From Cooperative Dual-Arm Manipulators to Cooperative Multi-Arm Manipulators—Where Are We Standing Today?

Abstract

This paper highlights the state of the art in Cooperative Dual-Manipulation (CDM) and Cooperative Multi-Manipulation (CMM), comparing advances in modeling, control, planning, sensing, vision, and end-effector technologies. Methods originally established in CDM have been extended or adapted to support higher complexity of CMM. A historical timeline visualizes the steady growth of cooperative manipulation (CM) and the recent acceleration of CMM driven by rising process complexity and the need for more flexible automation strategies. CM is becoming increasingly relevant as industrial processes demand higher payload capacity, larger workspaces, and greater flexibility. In addition, this paper categorizes existing applications by cooperation type and application domain. Here, a clear dominance of simultaneous object manipulation tasks is visible (fixation-fixation). However, fixation-tooling tasks, where one manipulator grasps the product while another performs a tool operation, and tooling-tooling tasks, where multiple manipulators perform tool operations simultaneously, remain significantly underrepresented. A similar imbalance is found for rigid/non-deformable object manipulation and flexible/deformable object manipulation, respectively. Based on this review, several research gaps are identified: (i) reliable flexible object manipulation methods; (ii) CM strategies for disassembly (e.g., battery pack deconstruction); (iii) complexity in control and planning for multi-manipulator systems; (iv) pathways to industrial deployment beyond laboratory demonstrators; and (v) task-specific tooling and end-effector innovation.

1. Introduction

As industries experience a continuous need for growth in productivity, it is essential for companies to improve efficiency by optimizing their processes and enhancing product quality [1]. However, most industrial sectors have difficulties finding employees with the appropriate technical skillset. In order to cope with the constant market pressure, the number of robotic integrations has increased across various industrial sectors. Globally, the number of industrial robots has grown by an average of 11% over the last five years, with a 9% increase in 2024 compared with 2023, as illustrated in Figure 1 [2]. As a result, existing operators need to be (re)trained in order to program, commission, and maintain robotic systems [1].

Robotic manipulation started with a patent of the world’s first industrial manipulator, developed by George Devol in 1956. This manipulator was developed for a simple pick-and-place task [3]. Together with Joseph Engelberger, the first company to manufacture industrial manipulators (Unimation) was established, producing its first manipulator (Unimate) in 1961. Subsequently, the concept gained momentum and several companies, including Ford and General Motors, among others, also adopted industrial manipulators to automate their production lines [4].

In the early years of robotization, the tasks performed by manipulators were mainly repetitive tasks for which a single manipulator was sufficient to solve the problem. It was only in the early 1970s that the complexity of the tasks to be performed increased. While many tasks were still manageable using single manipulators, certain challenges proved too complex. These included scenarios such as transporting heavy loads that exceeded the payload capacity of a single manipulator [5], assembling intricate components that required coordinated movements [6], and manipulating flexible objects that demanded advanced dexterity [7]. To solve these complex problems, multiple manipulators, also known as cooperative manipulation (CM), work together in the same work environment. For multiple manipulators to be cooperative, the manipulators need to perform tasks on the same object in the same motion environment. With regards to handling objects, it is called simultaneous coordinated tasks. When operations or tasks are performed on the object, they are referred to as task-coordinated tasks [8]. In addition, a distinction is made in the number of manipulators. When two manipulators perform tasks on an object, it is referred to as Cooperative Dual-Manipulation (CDM). When more than two, hence multiple, manipulators perform tasks on one and the same product, it is referred to as Cooperative Multi-Manipulation (CMM). Dual-manipulation can thus be seen as the simplest variant of multi-manipulation. From 1980 to 1990, the core of development within robotics still laid within single manipulation. Due to this strong increase in development, through new concepts within kinematics, dynamics, and control systems, more research towards CM became more and more common [7,8]. In 1987, a cooperative work environment between robots and the objects to be handled, using a hybrid position/force control method, was developed [9].

It was in the 1990s that research into CM rapidly accelerated. The most researched topic is the control between different manipulators when they are handling objects. The different control systems that were investigated, were force/motion control [10], impedance control [11], model-based coordinated control [12], adaptive control [13], kinematic control [14], and task-space regulation [15]. These control techniques were developed mainly on objects that are considered rigid themselves while in a rigid grasp. Flexible object handling was much less of a concern at that time and is still a research topic today [7]. At the same time, research was performed with the focus on the synchronization of multiple cooperative manipulators. In 1994, the first attempt was made to have two manipulators perform a synchronized movement. It involved a system where 21 axes were controlled [4].

Subsequently, the research focus shifted towards intelligent control. This can be achieved with advanced computations, logical reasoning, deep learning, complex path planning strategies, and collaborative behavior [4]. The availability of novel, more advanced control methods has led towards an increase in the number of applications within CM over the years. Most prior research is based on CDM due to the higher complexity of CMM applications. However, with the continuous emergence of innovative technologies, the proportion of CMM applications has increased. This work applies the CMM applications in the domain of CM and compares with CDM-based approaches.

2. Hierarchical Divisions

Before heading to the state of the art, some clarifications and definitions within the CM topic need to be made. Since this topic can be widely discussed, it is important to establish the boundaries of what will be covered in this paper. In general, CM is one of the six main topics within the overall multi-robotic application domains [16]. The six general domains are: (i) surveillance and search and rescue, (ii) foraging and flocking, (iii) formation and exploration, (iv) CM, (v) team heterogeneity, and (vi) adversarial environments [16]. Within the CM topic, different divisions can also be made. An object can be handled cooperatively with multiple manipulators [17] or with multiple fingers [18]. Dexterous manipulation is a topic on its own and multi-fingers systems are often categorized as end-effectors. The end-effectors, used for CM applications, and the principles of grabbing an object, will be discussed without considering all the topics of dexterous manipulation. The focus will be on cooperative manipulators, starting from the work on dual-manipulators and expending to multi-manipulators. Within this work, a manipulator is defined as: “A robotic device with x DOFs (1 or more) and a single end-effector that can physically and directly manipulate the environment by using joints and motors to control its movements, such as revolute joints for rotation and prismatic joints for back-and-forth motion” [19]. In opposition to common understanding, manipulators are not exclusively robot arms. A single linear z-translation such as picking up a product with a suction cup, could also be defined as a manipulator.

