A Survey of Wheeled Mobile Manipulation: A Decision-Making Perspective

For over a half-century, robotic systems have been deployed to extend the reach of humans in performing dumb, dull, dirty, and dangerous tasks in numerous application settings. Early deployments featured largely static, fenced, and permanently integrated solutions, customized for particular operations. However, in the industry 4.0 era, robotics has evolved to provide flexible, reconfigurable, intelligent mobility, and manipulation capabilities while operating close to humans.

Mobile manipulators offer an evolution in robotic system architectures [1] that facilitates their graduation from mere automated systems to autonomous ones. Such systems merge mobility (offered by the mobile platform) with the manipulation capabilities (afforded by the mounted articulated arm) to create an unlimited manipulation workspace. Mobile manipulators have become popular in numerous deployment verticals: industrial (factories, warehouses), domestic (household), and outdoor field settings (highway maintenance, earthmoving/excavation to free-flying satellite repair) due to the flexibility they provide in undertaking and assisting a variety of tasks [2–9]. Building on the merger of mobility and manipulation, the abstracted tasks comprise relatively broad categories such as inspection, part-picking, transportation, assembly, site-tending, and more. Ultimately, the intersection of deployment application verticals/abstracted tasks in autonomous, semi-autonomous, or completely teleoperated operational contexts results in the seemingly infinite diversity of intelligent mobile manipulation examples.

The application of the robotics (Sense-Think-Act in real time) paradigm forms the underlying basis for an embodiment of intelligence in mobile manipulators. Sense encompasses capabilities that help in sensor-based perception of the robot and the surrounding environment; Think involves information-processing and decision-making capabilities required for a particular task, and Act pertains to the physical realization of intelligent mobile manipulation within the environment. In all cases, mobile manipulation entails an underlying cyber-physical articulated mechatronic system to realize intelligent physical interactions with the environment. Depending on the engendered timescales, this intelligence capability has to be realized via various decision-support methodologies/frameworks categorized as planning (offline, deliberative), control (online, reactive), and human–robot interaction (interactive) frameworks. Ultimately, it is required to close the loop in real time and the technological constraints (on computation/communication/miniaturization) that enable (or limit) the realizable intelligent behaviors.

Many variants of the mobile manipulator architecture are possible, based on the nature of the mobile bases (rail or XY-gantry mounted system, wheeled/tracked platform, and more) and the nature of the mounted manipulator (number and type of articulations and their actuation, e.g., serial versus parallel).

Wheeled mobile manipulators are composed of a wheeled mobile robot base with one or more mounted manipulator arms [6,10–12]. They are an important and popular subclass configuration of mobile manipulators. They offer ungrounded, easy-to-use, adaptable architecture suited for numerous applications. They inherit many typical benefits of redundant manipulators such as expanded workspace (both conventional and dexterous), reconfigurability, improved disturbance rejection capabilities, and robustness to failure. In addition, they also can (i) accommodate changes in the relative configuration (using the excess degrees-of-freedom (DOFs)); (ii) detect such changes (using sensed articulations); and (iii) compensate for disturbances (using the redundant actuation within the chain), while interacting with the environment. Figure 1 depicts a wheeled mobile manipulator modulating its end-effector motion/force interactions with the external environment while using surplus internal degrees-of-freedom within the system to control the relative configuration/pose. Ultimately, this redundancy proves crucial to enhancing their ability to perform autonomously or semi-autonomously, for example, assisting a human operator during dynamic interactions with the world. Key factors/considerations to realizing superior performance in mobile manipulators include

1. Design architectures: The underlying cyber-physical system architecture, with introduced design-freedoms, is crucial—the type, number, location, and actuation of the wheels/articulations within the mobile manipulator modules play a critical role in determining the performance capabilities. Careful selection of the topology, dimensions, and configuration; hence, it is necessary in determining the capabilities and overall system performance.

2. Planning-control frameworks: A planning-control framework is required to exploit the “design-freedoms” for realizing superior intelligent behaviors. Advanced planning/control frameworks are needed to overcome environmental constraints (e.g., contacts, nonholonomic), resolve redundancy (kinematic and dynamic), minimize singular configurations of the system, modulate physical power exchange with environment (motions and forces) during task performance, and improve robustness to local controller lapses and environmental disturbances.

3. Human–Machine Interfaces: Mobile manipulation offers new capabilities to extend human reach. However, the complexity of the underlying mobile manipulation system and physical interactions with the environment require new operational paradigms. Human–robot teaming frameworks with probabilistic sensor fusion (of distributed heterogeneous spatiotemporal data) with the computational model-based frameworks (both mechanistic- and learning-based) have now emerged. This now affords intelligent decision-support strategies, exploiting the strengths of both humans/robots to improve shared decision-making, control autonomy, and transparency with ever-changing, uncertain conditions.

Several surveys on topics related to mobile manipulation have emerged in recent years. Bostelman et al. [13] present a literature survey of mobile manipulator research with examples of experimental applications. It also provides an extensive list of planning and control references as this has been the major research focus for mobile manipulators, while also factoring in performance measurement of the system. Song et al. [14] discuss the development of mobile manipulators in the aspects of motion planning, coordinating control, and multiple mobile manipulator coordination and present various problems involved and the methods to deal with them. Youakim and Ridao [15] present a survey of motion planning algorithms like sampling-based, search-based, and optimization-based motion planners for underwater autonomous mobile manipulators and attempted to answer the tough question “Which planner to choose?”. The state of the art of the most common approaches are discussed, and a set of benchmarks are presented with the aim to provide a comprehensive review as well as a qualitative/quantitative comparison of the algorithms. Moreover, aerial manipulation has been surveyed in Ref. [16]. Li [17] in his textbook extensively explores the methods of modeling, planning, and control for wheeled mobile manipulators, but mostly in the context of multirobot cooperative control [18]. Other recent surveys provide high-level insights into the practical aspects of wheeled mobile manipulation. For example, a survey of the recent advancements in mobile manipulation state of the art and technical challenges with the RoboCup design competition as the focal point was presented in Ref. [19]. A review of the system architecture and application space of collaborative mobile manipulators was put forth in Ref. [20]. Although these surveys reference contemporaneous mobile manipulator research, they do not discuss or categorize the methodologies themselves, a task that we will attempt here.

However, given the vast amount of literature, we restrict our focus to solely surveying the decision-support methodologies, i.e., planning-control and human–robot interaction in the context of wheeled mobile manipulation. While design architectures, perception, learning, and so on are important, we succinctly discuss these topics in the rest of this section (and offer suitable further references for detailed follow-ups).

From the perspective of design architectures of mobile manipulators, it is important to factor the contributions of both the wheeled mobile base and the manipulator arm. The mobility, steerability, and controllability of the overall wheeled base depend largely on the type, nature, and locations of the attached wheels [21]. Hence, the decision-making aspects of such vehicles need to be explicitly taken into account the maintenance of the kinematic compatibility conditions. Recent trends highlight a renewed interest in addition of active or passive articulations between the wheels and chassis to ensure kinematic compatibility [22–25]. From a manipulator perspective, a careful selection of the topology (serial versus parallel, numbers of joints) as well as the structural parameters (link-lengths) and configuration parameters (joint angles) all affect the ultimate performance of the manipulator individually (as well as the part of the mobile manipulator system) [26,27]. Further readings on design architecture topic are available in previous articles [28–32].

Perception is another important dimension to harnessing intelligence in both mobility and manipulation tasks – traditionally, a multistep process (of sensing, signal processing, and cognition) underpins the perception module. Navigation in both indoor and outdoor settings, as surveyed for wheeled mobile robots [33], is relevant here—as are surveys of further cognitive-processing to develop situational awareness from perception [34]. Worthy of special mention are the adaptations of the active perception paradigm in the context of mobile manipulation in the form of vision-enabled manipulation including object recognition, grasping, pose estimation, etc. Detailed discussions on vision technologies necessary for 2D and 3D manipulation for robot manipulators can be found in Refs. [35,36]. Deep learning and other data-driven methods are extremely popular in perception as they provide an intelligent methodology to overcome several key problems in extracting information from sensing. These methods help overcome common disturbances and noise that a robot could encounter in processing visual information in its surroundings, as discussed in Refs. [37,38, 38]. Finally, visual-servoing is often pursued to develop a direct tie between the perception and decision-making processes for robots [3,39–41]. Specific works addressing the safety interwoven into the perception problems when the robot is operating among humans are discussed in the survey [42]. Although most perception algorithms are largely architecture agnostic, literature surrounding perception as applied to mobile manipulators can be found in Refs. [41,43–46].

Section 2 presents a description of applications that have the potential to be large beneficiaries from mobile manipulation capabilities while also attempting a categorization based on their dependence on key decision-making capabilities. Section 3 presents each capability that forms the tenants of the decision-making hierarchy that enables mobile manipulation. Finally, in Secs. 4 and 5, the future research directions and emerging opportunities for mobile manipulation are briefly presented.

