Progress in Construction Robot Path-Planning Algorithms: Review

Abstract

Construction robots are increasingly becoming a significant force in the digital transformation and intelligent upgrading of the construction industry. Path planning is crucial for the advancement of building robot technology. Based on the understanding of construction site information, this paper categorizes path-planning algorithms into two types: global path-planning and local path-planning. Local path planning is further divided into classical algorithms, intelligent algorithms, and reinforcement learning algorithms. Using this classification framework, this paper summarizes the latest research developments in path-planning algorithms, analyzes the advantages and disadvantages of various algorithms, introduces several optimization strategies, and presents the results of these optimizations. Furthermore, common environmental modeling methods, path quality evaluation criteria, commonly used sensors for robots, and the future development of path-planning technologies in swarm-based construction robots are also discussed. Finally, this paper explores future development trends in the field. The aim is to provide references for related research, enhance the path-planning capabilities of construction robots, and promote the intelligent development of the construction industry.

1. Introduction

The data indicate that the aging population in China significantly impacts the age distribution of construction workers. As workers grow older, many find themselves unable to perform high-intensity physical tasks. This prevalent aging trend in the workforce results in a depletion of human resources and intensifies labor shortages within the industry [1]. In response to the challenges faced by traditional construction practices and to align with contemporary trends, the Ministry of Housing and Urban–Rural Development (Beijing, China), along with other relevant agencies, has advocated for the enhanced utilization of construction robots. This initiative aims to foster the intelligent advancement of the construction sector and facilitate its transformation and upgrading [2]. Currently, robotic technologies find extensive applications across various engineering disciplines, including coal mining [3,4,5], agriculture [6,7], healthcare [8], military [9], and autonomous driving [10]. With the rise of intelligent construction technologies, the integration of construction robots has emerged as a promising avenue for the sector’s evolution. Intelligent construction seeks to achieve automation, intelligence, and digitization in production and construction processes, necessitating the use of integrated intelligent equipment such as robots. In recent years, mechanized and automated construction models have been progressively adopted in the industry, leading to the development and deployment of specialized construction robots. These robots are defined as machines capable of executing construction tasks either autonomously or semi-autonomously, following pre-defined programming. By addressing labor shortages, construction robots not only enhance operational efficiency but also improve the quality of work within the construction industry [11,12]. Compared to industrial robots, the history of construction robots is relatively short. The development of construction robots originated in Japan in the 1970s. In 1982, a fire-resistant material spraying robot (Shimizu Corporation, Tokyo, Japan) developed by the Japanese Shimizu Corporation (Tokyo, Japan) was successfully used on a construction site, marking the world’s first construction robot to be successfully applied. From the 1980s to the 1990s, with the development of robotic technology and increased investment from the United States, Europe, and others, the types of construction robots began to diversify and evolve towards automation and intelligence. In the 21st century, the scope of work for construction robots has expanded, gradually moving from traditional construction and construction fields towards demolition, building maintenance, and other areas. After 2010, the development and application of technologies such as artificial intelligence, machine vision, and 3D printing significantly advanced the field of construction robotics. This progress led to the emergence of new types of construction robots capable of autonomous navigation, decision-making, and precise construction tasks—for example, the exterior wall cleaning robot called GEKKO (GEKKO GmbH, Ostfildern, Germany) by the (Ostfildern, Germany). The development and application of these construction robots have made it possible for humans to be relieved from hard and dangerous construction work. Figure 1 shows the development process of construction robots.

Construction robots are divided into two categories: construction robots in the field of civil engineering (CRCE) and construction robots in the field of building construction (CRBE). The primary difference between the two lies in their construction scenarios. CRCE is mainly used for infrastructure projects, such as roads, bridges, tunnels, and dams. In contrast, CRBE is applied to housing construction, including residential buildings, office buildings, and commercial complexes.

CRCE typically handles large-scale operations, such as earthmoving, foundation treatment, concrete pouring, and steel structure installation. These robots require powerful performance and stability to operate in complex terrains and harsh environments. However, the degree of automation in CRCE is relatively lower compared to CRBE, often relying more on manual control. Common examples of CRCE include excavation robots, concrete spraying robots, and tunnel boring machines.

On the other hand, CRBE focuses on indoor precision tasks, such as bricklaying, wall treatment, decoration, and pipe installation. These robots demand high precision and flexibility to adapt to complex building structures and diverse construction needs. As a result, CRBE often achieves fully automated operations. Typical examples include bricklaying robots, wall spraying robots, and 3D printing construction robots.

Path planning is crucial for advancing autonomous navigation technology in construction robots and has become a significant topic in academic research. When constructing roads, bridges, and other large structures, CRCE robots often operate in environments with narrow roads, steep slopes, and many curves. Additionally, many remote areas have weak or no mobile communication signals, although the on-site environment remains relatively stable. Therefore, these robots can rely on offline global path-planning algorithms, such as A* and Dijkstra, for autonomous navigation.

In contrast, CRBE robots working in residential and office buildings face different room layouts, significant structural variations, and the presence of moving workers. These factors make it challenging to obtain a stable and reliable map in advance. However, since these work sites are usually located near urban areas, mobile communication resources are readily available. As a result, CRBE robots can use sensors to gather real-time environmental information and plan their next moves using local path-planning algorithms.

Local path-planning algorithms are further divided into classical and intelligent algorithms based on their principles. Classical algorithms, such as the Artificial Potential Field (APF) and Dynamic Window Approach (DWA), use mathematical models or predefined rules for path planning. These algorithms are suitable for helping construction robots complete simple path-planning tasks.

Intelligent algorithms, compared to classical algorithms, integrate technologies related to artificial intelligence and machine learning. These algorithms offer superior learning and adaptability, making them suitable for high-dimensional and complex environments. However, they are more complex to implement and require significant computational resources. Common examples include Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO).

Alternatively, Deep Reinforcement Learning (DRL) techniques can be employed, enabling robots to learn and adapt to their surroundings autonomously without the need for a pre-defined map. These algorithms, however, demand a large amount of raw data and substantial computational power. Examples of DRL algorithms include Q-learning and Monte Carlo Methods.

This article categorizes the surveyed path-planning algorithms, as illustrated in Figure 2, based on the aforementioned rationale. It reviews commonly used algorithms in the field of path planning from recent years, analyzes their advantages and disadvantages, and summarizes improvement strategies. Additionally, it provides insights into future trends in building robot path-planning. The aim of this study is to offer a comprehensive reference for related research, enhance the path-planning capabilities of construction robots, and promote the intelligent development of the construction industry.

The data sources for this article include the library database of Beijing University of Civil Engineering and Architecture and Google Scholar. The search keywords were: “algorithm name”, “robot path-planning”, and “specific improvement strategies” (e.g., path quality, obstacle avoidance), with a publication year range of 2018–2024. The primary goal of construction robot path planning is to ensure safety, encompassing both the robot itself and personnel, followed by improving work efficiency to reduce costs.