2.1. Task Divisions

In addition to the localization of CM within general robotic application domains, two different types of cooperation tasks appear within CM:

• Simultaneous coordinated: different manipulators are performing the same task on the same object. For example two or more manipulators are transporting an object from position A to position B;
• Task coordinated: different manipulators are performing a different task on the same object within the same time slot. For example, one manipulator holds a cup while another fills the cup with tea [8].

• Additionally, mobile robotics are not considered to contribute to the number of manipulators. Two examples provide more explanation:

• In some applications, manipulators are placed on a mobile system, giving them more freedom to operate in large areas. As shown in Figure 2, three manipulators are placed on their mobile platform to transport an object from position A to position B. This is referred to as CMM because three manipulators are simultaneously transporting the object. The mobile platforms, supporting the manipulators, are not considered in the number of manipulators. It is a cooperative multi-manipulator system with three manipulators [20];
• The second example involves an application where the manipulators are positioned on the same linear track. The two manipulators can move along the track to take the product and transport it from point A to point B. In this case, the linear track causes the movement of the manipulators but does not support the object. This system is called a CDM system with transportation on a linear axis (shown in Figure 3) [21,22].

2.2. Types of Cooperation

In addition to the different tasks that can be done with CM, a division can also be made into the different types of cooperation when tasks are performed by several manipulators. The following subdivision is made, and examples of each type are provided in Figure 4:

• Fixation-fixation: several examples of this are:
  • object transportation from point A to point B (Figure 4a) [20,23];
  • opening objects. For instance: one manipulator holds a bag and the second manipulator pulls the bag open, holding it in the opening position;
  • assembly of two or more objects: both objects are fixated and moved into each-other [6].
• Fixation-tooling: an example of this is cooperative a grinding application. One manipulator holds the object in place while the second manipulator performs the grinding operation (Figure 4b) [24];
• Tooling-tooling: car painting is an example of this. Each manipulator has a tool to spray paint the car, and none of the manipulators handle the car itself (Figure 4c) [25].

• In this context, fixation means that the manipulator is handling an object, while tooling means that the manipulator is performing operations with its tool on the object.

  Figure 4. Different types of cooperation: (a) is the fixation-fixation cooperation type where two manipulators are holding an object simultaneously, reprinted from Ref. [26], (b) is the fixation-tooling cooperation type where one manipulator is holding the object and the other one is doing an grinding operation on the object, adapted from [24], (c) is the tooling-tooling cooperation type different manipulators are spray painting a car. None of the manipulators are holding the car, reprinted from Ref. [25].

  Figure 4. Different types of cooperation: (a) is the fixation-fixation cooperation type where two manipulators are holding an object simultaneously, reprinted from Ref. [26], (b) is the fixation-tooling cooperation type where one manipulator is holding the object and the other one is doing an grinding operation on the object, adapted from [24], (c) is the tooling-tooling cooperation type different manipulators are spray painting a car. None of the manipulators are holding the car, reprinted from Ref. [25].

3. State of the Art

The following section gives an overview of the state of the art in CDM and CMM. Two different approaches appear in the discussion. The first approach discusses the state of the art of CM in general. The reason for this is that both CDM and CMM lay in the simple ground principles. The second approach compares both CDM and CMM with each other to identify where the differences are. Figure 5 gives an overview of the different topics discussed in this section.

(i) The section starts with an historical timeline of the different applications placed in chronological order (Section 3.1). (ii) Modeling is discussed for the CM domain (Section 3.2). (iii) Control methods and architectures are compared between CDM and CMM (Section 3.3). (iv) Planning strategies are mentioned for both CM domains (Section 3.4). (v) Different manipulators setup are explained in Section 3.5. (vi) A division in types of cooperation is discussed in Section 3.6. (vii) Next, the sensing part is discussed (Section 3.7). (viii) The vision topic is explained in Section 3.8. (ix) Different end-effector systems are explained in Section 3.9. (x) A comparison between the application domains is made in Section 3.10.

3.1. Historical Timeline

To contextualize the evolution of cooperative manipulation research, it is useful to examine how publications in CDM and CMM have developed over time. The historical progression of these two approaches reflects broader technological advances, shifts in research priorities, and the increasing complexity of robotic tasks. The following

Table 1. Overview with the different publications, in CDM and CMM, respectively, divided in time periods.

Period	CDM	CMM
[1980–1990[	[27]
[1990–2000[	[4,28,29,30,31,32,33,34,35,36]	[12,15,34,37,38,39,40,41,42,43,44]
[2000–2010[	[20,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73]	[70,74,75,76,77,78,79,80]
[2010–2020[	[21,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135,136,137,138,139]	[127,131,132,133,140,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155]
[2020–now[	[5,6,7,22,24,156,157,158,159,160,161,162,163,164,165,166,167,168,169,170,171,172,173,174,175,176,177,178,179,180,181,182,183,184,185,186]	[25,174,175,186,187,188,189,190,191,192,193,194,195,196,197,198,199,200]

provides a structured overview of key publications, grouped by decade, highlighting how interest in CDM and CMM has emerged, expanded, or stagnated during different periods.

As explained in the introduction, during the 1980s the primary focus in robotics was the development of single manipulators. However, during the 1990s, several control algorithms were introduced that enabled not only the control of individual manipulators but also the control of multiple manipulators performing cooperative tasks. This shift led to a noticeable increase in research related to both CDM and CMM.

During the 2000s, an increase in CDM applications became visible, while research on CMM applications slightly decreases. This can be attributed to the higher complexity associated with CMM, in terms of control, planning, modeling, etc. Moreover, CDM forms the conceptual foundation for CMM, as it represents the simplest form of CM. Consequently, CDM required a good foundation before more advanced research in CMM could progress.

In the following decade (2010–2019), research in CDM expanded even more rapidly. During this period, a clear shift occurred from fundamental CDM research toward more applied studies and laboratory demonstrators [8]. As a result of this evolution, CMM applications also became more relevant.

