Survey of Methods Applied in Cooperative Motion Planning of Multiple Robots

Abstract

Recent advances in robotics, autonomous systems, and artificial intelligence (AI) enable robots to perform complex tasks such as delivery, surveillance, inspection, rescue, and others. However, they are unable to complete certain tasks independently due to specific restrictions. In the last few years, researchers are keenly interested in deploying multi-robots for such tasks due to their scalability, robustness, and efficiency. Multiple robots; mobile robots, unmanned aerial vehicles (UAVs), unmanned ground vehicles (UGVs), and unmanned underwater vehicles (UUVs); are gaining much momentum and versatility in their operations, whereas cooperative motion planning is a crucial aspect of incorporating these robots into boundless applications. The purpose of this review chapter is to present an insightful look into the problem of cooperative motion planning with its solution and a comprehensive assessment of various path-planning techniques, task-based motion planning techniques, and obstacle avoidance protocols. It further explores the role of AI in the motion planning of multi-robots. Moreover, it highlights various applications and existing issues in these applications that require future consideration. This review chapter implies that researchers, industries, and academia should aspire to cooperative motion planning for robotic expansions.

1. Introduction

Recent advances in artificial intelligence (AI), robotics, communication, and other technologies are modifying the characteristics of robots. Different robots are deployed for different environments; for example, unmanned aerial vehicles (UAVs) are used for aerospace, unmanned ground vehicles (UGVs) are employed for ground, and unmanned underwater vehicles (UUVs) are applied for underwater. However, in complex scenarios, an individual robot is unable to reach its destination while avoiding collision with any obstacle and fulfilling its task in complex missions [1]. Therefore, a group or swarm of inexpensive and small robots are deployed that coordinate with each other and accomplish a common goal efficiently, rapidly, and robustly. Multi-robots are widely employed for various complex and challenging tasks such as inspection, monitoring, search and rescue, navigation, and security in disaster, marine exploration, manufacturing industries, smart agriculture, military, and other fields.

The cooperation among multiple robots occurs using the available information from the network. Hence, the accurate measurement of a robot’s position in correspondence to other robots and the environment provides assistance in avoiding collisions with other robots and obstacles [2]. Researchers are engaged in developing strategies for cooperative motion planning of robots. Coordination may be dynamic or static according to the environment. For cooperative motion planning, the shortest distance, safety distance from obstacles, trajectory smoothness, and computational time are the prime factors that must be considered [3].

Motion planning and path planning are closely related, and motion planning in many places is labeled as path planning. However, in motion planning additional dynamic properties such as velocity and acceleration are taken into consideration. Thus, motion planning is the subset of path planning [4]. For reliable operations of multi-robots, path planning aims to determine a collision-free path in the shortest time. Multiple robot path planning shows high computational complexity and offers optimal solutions.

To better deploy multi-robots, other research areas are task-based motion planning and obstacle avoidance protocols [5]. These approaches assist multiple robots to have a better knowledge of the environment while planning motion and executing tasks so that collision with the obstacles or other agents does not occur. Furthermore, the role of AI techniques in motion planning for operating multi-robots securely and intelligently is evident from their application in diverse studies [6]. Besides all these advances, limitations still exist for future considerations. Therefore, a proper cooperative motion planning approach has always been an non-deterministic polynomial-time hardness (NP-hard) problem, and developing efficient schemes has always been a crucial research topic in the last decades.

The aforementioned discussions have inspired us to examine and compare the widely applied algorithms and methods in the recent literature for carrying out the cooperative motion planning of robots. The novelty of this review chapter lies in covering literature on all robots such as mobile robots and unmanned vehicles; UAVs, UGVs, and UUVs, and presenting it in one place. The main contributions of this chapter are threefold,

• Presenting a review of breakthrough results in the context of motion planning for multi-robots.

• Comparing and analyzing the recent path-planning techniques, task-based motion planning approaches, and obstacle avoidance protocols for multiple robots.

• Exploring the role of AI in motion planning.

• Evaluating various applications of multiple robots.