The inclusion criteria for this study are as follows: Articles published between 2018 and 2024 that focus on improving mobile robot path-planning algorithms in terms of search speed, path smoothness, and obstacle avoidance. Additionally, articles addressing inherent defects of the algorithms were included, with preference given to those incorporating mathematical foundations and experimental results. Comparative studies were prioritized during the selection process.

The screening process involved two main steps. First, relevant research papers were collected through keyword searches. Second, the articles were screened based on the aforementioned criteria, prioritizing those with higher impact factors or those utilizing a combination of multiple algorithms. The final selection was based on full-text reading to ensure relevance and quality.

The structure of this article is as follows: Section 2 introduces common methods for environmental modeling and path quality evaluation criteria. Section 3 presents global path-planning algorithms, and Section 4 introduces local path-planning algorithms. Section 5 summarizes the algorithms mentioned above in a table, Section 6 discusses the future development of path-planning technology in swarm construction robots, and Section 7 is the conclusion.

2. Environmental Modeling and Path Quality Assessment Standards

Mobile construction robots typically follow these steps during autonomous navigation:

• Environmental Modeling:

For CRBE (Construction Robots in the field of Building Construction), environmental modeling is usually based on the Building Information Model (BIM) of the construction robot [13,14,15]. In contrast, CRCE (Construction Robots in the field of Civil Engineering) builds models based on known maps and various sensors to obtain environmental information.

2.

Decision Making:

After acquiring environmental information, the navigation system of the construction robot initiates the decision-making process to determine the optimal movement path. This process typically employs path-planning algorithms to analyze the environmental model and the robot’s state, thereby generating optimal action decisions.

3.

Motion Control:

The control module of the construction robot converts the generated decisions into motion instructions, enabling the robot to move along the planned path.

2.1. Environmental Modeling

The core of environmental modeling lies in the systematic organization and application of the known environment. A well-designed environmental model can significantly aid mobile robots in pathfinding. Typically, there are four main methods for constructing environmental models: geometric models (GM), topological maps (TM), grid cell models (GCM), and metric representations (MR) [16,17,18,19,20].

2.1.1. Grid Method (GM)

In 1968, Hodan proposed the concept of the Grid Map (GM), which divides the operational space of mobile robots into uniformly sized network cells that represent binary information. When the grid is devoid of obstacles, it is classified as a free grid, permitting unrestricted movement for the mobile robot. In contrast, the presence of obstacles designates it as an obstacle grid. If the obstacles do not completely fill the grid area, an expansion technique is applied to ensure that they occupy the entire grid space. The dimensions of each grid cell are determined by the actual size of the construction robot.

2.1.2. Topological Methods (TM)

The TM environmental modeling method uses nodes to represent specific locations and edges to show connections between these points. This method disregards the precise shape of the environment, instead using points and lines to depict important locations in space and their interconnections. By focusing on the connectivity of paths, this approach is particularly suitable for complex, large-scale environments, such as mountainous terrain [21].

Voronoi diagrams are commonly used to depict this type of environmental model. A Voronoi diagram divides the plane into different regions, each of which is called a Voronoi cell or Voronoi polygon. Each Voronoi cell is associated with a particular point, known as a seed point or generator. Any point within the cell is closer to its corresponding seed point than to any other seed point. See Figure 3.

2.1.3. Geometric Characteristic Method (GCM)

GCM is a technique that enables robots to derive abstract geometric features from the observed environment, including elements such as lines, arcs, and circles, to characterize the surroundings. This methodology results in a relatively compact representation, facilitating the process of position estimation. Nonetheless, the challenge of reliably extracting stable features from raw data is considerable, which limits the applicability of this approach primarily to structured operational environments.

2.1.4. Mixed Representation (MR)

The earliest hybrid representation was introduced by Thrun as a technique for extracting topological features from a comprehensive metric map with the aid of sonar sensors, facilitating the acquisition of feature data during robot navigation. A new node is created when the sonar-derived feature data significantly differs from those of established nodes. Conversely, if the feature data closely match that of known nodes, probabilistic localization is executed through Markov decision processes, concurrently generating a topological representation of the environment. Nevertheless, this approach encounters challenges in creating topological maps in areas where the obstacle feature data are ambiguous. To address this, Yeap and Jefferies put forth a method for deriving a global topological map from local metric maps. In this methodology, the local map is depicted using a grid and a collection of edges that resemble a topology that connects various existing local maps. A dual-level planning strategy, incorporating both local and global planning, is utilized for path planning.

2.2. Path Quality Assessment Criteria

In order to objectively assess the effectiveness of various search strategies and quantify their performance, this paper presents a concise overview of seven evaluation criteria: average path-planning time (Ta), relative standard deviation of average path-planning time (Rt), average path length (Dm), relative standard deviation of path length (Rd), path curvature (Rc), estimated path tracking time (Et), and success rate (Sr) [16], as shown in Figure 4.

3. Global Path-Planning Algorithms

3.1. A* Algorithm

The A* algorithm is a map-based search method widely used in various robots. In the construction field, it can be applied to construction robots working in places with weak or even no mobile communication signals, such as material transport robots and earthmoving robots working in mining areas. The algorithm was initially proposed and elaborated by researchers Peter Hart, Nils Nilsson, and Bertrand Raphael in 1968 [22]. The core concept of the algorithm is its heuristic evaluation strategy, where the heuristic function is defined as $f\left(n\right)=h\left(n\right)+g\left(n\right)$, $g\left(n\right)$ represents the actual cost incurred from the start node to the current node, and $h(n)$ denotes the heuristic estimated cost from the current node to the goal node. The total estimated cost from the start node to the goal node is represented by $f\left(n\right)$. By evaluating the potential cost of each node on the path, the search direction is guided to reach the goal node with the shortest path or optimal cost. The estimated cost is commonly calculated using Euclidean distance or Manhattan distance. This article uses Manhattan distance as an example, and Figure 5 illustrates its working principle.

However, the A* algorithm faces limitations such as long computation times, low efficiency, and rough path planning. To address these challenges, researchers have proposed various improvements.

Wang F. adopted the Octile Distance formula, replacing traditional Manhattan or Euclidean distances, which better adapts to complex paths [23]. Huang H. utilized the Jump Point Search method to identify and skip unnecessary nodes, reducing the number of nodes expanded during the search process and improving the algorithm’s efficiency [24].

Xiang D. combined the A* algorithm with a greedy algorithm, achieving a multi-objective path-planning strategy. This approach also enables the algorithm to find multiple high-quality paths that meet the problem requirements in complex environments [25].

Wang H. proposed the EBS-A* algorithm by incorporating extended distance, bidirectional search, and path smoothing techniques, significantly improving search speed [26]. Fu B. integrated the A* algorithm with local path-planning technology, enabling dynamic path adjustment and obstacle avoidance, thereby enhancing the practicality and safety of path-planning [27].

