible proportion of input reduction, desired output increase, and undesired output decrease.

Based on the Luenberger productivity index form (Chambers et al., 1996) [56], the $GTFP$ for period $\mathrm{t}+1$ is defined as follows:

$$GTFP=\overrightarrow{D^o}\left(x^t,y^t,b^t;g^t\right)-\overrightarrow{D^o}\left(x^{t+1},y^{t+1},b^{t+1};g^{t+1}\right)$$

(8)

The Non-radial Directional Distance Function (NDDF) framework is instrumental in gauging city-level GTFP in China, tracking each city’s deviation from the production technology frontier. In this framework, a GTFP value exceeding 0 indicates that a city in the period $\mathrm{t}+1$ is nearer to the frontier compared to the base period $\mathrm{t}$, indicating an improvement in GTFP. Conversely, a GTFP value lower than 0 suggests a regression in GTFP. In this study, the input factors include capital stock (K), labor (L), and total energy consumption (E), with GDP (Y) as the expected output, and CO₂ emissions and PM2.5 concentration (PM) as undesired outputs. Weights reflect policymakers’ priorities in adjusting variables. While different weights can be assigned based on research needs, many scholars agree that equal weighting of inputs and outputs is reasonable when there is no prior information [57]. This paper assumes that inputs, desired outputs, and non-desired outputs are of equal importance. In particular, the current Chinese macro-policy level emphasizes pollution reduction, carbon reduction, and green expansion and growth; therefore, each of the three is given an equal weight. Following the approach of Zhang et al. (2013), equal weights of 1/3 were assigned to desired outputs, undesired outputs, and input factors, respectively [58], a method that has been widely used in China’s environmental efficiency analysis. This approach aims to avoid subjective weighting biases while ensuring that the model balances economic growth objectives, resource utilization efficiency, and environmental constraints in the optimization process, consistent with the concept of sustainable development. The weights were then evenly distributed among the specific types of desired outputs, undesired outputs, and input factors, considering the synergistic impact of capital, labor, and energy on urban development. Thus, the weights for input and output factors were set as $w={\left(w_l,w_k,w_e,w_y,w_{co2},w_{pm}\right)}^T={\left(1/9,1/9,1/9,1/3,1/6,1/6\right)}^T$, which remains constant throughout the analysis, ensuring consistent optimization trajectories.

(3)

Control Variables

We added a set of control factors in conjunction with the existing literature. Specifically, (1) openness level (FDI), measured by the logged foreign investment inflows; (2) industrial structure (STR), using the secondary industry’s value-added ratio; (3) market dynamism (MAR), indicated by the ratio of urban entrepreneurs to employees; (4) digitalization level (DIGITAL), captured by the logged number of internet users; (5) tax burden (COST), represented by the VAT burden on industry; (6) degree of government intervention (GOV), calculated as fiscal expenditure per capita.

(4)

Spatial Weight Matrix

Furthermore, taking into account the fact that the spatial correlation of GTFP in cities is not only associated with geographic distance but also more strongly with economic development, this study creates an economic–geographical nested weight matrix, as shown in Equation (9):

$$W_{ij}=\left\{\begin{array}{c}\alpha \times W_{ij}^d+\left(1-\alpha \right)W_{ij}^e i\ne j\\ 0 i=j\end{array}\right.$$

(9)

where $W_{ij}^d$ represents the inverse distance weight matrix, $W_{ij}^e$ represents the economic distance weight matrix, calculated based on the average GDP growth rate during the observation period, and the value of $\alpha$ is set to 0.5.

(5)

Daa Description

This study measures city-level industrial robot penetration through three key steps. First, we calculate sectoral robot penetration rates using national manufacturing employment data from the China Stock Market & Accounting Research Database (CSMAR) and robot stock statistics from the International Federation of Robotics (IFR). Second, city–industry weights are constructed based on the distribution of urban manufacturing employment derived from the China Industrial Enterprise Database. Finally, these components are integrated to derive city-level penetration rates through a shift–share aggregation method.

The IFR data, compiled from global robot manufacturers, provide authoritative country–industry–year statistics that address coverage gaps in the imported data and ensure the accuracy of our core explanatory variable. We manually align two-digit industry codes to resolve discrepancies between the International Standard Industrial Classification (ISIC) and China’s 2002 National Economic Industry Classification. National subsector employment data for the base year (2006) are extracted using IFR-compatible industry coding rules. For city-level estimates, we reconcile firm-level employment records (China Industrial Enterprise Database) with aggregated totals from the China City Statistical Yearbook. This dual-source approach addresses the lack of official city–sector employment statistics while ensuring consistency with macroeconomic trends. As shown in Figure 2, the derived penetration index aligns closely with China’s industrial automation trajectory, validating the robustness of the methodology in this paper. In the empirical analysis, this variable underwent a logarithmic transformation.

Following the approach of Wu et al. (2014) [59], energy data were inferred based on provincial-level energy consumption data and nighttime light data. The formula is $E_{it}=k_{t}D{N}_{it}$, where $E_{it}$ is the total energy consumption of province i in year t and $D{N}_{it}$ is the sum of the grayscale values of all grids in province $i$ in year $\mathrm{t}$. CO₂ emission data were sourced from the Center for Global Environmental Research website, with city-level data extracted based on grid information. PM2.5 concentration data were obtained from Washington University in St. Louis, with city-level panel data on PM2.5 concentrations obtained by cutting and summarizing grid data within China. The remaining data were sourced from the China City Statistical Yearbook, resulting in balanced panel data for 273 Chinese cities from 2007 to 2019. Specific variable names and descriptive statistics are shown in

Table 1. Descriptive statistics of variables.