The structure of this chapter is organized as follows: Section 2 presents an analysis of related work. Section 3 defines the problem of cooperative motion planning and proposes a solution for it. Section 4 evaluates multi-robot path-planning techniques. Section 5 describes task-based motion planning. Section 6 covers optical avoidance protocols for multi-robots. Section 7 explores the role of AI in motion planning. Section 8 presents the applicability of cooperative motion in various fields. Section 9 discusses the conclusion of this chapter and future issues in cooperative motion planning.

2. Related work

Several literature reviews are conducted for the motion planning of multiple robots. The widely focused are mobile robots, and then followed by UAVs and autonomous underwater vehicles (AUVs). All the papers have discussed path-planning strategies from classical to emerging AI techniques, such as reinforcement learning (RL) and machine learning (ML).

Different motion planning techniques are analyzed with a prime focus on highway planning and UGVs [7]. Decision-making and path-generation concepts are discussed to elaborate motion planning. Findings reveal that a huge number of algorithms are reviewed in this chapter. This study includes not only state-of-the-art work but also suggests decomposition methods for highway motion planning and encourages autonomous driving. Various methods developed on motion planning policy are reviewed [8]. This chapter is focused on mobile robots in an unstructured environment but has explored some studies on UAVs. The conventional and emerging deep reinforcement learning (DRL) methods that involve multi-robot systems, meta-learning, and imitation learning are enlightened. Restricted theoretical development along with the low interpretability is suggested to be the main reasons that hinder their real-time applications.

Researchers survey studies that applied ML for control and motion planning in the navigation of mobile robots [9]. They compare and contrast ML approaches with classical approaches in the context of navigation. Findings reveal that classical navigation issues are required to be examined with an ML perspective. It further evaluates that despite advances, classical approaches are unable to solve navigation problems. A comprehensive and clear understanding related to opportunities, limitations, relationships, and the future of different motion planning algorithms is presented [3]. Traditional algorithms to policy gradient reinforcement learning algorithms are discussed for intelligent robots. This study paves the way for improved motion planning algorithms. Optimization methods are discussed in Ref. [10] for motion planning of UAVs. Findings reveal that swarm-based optimization approaches are preferred by researchers due to their exceptional ability in complex scenarios.

A survey is carried out to evaluate various methods of motion planning and task planning for the cooperative working of multiple mobile robots [6]. A taxonomy based on system capabilities is proposed in this study that applies to single-robot systems and multi-robot systems. Various motion planning methods, from classical to reinforcement learning (RL) approaches, are reviewed for single-robot and multi-robots [11]. It covers different types of robots such as UAVs, wheeled mobile robots (WMRs), AUVs, etc. It concludes that motion planner based on RL is model-free and achieves unification of the local planner and the global planner but shows various limitations that hinder its real-time applications.

3. Problem definition and solution proposed

Motion planning is to determine a sequence of feasible robot configurations called trajectories that allow moving a robot from its initial stage to its final destination with collision avoidance and obstacle avoidance for completing a given task. It involves various variables such as robots’ dynamics, kinematics, environment, and task constraints. Motion planning optimizes a robot’s motion by enhancing its throughput and minimizing its cycle time. It can be applied to verify a process’s feasibility and to estimate potential problems before deploying robots. The major problem in the development of robots especially autonomous vehicles is to devise a way in which they are capable enough to make their plans in different situations. Therefore, motion planning is essential in the deployment of multiple robots in an environment consisting of obstacles. The degree of motion planning problem varies according to a couple of factors whether all the obstacles’ information regarding their locations, sizes, and motion, is known before the deployment of the robot or whether the obstacles are stated or dynamic in an environment.