Up and until today (early 2026), both research domains continue to grow in terms of innovative developments. CMM now appears in more research applications than in previous decades. As manipulation tasks become increasingly complex, CMM is emerging as a viable and often necessary approach alongside CDM. This trend confirms that CMM is gaining ground within the field of CM, although CDM remains an essential foundation for future advancements. In addition, demand for industrial applications is expanding. The shift from lab demonstrators to industrial setups within CM is becoming a reality in the near future. A visualization of the publication trends is provided in Figure 6.

3.2. Modeling

In the context of robotic modeling, the objective is to develop mathematical representations capable of predicting manipulator behavior. The analysis of such models enables efficient system design and effective control methods and architectures. Modeling involves several components, among which the kinematics and dynamics of the manipulator are the most fundamental. Kinematic analysis focuses on determining the position, velocity, acceleration, and jerk of each joint during motion from point A to point B, as well as the kinematic links between the joints. A separation can be made between joint-space modeling and task-space modeling. As expected, increasing the number of joints within a kinematic chain leads to a corresponding increase in system complexity. Dynamic analysis, in contrast, examines the forces and torques generated during motion. In addition, the payload of the object in CM is important. Differences in center of mass and dimensions can influence the behavior and dynamic modeling drastically. Control methods and control architectures also constitute essential aspects of the modeling process and will be discussed in more detail in Section 3.3 [26].

Naturally, this all started with single manipulators. It took until the 1980s for CM to really gain traction. As CM became more prevalent, another problem arose. While it was relatively simple for single manipulators to handle objects, this was not the case for CM. It is complex to handle an object simultaneously, because of the closed kinematic chain that is formed in CDM and CMM [26]. Adding more manipulators obviously increases the complexity of CM and applies more computational power to the system, because the manipulators must receive feedback from each other in order to transport the object or perform an operation on the object [8]. An illustration of a closed kinematic chain is given in Figure 7 [26].

In order to effectively manage the complexities of object handling, different approaches were suggested and developed to combat this challenge. In CM, objects can be handled with fixed/rigid gripping methods [46] or dexterous hands and/or compliant soft structures [6,201]. Additionally, a division can be made based on the contact surface between the grippers and the object. This division includes three categories: (i) rolling contact [72], (ii) point contact, and (iii) soft contact [202,203]. Further distinctions can be made within point contacts: (a) common point contact without further constraints [48], (b) frictionless point contact [202,203], and (c) frictional point contact [204]. Closely related to this paragraph, additional information about different end-effectors used in CM, can be found in Section 3.9.

3.3. Control Methods and Architectures

In CM systems, the second major component concerns the control of manipulators. This aspect has been widely investigated due to the large variety of available control approaches. In the literature, a distinction is typically made between the control method applied and the control architecture employed. A comprehensive overview of control methods and architectures for CDM applications is provided by [8].

In the present work, this overview is extended by incorporating the corresponding control methods and architectures used in CDM and CMM applications.

Table 2. Overview control methods divided in non-adaptive and adaptive control.

Control Method	Non-Adaptive	Adaptive
Hybrid force/position control	x	x
Impedance control	x	x
Position-based control		x
Admittance control		x
Vision-based control		x

summarizes the control methods applied across CM, while the subsequent discussion compares the control architectures used in CDM and CMM to highlight their similarities and differences.

Regarding control methods, a general separation can be made into two domains: (i) non-adaptive control methods and (ii) adaptive control methods. The difference between these methods is that non-adaptive systems know the exact dynamics, parameters, and disturbances in the initial state of the manipulator, assuming no wear of the components. These systems always assume the factory-new state of the manipulators and do not adapt over time. Adaptive control methods, on the other hand, can recognize the wear of components and adapt their control system accordingly. One of the first developments within non-adaptive systems is mentioned in [12], while the first developments within adaptive control methods were made in [103].

Within non-adaptive control methods, a split into two domains can be made: (i) hybrid force/position control and (ii) impedance control. Both methods were investigated, starting in the early 1990s, with the first non-adaptive impedance control system described in [11] in 1992 and the first non-adaptive hybrid force/position control in [12] in 1993. In adaptive control methods, a division can be made into five different methods: (i) hybrid force/position control [103], (ii) impedance control [115], (iii) position-based control [205], (iv) vision-based control [206], and (v) admittance control [142].

In more recent developments, intelligent control architectures, such as neural networks and fuzzy control, have been implemented to make adaptive control systems more intelligent, more efficient, and adaptive to dynamic environments. The goal of these systems is to address the lack of information about the dynamic model mapping [8,17,26]. Intelligent control methods are combined with adaptive control methods due to their overlapping characteristics. Table 2 provides the different control methods.

3.3.1. CDM Control Architectures

Every control method comes with a control architecture or a combination of different control architectures. In non-adaptive control methods, several control architectures were used, including hybrid position/force control and impedance control. If a manipulator uses a hybrid position/force control method, two types of control architecture can be employed: computed torque [12] and feedback linearization [31]. If a manipulator uses the impedance control method, four different architectures can be used: computed torque [11], proportional-integral-derivative (PID) [46], proportional-derivative (PD) [117], and sliding mode control, which is used in several object handling applications [164].

In contrast to non-adaptive control methods, where control architectures are rather scarce due to less further developments, there is a wider range of recent developments within the control architectures of adaptive control methods. As mentioned above, combinations between adaptive control systems and intelligent control architectures are possible. Examples of control architectures for adaptive hybrid position/force control modes include:

• Function approximation technique-based control combined with linear observer-based control [161];
• Adaptive proportional derivative control [104];
• Adaptive backstepping control [101];
• Adaptive fuzzy control combined with backstepping control [120];
• Fuzzy neural network control [143];
• Kalman filter combined with iterative learning control [106];
• Radial basis function neural network control combined with model-based control architecture [134];
• Synchronous sliding mode control combined with radial basis function neural network [160];
• Passivity-based hybrid force/position control [35];
• Vibration suspension control [50];
• Leader-follower force control [185].