Variable	Obs	Mean	Std. Dev.	Min	Max
$GTFP$	3549	0.0516	0.2037	−2.6206	2.8475
$Robot\_exposure$	3549	2.7568	1.4954	−1.6286	6.8286
FDI	3549	9.7880	2.0040	−3.9453	14.9413
STR	3549	3.8419	0.2403	2.4596	4.5105
MAR	3549	9.1401	0.6051	6.2509	12.0518
DIGITAL	3549	13.0349	1.0929	5.4681	17.7617
COST	3549	13.0273	1.1888	6.0000	16.1795
GOV	3549	8.6576	0.7181	6.5769	11.6026

.

4. Baseline Regression: Validating the Existence of Spatial Effects

4.1. Spatial Correlation Test

As shown in

Table 2. Results of the core variables for Moran’s I and Geary’s C.

Variables	Moran’s I			Geary’s C
	I	z	p-Value	C	z	p-Value
$GTFP$	0.065	12.245	0.000	0.912	−6.414	0.000
$Robot\_exposure$	0.804	149.347	0.000	0.190	−138.624	0.000

, this study first examines the spatial correlation of the core explanatory and explained variables. The significance of Geary’s C index and Moran’s I index is determined by p-values and z-scores. Both indices for the core explanatory and explained variables are significantly greater than zero and less than one, respectively, indicating spatial correlation.

Considering the potential subtleties in observing spatial autocorrelation related to economic–geographical distance, the study further examines these variables using multi-stage, year-by-year Geary’s C and Moran’s I indexes.

Table 3. Local spatial correlation test results of green total factor productivity.

Year	2007	2008	2009	2010	2011	2012	2013	2014	2015	2016	2017	2018	2019
Moran’s I	0.021	0.003	0.023	0.002	−0.001	0.007	0.048	−0.007	−0.040	0.048	0.030	0.199	0.128
P	0.091	0.356	0.073	0.377	0.437	0.265	0.002	0.431	0.021	0.001	0.036	0.000	0.000
Geary’s C	0.948	0.968	0.954	0.994	0.972	0.914	0.975	1.005	1.053	0.851	0.921	0.794	0.845
P	0.034	0.204	0.074	0.434	0.323	0.063	0.255	0.429	0.112	0.001	0.002	0.000	0.000

shows that Geary’s C index for green total factor productivity is less than 1 for most years, with significance at the beginning and end of the sample period. Moran’s I index is generally positive, indicating varying degrees of significance, especially in the last five years of the sample period, though some years, like 2008 and 2011, are not significant. These results largely confirm the existence of spatial autocorrelation in city-level GTFP, suggesting spatial clustering (Ma et al., 2021) [60]. In fact, non-significance in Moran’s I index doesn’t entirely negate spatial autocorrelation, which may be region-specific or negated by opposing effects, necessitating further investigation using local Moran’s I index.

The scatter plot of local Moran’s I index illustrates local spatial correlations, indicating the degree of similarity or difference in GTFP between a city and its neighboring cities. The scatter plots for the first and last four years of the sample period are shown in Figure 3. The spatial autocorrelation of GTFP shows notable changes over the sample period, with high–high (HH) and low–low (LL) clustering characteristics. This pattern underscores a marked spatial autocorrelation of GTFP at the city level in China. Consequently, it becomes imperative to employ a spatial econometric model to analyze the spatial spillover effects of industrial robot adoption on GTFP.

4.2. Spatial Spillover Effect Test

While the SDM exhibits greater generality compared to the Spatial Lag Model (SLM) and Spatial Error Model (SEM) within the realm of spatial autocorrelation econometric models, rigorous verification is essential. To this end, we conducted a series of tests, including the Lagrange Multiplier (LM) test, the Likelihood Ratio test (LR test), and the Hausman test. As presented in

Table 4. Adaptability test results.

Test	Statistical Value	p-Value
LM_Spatial lag	71.856	0.000
LM_ Spatial error	68.775	0.000
Wald_Spatial lag	12.48	0.0859
Wald_Spatial error	18.06	0.0117
LR_Spatial lag	12.48	0.0860
LR_Spatial error	18.62	0.0095
Hausman	20.60	0.0044
Jointly significant test of spatial-fixed effects	26.91	0.0295
Jointly significant test of time-fixed effects	170.66	0.0000

, the first selection of the Spatial Durbin Model (SDM) results from rejecting the null hypothesis at the 1% level by both LM-Lag and LM-Error, implying the simultaneous effects of spatial lag and spatial error. Wald and LR tests were used to see if SDM could be reduced to SLM or SEM models, since SLM and SEM are particular examples of SDM. At various degrees of significance, the null hypothesis is rejected by both tests. The study selects the double-fixed SDM for further analysis based on the Hausman test, which is significant at the 1% level with a statistic of 20.60, taking into account the combined results of the spatial and temporal fixed effects tests.