The solution to cooperative motion planning is an integrated architecture for multi-robots. Figure 1 shows this architecture is comprised of three layers: the task assignment layer, the past planning layer, and the collision avoidance layer. The task assignment layer must receive the cost which is a distance function for reaching the final destination. This cost is received when communication occurs between the task assignment layer and the path-planning layer. The path-planning layer is called again and again until a feasible solution is obtained for task assignment. Then, the path-planning layer uses the information about obstacles that are known before the mission starts and computes an original collision-free path. When any unmapped obstacle appears, the collision avoidance layer determines a collision-free path efficiently locally detouring from the original path. It also prevents collision between two agents, when an agent reaches a target point during path following. If the agent cannot reach the required target points promptly, then the task assignment layer redistributes the task among other agents in the neighborhood. It is duly considered that by explicitly considering the characteristics of a vehicle or robot, this solution can be further refined.

Figure 1.

An integrated architecture for cooperative motion planning.

4. Multi-robot path planning

Various path-planning algorithms have been proposed and grouped on different criteria, such as a study categorizes the path-planning algorithms on coordination criteria into deliberative approaches and reactive approaches [12]. Deliberative approaches are further classified into evolutionary algorithms and road map-based algorithms, whereas the reactive approaches are further grouped into potential field algorithms and modern predictive control approaches. Another study classifies the path-planning algorithms according to their application in static and dynamic environments [13]. The main categories are classical algorithms and soft computing techniques. Classical algorithms are comprised of cell decomposition, road map, Voronoi-diagram, and potential field. On the other hand, soft computing techniques include artificial neural networks (ANN), fuzzy logic, and hybrid and evolutionary techniques. According to a recent study on multi-robot, path-planning approaches are grouped into classical approaches, heuristic algorithms, AI techniques, and bio-inspired algorithms [14]. Figure 2 shows classification of these path-planning algorithms.

Figure 2.

Path-planning approaches for multi-robots.

4.1 Classical approaches

Classical approaches usually involve a predefined graph that requires high computational space and time. These techniques do not ensure completeness and are not capable of re-plan the path in the application. These approaches are classified into graph-based approaches, artificial potential field (APF), and sampling-based approaches. A study proposed APF for AUV flocks that are enabled with software-defined networking (SDN) [15]. Results show that the suggested path-planning scheme allows efficient path planning.

4.2 Heuristic approaches

The heuristic approaches solve the problems that cannot be addressed by other approaches and estimate an approximate solution rather than an exact solution. Therefore, these algorithms are also called approximation algorithms. These algorithms are easily applicable and develop cost functions to evaluate the path. It searches the subspace of the search space and generates only near-optimal results. Moreover, they require lower space and runtime. A-star (A*) search algorithm and D* algorithm are extensively applied through heuristic algorithms. The A* algorithm with a distributed velocity perception strategy is proposed for the cooperative motion planning of multiple UAVs in three-dimensional (3D) [4]. Simulations suggest that the applied algorithm enables UAVs to take less time and shorter paths for reaching their destinations safely.

4.3 Artificial intelligence techniques

Intelligent systems have gained more attention these days. AI techniques are developed to overcome the limitations of traditional reinforcement learning. AI-based algorithms and models possess self-learning abilities and have completed characteristics for the path planning of multiple robots with faster convergence. Researchers have focused more on machine learning (ML) algorithms, reinforcement learning (RL), neural networks (NN), fuzzy logic, etc. A study has developed a new mobile edge computing (MEC) platform for multi-UAVs and has suggested an RL framework for path planning [16]. Simulations show the feasibility and effectiveness of the platform.

4.4 Bio-inspired algorithms

Bio-inspired techniques are inspired by the behavior of animals and use particles to generate paths. They are primary algorithms for the path planning of multiple robots because they show computational efficiency and have powerful implementations. Bio-inspired techniques include particle swarm optimization (PSO), pigeon-inspired optimization (PIO), ant colony optimization (ACO), genetic algorithm (GA), gray wolf optimizer (GWO), and other bio-inspired techniques. A novel switching delayed particle swarm optimization (SDPSO) algorithm is proposed for UAVs [17]. Simulation results evaluate that the proposed technique shows robustness and quickly plans paths of high quality. Table 1 presents a comparative analysis of path-planning algorithms applied in different studies.