3.2. Dijkstra

The Dijkstra algorithm is a graph-based data structure algorithm primarily used to solve the shortest path problem in maps. It was proposed by Dutch computer scientist Edsger W. Dijkstra in 1959. The algorithm works by progressively expanding the node with the known shortest path. Through this process, it finds the shortest path from the starting node to all other nodes. In large-scale construction, the Dijkstra algorithm can assist construction robots in obtaining the shortest path from one construction point to another. For example, it can help mobile concrete robots quickly distribute concrete to various target points during large-scale construction projects.

The Dijkstra algorithm often faces challenges in dynamic environments, primarily due to low efficiency, high computational resource consumption, and the inability to handle negative weight edges. To address these issues, experts have proposed numerous improvements.

Alshammrei S designed an optimal collision-free offline Dijkstra algorithm. In this approach, robots follow the algorithm to generate the shortest path from the starting point to the destination. At each node, the robot stops to test if there is an obstacle between the current node and the next node. If an obstacle is detected, the robot excludes the corresponding node and updates its directed graph. This process repeats until the target point is reached. The improved Dijkstra algorithm can effectively solve the optimal path-planning problem in environments with obstacles [28].

Bing H proposed a method to reduce computational effort by considering the distance to the destination while calculating the shortest distance from a node to the starting point. This causes the exploration process to generally progress towards the destination. By comparing the computational complexity of traditional algorithms, the effectiveness of the improved method was proven, reducing the complexity of the Dijkstra algorithm [29].

Li X solved the problem of planning a collision-free and approximately optimal path for mobile robots in large-scale and densely obstacle-filled environments. This was achieved by combining the Dijkstra algorithm with four-neighborhood search, eight-neighborhood search, and reverse search algorithms [30].

Szczepanski R proposed a hybrid path-planning method based on the artificial bee colony algorithm and the Dijkstra algorithm. This method combines the advantages of offline global path optimization and online local path planning, successfully reducing the time for robot path planning while ensuring path quality [31].

Sundarraj S introduced an inertia weight to control the global and local search capabilities of the particle swarm, optimizing the path-planning problem. This method significantly enhanced various capabilities of the Dijkstra algorithm, such as obstacle avoidance, path-planning speed, and computation time [32].

A system combining the Dijkstra algorithm and the Dynamic Window Approach (DWA) was proposed, enabling intelligent vehicle robots to navigate autonomously and avoid obstacles in unknown indoor environments [33].

An improved Dijkstra algorithm was proposed for the path-planning of track-based logistics robots. This algorithm introduced a real-time node occupancy factor into the path weight function and established a time window model within the algorithm to improve computational efficiency. Additionally, the improved algorithm can solve the problem of robot conflicts in multi-robot systems [34].

3.3. RRT Algorithm

Rapidly Exploring Random Tree (RRT) is a sampling-based global algorithm with powerful global search capabilities. It explores space through random sampling and gradually builds a tree to find a path from the start to the goal. This method is particularly suitable for dealing with non-convex and high-dimensional search spaces. For example, RRT can help construction robots with robotic arms or humanoid robots quickly find a path to a target point.

The algorithm starts by initializing a tree that includes the start point. It then randomly generates a target point within the search space. Next, it finds the node in the tree that is closest to the target point and attempts to connect it to the target point, thereby extending the tree towards this point. During the extension process, the algorithm checks whether the path from the current node to the target point intersects with obstacles. If an intersection is detected, resampling is performed. If the connection is successful, the target point is added to the tree. This cycle repeats until the goal is reached. See Figure 6.

The RRT algorithm is characterized by its ability to quickly find a feasible path but not necessarily the optimal path [35]. Xin P and Zhang W introduced a heuristic function to guide the selection of sampling points, reducing invalid sampling and thus improving search efficiency [36,37].

Connell D implemented a two-way RRT method, where the algorithm draws samples from both the starting and ending points simultaneously, aiming to link the two trees together in the center. This bidirectional extension significantly improves search coverage and speed, especially in high-dimensional spaces [38].

Qi J developed a modified version of the Rapidly Exploring Random Tree Star (RRT*) algorithm to establish an initial path and create a state tree structure. Furthermore, he introduced a rapid optimization technique for this initial path. This modified algorithm stands out from other static planning methods due to its ability to produce higher-quality initial paths, enhancing the effectiveness of path planning in various applications [39].

In a parallel effort, Jun G proposed a data-driven approach known as UPPHE, which incorporates the Feedback Rapidly Exploring Random Tree Star algorithm (FRRT*). This innovative method features a risk network with feedback modules, allowing it to leverage information derived from case data. By utilizing this data, FRRT* effectively constrains the expansion of the random tree and makes adjustments skewed towards optimizing planning efficiency, significantly enhancing the traditional RRT* approach [40].

Wang H recognized the unpredictable nature of tree growth in the original RRT* algorithm and sought to improve its performance by integrating the Attractive Potential Field (APF) method with RRT*. This combination refines the targeting of the sampling process, making it more focused. Additionally, in scenarios where RRT* encounters narrow channels or dense environments, Wang H’s approach employs adaptive step size adjustments facilitated by fuzzy control mechanisms, enhancing search efficiency in challenging areas [41].

Qie T introduced a refined sampling and tree growth mechanism that concentrates on the forward area near the target point during the sampling and growth phases. By restricting the sampling and tree development to this region, the method significantly accelerates the path-planning process, offering a more efficient solution to navigation problems in complex environments [42].

Chen J introduced an enhanced RRT-Connect algorithm for mobile robot path-planning, generating a third node within the space to allow for the algorithm to greedily expand with a quadtree [43]. Cui Xi proposed a sparse sampling mechanism to improve global search efficiency by reducing redundant sampling. Additionally, Cui proposed a dynamic target bias strategy that provides directional guidance for the expansion of the random tree [44].

Liao B optimized the cost of each sampling point using the triangle inequality, which outperforms normal algorithms in terms of both the quality of the initial solution and the speed of convergence [45]. Zhang W introduced a memory strategy to avoid over-sampling or getting stuck in local minima during the expansion of the random tree in non-convex environments [37]. Li Y combined the advantages of P-RRT* and Quick-RRT*, ensuring the rapid capture of optimal solutions and the formation of more effective initial solutions [46].

4. Local Path-Planning Algorithms

Local path planning does not use pre-prepared map data; instead, robots typically use various sensors to perceive the surrounding environment, generate corresponding map data on-site, and plan paths based on this data. Commonly used sensors include Laser Radar Sensors (LRS), Visual Sensors (VS), etc. [16].

4.1. Environmental Monitoring

4.1.1. LiDAR Sensor (LRS)

LRS is a key component of autonomous navigation perception systems and is widely applied in LRS mobile robots and autonomous driving technology. LRS precisely measures distances by emitting and receiving laser pulses, utilizing the principle of light reflection to generate detailed three-dimensional models [47,48,49].