• For impedance control, a wide variety of architectures can be found in the literature. These include:

• Adaptive impedance combined with a leader-follower control method [108];
• Proportional derivative control combined with gravity compensation [115];
• Adaptive variable impedance control combined with sliding mode control [165];
• Sliding mode control combined with adaptive radial basis function neural network [24];
• Adaptive hybrid impedance control [207];
• Impedance control combined with trajectory coordination [182];
• Impedance control to achieve compliant control [184].

• Besides these combined architectures, two individual methods were discussed: adaptive neural networks [159] and adaptive fuzzy control architectures [98,135]. Different position-based control architectures include:

• Adaptive proportional derivative architecture [102];
• Fuzzy sliding mode control [208];
• Adaptive radial basis function neural network combined with sliding mode control [99];
• Adaptive radial basis function neural network combined with backstepping control and sliding mode control [100];
• Dynamic surface control combined with radial basis function neural network [158].

• The fourth control method in adaptive control systems is admittance control. In admittance control methods, three different control architectures exist:

• Adaptive backstepping [187];
• Adaptive variable admittance [107,109];
• Adaptive neural networks [142,209].

• The last main control structure in adaptive control systems is the vision-based control method. Further information about vision is mentioned in the Section 3.8. In [8], a brief overview of flexible manipulators’ control methods and architectures is given. The reason these architectures are not mentioned is that they are less relevant for CMM research. Flexible manipulators, such as product manipulation with dexterous fingers, are topics on their own.

As mentioned in Section 2.1, manipulators can also be mounted on a mobile base. This configuration provides greater workspace flexibility compared to fixed-base manipulators. For CM on mobile platforms, several control approaches have been explored in the literature. In [71,72,127,128] adaptive coordinated control architectures are proposed to handle manipulators with relative motion, enable interaction with rigid environments, and transport objects between locations, respectively. A related strategy is formation control, as discussed in [133]. Neural network-based telemanipulation is used in [132] to control a single master device that commands multiple slave manipulators to cooperatively manipulate an object. Many CM approaches rely on force/torque measurements to coordinate multiple manipulators. However, Ref. [177] applies prescribed-performance control to manipulate various objects without requiring force/torque sensing. In [139] a crawler robot uses a compound manipulation control approach to improve the reach of end-effectors to grasp rigid objects. CM is not limited to ground robots. Aerial and underwater cooperative manipulation are also active research domains. In [130], multiple aerial vehicles manipulate an unknown object using a constraint-based control architecture. Distributed dynamic control and task-priority control are applied to underwater manipulators for object manipulation and transportation, as presented in [131].

3.3.2. CMM Control Architectures

For CDM, the control aspect is an important factor to determine. The same applies to CMM systems. The control architectures of CDM are the global base to build upon towards CMM applications. Similarly, the proportion of non-adaptive control methods is lower in comparison with adaptive control methods. As the control architectures of CMM are less investigated, a portion of the literature papers has an introduction and modeling section about CMM. However, the results are more related and demonstrated to CDM. These control architectures are the following:

• Computed torque control [12];
• Fuzzy neural network position/force hybrid control [143];
• Dynamic surface control [74];
• Adaptive backstepping admittance control [187];
• Adaptive fuzzy backstepping control [146];
• Adaptive coordinated control [127].

• Several publications have introduced control architectures specifically for CMM. Many of these approaches overlap with those used in CDM, although some techniques are more prominent in CMM. Adaptive control clearly dominates in CMM applications. An observer-based adaptive sliding mode control combined with proportional derivative, to grab an object with three manipulators, is discussed in [140]. An adaptive fuzzy hybrid intelligent position/force controller is proposed in [75]. Adaptive synchronized control is used in [76] to guide an object synchronously during assembly tasks, while an admittance-based adaptive cooperative control scheme is presented in [142]. Another adaptive method is adaptive consensus control [77], a multi-agent technique that does not appear in the CDM literature. Sliding mode control is another commonly used architecture in both manipulation systems. For welding applications, sliding mode control is applied to cooperative multi-manipulators in [149]. Incremental motion control is used for the simultaneous transport of large objects in [41]. In human-robot interaction, a CMM system is developed to transport large objects together with human operators. Therefore, impedance control combined with leader-follower control is proposed in [155]. Another leader-follower application is proposed in [186] where multiple mobile manipulators transport a rigid object simultaneously. As mentioned in CDM, CM is also performed in underwater environments [131]. The control architectures used there can be extended to CMM systems. In multi-manipulator assembly systems, an agent-based control method is used to coordinate three manipulators operating in the same workspace [42]. A distributed neural network for solving the motion generation problem in multi-manipulator systems is presented in [198]. In telemanipulation, neural network-based control [132] and fuzzy control [152] are proposed for cooperative object transport with multiple manipulators. Motion control approaches using the Open Motion Planning Library and networked mobile platforms are discussed in [153,192]. As mentioned in CDM, formation control and observer-based control are also relevant and can be extended to CMM. Solutions are provided in [133,154].

  Table 3. Summary control architectures in CDM and CMM applications.

  Control Architecture	CDM	CMM
  Non-adaptive
  Computed torque	[11,12]	[12]
  Feedback linearization	[31]
  PID control	[46]
  PD control	[117],
  Sliding mode control	[24,164]
  Adaptive
  PD control	[102,104,115]	[140]
  KF control	[106]
  Sliding mode control	[99,100,160,208]	[140,149]
  Function approximation technique based control	[161]
  Admittance control	[107,109,187]	[142,187]
  Impedance control	[108,165,182,184,207]	[155]
  Radial basis function neural network	[24,99,100,134,158,160]
  Model-base control	[134,165]
  Iterative learning control	[106]
  Neural network	[132,143,159]	[132,143,198]
  Dynamic surface control	[142,158,209]	[74]
  Backstepping	[100,101,120,187]	[146,187]
  Passivity based control	[35]
  Vibration suspension control	[50]
  Fuzzy control	[98,120,135,143,208]	[75,143,146,152]
  Leader-follower control	[108,185,186]	[155,186]
  Compound manipulation control mode	[139]
  Gravity compensation	[115]
  Observer-based control	[161]	[140,154]
  Coordinated control	[71,72,127,128]	[127]
  Formation control	[133]	[133]
  Prescribed performance control	[177]
  Constraint-based control	[130]
  Task priority control	[131]	[131]
  Distributed dynamic control	[131]	[131]
  Synchronized control		[76]
  Consensus control		[77]
  Motion control		[41,153,192]
  Agent-based control		[42]

  provides a comparison between CDM and CMM control architectures.