As shown in

Table 5. Results of spatial Durbin model testing.

Variables	SDM
$Robot\_exposure$	0.0277 **
	(0.0141)
$\mathrm{W}\times Robot\_exposure$	0.0352 *
	(0.0214)
Direct effect	0.0300 **
	(0.0139)
Indirect effect	0.0421 *
	(0.0231)
Total effect	0.0721 ***
	(0.0209)
Spatial rho	0.1237 ***
	(0.0317)
Variance sigma2	0.0387 ***
	(0.0009)
Controls	YES
Fixed time	YES
Fixed city	YES
R-squared	0.0062
N	3549

Note: *, **, and *** indicate significance levels of 10%, 5%, and 1%, respectively; standard errors are reported in parentheses.

, the coefficient for industrial robot adoption is statistically significant and positive. The results reveal that the spatial autoregressive coefficient rho is significantly positive, indicating a general spatial clustering of city-level GTFP, aligning with earlier observations from Moran’s I index. Crucially, the interaction term between industrial robot penetration and the spatial weight matrix ($\mathrm{W}\times Robot\_exposure$) demonstrates a positive and statistically significant coefficient at the 10% level. This finding provides robust evidence of spatial spillover effects, indicating that industrial robot adoption in one city positively influences GTFP in geographically or economically connected neighboring regions. To further disentangle these spatial dynamics, the direct effect is significantly positive at the 5% level, suggesting that increased local industrial robot penetration improves local GTFP. The indirect effect is significantly positive at the 10% level, suggesting that local industrial robot penetration positively influences GTFP in spatially connected regions. Overall, industrial robot adoption not only significantly enhances local GTFP, but also has a noticeable spatial spillover effect, facilitating improvements in GTFP in connected regions.

5. Mechanism Study: The Role of Knowledge Flow

5.1. Construction of Spatial Social Networks

During the latter half of the 20th century, technological innovation transitioned into a networked form, facilitating extensive global knowledge flows [61]. This shift presents new opportunities and challenges for technological innovation. Scholars have recently explored the network structures and impacts of knowledge flow across various spatial scales, from local to global contexts [62]. Incorporating green technological progress into green total factor productivity, this study employs a gravity model to measure knowledge flow networks, supported by city-level green patent data. The gravity model, extensively used in international trade, regional integration, and knowledge flow, is adapted, in this study, to assess knowledge flow network interactions at the spatial level, as detailed in Equation (10):

$$R_{ijt}=G_{ijt}{\displaystyle \frac{\sqrt{G{P}_{it}\times K_{it}}\times \sqrt{G{P}_{jt}\times K_{jt}}}{D_{ij}^2}}$$

(10)

where $R_{ijt}$ denotes the strength of knowledge flow association between cities $i$ and $j$, $G{P}_i$ and $G{P}_j$ denote the number of green patent applications in city $i$ and city $j$, and $D_{ij}^2$ is the geographic distance of the center location between city $i$ and city $j$. $K_i$ and $K_j$ denote the knowledge stock of city $i$ and city $j$ and $G_{ijt}$ is the modified gravity coefficient, which is constructed using the city knowledge stock, drawing on the practice of Ma et al. (2022) [63]. The formula is $G_{ijt}={\displaystyle \frac{K_{it}}{K_{it}+K_{jt}}}$, which is calculated to reflect the contribution of city $i$ in the knowledge stock linkage between the two cities. Among them, the knowledge stock is measured by the perpetual inventory method. The specific equation is as follows:

$$K_{it}=\left(1-\delta \right)K_{i,t-1}+I_{it}$$

(11)

$$K_{i0}=1/2\left[{\displaystyle \frac{I_{i0}}{g+\delta }}+{\displaystyle \frac{I_{i0}\left(1+g\right)}{g+\delta }}\right]$$

(12)

where $K_{it}$,$K_{i,t-1}$ is the stock of knowledge of city $i$ during period $t$ and period $t$−1, $\delta$ is the depreciation rate, which is taken as 15% here, and $I_{it}$ denotes the real science and technology expenditure of city $i$ during period $t$, which is calculated from the nominal science and technology expenditure using the CPI price index with 2006 as the base period. $K_{i0}$ is the stock of knowledge during the base period, $I_{i0}$ is the real science and technology expenditure during the base period, and g is the average growth rate of the real science and technology expenditure in the examination period.

The study constructs a network matrix of asymmetric knowledge flow connections between cities based on green patents and knowledge stock which serve as pivotal indicators of urban knowledge creation and accumulation. Social network analysis of this asymmetric network is used to gauge the cities’ capabilities in generating knowledge flows. Within this framework, the degree centrality indicates whether a city is at the core of the spatial network. Knowledge aggregation and spillover are two fundamental aspects of knowledge flows within the network. By measuring in-degree and out-degree, we can directly observe the direction and quantity of knowledge flows between cities. The in-degree reflects a city’s ability to aggregate knowledge from other cities in the spatial network, indicating its knowledge aggregation ability. Conversely, the out-degree reflects a city’s ability to disseminate knowledge to other cities, indicating its knowledge spillover ability [64].