References	Deployed platform	Path planning	Evaluation	Advantages
[4]	Multi-UAVs	Velocity aware A*	Simulations	Produces shorter paths Assists in reaching the destination in less time
[15]	AUV flock	APF-SDN	Simulations	Allows efficient path planning
[16]	Multi-UAVs	MEC-RL	Simulations	Provides flexibility and effectiveness
[17]	UAVs	SDPSO	Monte-Carlo simulations	Shows robustness Quickly plans paths of high quality

Table 1.

Comparative analysis of path-planning algorithms.

5. Task-based motion planning

For performing a task, a robot has to determine feasible paths then while executing the plan, it has to consider the complexity of the environment before stepping ahead to accomplish its goals. This whole procedure is classified into task planning and motion planning. Task planning is referred to computing solvable plans for decomposing a long-horizon task into several elementary subtasks in a discrete space, whereas motion planning is referred to transforming a subgoal into a set of perimeters for achieving that subgoal in a continuous space. As both concepts share a few similar designs, task-based motion planning is introduced for integrating the discrete planning methods with continuous planning methods. Researchers referred to it as task and motion planning (TAMP) [18]. TAMP chooses the chronology of high-level actions for the robot. Then, it selects the hybrid parameter values that find a way for performing these actions. Finally, it prefers low-level motion for executing these actions safely. However, solutions to this merger vary in the manner in which their joint search space is explored. Figure 3 illustrates TAMP approaches are classified into classical approaches, learning approaches, and hybrid approaches [19].

Figure 3.

Classification of task and motion planning approaches.

In certain cases, various tasks are assigned to multiple robots in chronological order of tasks. Once the multi-robots are deployed, the dynamic adjustment of tasks leads to difficulties in distribution and path re-planning. An optical reference point built on an improved fruit fly optimization algorithm (ORPFOA) is utilized to solve this path-planning issue with changing tasks in the 3D oilfield inspection by cooperative multi-UAVs [20]. Results show that the proposed algorithm solves not only the path-planning issue but also the assignment issue of the initial and new tasks. Two algorithms based on task assignment: The marginal-cost algorithm (MCA) and regret-based MCA (RMCA) are developed to carry out task assignment simultaneously with path planning for multi-agent robots [21]. Large neighborhood search (LNS) is also added that enhances solution sustainability. The applied strategies allow each robot to carry multiple packages. Local motion planning (LMP) is integrated with a robust H∞ decentralized feedforward reference tracking fault-tolerant control (FTC) for a hybrid team system comprised of UAVs and biped robots [22]. The suggested strategy shows effectiveness in the search and rescue tasks.

However, instead of AI technologies, bio-inspired algorithms are suggested widely to enable real-time applications [23]. Table 2 highlights that most of the studies applied bio-inspired algorithms for task-based motion planning in simulations to assure their significance for real-time applications. Cooperative UUVs are used for target detection task scheduling merged with an underwater acoustic environment [24]. In this study, task scheduling is built on the GA algorithm. Findings reveal that it refines the cooperative detection ability at the task level and solves the optimization issue. A hybrid algorithm is suggested with the estimation of the distributed algorithm (EDA) and the GA [1]. Simulations show EDA-GA plans stable and quality paths. An online adjustment strategy is also proposed that maintains complete coverage and reduces the effect on circular paths.

References	Deployed robot	Task-based motion planning technique	Focus	Environment	Performance
[1]	UAVs and UGVs	EDA-GA with online adjustment strategy	Surveillance task	Simulation	Plans stable and quality paths The online strategy maintains complete coverage and reduces the effects on circular paths
[20]	Multi-UAVs	ORPFOA	Oilfield inspection	Simulation	Solves path planning faster Allows higher optimizing precision
[21]	Multi-agent robots	MCA and RMCA with LNS	Pickup and delivery	Numerical simulation	Performs task assignments simultaneously with path planning Allows each robot to carry multiple packages
[22]	UAVs and biped robots	LMP-H∞ decentralized feedforward reference tracking FTC	Search and rescue	Simulation	Shows effectiveness in task and motion planning
[24]	UUVs	Task scheduling based on GA and integrated with an underwater acoustic environment	Target detection	Simulation	Refines the cooperative detection ability at the task level Solves the optimization issue

Table 2.