3.4. Planning Strategies

Besides modeling and control of the manipulators, planning is a third relevant topic to discuss with regards to CM. Without a motion planning method, manipulators will not know how to reach the object without colliding with other known objects or with each other. A first division can be made for planning algorithms where manipulators plan their paths separately in a specific order, timing, or with the use of a generated road map or topological graph to avoid collisions, also called decoupled planning. The different path planning strategies based on the planning strategy are called: (i) prioritized planning, (ii) fixed-path planning, and (iii) fixed-road map coordination. This division was made by Lavalle in [210]. A comparison in planning strategies for CDM and CMM is given in

Table 4. Comparison planning strategies in CDM and CMM applications.

Planning Strategies	CDM	CMM
APF planning	[105,211,212,213]	[79]
Sampling-based planning	[89,136,167,210]	[175]
Optimization planning	[22,114,137]
Neural network and reinforcement learning-based planning	[214]	[192]
Time optimal planning	[137]	[200]
Optimal energy consumption planning	[213,215]
Welding trajectory planning	[126]	[151]
Incremental displacement planning		[41]
Agent-based planning		[42]
Elastic-strip planning		[79]
Strategy-based planning		[79]

.

3.4.1. CDM Planning Strategies

A more recent division of planning strategies for CDM is mainly discussed in [8]. Four different planning algorithms are mentioned: (a) potential field (APF) based methods [105,211,212,213], (b) sampling-based methods (based on the Rapidly-explored Randomized Tree (RRT)) [89,136,167,210], (c) optimization methods [22,114,137], and (d) neural network and reinforcement learning-based methods [214]. Path planning techniques were mostly developed with the main objective of achieving the correct pose that the manipulator needed to move to. Most movements were not the shortest trajectory to reach a specific point, such as going through singularities when moving from point A to point B [17]. An industrial application where singularities can occur is welding. Here, a specific welding trajectory is specified and singularities are sometimes unavoidable. However, CDM welding applications give more freedom to operate in a fixed-tool configuration. One manipulator can rotate/position the object in the correct configuration to avoid singularities during the execution of the welding trajectory of the other welding manipulator. An offline motion planning strategy is discussed in [126] for planning of CDM welding applications.

Over the years, more strategies have been developed with the intelligence to consider the most efficient way to achieve a pose with specific end-effector positioning, while also minimizing energy consumption. These techniques result in higher productivity and lower energy consumption, respectively. These two strategies are called time-optimal path tracking [137] and energy consumption path tracking [213,215]. Time-optimal strategies are more advanced techniques based on the previously cited strategy of fixed-path planning. Examples of fixed-path planning from the 1990s are mentioned in [216,217].

3.4.2. CMM Planning Strategies

On the other hand, planning strategies are also necessary for CMM. Manipulators need to reach the end location without colliding with each other or with surrounding objects in the shared workspace. In [192], task planning strategies are generated using reinforcement learning. The proposed planning algorithm used in [41] is an algorithm that uses incremental, linear, and angular displacements. These parameters are calculated using homogeneous transformations matrices to achieve the desired motion of the manipulated object. A motion planning strategy for cooperative multi-manipulator systems to assemble different parts is mentioned in [175]. For manufacturing applications, a novel non-collision sample-based path planning strategy is developed for CMM [175]. As mentioned in Section 3.4.1, a control method for an agent-based manipulator system is used. The agent-based path planning algorithm is also discussed [42]. In [79], three planning strategies are implemented for CMM applications, namely: (a) the elastic-strip planning method, (b) the strategy-based planning method and (c) the potential field planning method. In welding applications, a path planning method is proposed to weld different parts using a combination of two welding manipulators and one fixation manipulator [151]. Finally, time-optimal path tracking is also performed for CMM [200].

3.5. Manipulator Setup

The manipulators also come in different forms when CM tasks. In general, four main categories can be given for both CDM and CMM, only the number of manipulators is different: (a) a fixed manipulator setup [21], (b) a mobile manipulator setup [176], (c) human-like manipulators [138], and (d) a torso/waist setup [151]. Fixed manipulators stand on a fixed base separated from each other and cannot move around. Meanwhile, mobile manipulators are installed on a base that can move around. Torso or waist setups are manipulators that are fixed on the same base, so not separated from each other. Human-like manipulators, as the name suggests, are manipulators that resemble a human. The different manipulators set-ups are demonstrated in Figure 8.

3.6. Division in Types of Cooperation

As mentioned in Section 2.2, three different types of cooperation exist: (i) fixation-fixation, (ii) fixation-tooling, and (iii) tooling-tooling.

Table 5. Overview about the classification of the cooperative multi-manipulator applications into the different types of cooperation.

Type of Cooperation	CDM	CMM
Fixation-fixation	[5,6,7,15,20,21,22,28,29,30,31,32,33,34,36,46,47,48,49,50,51,52,53,54,55,56,58,59,60,61,62,63,64,65,66,67,68,69,70,71,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,103,104,105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,125,127,128,129,130,131,132,133,139,156,157,158,159,160,161,162,163,164,165,167,168,170,171,172,173,174,175,176,177,180,181,182,183,184,185,186,218]
Fixation-tooling	[24,55,58,85,124,126,199]	[76,78,145,149,151,191]
Tooling-tooling	[57,166,169,174]	[25,76,148,174,191,194]

gives an overview of the applications sorted in the three different types of cooperation for CDM and CMM, respectively.

The results from Table 5 are further visualized in Figure 9. In general, it can be concluded that the fixation-fixation type dominates within the types of cooperation. In total, 87.63% of the publications fall under the fixation-fixation type, of which 63.98% are CDM applications and the remaining 23.66% are CMM applications. The two other types of cooperation account for the remaining 12.37%. Of this, 6.99% are fixation-tooling applications, with 3.76% in the CDM domain and the other 3.23% in the CMM domain. Finally, the tooling-tooling type represents only 5.38%, of which 2.15% and 3.23% are CDM and CMM applications, respectively.