This knowledge network model integrates geographic proximity and economic linkages. Geographically, distances and transportation links between cities are calculated using ArcGIS. Economically, cross-regional patent data are weighted. This hybrid measure inherently reflects spatial relationships, showing both geographic proximity and innovation-driven linkages. Some scholars have argued that social network variables can replace the traditional spatial weight matrix and can effectively characterize spatial associations in traditional panel models [64,65]. Therefore, in order to explore how the industrial robot adoption affects GTFP by influencing knowledge flows, in turn leading to changes in urban innovation factors, we constructed the following panel model to test it1:

$$flo{w}_{it}={\alpha }_0+{\alpha }_{1}Robot\_exposur{e}_{it}+\gamma contro{l}_{it}+{\delta }_i+{\mu }_t+{\epsilon }_{it}$$

(13)

where $\mathrm{flow}$ denotes the urban knowledge flow capacity, which contains two channel variables, namely knowledge agglomeration capacity ($Gatherin{g}_{it}$) and knowledge diffusion capacity ($Diffusio{n}_{it}$), represented by in-degree and out-degree in the social network of knowledge flow, respectively, and log-transformed in the empirical analysis. The term ${\alpha }_1$ reflects the effect of industrial robot adoption on the city’s knowledge agglomeration capacity or knowledge diffusion capacity. The green patent data used in this section were obtained from the State Intellectual Property Office (SIPO), and the green patents were screened according to the Green Patent List provided by the World Intellectual Property Office (WIPO) which contains seven categories of green technologies: transportation, waste management, energy conservation, alternative energy production, administrative regulation and design, agriculture and forestry, as well as nuclear power. Given the strong interaction between cities in this gravity model, we used provincial-level clustered robust standard errors to correct heteroskedasticity and correlation issues caused by intercity interactions.

5.2. Mechanism Testing Regression Results

The estimated coefficients of industrial robot penetration, as indicated in columns (1) and (3) of

Table 6. Results of mechanisms testing for the role of knowledge flow2.

Variables	(1) OLS	(2) 2SLS IV	(3) OLS	(4) 2SLS IV
	Knowledge Aggregation Capacity		Knowledge Diffusion Capacity
$Robot\_exposure$	0.5574 ***	0.6589 ***	0.3859 ***	0.4553 ***
	(0.1517)	(0.1730)	(0.1187)	(0.1369)
Controls	YES	YES	YES	YES
Fixed time	YES	YES	YES	YES
Fixed city	YES	YES	YES	YES
Cons	4.6747 **	6.4513 ***	4.9675 **	7.9904 ***
	(2.1318)	(2.2814)	(2.0800)	(2.1379)
Kleibergen-Paap rk LM statistic	/	Chi-sq(2) = 19.96 p-val = 0.0000	/	Chi-sq(2) = 19.96 p-val = 0.0000
Stock-Yogo weak ID test critical values	/	The critical value for a 10% significance level is 19.93	/	The critical value for a 10% significance level is 19.93
N	3549	3276	3549	3276
R-squared	0.2955	0.6789	0.1880	0.674

Note: *, **, and *** denote significance levels of 10%, 5%, and 1%, respectively; clustered robust standard errors at the provincial level are presented in parentheses.

, are all significantly positive at the 1% level when knowledge agglomeration and diffusion capacity are used as explanatory variables. This suggests that the higher the penetration of industrial robots, the more conducive to the enhancement of the city’s knowledge agglomeration and knowledge spillover capacity. Adoption of industrial robots can result in knowledge accumulation, application, and dissemination, a crucial indicator of technological advancement and shifts in productivity patterns which can greatly increase the productivity gains of green total factors and technological innovation in nearby areas [35].

Additionally, firms or regions with stronger green knowledge agglomeration and diffusion capabilities may prefer to invest in technologies with green production capabilities, including industrial robot adoption, in production processes. To mitigate the endogeneity problem of the model caused by possible bidirectional causality, this paper follows the approach of Zhang et al. [19] and selects the U.S.-level industrial robot penetration data with lagged one-period explanatory variables as the instrumental variables to conduct two-stage OLS of the instrumental variables. The U.S.-level industrial robot penetration is calculated as shown in model (6), with data from the U.S. Bureau of Statistics. Based on the explanatory variables of knowledge agglomeration capability and knowledge diffusion capability, the estimated coefficients of industrial robot penetration are both significantly positive at the 1% level in columns (2) and (4), verifying the robustness of the previous conclusions.

5.3. Heterogeneity Analysis

(1)

Urban Location Factors: Central-Peripheral Cities

In August 2019, the fifth meeting of the Central Finance and Economics Commission (CFEC) featured a special discussion on regional economic development and put forward a new idea of regional development: central cities and city clusters are the main forms of regional development. Studies have identified a spatial polarization between central and peripheral cities, with significant disparities in market opportunities, infrastructure, public services, and quality of life [66]. This study examines the differences in the impact of industrial robot adoption and knowledge flow in various major urban clusters, including the Chengdu–Chongqing economic circle, the Yangtze River Delta city cluster, the Pearl River Delta city cluster, and the Beijing–Tianjin–Hebei city cluster. Based on the National New Urbanization Plan (2014–2020), we identify the central cities in these urban clusters3.