The application perspective offers a convenient approach to developing key performance indices. Identification of quantitative measures and criteria for system-level performance provide key insights that aid with the choice of decision-making strategies. Our emphasis will be on abstracting the performance requirements across applications. We will focus on four representative application verticals that are common deployment venues for wheeled mobile manipulation. Representative mobile manipulators exemplars deployed in various applications discussed in this section are highlighted in Fig. 3. In addition, a detailed graphical mapping of the literature that specifically addresses all the application verticals discussed in this section is presented in the form of a flowchart in Fig. 2. Further, this flowchart also classifies the literature in the laterals of the particular decision-making aspect addressed, namely, planning, control, and human–robot interaction.

Wheeled mobile manipulators are widely used in warehouses and shop floors for object (e.g., tools, parts, packages) transportation, which involves picking up objects and dropping them off at an appropriate location or handing it off to a human worker. In addition, they are also deployed for machine tending [47,48], which can be defined as material handling such as loading or unloading and part feeding for machines (e.g., computer numerical control (CNC) machines). The use of fixed-base manipulators (in small fenced cells or as part of a production line) raises the infrastructure costs and creates challenges for line balancing. The use of the ungrounded mobile manipulator configuration to empower flexible production lines, hence becomes very attractive. In this section, we explore the requirements of a mobile manipulator employed in the context of this industrial operation.

Several mobile manipulators have been used for transportation of objects [49–59]. Transportation of small objects is typically done either by holding the manipulator throughout the entire operation or by securing the object onboard the mobile base. The robot’s decision-making is targeted toward only one subsystem at a time, driven by the emphasis on modularity/vendor diversity for the arm and the base. Literature on independent motion planning for the arm and base can be found in Refs. [60–67] and Refs. [60,61,65,68–73], respectively. On the flip side, in the case of transporting larger objects, as shown in Fig. 4, the base needs to be aware of the obstacles in the vicinity of the arm and subsequently the object. This creates a need for an integrated/whole-body approach toward decision-making. In Ref. [74], for example, a sampling-based method is deployed for the planning of nonholonomic mobile manipulators. In general, any high degree-of-freedom motion planners [60,61,65–67] can be used to determine the solution; however, specific computational speedups would be crucial to support real-time deployments.

In some cases, whole-body decision-making is also required for manipulating small objects during their transport phase, e.g., preventing sloshing in open liquid containers, maintaining task-space orientation of manipulated parts during transport, and so on. Researchers have handled these edge-cases by making the planner implicitly also determine the waypoints for the transported object [75–78].

Bi-manual mobile manipulators that mimic humans in mobile manipulation scenarios can use their excessive redundancies to hold an object in place, while the base can plan its motion independently [60,61,65–67]. Optimization-based motion planning [77] are often deployed for whole-body planning, including for constrained motion planning subproblems such as the piano movers problem [79].

The industrial applications also motivate an important class of problems pertaining to enhancing manipulator/gripper access while picking parts. Grasp locations on the object can be determined using several existing methods [80–88]. Motion planning has been enhanced for picking up objects with moving mobile bases [71,89], while factoring in object-pose uncertainty by modulating the gripper speeds [78]. Accurately executing these trajectories where the mobile base and the manipulator move concurrently has led to the exploration of the multiplicity of control approaches [90–94].

Material handling refers to loading or unloading completed parts from machines such as 3D printers or CNCs as shown in Fig. 3 (in image 3). Ensuring safety in an unstructured environment, ease-of-use and robustness have been considered as important criteria to apply material handling in real-world applications. Usually, material handling consists of a small set of skills, including but not limited to transportation and pick-and-place. In some cases, it includes skills like opening/closing doors, pressing buttons, and turning knobs. Breaking down the task into unit skills allows for reutilization at the task level.

In part feeding applications, the usual process is picking up the component, moving toward the corresponding feeder location, and serving it to the feeder. The minutia of the task itself may differ based on the characteristics of the feeder that is to be serviced. Unlike the relatively independent material handling machines, part feeding machines work together, sometimes as part of the same assembly line, to produce a product. Thus, coordination between multiple feeders and mobile manipulators, efficient task sequencing, and scheduling are important requirements in part feeding applications. The task-level programming method that addresses tasks, skills, and device primitives individually has shown promise [51,95,96]. Other aspects such as scheduling and specific machine minutia have been discussed in Refs. [97,98].

Constrained collision-free motion planning and control often play an important role in this application vertical given the robot’s proximity to other machining tools, human workers, etc. The mobility constraints that are imposed by the limited knowledge on the kinematics and dynamics of the mobile bases impose a complicated mobility control problem, further enhanced by supplemental safety tasks such as safe and guaranteed dynamic-obstacle avoidance. In Ref. [99], a controller is introduced that consists of a behavioral learning layer that redirects the desired trajectory while transporting an object by predicting the behavior of the dynamic obstacles.

The pick-and-place control as a whole for mobile manipulators has been discussed in Refs. [100,101], where factors such as grasping, transportation, obstacle avoidance, and stability have been addressed as a whole. Often this problem is studied in a simplified context by holding the mobile base stationary. Control aspects for subproblems such as fast pick-and-place, contact dynamics, grasping, etc., have also been explored in the form of model-based control (MBC) (kinematic/dynamic) [102–104] as well as model-free control (MFC) approaches [105]. Auxiliary tasks such as intelligent obstacle avoidance, for instance, manipulating doorways, have been studied from a control perspective in terms of architecture and implementation in Ref. [106]. The existence of dynamic uncertainties on the manipulator’s part while carrying a variable payload adds robust stability control requirements for the system. Further, to handle nonlinearities and uncertainties such as unknown dynamics, parameter errors [107], and so on, adaptive control techniques are often required for such applications as demonstrated by Dong [108] and Ge et al. [109]. In Ref. [110], the authors have addressed the problem of motion control with obstacle avoidance for transportation jobs while carrying the maximum payload, which can cause significant tip-over stability concerns. Some innovative control methods for heavy material handling have been discussed in Refs. [111,112], where additional mechanical systems such as rollers and outriggers are added to the mobile manipulator, and the control is implemented on the robot’s side to accommodate for the mechanical add-on. Additional nuances when handling mechanical add-on in the obstacle avoidance and motion control aspects to accommodate for force interactions that are unmodeled in the control design are highlighted by Balatti et al. [113].

Mobile manipulators provide the desired task flexibility and mobility for assembly operations involving large parts. For example, a mobile manipulator can go to a work area to get a part and perform assembly operations on a moving conveyor or build a pallet by stacking widgets in the desired pattern. Several works discuss the specifics of autonomous constructive assembly ranging from assembling furniture to load palletizing to wall building [114–117].

Assembly introduces manipulation challenges that involve highly reactive, fast, robust, and micromanipulation/fine-manipulation control. The challenge of integrating high-level strategies with low-level controllers requires dealing with uncertainties. A lot of work has been done specifically in the control domain of mobile manipulation, particularly applied to the assembly application [118–121]. Some modern approaches have decided to forego the complex process of modeling such minute aspects and instead have deployed data-driven approaches for skill acquisition [122].

Additive manufacturing is a process in which parts are built by adding material layer by layer [123–126]. It has numerous advantages, such as the capability to build complex parts, reduced tooling, quick prototyping, less material wastage, and more. However, one of the major limitations of additive manufacturing is the trade-off between the part size and the setup cost [127]. As the part size grows, the setup cost also increases, which usually puts a restriction on the part size. Thus, there is a need to develop a cost-efficient setup to perform additive manufacturing at a large scale. Mobile manipulators have a theoretically unlimited workspace and hence could serve in this space effectively. They can be used in additive manufacturing to build large-scale parts without any size restrictions [128,129]. Most importantly, the mobile manipulator setup cost will not increase with the size of the built part. Literature that highlights the capabilities of mobile manipulators in this space shows their application in facets such as building large-scale parts [130], building a pipe of infinite length [129], and so on.

The importance of high-performance control, estimation, and planning when it comes to deployments in this space is highlighted in Refs. [126,131]. It is essential for the robot manipulator to perform robust disturbance rejection control [132–134] based on the feedback from the build process and the deposited material. In addition, the motion control itself is extremely time sensitive, and hence, a very reactive and dynamic trajectory control algorithm is required to track the spatiotemporal references. Due to the high interest in the advantages robotics can bring to the additive manufacturing space, there have been several works published in this area [127,135–138]. Literature reveals that planning challenges such as continuous replanning, optimal mobile base placement, and constrained motion planning also play an important role in this application vertical.

The usefulness of mobile manipulators is quite apparent in service tasks in home and hospital environments. Several applications such as disinfection [139–141], providing therapeutic assistance [142], administering medications [143], elderly care [144,145], disability assistance [146], and so on can be envisioned. Utilizing both mobility and dexterity, they can take over or assist humans with day-to-day tasks, similar to how humans perform them.