Review of various task-based motion planning techniques applied on different multiple robots.

6. Obstacle avoidance protocol

Multi-robots are destined to explore more distance in less time with collision avoidance. When multiple robots are deployed, they sense other agents and obstacles in the path, avoid them, and continue their motion. Therefore, collision avoidance is an integral part of part planning of robots. Collision avoidance protocols are categorized into two main approaches: centralized methods and decentralized methods. The centralized methods are efficient for smaller numbers of robots, whereas the decentralized methods are less expensive computationally and more effective for large groups of robots. A decentralized architecture namely, spot auction-based robotic collision avoidance scheme (SPARCAS) is applied for collision avoidance and M^* is used for path planning of UGVs [25]. It results in prioritization and dynamic handling.

In various robots, the collision avoidance control system is either deliberate or reactive. Deliberate models are plan-driven and computationally expensive as it requires a priori data of the environment. Contrary, reactive architectures are faster and require only real-time sensor information. Such as a deep RL method, policy proximal optimization (PPO) with curriculum learning is used for AUVs [26]. The applied methodology effectively plans paths with collision avoidance in 3D.

Several studies indirectly address the problem of path planning with a collision avoidance perspective for multiple robots. Such as an APF function (APFF), which uses position information, is proposed for multi-USVs during navigation [27]. It improves safety while ensuring collision avoidance and obstacle avoidance. Similarly, the max-min ant colony optimization (MMACO) approach with the Cauchy mutation operator (MMACO-CM) and MMACO approach with differential evolution (MMACO-DE) are suggested in Ref. [23]. Findings show that MMACO-DE reduces distance and saves time by taking lesser turns in a 3D complex environment. Optimized-weighted-speedy Q-learning (OWS Q-learning) algorithm and a collision avoidance cooperation method are suggested for multiple UGVs [28]. Table 3 summarizes the abovementioned literature in tabular form and provides a comparative analysis of the deployed platforms, applied obstacle avoidance protocols, considered obstacles and evaluation indexes, and finally the resulting performance in each study.

References	Deployed platforms	Obstacle avoidance protocol	Obstacles	Evaluation indexes	Performance
[23]	UAV flock	MMACO-CM; MMACO-DE	Tornados and mountains	Convergence time and route taken	MMACO-DE takes lesser time Reduces distance and saves time
[25]	Multiple robots	SPARCAS-M^*	Obstacles	Path length, average payment, average waiting, and time	Allows prioritization and dynamic handling
[26]	AUVs	PPO-curriculum learning	Obstacles or ocean current	Collision rate, success rate, and average tracking error	Effectively plans path with collision avoidance in 3D
[27]	Multi-USVs	APFF	Obstacles	Time, route length, path length	Improves safety
[28]	Multi-UGVs	OWS Q-learning with collision avoidance cooperation method	Green belts and buildings	Average shortest path, path steps, average reward, and time	Effectively solves the coordination problem Achieves the convergence effect and collision-free path in the least time

Table 3.

Comparative analysis of obstacle avoidance protocols.

7. Role of artificial intelligence in motion planning

Motion planning is an integral part of robots. AI-based motion planning is revolutionizing the robotics and autonomous system fields. Due to this, these robots and autonomous systems have become capable to navigate themselves autonomously while avoiding collisions and dynamic and uncertain obstacles to perform any task. The flourishing AI techniques: deep learning (DL) algorithms and reinforcement learning (RL) give better performances in handling nonlinear problems with complexity [3]. The complexity includes incompleteness, ambiguity, and the most challenging one uncertainty. As classical ML algorithms do not address time sequential planning problems, they are modified into optimal value RL and policy gradient RL. Optimal value RL applies Q learning and transforms it into deep Q-learning network (DQN), whereas policy gradient RL uses policy gradient and actor-critic algorithm and forms asynchronous advantage actor-critic (A3C), trust region policy optimization (TRPO), and deterministic policy gradient (DPG) [29].