The knowledge flow intensity diagram for these city clusters shows significant regional variations, as shown in Figure 4. Central cities in these clusters often have stronger knowledge flow connections and act as significant sources of knowledge dissemination. For example, in the Chengdu–Chongqing economic circle, Chengdu and Chongqing have strong interactions and significant “radiation effects” on surrounding cities. In the Yangtze River Delta city cluster, there is a strong knowledge flow relationship between central cities. The Pearl River Delta city cluster shows more obvious internal differences, such as Dongguan City, Shenzhen City, and Guangzhou City whose intensity is greater, and the dispersion effect is more obvious. The Beijing–Tianjin–Hebei city cluster “double core” development trend is obvious. In recent years, the relevant top-level design has been strengthened, and regional coordinated development has entered the fast lane.

To quantify the differences in knowledge aggregation and diffusion capabilities between central and peripheral cities within urban clusters, we constructed a dummy variable (center) which takes the value of 1 for central cities. The interaction of this dummy variable with the core explanatory variable (industrial robot penetration) was then analyzed for causal reinforcement. As shown in

Table 7. Heterogeneity analysis of the role channels of peripheral cities in various urban agglomeration centers.

Panel A: Chengdu–Chongqing Economic Circle
	(1)	(2)	(3)	(4)
Variables	Knowledge Aggregation Capacity		Knowledge Diffusion Capacity
$Robot\_exposure$	−0.1488	−0.1315	−0.1294	−0.1206
	(0.1546)	(0.1303)	(0.1027)	(0.0893)
$center\times Robot\_exposure$	0.4031 ***	0.4258 ***	0.2749 ***	0.2903 ***
	(0.0320)	(0.0389)	(0.0216)	(0.0277)
Cons	0.0913	−1.358	0.0863	−0.8474
	(0.1713)	(1.3879)	(0.1155)	(0.9098)
N	208	208	208	208
R-squared	0.6753	0.7179	0.6307	0.6661
Panel B: Yangtze: River Delta City Cluster
	(1)	(2)	(3)	(4)
Variables	Knowledge Aggregation Capacity		Knowledge Diffusion Capacity
$Robot\_exposure$	0.4158	0.3060	0.4333	0.2613
	(0.4869)	(0.3845)	(0.4173)	(0.3670)
$center\times Robot\_exposure$	0.5460 ***	0.4714 ***	0.7114 ***	0.6226 ***
	(0.1147)	(0.1084)	(0.1549)	(0.1257)
Cons	−0.4677	15.2726 **	−0.5124	16.9859 **
	(0.5131)	(7.4456)	(0.4949)	(7.9153)
N	351	351	351	351
R-squared	0.7934	0.8366	0.6862	0.7368
Panel C: Pearl River Delta City Cluster
	(1)	(2)	(3)	(4)
Variables	Knowledge Aggregation Capacity		Knowledge Diffusion Capacity
$Robot\_exposure$	3.3195	2.0992	1.8435	1.0561
	(1.8795)	(2.1233)	(1.5032)	(1.9359)
$center\times Robot\_exposure$	0.6963 **	0.7503 *	1.1505 ***	1.2749 ***
	(0.2462)	(0.3115)	(0.1281)	(0.1508)
Cons	−2.8548	−0.7532	−1.7920	−5.1455
	(1.5050)	(17.9604)	(1.3149)	(15.0601)
N	91	91	91	91
R-squared	0.9017	0.9257	0.8787	0.9139
Panel D: Beijing–Tianjin–Hebei City Cluster
	(1)	(2)	(3)	(4)
Variables	Knowledge Aggregation Capacity		Knowledge Diffusion Capacity
$Robot\_exposure$	−0.0599	0.0343	−0.0248	−0.00398
	(0.176)	(0.127)	(0.0344)	(0.0617)
$center\times Robot\_exposure$	1.1936 ***	1.2233 ***	1.2995 ***	1.3579 ***
	(0.0892)	(0.1152)	(0.0697)	(0.0449)
Cons	−0.0167	−7.0872 *	−0.0441	−3.7794 *
	(0.1436)	(3.4655)	(0.0755)	(1.9400)
N	182	182	182	182
R-squared	0.9216	0.9314	0.9697	0.9715
Controls	NO	YES	NO	YES
Fixed time	YES	YES	YES	YES
Fixed city	YES	YES	YES	YES

Note: *, **, and *** denote significance levels of 10%, 5%, and 1%, respectively; clustered robust standard errors at the city level are presented in parentheses.

, the coefficients of the interaction terms for each city cluster are significantly positive at the 1% level, indicating that the enhancement of knowledge aggregation and diffusion capabilities by industrial robot adoption is more significant in the central cities. It is worth noting that there is significant heterogeneity in the direction and degree of knowledge flow by industrial robot adoption in central cities among city clusters. While geographic proximity is important, factors such as the level of economic development, industrial structure, policy support, technological complementarities, and regional resilience within clusters are more critical to the facilitation of knowledge flows and technological innovation. These variations may lead to differences in the green knowledge clustering and diffusion effects generated by industrial robotic adoption. For instance, in the regression analysis of the Chengdu–Chongqing economic circle, the interaction term coefficient is relatively smaller than that of other city clusters. This is due to the fewer and lower-tier large- and medium-sized cities within this economic circle. Compared to the Yangtze River Delta, Pearl River Delta, and Beijing–Tianjin–Hebei clusters, the Chengdu–Chongqing economic circle has a weaker foundation for technology integration and green industrial development. Hence, the impact of industrial robots on green knowledge in this region is less significant.