4.1.2. Visual Sensor (VS)

A Visual Sensor (VS) is a passive sensor that collects information by capturing the light reflected from the surface of objects. Compared to a Laser Radar Sensor (LRS), a VS offers a broader detection range and can capture a large amount of image data. Depending on the number and capabilities of VS, they can be classified as monocular vision detection, binocular vision detection, and omnidirectional vision detection [50].

4.1.3. Multi-Sensor Information Fusion (MIF)

MIF enhances the accuracy and reliability of information by integrating data and information collected from multiple sensors [47]. In fields such as autonomous driving and robotics, the application of multi-sensor information fusion algorithms and models is becoming increasingly widespread [51,52].

4.2. Classic Algorithms

4.2.1. Artificial Potential Field Method (APF)

In 1986, Oussama Khatib proposed a path-planning method based on artificial potential fields. The core idea of this method is to consider the target point as an attraction source and obstacles as repulsion sources, thereby creating an artificial potential field similar to an electromagnetic field. As the robot moves within this field, it is simultaneously affected by the combined force of the target’s gravitational attraction and the repulsive force from obstacles. This enables the robot to effectively avoid obstacles and move towards the target point [53,54]. The focus of algorithmic research is on how to use “attraction” and “repulsion” to guide the robot. The APF algorithm is illustrated in Figure 7.

APF is an algorithm widely used in the field of autonomous navigation. In the field of construction robotics, it can help robots such as tile-laying robots, floor treatment robots, and on-site logistics robots to effectively avoid obstacles, ensuring their safety during operation. However, APF has several obvious drawbacks. Firstly, it tends to fall into local optimal solutions. Secondly, when the target point is surrounded by obstacles or the distribution of obstacles is uneven, the algorithm may become stuck. Thirdly, in narrow passages and similar situations, APF may cause path oscillation. Construction robots frequently encounter scenarios with these conditions when working indoors. These issues can significantly impact the efficiency and safety of the robots.

Hanli Y proposed an improved repulsive force function to dynamically adjust the direction of the repulsive force, a method that can reduce the failure of path-planning caused by complex obstacle shapes [55].

Cheng C proposed incorporating the relative velocity approach into the enhanced APF. This involves the real-time monitoring of dynamic obstacles’ motion states, which allows for the adjustment of the robot’s path-planning strategies. This innovation increases the efficiency of obstacle avoidance and strengthens the robot’s adaptability in intricate environments [56].

Wu Z introduced a temperature parameter into the potential field function, which can effectively prevent the robot from being trapped by local minima. This method allows the influence range and direction of the potential field to be dynamically adjusted, thereby improving the flexibility and adaptability of path planning [57].

Wenlin Y addressed the issue of local optima by adding virtual targets and evaluation functions to change the direction and influence range of gravity and resolved the situation of robot deadlock by increasing the gravitational force [58]. Szczepanski R employed an adaptive local path-planning based on the Widrow–Hoff rule, which enables robots to maintain a safe distance from obstacles during movement and achieve smooth and collision-free path-planning by limiting the scaling factor of the repulsive potential field. This adaptive algorithm performs well in dynamic environments [59].

Xing T combined deep learning with APF, using a feedback mechanism to continuously optimize the reward function, thereby finding more efficient paths and enhancing the robot’s adaptability to changing environments [60]. Zhang Y proposed an improved APF (FV-APF) method, specifically designed for corridor environments, which better adapts to the needs of specific scenarios [61].

4.2.2. Dynamic Window Approach (DWA)

The Dynamic Window Approach (DWA) is a widely used technique in mobile robot path planning, which performs well in real-time dynamic environments, especially when dealing with dynamic obstacles. The core idea of the algorithm is that it acquires the current state of the robot, such as direction, speed, acceleration, etc., and generates a dynamic velocity window to predict all possible movement trajectories of the robot at the next moment. Then, it selects the path that best suits the mobile robot [62]. This algorithm has good application in construction scenarios with heavy personnel traffic, such as material handling robots, bricklaying robots, interior decoration robots, etc. However, DWA is prone to local optimal paths and requires significant computational resources during actual operation, especially in dynamic environments.

Gao Y combined DWA with D*Lite, making the algorithm’s performance in local path-planning superior to general algorithms in intricate scenarios [63]. Mai X introduced an innovative density-related evaluation factor within the path evaluation function. This new factor significantly enhances the robot’s ability to perceive and assess the distribution of densely located objects ahead of time. Through experimental studies, it has been demonstrated that this enhancement has led to a notable increase in the efficiency of the improved Dynamic Window Approach (DWA), with an impressive rise of 25%. This advancement suggests that incorporating density awareness into robotic navigation can greatly improve performance in complex environments [64].

Liao T introduced a novel way that mixes the modified A* with an enhanced DWA. This technique involves the incorporation of trajectory estimation function into the evaluation process of the algorithm. Additionally, the resulting paths are further refined and optimized through the application of the B-spline curve method, allowing for smoother and more efficient trajectories [65]. Li Y proposed an algorithm that combines the improved A* algorithm with DWA, and the hybrid algorithm successfully integrates the advantages of both, enabling robots to dynamically avoid obstacles while navigating globally [66].

Zhou R introduced an innovative way that integrates three algorithms. In this enhanced approach, the traditional obstacle distance evaluation function of the dynamic window algorithm was substituted with an artificial repulsive field function, effectively improving the algorithm’s response to nearby obstacles. Additionally, the introduction of an end distance evaluation function further refines the path optimization process. Furthermore, the improved dynamic window method was synergistically combined with a gradient descent-smoothed version of the A* algorithm. This integration effectively tackles the limitations associated with poor global planning found in conventional algorithms, ultimately resulting in paths that are not only safer but also smoother [67].

4.3. Intelligent Algorithm

4.3.1. Genetic Algorithm (GA)

A genetic algorithm (GA) is an optimization technique inspired by the principles of natural biological evolution. This algorithm improves the quality of the population, which is equivalent to the quality of problem solutions, by simulating the biological evolution process. The process includes three key steps: selection, crossover, and mutation. GA eliminates less fit individuals and evolves better ones, ultimately converging to the optimal or near-optimal solution. The evolutionary process of GA is illustrated in Figure 8.

Genetic algorithms possess powerful global search capabilities, effectively exploring the solution space, and are capable of parallel searching. These features make GA algorithms highly efficient in searching. GAs have applications in a variety of problems, such as operations management, multimedia, and more [68]. Genetic algorithms are also extensively used in autonomous robot navigation. For example, in construction sites, multiple material transport robots and structural assembly robots need to cooperate to complete tasks such as hauling and assembly. Genetic algorithms assist in task allocation while helping robots with path planning, enhancing overall efficiency. However, genetic algorithms also have significant drawbacks, such as slow convergence speed, premature convergence, and low precision.