Within the fixation-fixation type, the focus is on handling both rigid/non-deformable and flexible/deformable objects. The reason this type is so prominently represented within the literature is that it often forms the basis on which fundamental research is conducted. Topics such as control and planning strategies typically start from the idea of grasping a rigid/non-deformable object, after which the developed architecture or planning strategy is applied. The other types of cooperation build on this fundamental research and apply it to examples that occur in practice.

Examples of fixation-tooling within the CDM domain include machining processes wherein one manipulator holds the object while the other manipulator, equipped with a tool mounted on the end-effector, performs the operation [24]. Additive processes also make use of fixation-tooling, with both CDM and CMM possibilities [126,149,151,191,199]. Another example is pouring a drink [55,85] or adding a sugar cube to coffee [58]. A further household example is a system in which two dual-manipulators make pancakes, which can be seen as a four-manipulator system [145]. Vision systems for grasping objects also fall within this category [124]. Finally, the assembly of components is also considered fixation-tooling, where the part to be assembled is regarded as the tool [76,78].

Within the tooling-tooling category, several examples can also be found. For instance, Ref. [57] presents a dual-manipulator system that plays music on a drum set. A vision system on each end-effector combined with a gripper to perform operations also belongs to the tooling-tooling type [166]. Simultaneous harvesting of fruits on the same tree is another example [169]. A range of tooling-tooling applications is proposed in [174], including automotive assembly lines where different parts are mounted together simultaneously. Additionally, additive processes with multiple manipulators, such as paint spraying [25] and 3D printing [148], fall into this category. On the other hand, task allocation and scheduling with collision constraints are discussed in [194]. Finally, Refs. [76,191] mention fixation-tooling applications, but they also include applications that belong to the tooling-tooling domain.

3.7. Sensor Systems

Within CM, a variety of sensors is used to receive information. This not only includes internal feedback information regarding the kinematics and dynamics of the manipulators but also external sensor feedback information about the surroundings and environments where the manipulators are located. The most common feedback sensors are force and torque sensors inside the joints or end-effectors of the manipulators. They can be used separately or simultaneously to provide feedback to the control system. Examples of force-only sensors are mentioned in [86,121,207]. Torque sensors used separately are mentioned in [17], and the combination of the two is discussed in [22,45,73,127].

To measure distance to an object, two different approaches are used. The first sensor is a structured light sensor, and the second sensor is a time-of-flight sensor [91]. To scan the environment, IR sensors are used, as mentioned in [87]. These three types of sensors also relate to the vision perspective, where these methods are used to measure depth to objects. Additionally, the position of the joint states of the manipulators must also be measured, mostly done by gyroscopic sensors [219]. Position sensors [62] and accelerometers [219] are integrated to determine the exact placement and the accelerations and decelerations of the manipulators. Recently, central IMUs are units where different sensing components gather information about the state of the manipulators in terms of motion and forces. Various examples of these are explained in [138].

3.8. Vision Systems

In addition to sensors, vision systems can provide feedback to manipulators when doing CM tasks, and provide the relative pose of the manipulator to the environment (visual servoing). Three general categories can be made within the different vision systems, and within those categories, a further division into three different topics each. The three main categories are: (i) different vision feedback techniques, (ii) the mounting position/location of the camera, and (iii) the number of cameras used in the applications [8]. A general overview of the main and subcategories is given in Figure 10.

Looking into the types of vision feedback techniques, a division into position-based, image-based, and a combination of the two techniques is made. The position-based technique uses an error signal, which is the difference between the position of the known place of the 3D object and the final known position used as a local camera frame. This error signal generates a control signal to adjust the position of the manipulator to the final pose. Examples of position-based techniques are presented in [73,220].

With the image-based technique, the error signal is calculated as the difference between the desired and actual position of a certain parameter within the parameter space of the camera. With this difference, a new control signal is calculated, which is closer to the end position of the manipulator. Image-based techniques are mentioned in [206,221], and different hybrid applications, which are a combination of both methods, are discussed in detail in [8,17].

The second main category is the mounting position of the cameras. The subdivision here can be made into end-effector mounted or eye-in-hand, fixed position in the workspace or eye-to-hand, and active robot heads. With eye-in-hand applications, the camera is directly mounted on the end-effector of the manipulator [81,222]. For the eye-to-hand applications, the camera is not directly mounted on the end-effector of the manipulator, but placed in its surrounding environment [73]. The last configuration is the active heads, where the cameras are placed on top of the manipulator’s torso to act like human eyes [61,223].

The third main category is the number of cameras that are used in the CM setup. The split is made into monocular, binocular, and multi-ocular. Monocular means that one camera is used. With binocular, two cameras are used in the cooperative application, and multi-ocular means more than two cameras. Examples of binocular applications are [116]. A binocular application is discussed in [224], and a multi-ocular system is developed in [225].

Visions systems are also used for deep learning processes. In [180], a camera on the robotic setup is used to perform goal-conditioned imitation learning. Here, the CDM system learns how to peel a banana by imitating first the human demonstration of it. A eye-in hand setup is used here. Another imitation learning application is mentioned in [181]. In this case, a eye-to-hand camera setup is used to learn different tasks, including a pick-and-place task of a common object.

3.9. End-Effector and in Hand Manipulation

End-effectors can be characterized in a continuous spectrum between rigid and soft grippers [226]. The more conventional grippers are typically more rigid with fewer degrees of freedom and less compliance. More compliant, soft, grippers typically have more degrees of freedom and, therefore, are often less precise.

A large part of the prior work on gripper technologies is on the aspect of grippers in the manufacturing industry. In this industry the shape, size, weight, nature and material of the objects that are manipulated are mostly known. Within these applications conventional, hard, grippers are mostly used. Conventional or hard grippers are often designed to grasp one or multiple specific objects. Four types of conventional grippers exist: impactive grippers, astrictive grippers, ingressive grippers and contigutive grippers. Impactive grippers are grippers which exert forces with impact against the object surface. Often an actuator is moving rigid fingers against the surface of the object. Astrictive grippers are grippers that generate a continuous holding force without exerting a compressive stress onto the object. Grippers that permeate the surface of an object to fixate the object, for instance with pins or needles are ingressive grippers. Lastly, when tension forces are used to adhere to an object are contigutive grippers. This tension can be generated by chemical or thermal adhesion [227]. The three most common examples of standard rigid/conventional tooling are: (i) parallel grippers, (ii) vacuum grippers, and (iii) magnetic grippers [26].