(2)

Human Capital

Regions with higher levels of human capital are usually better able to adapt to new technologies, including industrial robots [67], a factor which helps to enhance human–machine collaboration and transfer positive externalities to the entire field, enhancing green knowledge and practices. The study uses the average educational years return coefficient to represent human capital levels. Cities were divided into high- and low-human-capital groups for regression analysis.

Table 8. Heterogeneity test of human capital.

Variables	(1) Low Group	(2) High Group	(3) Low Group	(4) High Group
	Knowledge Aggregation Capacity		Knowledge Diffusion Capacity
$Robot\_exposure$	0.3529	0.8027 ***	0.2431	0.6014 ***
	(0.2293)	(0.2101)	(0.1822)	(0.1780)
Controls	YES	YES	YES	YES
Fixed time	YES	YES	YES	YES
Fixed city	YES	YES	YES	YES
Cons	3.7364	4.4336	4.0046	6.0595 *
	(2.9786)	(3.2864)	(3.0260)	(3.3966)
N	1774	1775	1774	1775
R-squared	0.1972	0.3954	0.1146	0.2703

Note: *, **, and *** denote significance levels of 10%, 5%, and 1%, respectively; clustered robust standard errors at the provincial level are presented in parentheses.

shows that, in high-human-capital cities, the coefficient for industrial robot penetration is significantly positive at the 1% level, suggesting that the higher the level of human capital in a city, the more pronounced the increase in knowledge agglomeration capacity and knowledge diffusion capacity due to industrial robot adoption.

(3)

Material Aspect: High-Speed Rail Operation

Geographical distance affects the flow of knowledge. One important advancement in transport infrastructure that has a big influence on innovation and economic growth is high-speed rail. This study explores the variability in the geographical spillover impact of industrial robots on GTFP with a focus on high-speed rail operations. For regression analysis, cities were categorized based on how well their high-speed rail operates. According to

Table 9. Heterogeneity testing for high-speed rail opening.

Variables	(1) Cities Without High-Speed Rail	(2) Cities with High-Speed Rail	(3) Cities Without High-Speed Rail	(4) Cities with High-Speed Rail
	Knowledge Aggregation Capacity		Knowledge Diffusion Capacity
$Robot\_exposure$	0.1789	0.6384 ***	0.0588	0.4636 ***
	(0.1216)	(0.1803)	(0.0518)	(0.1467)
Controls	YES	YES	YES	YES
Fixed time	YES	YES	YES	YES
Fixed city	YES	YES	YES	YES
Cons	−0.0915	5.8677 **	−0.0668	6.2788 **
	(0.7183)	(2.6717)	(0.2509)	(2.5770)
N	832	2717	832	2717
R-squared	0.1530	0.3318	0.0870	0.2177

Note: *, **, and *** denote significance levels of 10%, 5%, and 1%, respectively; clustered robust standard errors at the provincial level are presented in parentheses.

, the predicted coefficients of industrial robot penetration in columns (1) and (3) are not significant, whereas those in columns (2) and (4) are considerably positive at the 1% level. This indicates that industrial robot adoption has a greater influence on improving knowledge agglomeration and diffusion capacities in high-speed rail-enabled cities than in non-high-speed rail cities.

(4)

Industrial Policy Support

Emerging industries are guided and supported, in large part, by industrial policies. Industrial robots are one of the most important scenarios for the application of AI technology. Therefore, this paper selects the strength of AI industrial policy support as a moderating variable. This paper takes the various types of industrial support in the government’s five-year planning report as the research object and uses the scoring method to give the corresponding score to measure the strength of AI industrial policy support. Based on the official government websites of provinces and municipalities directly under the central government, as well as the 11th Five-Year Plan, 12th Five-Year Plan, and 13th Five-Year Plan artificially collected from websites such as Beida Fabulous, this paper categorizes, labels, and artificially assigns values to AI-related industries which specifically include the name of the subindustry, the attitude of the policy (categorized into neutral and encouraging, assigning 1 and 2 points, respectively), and the key support (categorized into not having and having support, assigning 1 and 2 points, respectively).

We conducted a regression analysis on the interaction between the Intensity of Spending Policy ($ISP$) in the artificial intelligence industry and the core explanatory variables, as shown in

Table 10. Heterogeneity test of industrial policy support.

Variables	(1)	(2)	(3)	(4)
	Knowledge Aggregation Capacity		Knowledge Diffusion Capacity
$Robot\_exposure$	0.2103 **	0.1759 **	0.1457 *	0.1161 *
	(0.0877)	(0.0800)	(0.0772)	(0.0687)
$ISP\times Robot\_exposure$	0.0203 ***	0.0204 ***	0.0142 ***	0.0143 ***
	(0.0035)	(0.0033)	(0.0033)	(0.0032)
Controls	NO	YES	NO	YES
Fixed time	YES	YES	YES	YES
Fixed city	YES	YES	YES	YES
Cons	−0.2895 ***	5.2641 ***	−0.2011 ***	5.3823 ***
	(0.0764)	(1.5596)	(0.0729)	(1.6224)
N	3549	3549	3549	3549
R-squared	0.3418	0.3809	0.2034	0.2521

Note: *, **, and *** denote significance levels of 10%, 5%, and 1%, respectively; clustered robust standard errors at the city level are presented in parentheses.