Researchers, technology giants, and institutions are engaged in introducing new motion planning approaches by employing AI techniques or integrating advanced ML techniques with traditional algorithms. These approaches have introduced us to the concepts of autonomous vehicles and unmanned vehicles [30]. Such as Google, Tesla, Baidu, Audi, and Toyota have developed autonomous vehicles (AVs) using AI technologies. These vehicles improve human safety, assist in environmental protection, and reduce emissions, crashes, and congestion.

AI-based motion planning gives reactive or instant responses by considering short-term optimal or suboptimal reactive strategies. Moreover, it achieves long-term optimal planning objectives such as path planning during robots’ interaction with the environment [7, 31]. Moreover, it streamlines various processes and replaces hardware with software. This results in improved machine performance, product quality and productivity, and efficiency, with reduced time to market and costs. The automation industry relies on AI-based motion control and planning. Particularly manufacturers trust AI for smart, safe, flexible, and productive decision-making in motion control processes. Particularly, AI reduces the time and costs for setting up the configuration of large plants and for hiring and training new employees and enhances productivity through automation [10].

Besides so many advantages, several limitations and challenges still exist. The promising ongoing research and development will overcome these hurdles and will produce more efficient AI-powered systems in the near future. Additionally, AI-based motion planning will gain more focus for development in the future of automation and robotics [6, 32]. This will enhance the quantity and applications of automation and robotics in the military, agriculture, and various industries, such as healthcare, manufacturing, transportation.

8. Applications of cooperative motion

Groups or swarms of robots are comprised of multiple heterogeneous and homogeneous agents. Cooperative motion enables these agents to achieve their single goal [33]. Earlier the multi-robot cooperative system was applied at a small scale and was relying on simple interactions and rules. Nowadays, such systems are developed for collaborative intelligence, large-scale, and cluster-scale operations. Diverse theories are introduced into these cooperative systems that need to progress in experimental research and theoretical exploration. Considering the example of a multi-vehicle cooperative system, this collaborative system involves cooperative motion planning, collaborative positioning, scheduling and allocation of intelligent tasks, traffic flow optimization, and collaborative planning control. These technologies enable them to analyze and make decisions according to the information achieved from interactions with other vehicles, networks, and roads [34].

Clusters of multi-unmanned systems are used in complex weather, terrain, and electromagnetic environments signifying the efficiency of cooperative motion. Various scientific research institutions have conducted a series of research on cooperative systems in unknown larger areas. For example, the US Office of Naval Research carried out research on clusters of unmanned combat vehicles. Great improvements in their capabilities were realized [35].

Besides military applications, mobile robots and unmanned vehicles play significant roles in the manufacturing industry, agriculture sector, service industry, hospitals, and in hazardous scenarios such as disaster management, firefighting, rescue and search operations, and so on. They are extensively employed to replace humans, reduce injuries, and improve efficiency in a limited time, such as JD Logistics employs the collaborative systems of multi-robots in its warehouse and distribution center for performing collaborative operations to sort express deliveries. In industries, cooperative systems work for assembling and disassembling numerous small parts of a machine [36]. Similarly, groups or swarms of UAVs are widely used for monitoring, disease and pest detection, harvesting, and other purposes. These collaborative systems render their rapid service in catastrophes scenarios, such as monitoring, search, and rescue, and delivering aid and medicine in disasters.

This section signifies that all frameworks of collaborative motion planning possess a high scope of application. However, the mutual obstacle avoidance protocol is susceptible to environmental interference and sensor interference, which leads to catastrophic situations. Therefore, the safety essentials of the multi-vehicle system are unable to take full advantage of cooperative motion and demand further research.