The home and hospital environments are typically more unstructured compared to an office or industrial setting, making it challenging for robots to navigate and perform manipulation tasks safely and efficiently. The dynamic and ever-changing environment needs to be handled in real time, which increases the need for reflexive and reactive behavior. Most everyday tasks in this space mainly involve pick and place operations and forceful interactions, which require capabilities such as target identification and localization, safe and fast navigation, and object manipulation. Mersha et al. [147], Stückler et al. [148], Mitrevski et al. [149], Jain and Kemp [56], Qi et al. [150], Mucchiani et al. [151], Yamamoto et al. [142,152,153], Choi et al. [154], Lu and Wen [155], and Zhou et al. [156] are some works that demonstrate mobile manipulator robotic platforms performing tasks within these environments. These article cover a variety of tasks done by assistive mobile manipulators such as fetch and carry, tidying-up, cooking, dancing, serving, and so on. Many of these works, experiments are being conducted in nursing homes, elderly care facilities, as well as regular homes, with the aged and disabled as beneficiaries. The goal is to provide support for independent living, remote care, and household management [142,151–155].

There are many complex tasks in these situations that do not fit into object pick-up and placement operations. Capabilities such as opening doors and drawers, cleaning surfaces, pushing, and pulling beds are fundamental for operating in these environments [157–160]. People envision mobile manipulators to do complex cleaning tasks in the house, which the present floor-cleaning mobile robots cannot perform. Mobile manipulators can use their arms to use human tools and perform dusting, scrubbing, wiping, mopping, sweeping of different surfaces, and objects within the house [161,162]. There have also been some works on using mobile manipulators in the cooking domain [162,163]. In Ref. [164], a look into force-inclusive motion control for cooking applications is explored. Mobile manipulators can also help in organizing different items in the house, such as shelves for libraries [165]. Force control is essential for these operations that comprise subtasks such as opening/closing doors/drawers, manipulating taps/switches, and so on [106,166].

Telenursing is an exemplar patient-caring task that [143,167–169] requires specific capabilities including safe mobility of the robot, capability of light- to medium-duty manipulation tasks, safety while in close proximity to and touching humans, facilitating audio/visual communication between a nurse and patient, etc. Tele-operated mobile manipulators are especially useful for nursing and telediagnosis of quarantined patients, which can limit health care workers’ exposure to dangerous microbes [170]. It can also help to expand the caregiving capacity of a hospital very fast whenever the need arises. Studies presented in Refs. [171,172] describe a smart hospital ward management system with a mobile manipulator, which can perform drug distribution and other tasks. Kapusta et al. [146] demonstrates a mobile manipulator, collaborating with a robotic bed to provide physical assistance to patients with disabilities. Several assistive robots are studied specifically for people with disabilities [154,155].

Apart from physical assistance, mobile manipulators are often deployed to be socially assistive robots as well, to provide nonphysical assistance to human users through social interactions. There are multiple applications for such robots, ranging from aiding children and elders with autism to rudimentary tutoring for special needs kids to help out teachers [173,174]. Safe human collaboration is a very important aspect here as tasks such as handing over things to humans are high-risk operations [175]. This raises extreme robustness requirements as any safety breach is potentially extremely harmful, as demonstrated by Park et al. [176]. Accurate end-effector force control continues to be extremely important here, despite challenges such as lack of tactile feedback in many cases [177]. Several aspects of control in regards to these constraints are discussed and studied in Refs. [146,178,179].

Mobile manipulator deployments to support precision agriculture practices have proven to be one of the most effective ways to significantly reduce the negative impact of farming on the environment [180]. Agricultural operations serve as a representative application arena to highlight beneficial deployments of mobile manipulators in an off-road/unstructured operational setting [181]. However, the applicability in other unstructured application environments such as nuclear inspection, disaster response/rescue, and other terrestrial/extraterrestrial fielded missions should be readily apparent. These environments are often unmapped (or poorly mapped with significant uncertainty), creating significant navigational challenges and necessitating advanced controls [182,183]. Oil/gas and nuclear power industries heavily rely on mobile manipulators for inspection purposes as the job is harmful to humans in many cases [184,185]. Insights into the system architecture and requirements in these specific subapplications can be found in Refs. [186,187].

A major role of mobile manipulators in agriculture arises in the areas of weed/pest killing, targeted spraying, and harvest picking. For example, weed control mobile manipulators perform targeted elimination of weeds using methods such as controlled spraying, electrocution, and laser targeting while covering large areas of farmland as fast as possible. Task-specific customization is often pursued—both in the initial design and in the online control—to overcome variability in the nature of fielded tasks as well as a need for computationally efficient yet robust decision-making processing. Similarly, one can find several architectures that are deployed and pursued based on which specific type of crop the mobile manipulator is designed to handle, as shown in Refs. [180,188,189,189–193,].

Deploying task-specific designs is not unique to the agricultural field—disaster response robots often need very specific mobile manipulators depending on the type of environment, operating conditions, manipulation target, etc. (see Fig. 3, in image 7c). Task-specific designs as it pertains to inspection and radiation detection in nuclear plants is a prime example of such a case [194,195]. Mobile manipulators used for rescue that perform tasks such as getting over unknown steps, navigating dense urban environments, opening doors, and operating in chaotic conditions are presented in Refs. [196–200].

The robot localization problem plays a very important role in aiding mobile manipulator decision-making hierarchy and is often solved with relative ease in indoor industrial applications using simultaneous localization and mapping (SLAM) [201] due to the ability to have some form of structure or control on the environment. However, in the case of outdoor settings, vision-based methods are challenging due to the where the complexity of problems such as lack of constants in the environment that could potentially serve as landmarks. Even though there have been studies that deployed SLAM in these applications [202], in outdoors, localization of the robot is performed using global positioning system [203] or advhoc sensor systems such as ultra-wide band [204], etc. The localization part is also complicated even in certain indoor settings such as agricultural greenhouses since it suffers from the same uncertainty problems that make vision methods difficult, but at the same time, satellite navigation techniques cannot be used. Hence, in many greenhouse settings, the mobility aspect of the mobile manipulators is handled through the use of rails [205,206]. There is also some work on wheeled mobile robots [207]. The initial step in every single task before the decision-making hierarchy of the robot takes over would be the vision problem of identifying and localizing objects of interest, which are addressed in works such as Refs. [208,209].

Safe navigation is essential in agriculture to minimize the impact on the soil and avoid crop damage since the error margins are very small [210]. The previous study [211] proposed a robust navigation technique based on cost maps and also presented a framework that can make the inculcation of such techniques to tailored environmental situations. The applicability of this framework using popular middleware such as robot operating system is also presented. An extension of this work [212] also went on to propose parameter tuning algorithms that close the gap of uncertainties in the robust navigation techniques by making them more adaptable to robot architecture and environmental condition changes. A review of mobile agricultural robotics covering several aspects of planning, manipulation, and design has been presented in Ref. [213]. To handle the ever-changing environment (growing plants, lighting, weather, etc.), Pretto et al. [214] studied an aerial-ground robot combination for deployment in fields wherein the aerial robot maps the field often, and the information is transferred to the ground robot. Similarly, Xiong et al. [215] present several aspects of vision and planning, which is further reinforced by application toward mobile manipulator-based strawberry picking application. Several other studies that can explored for safe and robust planning can be found in the literature [216].

In terms of the control capabilities for outdoor settings, the biggest focus is on robust control of both the base and the arm to deal with terrain irregularities, deformations, slips, dynamic obstacles, and sensor noise. In agriculture, any error in its trajectory tracking could lead to regular damage of crops, which in summation could lead to a big loss to the farmers. The work presented in Ref. [217] explores a systematic design and control architecture development for heavy material handling mobile manipulators that can handle parametric perturbations and uncertainties. Performance comparison of control strategies for existing agricultural mobile manipulators based on various sensors is presented in Ref. [218]. In general, control strategies focused on picking objects for mobile manipulators [43] and off-road navigation of unmanned ground robots [219] can be translated to agricultural applications as well. A comprehensive review of agricultural robots in field operations can be found in the two-part study presented in Refs. [220,221].

Given the variability in design, it is often difficult to utilize general off-the-shelf solutions for such applications. For blank slate design, there have been some studies that have addressed the question “Which abilities are necessary to solve particular tasks autonomously?” and discussed how they could be converted into a control architecture in general [222] and in particular to mobile manipulation [223]. Systems that are goal oriented and the structure of their control architecture differ based on their expected behavior and/or the environment they operate in, as demonstrated by Pin et al. [224] and Krotkov et al. [225]. Control strategies such as dynamic-obstacle avoidance [226,227] need to be adapted for off-road operation in unstructured environments, such as by inclusion of wheel slip [228] and nonrigid terrain adaptation [229].

Decision-making helps to close the loop between the sensing/perception and action in intelligent robotic systems with various associated nomenclature at the various deployment stages. Coupled with extent of feedback between sensing and action, this continuum ranges from motion planning/open-loop control methodologies used primarily in offline settings to various forms of online-replanning/closed-loop control suitable for deployments in online settings to address dynamical ego-system and environment.