. When the dependent variables are knowledge agglomeration capability and knowledge diffusion capability, the estimated coefficients of the interaction terms are all significantly positive at the 1% level. This suggests that, as the support for spending policies in the artificial intelligence industry strengthens, the impact of industrial robots on the improvement of knowledge agglomeration and knowledge spillover capabilities becomes more pronounced.

6. Discussions

(1)

About The Spatial Spillover Effect of Industrial Robot Adoption on GTFP and The Mechanism of Knowledge Flow

Spatial spillover is an important channel for industrial robot adoption to promote GTFP in neighboring places. The industrial robot adoption not only enhances local GTFP, but also has a positive spatial spillover effect, which significantly enhances GTFP in the associated places. This study first validates the positive impact of industrial robot adoption on green development. Although some scholars have expressed concerns about the potential increase in energy consumption due to the intensive development of industrial robots in the short term [32], in the long run, the maturation of industrial robot technology, improvements in energy efficiency, and transformation of energy structures will effectively counteract the rebound effect, facilitating a green transformation in societal production [17]. This research also provides a comprehensive assessment of the green effects of industrial robot adoption from a city-level perspective.

Moreover, these findings suggest that the digital transformation of the manufacturing industry driven by industrial robot adoption is not only beneficial to a single firm or city but may also promote the optimization of the spatial pattern of innovation and the synergistic development of regions [32]. The adoption of industrial robots enhances intra-regional knowledge aggregation and frequent interactive exchanges. Furthermore, robots facilitate heterogeneous knowledge flow across regions, breaking down spatial boundaries and overcoming distance friction, particularly with the application of big data and cloud computing which further diminish the distance, resulting in a stronger diffusion effect of knowledge [68]. Knowledge flow, particularly the flow of green technology and practices, plays a key supportive role in promoting innovation levels, improving the efficiency and sustainability of production processes, reducing waste, lowering energy consumption, and minimizing environmental impacts [69]. This contributes to improving GTFP, ultimately.

Overall, industrial robot adoption is conducive to both knowledge agglomeration and knowledge diffusion in cities where manufacturing firms are located. The former can be interpreted as a buzzing knowledge flow pattern, i.e., the industrial robot adoption is conducive to strengthening knowledge aggregation and frequent interactions within the region. The latter can be understood as a pipeline knowledge flow mode, that is, the industrial robot adoption helps realize the heterogeneity of knowledge flow between different regions, strengthening the knowledge diffusion effect. The enhancement of knowledge agglomeration via industrial robots arises from three key sources. (1) They facilitate advanced R&D, including eco-friendly tech and processes [70]. (2) Robot adoption attracts highly skilled personnel, making robot-advanced regions appealing to green-tech professionals [71]. (3) Robotics integration fosters cross-industry collaboration, aiding holistic green innovation. As for knowledge diffusion, industrial robots advance data and knowledge integration, enhancing sharing within and across organizations [72]. Additionally, firms adopting green production through robots can serve as models for others.

(2)

About The Heterogeneity of The Mechanisms

The analysis also elucidates the diverse mechanisms by which industrial robot deployment facilitates spatial spillover effects on GTFP. Examining urban locational attributes reveals notable disparities in how industrial robot infiltration impacts knowledge flows within central and peripheral cities in urban clusters. For central cities, these are usually the economic engines of the region or cluster, typically possessing more mature industrial systems, advanced infrastructure [73], and robust technological foundations, all of which are conducive to the integration and utilization of advanced technologies like industrial robots. These cities, being hubs of human capital and expertise, attract a wealth of technical professionals, researchers, and high-tech enterprises [74]. This confluence of skilled talent and the structure of the innovation ecosystem plays an instrumental role in the adoption of green technologies and in facilitating the dissemination and proliferation of green knowledge [75]. Conversely, peripheral cities frequently encounter challenges in attracting a similarly diverse and skilled workforce. They are typically characterized by lower levels of market capital and policy support, factors that can hinder knowledge exchange and attenuate the green knowledge aggregation and spillover effects engendered by industrial robot integration. In addition, there are equally significant differences between city clusters, with different levels of development in terms of knowledge pooling, which, in turn, provides new positioning ideas for different city clusters to promote industrial robotics adoption.

Additionally, human capital enhances a city’s capability to absorb technology effectively. This is primarily due to the high human capital level reflecting educational training, diverse perspectives, and adaptability to new skills and environments [76]. Cities with high human capital levels tend to be more proactive in dealing with future demands and embracing disruptive technologies, such as robots. They are more likely to see industrial robots as useful tools that can help unleash human potential rather than as threats to employment. This leads to broader adoption and better utilization of these technologies, thereby enhancing productivity, innovative capabilities, knowledge aggregation, and the spillover effect.

The progress of high-speed railways enhances interregional connectivity and mobility, making the area more attractive to professional technical personnel, researchers, and investors [77]. When the high-speed railway connects the main central cities with surrounding cities, the knowledge and technology of the central cities, including innovations and efficiency improvements brought about by the application of industrial robots, will spread faster and wider, forming a significant knowledge spillover effect. In addition, this connectivity not only narrows spatial distance, but also contributes to market integration [78]. These market-active elements can enhance the green impact of knowledge and technological enhancement brought about by the application of industrial robots.