9. Conclusion and future issues

This study has conducted a review of breakthrough results and existing future challenges of motion planning for diverse categories of robots. The main objective is to take an insightful look into the problem of cooperative motion planning with its solution, and a detailed analysis of various path-planning techniques, task-based motion planning techniques, and obstacle avoidance protocols and to present it in one place. A comparison with other review papers shows that most of the papers surveyed different classical methods and advanced methods, while most of them are focused on RL methods, UAVs, and mobile robots. The road map approach, the cell decomposition approach, and the potential field approach are observed to be advantageous solutions for the problem of cooperative motion. The path-planning algorithms allow flexible, efficient, high-quality, more traversable paths and cover more distances and show robustness for UAVs, AUVs, and UGVs in simulations and real-time experiments. Bio-inspired algorithms are widely employed for improving the cooperative motion planning of robots and UVs at the task levels, planning faster and more stable paths, and allowing each agent to accomplish its task. Various obstacle avoidance protocols are discussed that effectively solve the coordination problem, and achieve the convergence effect with safety and collision-free paths for multiple UVs and robots in different terrains. The significant role of AI in motion planning is also elaborated. The research and diverse application areas of these intelligent robots and intelligent vehicles show promising outcomes. In general, the analysis reveals that bio-inspired and RL methods are extensively studied and UAVs are commonly deployed in most of the studies. The studies have conducted these algorithms mostly in simulations. Real-time applications are rarely considered.

Besides advantages, some future issues of the analyzed methods are also evaluated in this section. As the integration of learning-based approaches into planning is essential for planners to work and reason with learned action models rather than human assistance or knowledge, task-based motion planning requires more incorporation of sampling and optimization methods. Further research is essential to plan more curvilinear paths and develop more adaptable obstacle avoidance protocols for complex scenarios with narrow paths and a large number of robots.

Considering the application domain, cooperative motion planning provides solutions for small instances only but at the scalability cost. Further issues that require future considerations include timeout failures, greedy assignments, and the least regard for their effects on the costs of overall solutions. Cooperative motion planning necessities more extension to realistic environments with deformable objects, dynamics, time, etc. These approaches need more capabilities to consider the uncertainty of present and future states with a prime focus on safety and performance.

Abbreviations

3D	three-dimensional
A3C	Asynchronous Advantage Actor-Critic
ACO	Ant Colony Optimization
AI	Artificial Intelligence
ANN	Artificial Neural Networks
APF	Artificial Potential Field
APFF	Artificial Potential Field Function
AUVs	Autonomous Underwater Vehicles
AVs	Autonomous Vehicles
DL	Deep Learning
DPG	Deterministic Policy Gradient
DQN	Deep Q-Learning Network
DRL	Deep Reinforcement Learning
EDA	Estimation of the Distributed Algorithm
FTC	Fault-Tolerant Control
GA	Genetic Algorithm
GWO	Gray Wolf Optimizer
LMP	Local Motion Planning
LNS	Large Neighborhood Search
MCA	Marginal-Cost Algorithm
MEC	Mobile Edge Computing
ML	Machine Learning
MMACO	Max-Min Ant Colony Optimization
MMACO-CM	Max-Min Ant Colony Optimization with Cauchy Mutation Operator
MMADE	Max-Min Ant Colony Optimization with Differential Evolution
NN	Neural Networks
NP-hard	Non-Deterministic Polynomial-Time Hardness
ORPFOA	Optical Reference Point-Fruit Fly Optimization Algorithm
OWS Q-learning	Optimized-Weighted-Speedy Q-Learning
PIO	Pigeon-Inspired Optimization
PPO	Policy Proximal Optimization
PSO	Particle Swarm Optimization
RL	Reinforcement Learning
RMCA	Regret-Based Marginal-Cost Algorithm
SDN	Software-Defined Networking
SDPSO	Switching Delayed Particle Swarm Optimization
SPARCAS	Spot Auction-Based Robotic Collision Avoidance Scheme
TAMP	Task And Motion Planning
TRPO	Trust Region Policy Optimization
UAVs	Unmanned Aerial Vehicles
UGVs	Unmanned Ground Vehicles
UUVs	Unmanned Underwater Vehicles
WMRs	Wheeled Mobile Robots