In cases where the objects are less known or defined, soft grippers play a more important role. Inspired by nature and the mechanical soft structures found in insects and animals, various soft grippers have recently been developed. The soft structures and mechanical compliance allow for superior manipulation, reduced control complexity and safer interaction with humans and natural environment. This compliance can be introduced in three ways: by controlling the actuation, stiffness or adhesion from the gripper [226,228,229,230,231,232].

Aside from the prehension type, grippers can be categorized by the type of manipulation as well. When the grippers are designed to fixate one or a set of objects without changing the orientation of the object with respect to the gripper after grasping, the gripper is considered fixed shape tooling. Examples of tools specifically developed for one application can be found in [37,233]. Alternatively grippers can be articulating. Grippers are considered articulating when grippers have multiple degrees of freedom and can perform precise and complex movements with the grasped object. Articulated tools are very human-like and allow for enhanced agility and adaptability, called dexterous grippers. Examples of articulated tooling are dexterous hands with several fingers [6,205]. Most of the grippers among the conventional and soft grippers are fixed-shape tooling.

As an alternative to articulated tooling, fixed-shape tooling can be exchanged to facilitate a suitable tool for each manipulation task. This can be done with tool changers or by mounting multiple grippers on the same robot flange. The tool changer allows the (fixed-shape) tools to be (dis)mountable from the end-effector [88,234]. While installing multiple grippers on the same robot flange allows direct switching between grippers by changing the robot pose. However, additional weight is applied to the flange and therefore the velocity of the robot is reduced.

Often articulated tooling allows for in-hand manipulation due to the extra degrees of freedom. Since this in-hand manipulation is moving the objects with different manipulators that are integrated within one end-effector, this could arguably be considered CMM on itself. However, within this field, various aspects have been investigated. Since this type of manipulation is very specific and on a smaller scale, the work regarding this topic is not considered to be within the scope of this article.

3.10. Application Domains

Various applications domains are targeted in CDM and CMM across the literature. The most investigated application domain is object handling, as this is the starting point for different research topics such as control, planning, etc. Within object handling, a division can be made, based on the type of object that manipulators are handling, namely: (i) rigid/non-deformable, and (ii) flexible/deformable objects. The ratio of applications is high, with rigid/non-deformable objects being more investigated than flexible/deformable applications. This is due to the added complexity brought on by their flexibility. Additionally, a higher computation power is needed for flexible object manipulation. This is due to high complexity of modeling, calculating, and estimating the object behavior. Examples of rigid/non-deformable objects are mentioned in [28,31,36,47,48,49,50,51,153]. A visualization of rigid/non-deformable object manipulation is given in Figure 11A. Examples of flexible/deformable objects, such as clothes folding or manipulation of flexible cables, are explained in [60,82,90,97]. Figure 11B shows the example of folding clothes. An overview of all the applications categorized in rigid/non-deformable, and flexible/deformable objects for CDM and CMM applications, respectively, is given in

Table 6. Object handling applications in CMM divide in the type of object that is handled.

Type of Object	CDM	CMM
Rigid/non-deformable	[5,6,7,15,20,28,29,31,32,34,36,47,48,49,50,51,55,56,58,62,64,65,68,69,70,85,87,89,91,92,93,94,95,96,99,100,101,103,104,105,106,107,108,109,110,111,112,115,116,117,118,119,120,122,123,124,125,127,128,129,130,131,132,133,139,157,158,159,160,161,162,163,164,165,167,169,170,172,173,174,175,176,177,184,185,186,218]	[12,31,34,37,39,40,41,43,44,74,75,76,78,80,131,132,133,141,142,143,144,145,146,147,151,152,153,154,155,174,175,186,187,188,189,190,192,195,196,197,198]
Flexible/deformable	[21,22,30,33,46,53,55,60,62,63,66,82,83,84,86,88,90,97,121,167,168,171,180,182,183]	[38,70,127,145,150,189]

.

The conclusion is that rigid/non-deformable applications have the upper hand compared to flexible/deformable object manipulation. This applies to both CDM and CMM applications. In addition, there are more CDM applications in both topics then CMM applications. Generally, flexible/deformable object manipulation can be investigated further in the near future.

A more specified list of application domains in CM includes: (a) manufacturing, (b) construction, (c) domestic/household, (d) healthcare, (e) agriculture, (f) aerospace, and (g) hazardous environments. The most common example of manufacturing applications is the assembly of different parts into one final part [6,64,81,175]. Additionally, a disassembly application is explained in [167]. This topic will grow exponentially in the near future, considering the required disassembly of all the car and bicycle batteries the upcoming years. Hence, the replacement of ten of thousands of solar panels on roofs and solar farms. In addition, there are additive manufacturing processes such as welding [126,174,199], spray painting, or 3D printing [174]. Substrative manufacturing is also performed in CDM such as grinding [24]. In construction applications, an automated scaffold applications is mentioned in [178]. Additionally, 3D-printing applications are used to print concrete buildings simultaneously [191]. Other CDM use cases can be found in the domestic/household domain. An example of replacing items in a bookshelf is mentioned in [168], clothing folding applications are explained in [60,63,86,88,90,97,183]. Even playing music on drums is done in [57]. Aubergine harvesting with the use of a dual-manipulator is an example in agricultural applications [8]. Different approaches within space applications are mentioned in [50,166,234]. Manipulation in space is considered a separate application topic, since the influence of the environment on the manipulation task is completely different. In [130,185], aerial applications where drones transport and object in the air are explained. The final group of applications is hazardous environments. A dual-manipulator performing manipulation tasks in hazardous environments using a teleoperation system developed at CERN is an example of this [179]. Additionally, a crawler robot is performing object manipulation tasks on rough terrain in [139]. However, exploration and doing tasks underwater is also hazardous [131]. Figure 12 provides examples of different CDM application domains.