Hu Y proposed an improved genetic algorithm aimed at the problem of mobile robots in unstructured random environments path-planning. This algorithm can accurately detect collisions and reflect the quality of the robot’s path. The algorithm can capture optimal paths in dynamic random environments [69].

Li Y proposed an algorithm mixing the improved genetic algorithm with the dynamic window approach, which enhances the dynamic obstacle avoidance capabilities for robots while quickly generating shorter and smoother paths [70]. Suresh K S introduced an innovative approach for mobile robot pathfinding, which is centered around a Multi-Objective Genetic Algorithm (MRPS-MOGA). This method utilizes several advanced techniques, including tournament selection, cyclic crossover, and adaptive bit-string mutation. These strategies work in conjunction to effectively identify the most optimal path for mobile robots. One of the significant advantages of this algorithm is its enhanced performance in situations where the computational requirements for determining the optimal path are minimal, thus making it particularly efficient in real-time applications [71].

Liu C proposed a multi-objective genetic algorithm based on an elitist strategy (MLRMOEGA) for the path-planning problem of multi-locomotion-mode robots (MLR). The algorithm adopts an elitist strategy to improve the global convergence of genetic algorithms by preserving the best individuals, thereby avoiding premature convergence. To address the premature convergence issue of genetic algorithms, map analysis operators and population diversity expansion operators were proposed. These operators enhance the algorithm’s global search capability. By changing the weights of optimization objectives, the multi-objective optimization performance of the algorithm was tested. The results indicate that the MLRMOEGA algorithm can optimize the value of specific objective functions according to the decision-maker’s intent. Additionally, it generates relatively balanced paths when considering multiple equal optimization objectives [72].

4.3.2. Particle Swarm Optimization (PSO)

The Particle Swarm Optimization (PSO) algorithm was initially proposed by electrical engineer Russell C. Eberhart and social psychologist James Kennedy in 1995. The algorithm is inspired by the social behavior of bird flocking, aiming to simulate how organisms find optimal solutions through group cooperation [72,73].

In 1998, Eberhart and Shi further extended and improved the PSO algorithm, introducing several enhanced versions, such as the inertia weight and constriction factor. These improvements significantly enhanced the algorithm’s convergence speed and search capabilities [74,75].

After entering the 2010s, research on the PSO algorithm deepened further, resulting in many variants and improved versions. These new versions often incorporate other algorithms, such as genetic algorithms, Ant Colony Algorithms, etc., to form hybrid optimization algorithms, improving performance and adaptability [73].

The fundamental idea of PSO originates from nature; when a flock of birds discovers a new feeding area, at the beginning, each bird heads towards places where it believes food might exist and notes the places where it sees the most food during the search process. When the flock comes together, they share the information they have found, allowing them to know where the most food currently is. Each bird then re-searches based on the information it has gathered and the information shared by the team. After several rounds of sharing, the flock can find the place with the most food, as shown in Figure 9.

PSO has strong dynamic environmental adaptability, which makes it widely used in path planning based on high-dimensional maps. In the field of construction, it can help with path-planning for inspection drones, fireproof coating robots, etc. PSO has some drawbacks, including getting stuck in local optimal solutions in multi-dimensional spaces, slow convergence speed in the late stages of optimization, and sensitivity to parameter settings. To address these issues, researchers have proposed many improved Particle Swarm Optimization (PSO) algorithms.

Yang Z proposed an Improved Particle Swarm Optimization algorithm (IPSO) that enhances the algorithm’s exploration ability and search accuracy by improving the velocity update method and inertia weight. It also employs a particle initialization strategy to increase population diversity and introduces strategies to deal with local optima, thereby shortening the convergence path length. Experimental data indicate that IPSO outperforms other algorithms in terms of average path length and shortest path length [76].

Yuan Q addressed the shortcomings of traditional Particle Swarm Optimization (PSO) algorithms in path planning, such as slow convergence speed and a tendency to get stuck in local optima, by proposing an improved Particle Swarm Optimization algorithm based on Differential Evolution (IPSO-IDE). This algorithm enhances the convergence speed of traditional PSO by introducing adaptive adjustments to the weights and acceleration coefficients. It also adds a leader particle to the swarm, selected through a voting mechanism to guide the group towards the optimal position. To overcome the deficiencies of traditional Differential Evolution algorithms, Yuan Q proposed an optimization method for the adaptive scaling factor F and the crossover probability factor CR to improve the convergence performance and iteration accuracy of the algorithm. Combining the improved Particle Swarm Optimization algorithm with Differential Evolution, a new hybrid optimization mechanism was proposed. This mechanism incorporates the concept of “mutual benefit and win-win”, meaning that the two algorithms are in a cooperative relationship, optimizing each other to enhance their performance, thereby improving the overall performance of the swarm [77].

Zheng L introduced an innovative Particle Swarm algorithm that incorporates the principles of artificial potential fields, referred to as the apfrPSO. This advanced method focuses on generating optimal planning paths for robotic systems by fine-tuning the inertia weight parameter. Additionally, it employs a technique for sorting the position vectors of particles, known as rPSO. Through these enhancements, the apfrPSO significantly improves the capacity for robots to adapt to different operating conditions while also increasing the precision and overall efficiency of their path-planning processes [78]. Zaman H R R modified the mutation and crossover operators of BSA by neighborhood to improve convergence rates. Additionally, a novel mutation factor was joined to enhance the convergence precision of avoiding local optima [79].

4.3.3. Ant Colony Optimization (ACO)

Ant colony Optimization is a heuristic optimization algorithm inspired by the foraging behavior of ants in nature. In nature, ants release pheromones to mark their paths, and when ants find food while walking along a path, they release more pheromones on that path. The concentration of pheromones decreases over time, creating a negative feedback mechanism. When choosing paths, ants tend to select those with higher concentrations of pheromones, thus forming a form of collective intelligence. This mechanism ensures that the algorithm explores new paths and avoids converging prematurely on suboptimal solutions, as shown in Figure 10. This algorithm has been applied in the field of path planning for construction robots.

When ants choose paths, they do not simply opt for the one with the highest pheromone concentration; instead, they make choices based on a certain probability. This probability is generally related to both the pheromone concentration and the heuristic information of the path. The Ant Colony Optimization (ACO) algorithm, which simulates evolutionary algorithms, has clear advantages over traditional algorithms, including strong robustness, good global optimization capabilities, efficiency when facing complex problems, good readability, and a wide range of applications. If, in large-scale construction projects, a large number of construction robots need to work collaboratively, the Ant Colony Algorithm can be used to optimize swarm control strategies, achieving efficient task allocation and path-planning and avoiding conflicts between robots. However, it also has some limitations: a lack of effective search strategies in the early stages of search can lead to significant blindness in complex problems; convergence speed is usually slow, especially in larger-scale problems; the algorithm needs multiple iterations to capture optimal solution; and ants can be trapped in local optima easily [80,81].