Industrial policy can effectively enhance enterprises’ propensity to adopt new technologies by offering financial assistance, tax benefits, R&D subsidies, and various other incentives which can mitigate the risks associated with technological adoption, promoting the exchange of and cooperation between knowledge and technology within the industry [79,80]. These initiatives will greatly facilitate the implementation and diffusion of advanced technologies, such as industrial robots, thus contributing positively to the GTFP of cities and regions.

7. Conclusions

7.1. Findings

Using urban panel data from Chinese cities spanning 2007 to 2019, this study discusses the spatial spillover effects of industrial robot adoption on green total factor productivity from the perspective of knowledge flows. It also explores the mechanisms of this impact through the dimensions of knowledge agglomeration and diffusion capabilities. The main findings are as follows.

(1) Green total factor productivity among cities exhibits a degree of spatial autocorrelation. According to the spatial Durbin model regression results, the adoption of industrial robots significantly enhances the GTFP in spatially associated areas, demonstrating a technology dividend phenomenon. (2) The use of industrial robots positively influences urban knowledge agglomeration capabilities, creating a “buzz” in the region, and simultaneously strengthens knowledge diffusion capabilities, facilitating the intercity flow of labor, capital, and technology, thereby enabling knowledge to spread across regions through a “pipeline” effect. (3) The effect of industrial robot adoption on knowledge agglomeration and diffusion capabilities is moderated by the city’s location. Central cities, with stronger knowledge flow linkages, often act as gravitational and radiating centers for knowledge, talent, and technology within urban clusters. There are differences in the state of knowledge flow between central and peripheral cities across different urban clusters. Central cities in each urban cluster have a more significant effect on enhancing knowledge agglomeration and diffusion capabilities compared to peripheral cities. Furthermore, the influence on knowledge diffusion capabilities varies across regions like the Chengdu–Chongqing economic circle, the Yangtze River Delta city cluster, the Pearl River Delta city cluster, and the Beijing–Tianjin–Hebei city cluster. (4) The impact of industrial robot adoption on knowledge agglomeration and diffusion capabilities is also moderated by urban resources. Areas with higher human capital levels see a more pronounced effect of industrial robot adoption on enhancing knowledge agglomeration and diffusion capabilities. The inauguration of high-speed rail services benefits the enhancement of these capabilities, and regions with stronger policy support for the AI industry witness a more significant enhancement due to industrial robot adoption.

7.2. Implications

This study advances theoretical debates in environmental economics, spatial dynamics, and innovation ecosystems. By integrating knowledge flow theory with spatial spillover analysis, this study demonstrates that industrial robot adoption realizes both economic and environmental benefits not only through increased efficiency but also through knowledge-driven aggregation and diffusion. This bridges the gap between automation studies and sustainability research, offering a unified framework to interpret GTFP. In addition, the empirical validation enriches spatial economics by revealing that technology-driven knowledge externalities transcend geographic and economic boundaries through spatial interactions.

The findings help companies adjust to changes in technology and direct policies that promote green growth. The exchange of knowledge, technology, and labor significantly boosts regional development. Firms should collaborate with research institutions and universities to share experiences in industrial robot adoption through joint R&D centers and technical forums. Open innovation platforms leveraging big data and the Internet can accelerate technological upgrades and market expansions while driving supply chain transformation. Governments should optimize the spatial layout of industrial robot adoption. The government should encourage central cities to focus on establishing innovation centers for industrial robots and green technologies, while peripheral areas should develop strategic partnerships to strengthen economic linkages. Industrial policy should adopt a tiered strategy. For example, central cities promote cutting-edge robotics research and development through green patent rebates, while peripheral cities subsidize small- and medium-sized enterprises that adopt energy-efficient robots. Governments must also remove barriers to the mobility of innovation factors to enhance regional agglomeration and diffusion, and educational reforms must consider vocational training to cultivate robotics-literate workforces, ensuring inclusive transitions in labor markets disrupted by automation.

7.3. Limitations

Despite these insights, this study primarily relies on relatively macroscopic indicators, such as in-degree and out-degree, for depicting knowledge flows. Although this approach outlines the general trend of knowledge flows, it fails to capture more subtle and complex phenomena, such as tacit knowledge flows. For instance, the knowledge transmission through the movement of senior corporate personnel and skilled workers, both of which are forms of implicit and informal knowledge exchange and flow, are not sufficiently addressed in the current research. Future research will focus more on details, aiming to introduce a richer and more refined set of data resources, such as green patent citation data, scientific collaboration data, paper citation data, and information on “three meetings and one layer” team members. Moreover, the characterization of knowledge flows in this study is based on patent data, which, although constituting an important manifestation of knowledge, do not fully explore the specific transmission mechanisms of knowledge flows. These mechanisms include pathways such as industrial chains, supply chains, and talent mobility, none of which are analyzed in detail due to the focus of this study and the lack of specialized data on talent mobility in the industrial robot sector. This paper combines spatial econometric and standard panel regression models. Spatial econometric models are highly promising for analyzing knowledge flows. Future research will expand in this direction to provide a more comprehensive understanding. This approach will allow for a more accurate portrayal and analysis of the complete picture of knowledge flows, providing deeper insights into the effect of industrial robot adoption on urban GTFP and knowledge flow.