Since its introduction in the 1980s, the configuration space [230] approaches have helped create the abstract notion of configuration/state of the system and associated choices of extended versus minimalistic representations. Notions of spatial contiguity (free/obstacle) and temporal continuity (continuous/discrete) ascribe additional qualitative aspects to the underlying state and motions defined as the sequence of states. Motion planning for the dynamical system (mobile manipulator) then becomes a computational problem of finding a sequence of actions to move from an initial state to a goal state in the presence of spatiotemporal constraints. Such spatiotemporal constraints arise naturally due to the dynamics within the environment and most certainly for the dynamical system engendered by the mobile manipulator. A wide diversity of constraints can ensue based on extent of mobile manipulator/environment coupling as well as spatial and temporal coupling (i.e., reducibile to only spatial or only temporal).

Despite the existence of these spatiotemporal constraints, there remains enormous redundancy and thence choice for selection of the motion plan—this paves the way for creation of decision-making frameworks. At their core, such decision-making frameworks are based on the “design-optimization paradigm”: systematic creation of choices, systematic evaluation of choices, and systematic elimination of choices. In particular, the addition of a parametric capability facilitates the leveraging of computational support to address problems of ever-increasing size and scope (as we will note). Further, varying grades of planning/control are possible based on the extent of reliance on sensing (ongoing information gathering within the environment) versus models (preexisting information about the environment).

There is a significant body of literature in motion planning for mobile manipulation that came to the forefront in the late 1990s and is captured in surveys such as Refs. [14,15,19,231] as discussed in Sec. 1.2. As a quick aside, a disambiguation is necessary between the often interchangeably used “motion planning” and “path planning.” Although path planning only generates a path within the configuration space, motion planning generates time-indexed motion trajectories. Inasmuch path-following only requires spatial feasibility (e.g., obstacle avoidance), while motion planning requires compatibility with spatiotemporal constraints (engendered in dynamics of both robot and environment). It is also noteworthy that ultimately any path planning effort requires a final time parameterization into a motion planning exercise before deployment.

In the couple of decades since the principal enhancements have come from multiple perspectives: (1) more complex mobile manipulators (larger configuration spaces), (2) deployed in more diverse application settings (greater spatiotemporal constraint complexity), (3) availability of new theoretical/algorithmic tools (sampling-based methods, kinodynamic planning), and (4) triple convergence of computing (GPGPUs) communication (distributed), miniaturization (embedded controllers) that has permitted the tackling of larger-scale problems in more real-time settings.

The combined controllable degrees-of-freedom within the kinematic-chain (from both mobile base and the articulated manipulator) presents the mobile manipulator design architecture the opportunity to address very complex tasks. However, resolving the redundancy (internal/external) is crucial to realizing this potential.

As the complexity of overall mobile manipulation process increases, a two-stage hierarchical approach is often pursued: (i) task planning/breakdown into a series of tractable motion planning subtasks and their sequencing; (ii) motion planning of the high degree-of-freedom mobile manipulator within each sequenced task. It is noteworthy that the two steps (task planning and motion planning) are closely coupled and should be solved concurrently but are addressed separately from a computational tractability perspective.

Such task and motion planning sequencing issues exist for any robotic manipulator (or subsystems) performing sufficiently complicated tasks over a sufficiently long time horizon. However, a breakdown along the lines of mobile manipulator subsystems (mobile base versus manipulator versus gripper or combinations) or along the nature of the manipulation task (transportation versus grasping) feature prominently in the literature. Some of the earliest efforts of Wolfe et al. [232] focused on a fixed hierarchical partitioning of task and motion planning for the mobile manipulator. In more recent times, a more flexible contingency-based partitioning of task and motion planning, combining perception systems with human knowledge, has been pursued by Akbari [233]. An important consideration in integrating task and motion planning is effectively utilizing symbolic and geometric representations [234,235]. Recent work in this area tries to reduce calls to the motion planner by taking into account spatial and temporal constraints and efficiently caching previously generated motion plans [236,237].

Coincidentally, such task- and motion-partitioning problems for mobile manipulators have also been studied in the context of the coordinated planning problem. For example, in Ref. [74], where the mobile manipulator (high dof) has to carry bulky objects like a long rod or a chair through narrow passages (large configuration spaces with significant spatial constraints over extended time horizon). Successful pursuit of task and motion planning will remain crucial even as various application deployments pursue the long-term autonomy goals.

As a fundamental functional capability needed for the effective operation of mobile manipulators, motion planning has been pursued with two broad flavors: point-to-point motion planning to move from a starting configuration to a goal configuration (avoiding static obstacles and other primarily spatial constraints) and constrained motion planning addressing various spatiotemporal constraints (primarily arising from the mobile manipulator). Motion planners for high degrees-of-freedom systems like Refs. [60,61,65,238–242] can be used for solving most motion planning problems; however, in the context of this article, we will focus only on those planners that are designed specifically for mobile manipulators.

Historically, point-to-point motion planning developed the earliest and often faced the challenges of computational complexity. Hence, early exemplars often showcased decoupled mobile base and the manipulator subsystems approaches. Doing so not only permitted modularity but also potentially reuse/retasking of generic mobile base and manipulator motion planners. However, given the dependence of the manipulator workspace on base motions, manipulators were required to be folded/homed to a standard position. In addition, while sequential and separate, the motion base and manipulator motion planning were dependent—mobile base motion depended on current (stationary) manipulator configuration while manipulator motion planning was dependent on the current (stationary) location of the mobile base. Several approaches use the same planner [243,244] for both mobile base and the manipulator planning sequentially. Similar strategies for the motion planning of mobile base and manipulator sequentially with the additional consideration of environment uncertainty are discussed in Refs. [72,73].

However, as noted earlier, spatiotemporal constraints arise naturally due to the dynamics within the environment and most certainly for the dynamical system engendered by the mobile manipulator. A wide diversity of constraints can ensue based on extent of mobile manipulator/environment coupling as well as spatial and temporal coupling (i.e., reducibile to only spatial or only temporal).

Constraint imposition may arise due to physics (e.g., contact) or due to heuristics (e.g., virtual fixtures), and these constraints need to be satisfied regardless of the motion-planning framework. From a physics perspective, the underlying mobile manipulator constraints may be imposed at the component level (kinematic actuator rates/dynamic actuation limits); subsystem level (holonomic versus nonholonomic wheeled bases, serial/parallel/hybrid manipulators); or the system level (closed-kinematic loops due to environmental contacts, other forms of end-effector motion/force constraints). The interplay of these constraints with the environment constraints (nonholonomic motions) creates a rich fertile problem space. The further addition of virtual constraints on both the mobile manipulator (e.g., preferred manipulability) and environment (safety buffers) adds to the overall problem complexity. For example, the task- and motion-level planning frameworks may be viewed as a form of ‘artificially constrained” motion planning within a higher dimensional space. Thus, in such a milieu, constrained motion planning offers a more general and certainly more contemporaneous approach but comes at the cost of increased computational complexity.

Optimization-based and sampling-based offer broad categorization of motion planning approaches deployed to address the challenges offered by (i) the need to resolve redundancy and down-select a plan and (ii) ever-increasing high-dimensional configuration spaces. In the context of mobile manipulators, parametric optimization-based methods like in Refs. [77,245] have been modified for the generation of constrained motions of the robot with one or more subsystems (mobile base, manipulator, gripper) moving concurrently. Additional variants [78] have encompassed aspects such as imposition of pose and velocity constraints (e.g., for grasping objects) as well as temporally fluctuating imposition of constraints. Such frameworks greatly facilitate intuitive inclusion of continuity requirements arising from the mobile manipulator dynamics/constraints. In Ref. [246], the planner simultaneously plans for the optimal location of the mobile base, as well as joint motions needed to traverse the desired end-effector path. The quadratic programming-based approach can easily incorporate additional constraints such as for wall painting, object sanding, or car washing operations using mobile manipulators. Alternately, a secondary optimization approach has been used to loosely influence the solution constraints such as end-effector tracking for a mobile manipulator while attempting to enhance tip-over stability margins [247].

In contrast, a bulk of the growth in motion planning for robotic manipulation has been in sampling-based approaches that have come to dominate the landscape. General planners like BIT* [61,248] can be extended to the planning of holonomic and nonholonomic mobile manipulators. In these methods, the shrinking/refinement of the focus region leads to progressively better solutions when: (a) determining the initial feasible path is easy and quick and (b) the length of the optimal path is relatively long, i.e., there is no maze-like structure in the configuration space. Pathological cases such as planning for nonholonomic mobile manipulators carrying large objects in cluttered environments do not satisfy either of these conditions and hence using these methods may not work. This is the focus of the planner presented in Ref. [74], where focus regions, hybrid sampling, and connection heuristics are used for accelerating the motion planning for mobile manipulators. Planners like TGGS [249] use separate planning roadmaps (for base/arm) and cannot be directly used coordinated manipulator/mobile base planning.

Other generic sampling-based methods like task-space regions, Tangent Space RRT, and Atlas-RRT [79] have been used for constrained motion planning for mobile manipulators. The sampling-based task-constrained motion planning approach [250] enforces constraints on orientation of the object being carried and are applied throughout the motion by a holonomic mobile manipulator. Random trees are grown in the high-dimensional configuration space, and samples are projected onto the appropriately constrained manifold via the mobile manipulator Jacobian. In other work, an end-effector capability map is used together with an online sampling-based method as a desired end-effector trajectory with holonomic mobile manipulators [251].