Luo Q proposed an improved algorithm aimed at optimizing issues such as low search efficiency, slow convergence speed, and the tendency to fall into local optimal solutions in the Ant Colony Algorithm. By providing unevenly distributed initial pheromones, it avoids the blind search of the Ant Colony Algorithm in its initial stages. By introducing a damping coefficient into the heuristic function, the convergence speed of the algorithm is accelerated. Combining the advantages of the MMAS algorithm model and the elite ant algorithm, the pheromone update rules were improved to prevent the algorithm from converging too early. A pseudo-random state transition rule is used to determine the next path, enhancing the algorithm’s global search capability and convergence speed [82].

Zhang S proposed an innovative method called the Improved Ant Colony Algorithm based on Population Information Entropy (AIACSE). This algorithm describes the diversity of the population during the iteration process through information entropy and constructs a non-uniform initial pheromone distribution to reduce the blindness in the early stages of search; it adopts a non-uniform distribution initialization strategy based on the A algorithm, avoiding the blindness in the evolutionary phase and improving the convergence speed. The adaptive parameter adjustment strategy based on population information entropy automatically adjusts the values of parameters according to the changes in population entropy, maintaining the balance between exploration and exploitation. Through comparative experiments, this improved algorithm performs excellently in terms of convergence speed and success rate [83].

Wu L proposed an improved ACO algorithm. First, a directional factor is added to the heuristic function to improve the convergence speed of the algorithm. Second, an improved function is proposed to reduce the frequency of turn in the planned path; then, a new state transition probability rule is added to the algorithm for increasing search efficiency and population diversity; finally, a method for inhomogeneous distribution of initial pheromone is proposed to avoid blind searching [84]. Miao C introduced an enhanced version of the adaptive Ant Colony Algorithm. This innovative algorithm incorporates angle guidance factors and obstacle exclusion factors into the transition probability of the traditional Ant Colony Optimization (ACO). Doing so significantly enhances the real-time responsiveness and safety of path planning. Moreover, the IAACO further refines the pheromone update rules within the ACO framework by including heuristic information, adaptive adjustment factors, and pheromone evaporation factors. This dual adjustment serves to create a more effective balance between the algorithm’s convergence speed and its ability to conduct a thorough global search, thus improving overall performance in optimization tasks [85].

Wang L proposed enhancements to the traditional Ant Colony Algorithm by implementing an improved pheromone update mechanism. This updated approach incorporates a safety value heuristic function that serves to optimize the overall efficiency of the algorithm. One of the significant advantages of this innovation is its ability to minimize the frequency with which the Ant Colony Algorithm becomes trapped in the local optima, a common challenge faced during complex problem-solving. Additionally, this improvement effectively decreases the search time required for 3D path planning, making the algorithm more robust and effective for such applications [86]. Hengyi Huang integrated ACO and DWA algorithms, proposing an algorithm for multi-robots that effectively avoids the potential collision between real-time path planning and obstacles for multiple moving robots [87].

Ronghua M enhanced the pheromone reinforcement for high-quality paths, making these paths more attractive to the Ant Colony in subsequent searches, and thus improving the quality of path-planning [88]. Wang H proposed an improved Ant Colony Algorithm. By combining APF attraction to enhance the heuristic function, this modification accelerated convergence. Additionally, during the pheromone update process, an extra pheromone increment was introduced, which enhanced the optimality of path length. [89]. Yu X combined ACO with A* to develop a two-layer algorithm, which enables AUVs to successfully navigate through multiple targets in environments with dense obstacles [90].

4.4. Deep Reinforcement Learning Algorithm

Reinforcement learning belongs to machine learning and is widely applied in path-planning for robots. The mathematical foundation of reinforcement learning is based on Markov Decision Processes (MDP). A MDP consists of a set of states, a set of actions, transition probabilities, and a reward function. Through MDPs, agents can choose actions in a given state and receive rewards, thereby learning and optimizing their strategies. The following is an overview of some commonly used reinforcement learning algorithms.

4.4.1. Q-Learning

Q-learning is a model-free reinforcement learning algorithm that is widely applied in the field of construction robotics. The algorithm aims to determine the optimal policy by learning the state-action value function (Q function). The Q function represents the expected return of taking a certain action in a particular state. The learning process mainly relies on the Bellman equation, which provides the theoretical basis for updating Q values. Through continuous exploration and exploitation of the environment, Q-learning can find the optimal policy within a limited time, thereby maximizing cumulative rewards. The algorithm has many advantages, such as simplicity, model independence, and wide applicability. However, its poor scalability, slow convergence, and the challenge of balancing exploration and exploitation have always constrained its development. Scholars have made many efforts to address these issues. Yu Z proposed the Double Q-learning method to address the slow convergence rate, which involves two independent Q-value estimations. Double Q-learning can more accurately estimate state values, thereby increasing convergence speed and learning efficiency [91]. Low E S used the Flower Pollination Algorithm (FPA) to improve the initialization of Q-learning, accelerating its convergence [92], and also utilized a modified version of Q-learning, the Improved Q-learning (IQL), to speed up machine learning [93]. Zhou Q introduced an enhanced version of the Q-learning algorithm, referred to as optimized Q-learning (O-QL). This refined approach incorporates a novel technique for initializing the Q-table, which plays a crucial role in setting the foundation for the learning process. Additionally, it presents a fresh strategy for selecting actions, aiming to improve decision-making during the learning phase. Furthermore, a redefined reward function is implemented to better align with the specific objectives of the task at hand. To optimize learning rate adjustments, the Root Mean Square Propagation method is utilized, ensuring that updates to the learning parameters are both effective and efficient. Collectively, these modifications lead to a significant acceleration in the learning process, enhancing the overall efficiency of path planning [94]. Gharbi A introduced a dynamic rewards factor to Q-learning for improving the efficiency of path planning and the capacity of obstacle avoidance [95]. Du H used artificial potential fields to prevent random exploration to avoid the slow convergence of Q-learning and provide basic information about the environment for mobile robots. Du also proposed a dynamic reward factor to accelerate algorithm convergence [96].