The same structure in CMM applications can be applied. A wide range of applications has been developed in additive manufacturing processes namely: (i) 3D-printing applications [148,191], (ii) spraying processes [25,76], and (iii) welding applications [149,151,188]. A range of assembly applications has also been mentioned in the literature of CMM over the past years [37,42,76,78,141,144,175,189]. In the automotive manufacturing industry, CDM is used to polish automobile parts [194]. In construction, 3D-printing applications are used to print concrete buildings simultaneously [191]. Household applications, such as handling common objects and a specific kitchen task (making pancakes), are also multi-manipulator applications [145]. Additionally, a clothing folding application is discussed in [38]. Four applications can be found in healthcare. Two studies involve a multi-manipulator for tooth arrangement in full denture manufacturing [141,144]. The other two applications are within the surgical sector: one publication discusses an endovascular intervention [150], and the other describes a rail-guided system for cross-scale targets [189]. In aerial applications, a three-manipulator construction is made for aerial transportation of objects [197]. Moving from air to water, underwater applications are mentioned in [131,196].

An overview of every application domain in CDM and CMM is given in

Table 7. CDM and CMM categorized by application domain.

Application Domain	CDM	CMM
Assembly	[6,64,81,175]	[37,42,76,78,141,144,175,189]
Disassembly	[167]
3D-printing	[174]	[148,191]
Spraying processes	[174]	[25,76]
Welding	[126,174,199]	[149,151,188]
Grinding	[24]
Polishing		[194]
Construction	[178,191]	[191]
Cleaning	[168]
Cooking		[145]
Folding	[60,63,86,88,90,97,183]	[38]
Playing music	[57]
Health care		[141,144,150,189]
Agriculture	[8]
Space	[50,166,234]
Aerial	[130,185]	[197]
Hazardous environment	[139,179]
Underwater	[131]	[131,196]

. Figure 13 provides examples of different CMM application domains.

4. Challenges and Future Work

CM comes with a lot of advantages in comparison with single manipulation. Different advantages are: (1) Greater adaptability: CM is more versatile, capable of managing complex tasks and adjusting to a variety of situations. This allows the manipulator system to perform well detailed and complex operations; (2) Enhanced load distribution: With dual- and multi-manipulators, the system can share the workload more efficiently, making it possible to handle heavier items; (3) Expanded operational range: These systems can operate over a larger area, making them ideal for tasks that require extensive reach and movement; (4) Increased reliability: In some cases cooperative manipulators add a layer of redundancy. If one manipulator fails, others can take over, ensuring continuous operation and minimizing downtime [235]; (5) Sometimes cooperative systems can perform simultaneous task handling: CM systems can execute several tasks at the same time, boosting overall productivity and efficiency [26].

However, while CM proves to be advantageous in several situations, CM comes with many challenges and drawbacks. The goal in the future is to address some challenges and turn them into advantages for specific applications. To start, CM can be used to handle complex parts or objects. However, these parts often require complex control and planning algorithms to handle the object correctly. The increased complexity is a bottleneck in the development of dual and multi-manipulation systems. Naturally, the complexity increases as the number of manipulators increases. This can also be seen in the number of applications currently available in the literature. More CDM then CMM is available.

Specifically looking at flexible object handling, research opportunities can be found. Only a few applications for both dual and multi-manipulation, compared with rigid object handling, are available. The reason for this is increased complexity. Further investigation is needed. Another challenge and potential area for future work is the extension of control methods/architectures and path planning strategies towards multi-manipulation. Communication overhead is a common problem in CM due to the exchange of large amounts of information, such as synchronization actions and collision avoidance. This can result in non-robust systems due to latency in communication signals [26]. Additionally, collision detection and synchronization between different manipulators are more complex within CMM because of the increased number of manipulators and DOF. While the knowledge and technologies are present within dual-manipulation, the extension towards multi-manipulation is still lacking. The literature shows that new methods are also used for CMM. Not all the CDM methods can be used in CMM.

From an application domain perspective, several assembly applications have been developed within CM over the years. While automated production processes for assembling parts are well-known, the disassembly process is often not considered. In the literature, only one example of a disassembly process is given in CM [167]. With the mass production of car batteries in mind today, this application domain surely has future work potential. For example, considering payload distribution is important because these batteries are heavy objects. Disassembly of components is certainly not the only example. Various industries with specific applications, such as construction and pharmaceuticals, can make use of CM.

When thinking of specific applications, tool design and modeling are essential. Most manipulators are equipped with simple end-effector grippers. More advanced task-related tool designs would be more effective in certain situations. Most cooperative manipulation applications remain highly task-specific. For each new problem, a dedicated robotic setup is typically designed, limiting the reuse of hardware and control architectures. Achieving generalization and flexibility that allow the same cooperative system to be deployed across different application domains therefore remains a significant challenge. Furthermore, cooperative manipulation involving multiple manipulators and human operators introduces additional complexities, including the need to predict human behavior and to ensure safety when humans and robots operate within the same workspace.

5. Conclusions

This paper mapped the state of the art of both CDM and CMM applications. Different fundamental topics in robotics, which involved, control, planning, modeling, sensing, vision, and end-effectors, are discussed. This results in an expansion of control and planning architectures from CDM to CMM. Additionally, different approaches for modeling, sensing, vision, and end-effectors are shown. The historical analysis highlights how interest in CM has grown over the recent decades. Industrial tasks are becoming more complex and more flexible solutions are needed. This gives new challenges and opportunities in CM. Overall provides CDM a technological foundation for CMM to expand further knowledge. Different types of cooperation and application domains are tackled to categorize every application. This revealed a strong dominance in fixation-fixation systems and a lack of fixation-tooling and tooling-tooling applications. Rigid/non-deformable object manipulation still dominates in object manipulation due to the high complexity of flexible object manipulation. Finally, the review identifies several research gaps: (i) flexible object manipulation; (ii) disassembly of components (e.g., batteries); (iii) complexity in control and planning for multi-manipulator systems; (iv) limited industrial deployment of CDM and CMM; and (v) the need for advanced, task-specific tooling and end-effector equipment.