Such Jacobian-based control can also be used to extend the trees for the nonholonomic mobile manipulator where the mobile base is constrained to move along a given path [89]. The sampling-based approach is used for unconstrained motions, and relevant nonredundant inverse kinematics solutions are used for the constrained motion sections. Another sampling-based method for constrained end-effector motion planning on a point cloud surface with a nonholonomic mobile manipulator is presented in Ref. [252]. In Ref. [71], the focus is on the planning for the mobile base so that the robot moves and picks up objects in a time-optimal manner. In Ref. [253], the requirement for the end-effector to reach a set of desired configurations requires multiple repositioning of the mobile base multiple times. First, the optimal manipulator configurations corresponding to desired end-effector poses are determined, and then successive optimal configurations are connected with minimal RRT generated paths.

Finally, it is also worth noting that several sampling-based approaches (e.g., RRT-star, A-star) can be enhanced by using physics-based measures (manipulability, etc.) as additional heuristic costs to guide the search (as discussed in the next section).

In the previous section, we noted that unique challenges emerge for specific mobile manipulator configurations, deployed tasks, and operational environments. While the arena of motion planning in high-dimensional configuration spaces can be (and is) well studied in an abstract setting, in what follows, we will highlight the specializations pursued to address specific mobile manipulation challenges. Our discussion around motion planning deployments will be focused along three distinct subproblems—mobile base placement, area coverage, and mobile manipulator grasping—that highlight some of the unique features and help justify the approaches

Partitioning of base/manipulator motion capabilities within a mobile manipulator is often justified from multiple perspectives: (1) modularity (any base can now be used with any manipulator), (2) architecture perspective (wheeled nonholonomic versus articulated serial), or (3) dynamics of manipulation (macro versus micro).

The mobile base placement (or sequence of placements) is crucial to not only ensuring feasibility of end-effector task but also its optimality. Other tasks that comprise continuous toolpaths that need to be executed on large workpieces require a sequence of mobile base placements instead of a single placement and minimization of base placements reduces the recalibration requirements. Zacharias et al. [254] used a reachability maps representation to find suitable mobile base placements so that the articulated arm can trace Cartesian space trajectories [255] including clustered regions [256]. The work done in Refs. [257,258] used the map to generate promising placements by sampling the workspace and improving reachability using gradient descent over the map. Other works build on computing and searching over an inverse reachability maps [259] to find suitable base locations. Reuleaux is an open-source library that also exploits inverse reachability maps to compute base placements [260]. A nearest neighbor search to obtain a set of feasible base placements together with a reachability score (for each query pose in the workspace), which can be maximized.

As in the traditional redundant manipulators cases, other end-effector criteria such as manipulability [261], end-effector stiffness [262], reachability [263], and uncertainty [253,264] also are necessary in the mobile manipulation space [,253,262,263].

Area coverage planning problems have been studied extensively in literature in the survey paper [265]. Area coverage is required for several robotic applications like lawn mowing [266], agricultural field plowing [267], bush trimming [268], spray painting [269], CNC machining [270], cleaning [271], etc.

When the mobile base and the manipulator move separately, coverage path planning for determining eddy currents in aeronautical parts as discussed in Ref. [272] using a zig-zag pattern can be used for a stationary mobile base. Similarly, in Ref. [273], a nonrandom targeted viewpoint sampling strategy for coverage planning to cover the entire area for camera-based inspection of large parts is implemented, which can be implemented for a stationary mobile base.

One of the effective methods for solving the coverage path planning problem is to formulate it as a traveling salesman problem (TSP) [274]. In Ref. [268], a bush trimming problem based on a manipulator fitted with a trimming tool at the end-effector was formulated as a TSP by using the mesh of the desired shape of the bush. Each triangle on the mesh was considered as a vertex for the TSP. Such an approach is effective only when there are not a large number of vertices. Further, they use a spline-based representation for optimization based on motion constraints for the robot such that the end-effector, i.e., the cutting tool, is constrained to go through the waypoints. In Ref. [275], a combined approach for robot placement and coverage planning for mobile manipulator, takes into account constraints like collision and stability, to determine appropriate base placements and solve the TSP for the end-effector.

The mobile base and the manipulator can move concurrently when planning for area coverage. A framework for 3D surface coverage by a redundant manipulator was implemented in Ref. [276], which can be easily extended for holonomic mobile manipulators. Here different inverse kinematic solutions for the robot are treated as individual nodes in a graph which is modeled as a generalized traveling salesman problem (GTSP). GTSP is where the nodes of a graph are subdivided into clusters, and at least one node in each cluster needs to be visited. This method can work for mobile manipulators; however, the nonholonomic nature of the mobile base can make it challenging to find feasible edges for the graph. Yang et al. [277] solved the problem of nonrevisiting coverage task with minimal discontinuities for nonredundant manipulators. Leidner et al. [278] examine cleaning and wiping chores with a redundant mobile manipulator (Rollin” Justin [279]), where the motion plan is generated, splitting the task into a high-level planning module and a specific control for the required cleaning action with an overall objective of optimizing the traversed end-effector Cartesian path length. This is also alternatively done in an end-to-end framework in Ref. [280]. In this light, Thakar et al. [281] seek an alternative approach to covering the area of a point cloud with spray disinfectant. The point cloud is first projected on a plane and a depth-first search based branch-and-bound method is pursued to cover the extracted planar polygon extracted with a combination of zig-zag and spiral segments.

Grasp planning for mobile manipulators is a challenging problem that has been dealt with in several ways in the literature. On the one hand, grasping requires coordination within a very challenging high-dimensional constrained configuration space (mobile base/manipulator/gripper). Further, grasping requires detecting object, constructing data-driven representation, determining the gripper approach-vector, and computing all the mobile manipulator’s plans in the presence of uncertainty.

Many of the traditional grasp planners (designed for stationary manipulators) can be used for mobile manipulators once the mobile base has been fixed. However, an opportunity exists for stationing the mobile manipulator at an appropriate position to ensure easy grasping which several approaches have exploited. In Ref. [86], the focus is on appropriately imaging an unknown object with a mobile manipulator for its accurate reconstruction, finding appropriate planes from segmentation to successfully grasp the object. Similarly, in Ref. [282], a stereo vision algorithm is developed for the object-pose estimation using point cloud data from multiple stereo vision systems. Coupled with ability to compute various grasp metrics (e.g., via GraspIt) coupled with manipulator end-effector metrics (e.g., manipulability), this can also set the stage for secondary optimization such as for base position and end-effector approach direction [282–284].

In addition, several of the model-based reinforcement learning (RL) approaches for robot motion and grasping have been adapted for mobile manipulation. Li et al. [87] proposed a RL strategy for manipulation and grasping of a mobile manipulator, which reduces the complexity of the visual feedback and handle varying manipulation dynamics. After several iterations of the RL and other methods, the bionic robot arm is able to reach the target position more accurately, even with unknown external perturbations. Other works [285] have also used RL for targeted grasping with active vision feedback geared toward mobile manipulators.

Dex-Net MM [88] is a deep learning framework for surface decluttering at homes where mobile manipulators with low precision can be potentially used for clearing and cleaning objects from the floor. This builds upon their previous frameworks on Dex-Net [80,286], where physics-based models are used to determine the grasping locations on CAD models of objects. These models are learnt using deep neural networks, which are then used for grasping unknown objects from a heap of unknown objects. In Ref. [88], the Dex-Net 4.0 [287] is modified to adapt for the parameters of a mobile manipulator because, due to inherent cost and weight limits, mobile manipulators have far lower precision in sensing and control than a fixed-based robot system.

Another data-driven mobile manipulator grasping of unknown objects is presented in Ref. [288]. The goal here is to autonomously scan the environment, model the object of interest, plan, and execute grasps in the presence of uncertainty in the pose of the mobile base. A single scan of the object may not reveal enough information for grasp analysis, and hence, the system autonomously builds a model of an object via multiple scans from different locations until a grasp can be performed.

All the methods mentioned earlier focus on manipulator motion and grasping with the mobile base stationary. However, a generic grasping pipeline is desirable, which achieves arm-base-gripper coordinated grasping given the information about object pose and the operating environment [289]. Such concurrent manipulator/mobile base motion approaches are being explored till grasping is successful [71,78,89] or at least until the gripper reaches the objects (and only manipulator moves for grasping) [46,290].

This may not be optimal as grasping can happen with the mobile base and the manipulator moving when the gripper is closing [89]. Hence, in Ref. [78], the explicit need for a stationary gripper is eliminated via a strategy for mobile base motion compensation by the manipulator arm, thereby keeping the gripper largely stationary during grasping.

Most higher-level planning algorithms tend to be computationally expensive and are therefore unable to replan fast enough for ensuring timely intervention in real time while navigating dynamic environments. Hence, control approaches serve a crucial role for the robot to achieve its desired goals in a safe, accurate, and repeatable manner. Some complex planning algorithms might be deployed progressively in real time [291], as computational resources are constantly evolving. Yet, real-time control aspects still remains the barrier for the avoidance of system failure.