4.4.2. Monte Carlo Methods

The birth of the Monte Carlo method is closely related to the Second World War. In the early 1940s, while working at the Los Alamos National Laboratory, John von Neumann and Stanislaw Ulam began to study how to make decisions under uncertain conditions. Ulam’s initial inspiration came from gambling, and he proposed an idea to solve complex problems using random sampling, thus naming the method “Monte Carlo” in honor of the famous casino in Monaco. The core idea of the Monte Carlo method is to use random number generators to estimate a mathematical function or simulate the behavior of complex systems through a large number of random samples. T. Dam proposed a new algorithm based on the Monte Carlo Tree Search (MCTS) algorithm for robot path planning, which increases the probability of the integrity of finding the best path and the convergence speed in complex environments [97]. Wang T proposed an improved Ant Colony Optimization (MC-IACO) based on Monte Carlo to address the issue of slow convergence at the initial stage of robots [98]. Castellini A introduced a novel approach that extends the principles of Partially Observable Monte Carlo Planning (POMCP) to address the challenge of regulating robot speed in environments reminiscent of industrial settings. These environments often present uncertain motion difficulties that can complicate the operational efficiency of robotic systems. By employing this method, it becomes possible to better navigate the complexities associated with varying levels of uncertainty in motion, ultimately leading to enhanced performance and reliability for robots functioning in such contexts [99]. A novel approach known as the sliding MCTS method was introduced by Xu P that builds upon the principles of the Monte Carlo Tree Search (MCTS) algorithm. In this innovative framework, Xu P adeptly achieved an effective equilibrium between the optimization and search functionalities inherent in the algorithm. This was accomplished through the implementation of a dynamic moving root node, which allows for more adaptable search strategies. Additionally, the method incorporates a mechanism for regulating the sampling duration, thereby enhancing the overall efficiency and effectiveness of the search process [100]. Mohseni A deployed a particle filter using the Monte Carlo localization method to resolve a Hidden Markov Process based on Recursive Bayesian Estimation using information entropy to identify outliers and probabilistic methods to remove detected outliers. Additionally, a new mutation process was added to the localization algorithm to actively detect high-likelihood areas using the posterior probability density function. The improved Monte Carlo localization algorithm was applied to mobile robots, and experiments demonstrated an improvement in the accuracy of path planning [101].

5. Summary

Architectural robots are now advancing to the forefront of digital transformation and intelligent upgrading in the construction industry. Path planning, as a core element in this technological evolution, has increasingly highlighted its strategic position. Based on detailed considerations of the information richness of construction sites, path-planning algorithms are divided into two major categories: global path-planning and local path-planning. Within this framework, this paper comprehensively reviews the path-planning algorithms applied in recent years and summarizes the advantages and limitations as well as the improvement directions of various algorithms, as shown in

Table 1. Summary of path planning algorithms.

Algorithm	Algorithm Type	Advantage	Limitations	Improvement
A* Algorithm	Global Path-planning Algorithm	Efficient: A* can find the shortest path quickly. Flexible: A* adapts to different problems by choosing different inspired functions.	Memory usage: A* In complex scenarios, it takes up a lot of memory. Sensitive: A* performance depends on the choice of the heuristic function.	Search speed: [23,24,26] Success rate: [27] Path smoothing: [24,25]
Dijkstra		Simple: the logic of the algorithm is relatively simple and easy to implement. Optimality: Dijkstra ensures that the shortest path from the start node to the target node is found. Adaptability: Dijkstra is able to accommodate the different models.	Computation complexity: the computational complexity of Dijkstra is relatively high. Low speed: Dijkstra search speed is low during run in dynamic environments. Inability to process the negative weight edges: there are negative weight edges in the graph, which may cause the algorithm to find the correct shortest path.	Computational complexity: [29,34] Path quality: [30,31,33] Optimality: [28,32]
RRT algorithm	Local Path-planning Classic Algorithm	Efficient: the RRT algorithm can quickly converge to the target. Flexible: RRT algorithms can easily be combined with other algorithms to form a more complex path-planning system. Adaptability: the RRT algorithm can adapt to the dynamic environment, that is, the path can be updated in real time when the environment changes.	Uncertainty: the randomness of the RRT algorithm makes it possible to get different results per run. Non-optimal paths: the paths generated by the RRT algorithm are not necessarily optimal, especially in complex environments.	Stochastic: [36,44,45] Non-optimal path: [37,39,46] Inefficiency: [38] Search speed: [40,41,42,43]
APF Algorithm		Fast: APF computational complexity is relatively low, suitable for application scenarios that require rapid response. Flexible: APF can adapt to different environmental conditions. By adjusting the parameters of the potential field, it can realize the effective navigation of various complex environments.	Suboptimal: APF falls into local minima. Success rate: APF depends on local information and will become locked in a complex environment, leading to failure of path planning.	Success rate: [55,60] Adaptation: [56,57,59,60] Suboptimal: [58] Flexible: [61]
Dynamic Window Approach		Fast: DWA can quickly calculate the direction of the robot’s motion. Flexible: DWA can adapt to different types of robots and variable environments. Barrier avoidance: DWA can dynamically identify and avoid obstacles.	Suboptimal: DWA falls into a locally optimal path. Low speed: DWA has great computational complexity during operation, especially in a dynamic environment.	Search speed: [63,65,66] Calculation efficiency: [64] Path smoothing: [67]
GA Algorithm	Intelligent algorithm	Fast: GA can find optimal or approximate optimal solutions to large-scale optimization problems in a short time. Adaptation: GA can deal with problems from different areas. Comprehensive: GA enables an efficient global search.	Search efficiency: GA may be less efficient in the process of finding solutions, especially when the initial population diversity is insufficient. Premature convergence: When the population diversity is poor, GA may converge prematurely to the local optimal solution.	Path quality: [69,71] Dynamic obstacle avoidance: [70,72]
PSO Algorithm		Global: PSO has excellent global search ability and can effectively find the global optimal solution. Efficient: PSO is computationally efficient and can handle large-scale and high-dimensional problems at low cost. Flexible: PSO can adapt to different types of problems; widely used in function optimization, combinatorial optimization, and other fields.	Slow convergence rate: PSO convergence rate is slow in the fine search phase. Sensitive: The performance of the PSO depends somewhat on the position of the initial particles. Improper initial location selection may result in inefficient search or poor convergent results. High-dimensional local optima: PSO easily falls into the local optimal solution in a high-dimensional space.	Search efficiency: [76,79] Search accuracy [77,78]
ACO Algorithm		High efficiency: ACO has high search efficiency in path-planning and can quickly find the optimal or near-optimal solution. Path smoothing: ACO generates path smoothing, especially in the environment of complex obstacles. Adaptation: ACO is highly adaptable to environmental changes.	Suboptimal: ACO easily falls into a local optimal solution. Computational complexity: ACO’s complexity of computational pheromone updating and path selection may be high in large-scale problems. Convergence rate: ACO’s convergence rate is slow and requires more iterations.	Search speed: [82,88] Suboptimal performance: [85,86] Path quality: [83,84,89] Damage avoidance: [87]
Q-learning	Reinforcement Learning-Based Path-planning Algorithm	Flexible: QL can work without a complete model of the environment. Multi-purpose: It can be widely applied in various scenarios. Optimal: Under certain conditions, QL can help to ensure convergence to the optimal policy.	Convergence rate: QL converges more slowly in the search for the optimal solution. Inefficient: QL is less flexible and efficient when facing complex or high-dimensional environments. Sensitivity: the performance of QL largely depends on the choice of hyperparameters, such as learning rate and discount factor.	Convergence rate: [91,92,96] Local optima: [94] Learning speed: [93,95]
Monte Carlo Methods		Comprehensive: MCM provides a more comprehensive risk assessment. Flexibility: MCM adapts to different models. Processing of complex data: MCM is able to handle complex probability distributions and multidimensional data through random sampling.	High cost: with the increase of dimensions, the number of samples required for MCM increases rapidly, and the computational time and resource demand will increase. Random: MCM relies on random sampling, and an insufficient sample size can bias the results. Sensitive: the effect of MCM depends on the accuracy of the selected model.	Convergence rate: [97,98] Path quality: [99,100,101]

.