As described earlier, in any robotic system with at least semi-autonomous capabilities, the overall software or functional architecture is broadly compartmentalized under the sense-think-act paradigm. In each of these divisions, the control algorithm or an extension of it plays a significant function. Sensing is of two types, namely, proprioception, the robot’s capability to identify its own states, and exteroception, its capability to identify its surroundings/environment. While exteroception is largely used in the planning and behavioral layers, proprioception appears in the robot control architecture in the form of state estimation formulations to eliminate uncertainties and noise introduced by the sensor suite available on the robot. Thinking involves the robot’s capability to break down its functions into requirements and consequently designing its own tasks to complete those requirements. Here, the role of control, generally called the upper-level controller, deals with the central problem of converting a desired path or trajectory that is computed by the planner to an acceptable and safe reference input (forces, torques, or angular rate) that an actuator can achieve. Finally, acting involves successfully and safely performing the tasks that the robot originally set out to achieve by coordinating and operating the actuators of the robot. The problem of control may also extend further into the aforementioned lower level as well, where the actuators are controlled to achieve the desired angular rate references that a higher-level control computes.

The process of generating the references for the upper-level controller have been discussed in Sec. 3.2, and in this section, we will explore the control methods for mobile manipulator previously developed and deployed in the literature.

In the following, criteria for the design of control architectures are broadly categorized in the context of mobile manipulator operation. This categorization follows the dominant selection in the literature as pointed out by other surveys [13,19,20]. In addition, Sec. 3.3.2 is introduced to capture several research contributions in niche topics within mobile manipulation that could not be categorized into the broader trends as observed in the literature. This section serves as an initial starting point for developing control architecture requirements and is not meant to exhaustive.

• Path and Trajectory Control

  The motion control framework consists of a feedback algorithm that computes appropriate actuator inputs for the motion of the arm or base. Here, path control is supposed to track global spatial/geometric references that can be specified by users or a planner. Path tracking controllers have been successfully applied to robots in general [292–295] as well as in the context of mobile manipulators [296–298]. When semi-autonomous or autonomous mobile manipulator tasks are time constrained, however, they ideally require spatiotemporal references. A higher-level motion planner generating references in the temporal space is therefore often distinguished as trajectory control [299–302]. Certain scenarios exist in which both strategies are deployed at the same time, i.e., the base being subjected to path tracking while the manipulator employs trajectory tracking (as demonstrated in Ref. [303]). It can be argued, however, that path tracking is advantageous due to its lack of unstable zero-dynamics (as present in trajectory tracking) when load-carrying applications subject to unpredictable dynamic interactions are concerned, e.g., in mobile manipulation [304,305].

• Whole Body and Modular Control

  In most control approaches to mobile manipulation, base and manipulator operation are strictly separated such that at any given time only one primary control objective is active. This separation principle is then augmented by a switching layer that determines the currently pertinent control objectives. Exemplary studies that specifically address this decentralized control problem in mobile manipulators can be found in Refs. [306–310]. The advantage of such a control formulation lies in its simplicity, i.e., priorities can be separated among the arm and the base with individually designed different control algorithms employed for each subsystem [311].

  Despite the disadvantage that unified control needs to adhere to a single control framework, it allows for the exploitation of mobile manipulation in the true sense of the term, wherein the manipulator and mobile base can be controlled at the same time. This can lead to several advantages during task achievement as discussed in Sec. 3.2 and makes the robot more dynamic in terms of its capabilities. The formulation for this type of control involves considering the onboard manipulator as an extended joint space of the mobile base, where the motion controller considers both the base and manipulator states. Here, the base control becomes a lower priority task and is hence projected into the null space of the manipulator control tasks, which is now of a higher priority. As a result, base control is completed simultaneously without affecting the performance of the end-effector manipulability. Several studies have been published in the whole-body coordinated control problem of mobile manipulators where the base and the manipulator move at the same time [312–316].

• Obstacle Avoidance

  In general, the problem of obstacle avoidance is largely handled at the planning level; however, since control algorithms operate much faster and largely deal with the safety of the system, several studies have included the obstacle avoidance problem (both for the base and the manipulator) at the control level. Obstacle avoidance essentially becomes one of the objectives in the robust path/trajectory control problem. Robustness plays an important factor, since a mobile manipulator can change its spatial profile drastically, large errors can cause collision with its environment. Several studies have proposed motion control algorithms for mobile manipulators that include a real-time obstacle avoidance scheme [317–320]. In most scenarios, obstacle avoidance is more efficiently enforced with the use of a unified/coordinated control algorithm that considers all the robot joints at any given time, as demonstrated in Ref. [321]. This goes back to the points discussed in Sec. 3.2, where there are several constrained obstacles such as tables, doorways, etc., that need to be considered as obstacles only for a subsystem of the robot, namely, base or manipulator. The only performance metric when it comes to obstacle avoidance at a control level is the clearance margins and robustness by which the robot performs against handling said obstacles.

• System Stability and Disturbance Rejection

  Stability is a large research area for mobile manipulators as it is a highly dynamic system that operates in complex environments where system failures such as tip-overs can lead to property damage or injuries. A mobile manipulator might not move at speeds that are comparable to on-road vehicles; however, since it often carries noncentric loads and goes through forceful interactions with its environment, the manipulator can dynamically disturb the base causing it to tip over. Due to the tendency of complex planning algorithms to often use low fidelity or even kinematic models during design to manage its computation loads, the references that it provides to the control level may not always lead to the safest operation of the manipulator. The purpose of these control algorithms is to act in coordination or as compensation to motion control algorithms to ensure that the system’s center of gravity remains within its stable region. Several works are available in the literature studying the stability control of mobile manipulators while in operation since this is one of the biggest avenues for system failure [247,322–327]. Another avenue of research with respect to system stability is the definition of stability margins to define the robustness parameters for the controller [328–330]. The control methodologies itself is explored in Sec. 3.3.2.

  Disturbance rejection as a field often proves to be hand-in-glove to the stability problem when it comes to mobile manipulation. The main factors that need to be addressed in disturbance rejection are disturbances on the base due to the ground, or the manipulator due to the end-effector interactions [133,331–333]. The need for manipulators to drown out vibrations, etc., caused by the interaction between the end-effector and the work surface is a well-studied problem in robotics [334–337].

• Off-Road Operation

  As discussed in Sec. 2, mobile manipulators are also heavily used in outdoor applications such as search and rescue operations, construction, agriculture, etc. This raises additional complexities not only in the perception capabilities of the robot but also in the control aspects [338]. Outside the space of mobile manipulation, navigation of ground mobile robots in off-road environments is an enormous field that consists of research spanning decades [339]. Even though mobile manipulators do not usually operate at extremely high velocities as some of the other robotics applications, the control challenges still remain significant. Researchers have studied the problem of mobile manipulation stability in unstable or deformable terrains [228,340–342]. In general, the robustness envelopes and constraints in the indoor environments, i.e., performing constrained motion control and dynamic-obstacle avoidance, have been proven to expand significantly in unstructured outdoor environments [343,344].

The control methodologies employed in any autonomous robotics applications and subsequently in mobile manipulation exist within a continuum whose extremities are model-based and model-free/data-driven controls. Traditional literature from the past tends to be model-based control. However, in the last decade, the neural network-based approach is conceived as a competitive way to control robot manipulators given the rise of computational capacity as well as significant interest in autonomy [345,346]. Although data-driven and completely model-free control methods show a lot of promise, problems such as lack of safety guarantees, adaptability, need for large amounts of data, etc. drive researchers toward adopting a hybrid method for control that explores the best of both worlds as illustrated in Ref. [347]. In this section, we will explore this continuum in the literature, starting with MBC and spanning through data-driven MFC approaches.

In regards to MBC, fundamental issues are kinematic and dynamic modeling, the control of nonholonomic systems, the hybrid motion/force, and hybrid position/force control. There are several works one can refer to for understanding the modeling aspects of a mobile ground system, compilations of which can be found in Refs. [21,348,349]. In several cases, the choice of base architecture can add significant additional control challenges as well. Researches have tried to introduce suspensions to the bases of mobile manipulators to help manage the dynamics exerted on the base by the manipulator as well as increase the traversability of the robot. This, however, leads to additional necessities such as suspension control, which by itself is a vast field of research [350]. Active motion control through the use of suspensions has been explored in the field of mobile manipulators [226,351]; however, the field is not well evolved in the research literature despite its promise. Some have tried to bypass the suspension necessities to instead use a self-balancing two-wheel robot as the base for their mobile manipulators [352]. These robots can adjust the angle of their pitch to adjust to the terrain, which increases operational capacities significantly. There is a unique challenge with this architecture, however, namely active stability control owing to their passively unstable nature. A thorough review of the advantages and state of the art in control for such a base as well as its mobile manipulator variants can be found in Ref. [353]. A preliminary look at the various architectures is presented in Fig. 5.

Similar to the modeling of mobile bases, the modeling of serial link manipulators and their variants is also a well studied and researched field since the dawn of robotics [354,355]. There are several other aspects to the analytical modeling of a manipulator that goes beyond the kinematics and dynamics itself, such as redundancy resolution, contact modeling, grasping, etc., which are also extremely important for the development of MBC algorithms. These aspects are well studied and explored, and a wealth of literature dealing with these topics exists, a compilation of which can be found in Ref. [356].

The decentralized or modular control methods discussed in Sec. 3.3.1 also extend to the modeling domain where the two subsystems can be treated separately in the control architecture. This provides some advantages since the mobile platform often has unique control properties when they are nonholonomic in nature, and it is often easier to model and control them separately for the sake of obstacle avoidance formulations, path control, etc. However, there is value in modeling the system as a whole which, despite increasing the complexities, can help exploit the value proposition of mobile manipulators completely, which is increasing the dynamic task capability of a robot. In Ref. [17], the authors derive and present the kinematics and dynamics of a differential-drive and Ackermann steered mobile platforms in great detail. In addition, Yamamoto and Yun’s body of work is often used as a popular source for extensive dynamic analysis, where modeling aspects such as dynamic task space, dynamic interaction between the manipulator and base, and so on, are thoroughly explored [11,12].

Differential-drive robots are the most common type of robotic base architectures that are used both in and outside the space of mobile manipulation. However, other platforms such as omnidirectional and skid-steered are also practically prevalent in this space, whereas the literature, on the other hand, is scarce in regards to the modeling of these systems. In Refs. [357–359], the complete kinematic and dynamic formulations are provided for an omnidirectional mobile manipulator, and a further inverse kinematics/dynamics control is developed for the trajectory tracking for the coordinated motion of the entire robot. A detailed modeling and analysis discussion was presented in Ref. [360] for skid-steered vehicles. Despite the control being a PD control (nonmodel based), the work itself is interesting due to experimental validation provided for the dynamic model using a skid-steer loader. Despite the fact that wheeled mobile bases with serial chain manipulators are not the only focus of [361], a complete look into a kinematic model-based control was presented with several citations to related works. The manipulability measure is a commonly used technique in the model-based control design of manipulators. Several works now use the same measure to gauge the performance of mobile manipulators as shown in Ref. [2], whereas some studies in the literature use the manipulability measure to design their controllers. The optimization of criteria inherited from the manipulability considerations (based on an analytical model) is used to generate the whole-body controls in Ref. [362]. In Ref. [321], singularities are avoided for the entire mobile manipulator by increasing the manipulability measurement. Several direct model-based active and reactive force/torque control frameworks are available in the literature [363–366].

There are several model-based control methods such as optimal control and feedback linearization that do not directly use strategies such as inverse dynamics yet use an analytical model as an assumption for system behavior in the control design phase. Optimal control has proven to be effective in several scenarios as shown by Avanzini et al. [367], Minniti et al. [368], Wang et al. [369], and Colombo et al. [370] where variants of model predictive control was successfully applied in navigation control for mobile manipulators. Further, given that the dynamics are known, feedback linearization techniques for mobile manipulators have attracted research interest in recent years. Input–output feedback linearization was applied to control the navigation of the mobile base such that the best manipulator position can be achieved at the preferred configurations (measured by its manipulability) in Ref. [315]. Similarly, in Refs. [11,371,372], decoupled control was performed using feedback linearization methods where force control of the end-effector was achieved while simultaneously compensating for the dynamic interactions between the base and the arm.

Compilations of the latest literature in classical model identification methods [373] and modern robot learning methods [374] often indicate a trend toward the deployment of data-driven methods to learn an input–output relationship for the robot. In addition, control of mobile manipulators with uncertainties is essential in many practical applications, especially for the case when the force of the end-effector needs to be considered. In the space of mobile manipulation, the method that seems to be prevalent in the literature for the handling of model or parameter uncertainties is robust and adaptive control techniques.

One such primary control method that has been extensively used in the field of mobile manipulators is variable structure control/sliding mode control (SMC), owing to its desirable tolerance to some degree of parameter uncertainty [375–377]. In such methodologies, a discontinuous control law is employed to guide the system to a switching surface wherein once the system is on the switching surface, the control law guarantees that the dynamics are insensitive to parameter variations. Further complex robust control techniques such as backstepping control [378] and its variants in combination with SMC [379] have been deployed. In Ref. [94], a nonlinear robust control technique was presented based on compensation for uncertainty, which was predicted/estimated using a disturbance observer. It is also noteworthy that some have designed robust control techniques based on dynamic compensation or torque compensation to eliminate the disturbance caused by the coupling of the base and manipulator [380,381].

Adaptive control methods can effectively cope with parameter errors as well as modeling errors by adapting controller gains. The feedback from the plant is designed in such a way that the end behavior is identical to the ideal system described by the analytical model. Several studies have used adaptive control directly for common control problems such as trajectory tracking, force control, and so on [109,297,382–384]. A variant of adaptive control is robust adaptive control wherein further error feedback is added for robustness, which has also been explored in the literature [132,385–388]. Also, advanced techniques can be combined to add robustness to adaptive control as shown in Ref. [389], where sliding mode control was used for disturbance suppression while adaptive control took care of uncertainties. Similarly, an adaptive force control technique was presented in Ref. [390] where robustness was added to attenuate disturbances.

In accordance with the discussion of the control methodologies being a continuum, a lot of model-free methods have risen in the recent past. Although rule-based approaches such as fuzzy logic controllers can also can be categorized as MFCs [391–393], intelligent approaches are pursed with a lot more interest. Methods such as RL bypass the need for model-based design by learning a feedback control law based on data and have already been used extensively in the mobile manipulation paradigm. An explanation of the fundamentals of these methodologies and references for the latest literature in the domain of static manipulators can be found in Refs. [345,394]. A pick-n-place motion tracking operation was controlled using the end-to-end deep RL technique in Ref. [101], wherein the base was controlled based on a controller that was learned using the manipulator end-effector task-space data. In support of the discussions presented in Sec. 3.3.1, to overcome the complexities of deriving an extended unified model, Ref. [395] introduced a whole-body control algorithm, learnt using RL, wherein direct visual sensor information was used as the training data. Similarly, a whole-body control algorithm based on RL was proposed in Ref. [396] in which the training was based on LiDAR scans of the robot. A hierarchical controller where the lower level joint space control alone is designed using RL was presented in Ref. [397], which serves as an adaptive feedback control law. On a different note, in Ref. [398], RL was used to learn the kinematic feasibility and manipulability of a mobile manipulator. The same is then used for dynamic motion generation for the robot to track the task-space references. Apart from RL, neural networks and their variants (recurrent, etc.) have also been used in the literature to map sensor data/output to a feasible input using a trained controller [399–402]. The problem of uneven terrain handling is also solved using a learning-based control algorithm in Ref. [403]. A similar application, robot parking among uncertainty and cloudy sensor data, has been solved using learning methods in Ref. [404]. A neural net-based tip-over avoidance controller that both detects and provides real-time mitigation control inputs to avoid tip-overs was proposed in Ref. [247]. Apart from directly just learning data-driven controllers, neural nets have also been deployed to solve nonlinear optimization problems. A robust zeroing neural network was proposed in Ref. [296] that deployed a successful robotic manipulator path tracking example in a noise-polluted environment to show its versatility.

A deduction can be made that MFC approaches are undeniably effective in handling the problems of mobile manipulation. However, the application scenarios and literature still raise questions about these methods in terms of repeatability, safety, guarantees, adaptability, etc. There have been hybrid methods in the literature that aim to combine the best-from MBC, and MFC approaches to truly expand the capabilities of mobile manipulators and other systems in general. Several published works in the mobile manipulator space focus on hybrid methods integrating system dynamics-based control for robustness and learning-based methods for adaptability [405]. Sliding mode control combined with intelligent methods has been explored in Refs. [406,407], where a robust controller is designed using the system dynamics to bring them to a sliding surface post, which a learned proportional control acts as compensation for adaptive operation. Alternatively, a combination of fuzzy control and neural networks have been presented in Refs. [408,409], wherein the fuzzy control takes care of generating control for navigation while the neural net acts as a robustness compensator. In a manner of perfect collision of both worlds, a model-based control that performs inverse dynamics for control is augmented using a neural net for robustness in Ref. [410]. Finally, in Ref. [411], a neural network was trained to provide feedforward torque, which was then layered with a fuzzy-based feedback loop. In addition, the weights of the neural network were continually updated using an extended Kalman filter in real time that estimates the best weights based on the system output measurement.

For quick reference, a graphical representation of the literature existing in the continuum discussed in this section is presented in Fig. 6.

Human–robot collaboration is an important aspect of mobile manipulators used in a variety of applications. In fact, most applications require some level of human control [154,155,168,412,413], and the interaction of the robots with humans is unavoidable [154,414–418] regardless of the autonomy level exhibited by the robot. Manual control of mobile manipulators can often be very challenging for humans due to the difficulties in perceiving the robot’s surroundings perfectly as the base or arm moves. Furthermore, the relationship between joint angles of a high DOF robotic arm and its resultant pose in Cartesian space is extremely nonintuitive for human operators. There needs to be a good balance between the level of autonomous decision-making and human oversight [419], so that the effort required by the human operator is minimized while improving robot performance. Moreover, robots often need to have physical interactions with humans in the workplace while assisting or collaborating on a task. This necessitates the robots to have robust decision-making capabilities to ensure safe and efficient interacts with humans despite being tele operated.