6. Path-Planning Technology in the Future Development of Clustered Construction Robots

6.1. Overview of Clustered Construction Robots

Swarm robotics is not merely the aggregation of multiple single-task robots in quantity. Instead, it involves selecting appropriate single-task robots based on actual construction task requirements to collaborate and achieve higher work efficiency or complete tasks that a single robot cannot. During the collaborative work of construction robots, construction workers can be considered efficient single-task robots. Therefore, the collaboration of swarm construction robots mainly falls into two categories: robot collaboration and human–robot collaboration.

Although collaborative work in construction is currently developing slowly, there have been attempts. For example, Tavares P used Building Information Modeling (BIM) to develop a human–robot collaborative structural steel manufacturing and welding system, effectively improving work efficiency and product quality [102]. Similarly, Tsuruta T developed an automated mobile robot system for marking free passage areas in construction sites [103]. Compared to traditional construction robots, swarm construction robots require a high level of autonomous decision-making and navigation capabilities. High-level autonomous navigation poses significant demands on path-planning algorithms.

6.2. Path-Planning Algorithms in the Future Development Direction of Clusterized Construction Robots

1. Construction robot swarms must be capable of both independent autonomous operation and the ability to form a group for collaborative work when necessary. This requires that the robot path-planning algorithms can respond quickly and are capable of calculating the movement paths of multiple robots simultaneously.

Additionally, if a robot in the swarm fails and cannot move, the path-planning algorithm needs to quickly remove it from the swarm. The failed robot should then be treated as an obstacle.

2. Construction robots need to move in changing environments, which requires the path-planning algorithm to have strong adaptability and obstacle-avoidance capabilities. It must ensure that robots within the swarm do not collide with each other while moving and that the entire swarm avoids collisions with static or dynamic obstacles.

3. Swarm construction robots work not only in planar maps but also in high-dimensional maps, such as in the air, in the ocean, or between high-rise buildings. This necessitates the ability of the path-planning algorithm to plan paths in high-dimensional space. Since different robots have different modes of movement, such as rolling, climbing, and walking, the suitability of the paths also varies. The algorithm must make path selections based on these differences in movement, which poses high demands on the intelligence and learning capabilities of the path-planning algorithm.

7. Conclusions

The primary goal of the path optimization of construction robot planning is to ensure safety and the second is to improve work efficiency to save cost. This paper reviews the commonly used algorithms in the path-planning of construction robots and collects the current optimization algorithms in the field of various robots, including ground robots, underwater robots, aerial robots, etc. The main optimization directions of the algorithm include search speed, path smoothing, and obstacle avoidance. At the same time, according to the mastery of the construction site information, the path-planning algorithm is divided into two categories: global path-planning and local path-planning. Local path planning can further be divided into the classical algorithm, intelligent algorithm, and reinforcement learning algorithm. Common environment modeling methods, path quality assessment standards, and common robot sensors are also introduced. Finally, we summarize the various algorithms, analyze their advantages and disadvantages, and look forward to the future development direction.

The A* and Dijkstra algorithms represent classic global search techniques. Their fundamental principle involves identifying the minimum estimated cost from the current node to the target destination. By utilizing random sampling, RRT can efficiently establish a path to the target point; however, ensuring path quality remains a challenge. The APF method employs a physical potential field to represent environmental data, assisting in obstacle avoidance and optimal path determination. Nevertheless, in intricate environments, it may easily become trapped in local optima. Drawing inspiration from natural evolutionary principles, the GA serves as a highly effective global search strategy, although its convergence rate is slower than that of alternative algorithms. The DWA integrates the robot’s motion model with its environmental representation, utilizing a window model to assess the robot’s motion states and evaluate viable trajectories. Both the ACO and PSO algorithms originate from the behavioral patterns observed in biological groups. ACO mimics the foraging behavior of ants searching for food, employing pheromones to direct the search process in order to ultimately find the best solution. Conversely, PSO reflects the foraging dynamics of flocks and schools, adjusting the positions of particles to seek the optimal outcome. However, like ACO, PSO can also fall prey to local optima. Reinforcement learning boasts substantial learning and decision-making prowess, enabling it to devise optimal paths even with limited information; furthermore, it displays considerable adaptability and flexibility within complex and uncertain environments.

Construction robot path planning is an important branch of research to improve its autonomy and has attracted the attention of the construction industry in the last decade. Although many path-planning algorithms are applied in practical scenarios, there are still many exploration directions:

• Due to the complex and changeable construction environment and the difficulty of construction tasks, the single path-planning algorithm and its improved model, although near-perfect, are still difficult to meet actual engineering needs. Therefore, most of the research directions began to turn to the combined algorithm. Compared with a single algorithm, the combined algorithm can well combine the advantages of each algorithm to make up for the defects of the individual algorithm itself, so as to better complete the construction task.
• The importance of path-planning technology based on reinforcement learning in the field of path-planning for architectural robots is gradually increasing. Based on the current development situation and the future development needs of the construction industry, the research direction of reinforcement learning technology in the future includes combining it with the traditional path-planning method and applying reinforcement learning to multi-agent collaborative path-planning. Meanwhile, deep learning has become a star in the realm of path planning. Deep learning can conduct autonomous information acquisition, information processing, and communication, and meets the needs of path-planning. Additionally, it can make the building robot with autonomous learning ability and gradually improve its ability to deal with the complex environment and construction efficiency.
• In the construction site environment complex, there are many uncertainties, often appearing as narrow spaces and obstacles in the area; for these cases, the sensor in the path-planning algorithm will become the future research hotspot. Compared with a single sensor’s uncertainty and information integrity, more sensors can be more accurate and more comprehensively detect and describe the environment.
• As seen in the research on the path-planning algorithm of high-altitude building robots and marine building robots, with the exploration of deep space and the ocean, the future of architecture is not confined to land. Compared with land, construction robots working in the ocean and deep air will receive stronger environmental factors, which puts forward high requirements on the performance of the algorithm. For example, marine building robots are disturbed by ocean currents and marine life, requiring path-planning algorithms to quickly search for complex environments while enabling dynamic obstacle avoidance.