There are many works focusing on the different human–robot interfaces (HRIs) for controlling a mobile manipulator. The work described in Ref. [420] presents a successful remote manipulation method using gestures, which can be employed by a nonprofessional operator that results in high accuracy tracking, showcasing the need for intuitive command techniques that require less effort from humans. Alternatively, the methodology reported in Ref. [155] uses a 3-DOF interface like a sip-puff or joystick to command end-effector for performing pick and place operations, to be used by mobility-impaired people and hence foregoing the need for advanced gesture-recognition techniques. However, advanced artificial intelligence techniques can significantly improve interpretation and skill on the mobile manipulator’s side as described in Ref. [419]. Voice commands have been explored as an useful alternative to both perception-based or joystick-based teleoperation for simple domestic tasks in Refs. [142,152,153].

In the case of the teleoperation of a mobile manipulator from a remote location, understandably, the human operator often feels a disconnect from the actual location where the robot is operating. To ensure transparency in teleoperation, work presented in Ref. [421] demonstrates an augmented reality (AR)-based system to give the operator the feel of the operation site of the robot, where they consider senses of sight, touch, and hearing. Another intuitive way is to integrate virtual reality interfaces within the control system of a mobile manipulator as demonstrated in Ref. [422].

As discussed previously, delivering an object to a human is a necessary general-purpose capability for an assistive mobile manipulator and is hence a major area of study due to the high safety standards. For example, control issues for service robots are addressed in Ref. [417] for delivering and handing over objects to a human. Further, Choi et al. [154] present a user study for object delivery to motor-impaired patients autonomously. Here, the direct delivery method (handing the object to the user) and the indirect delivery method (placing the object on a nearby table) were compared, and it was established that the indirect delivery approach was more robust and reliable with high user satisfaction.

Assistive tasks often require tactile and pressure sensing capabilities. To that end, a pressure-sensitive “robot skin” was utilized in Ref. [423] for safe physical human–robot interaction. Further, Dean-Leon et al. [424] present the mechatronic design of their tactile omnidirectional robot manipulator (TOMM) with robot skin and methodologies that utilize the multimodal tactile information.

In addition, functional capability of a mobile manipulator needs to consist of mechanisms to ensure the safety of the people around while operates in an environment with humans. The autonomous mobile manipulator in Ref. [425] uses heat map-based social density monitoring for selective cleaning while maintaining social distancing with surrounding humans. Needless to say, telenursing robots like TRINA [168] require safety in close proximity humans.

Apart from just interaction, it is also important to focus on the ease of human operators to assign tasks to mobile manipulators. The operators should ideally not be expected to have any acquired skill to operate a robot to ensure ease of deployment. However, this can often be an issue especially in a complex system such as mobile manipulators. Huang and Cakmak [426] describe Code3, which is a system for user-friendly, skill-irrelevant, rapid programming of mobile manipulators. Similarly, Schou et al. [427] described a human–robot interface based on task-level programming and kinesthetic teaching to simply programming specifically for industrial tasks. The interface was assessed by people with varying robotics experience, and the goal is to make it available for production floor operators in industrial settings.

Collaborative tasks with humans can be especially challenging since the robot requires sufficient adaptability and safety measures. This issue was highlighted in Ref. [428], which deals with active cooperative tasks between a mobile manipulator and a human, such as carrying a long object together. They demonstrate an intention recognition capability in the robot based on the search for spectral patterns in the force signal measured at the arm gripper. An architectural discussion for a robot involved in human interaction setting was discussed in Ref. [416]. Here, a human -aware manipulation planner was dedicated as a supervision system during collaborative task achievement thus ensures safety. An intuitive control strategy was presented in Ref. [412] for human–robot collaboration involving mobile manipulators.

A mobile manipulator can be used for a wide variety of tasks since the environment where they need to work is designed for humans, and the structures of mobile manipulators resemble human anatomy, i.e., they have a robotic arm on a mobile platform. However, to make them more useful, they need to possess capabilities to take over tasks that are challenging or ergonomically infeasible for humans. With this in mind, here are some of the potential future challenging problems that one can work on.

Grasping or manipulating an object while the mobile base moves at high speed can have several benefits. It can optimize pick-up and transportation tasks. However, it is a challenging task due to the precise coordination requirement between the arm and the mobile base. For example, grasping an object while moving requires high-speed perception, precise control, and appropriate motion planning of the mobile manipulator. Even though grasping moving objects with a stationary manipulator [429] has been explored, the antithetical problem of grasping an object while mobile base is in high-speed motion is relatively unexplored. Further, uncertainty in the pose of the object can add further complications to the problem. Hence, a systematic study that involves coordinating the perception, planning, and control of the robot is necessary. Hence, demonstrating dynamic mobile manipulation at high speeds is a necessary future step.

Another area of research that can be explored is to utilize a sensor-equipped manipulator to assist in the navigation of the mobile base. For example, in tight spaces where the uncertainty in the pose of the base results in a high probability of collision, a camera with vision sensors can dynamically move and image parts of the base close to the obstacle. This is a useful intuitive way to safely help the mobile base navigate in low-tolerance environments. Furthermore, in cases where the vision system is turning out to be unreliable in depth perception, the tactile sensors and thereby impedance control on the manipulator can be employed to get a sense of the distance to obstacles.

Mobile manipulator design plays an important role in the accuracy of execution of the computed trajectories. A more dynamically stable manipulator will be accurate in performing high-speed tasks. Currently, a mobile manipulator is constructed by integrating a commercially available base and a manipulator. The integration requires designing a structure on the base to support the manipulator, controller, and other equipment. Lack of design principles will often lead to an unstable system and a higher center of gravity of the mobile manipulator. Careful consideration of the weights and balance of equipment on the mobile manipulator can be given to improve the accuracy. Research work can be done in this area to study how the construction of the mobile base influences the trajectory execution. The structured design of the robot will also generate a known dynamic model that can be used during planning.

The simplest version of a mobile manipulator, i.e., with one arm, is well studied. But a mobile manipulator with one arm is not always the best solution for an application. A few applications such as machine tending, assembly, hospital tasks, etc. are better performed by a mobile manipulator with two or more arms. This is often not pursued because two or more arms add significant complexities to the planning problem. For example, in task planning, the number of agents increases with the number of arms, and these agents need to communicate for efficient task execution. In the case of motion planning, multiple arms serve as dynamic obstacles for each other, which leads to a significantly overconstrained problem that might lead to infeasible solutions. Also, there are applications that require multiple arms to move in a synchronous manner. These are significant yet necessary problems that need to be solved for the advancement of this field.

Planning algorithms and navigation strategies have almost entirely aimed at avoiding collisions, with the obstacles and workspace boundaries serving to impose constraints on robot paths. In the past, robots’ structural and internal components were expensive, fragile, and in general not considered dispensable, which clearly motivated the urge for collision avoidance. Today, however, new methods for manufacturing have enabled the development of a different class of robots that can either withstand collisions or be considered dispensable when they do not always survive them. There has even been evidence that robotic missions can benefit from contact with the environment. Yet efforts to understand, implement, and exploit such behaviors are still in their infancy. Early work demonstrates the potential for physical robot–environment interaction to contribute to new strategies for motion planning. Mobile manipulators offer more room for the designer to make them robust to collisions, add shielding, and even employ mechanisms to exploit boundary interactions. Mathematical models that capture impact behavior and planners that intentionally incorporate these behaviors are currently underdeveloped.

Mobile robots often are employed in the transportation or manipulation of deformable objects. For example, industrial applications often require transporting flexible materials from one factory station to the other. The materials need to be handled by a collaborative effort between multiple human operators leading to misuse of valuable human labor. Advances in the planning of synchronized trajectories for multiple mobile manipulators to carry and transport flexible materials is a significant control task. The fact that the materials themselves could also required to be manipulated and stored adds further complexities. For instance, composite industries require transportation and manipulation of large flexible composite sheets from the material station to the mold on which they are formed. Planning synchronous trajectories and highly coordinated reactive control for multiple mobile robots poses algorithmic challenges that need to be addressed more.

Mobile manipulators, when deployed, need to be energy efficient to be economically feasible. If the mobile manipulator needs to charge while performing a task, it slows it down and adds downtime. Arguments can be made that technologies like fast charging and batteries with higher charge density could solve those issues. However, consideration of factors like efficient task and motion planning during the algorithm development phase can also be very useful. In the case of multiple mobile manipulators, the tasks should be planned such that the overall energy utilization is low and the energy usage is equally distributed between multiple mobile manipulators. Motion planning and control can be done in a way that the mobile base uses the least amount of energy, and the manipulator does not need to perform inefficient manipulations. This area of research will be very useful for industries aiming to mass-adopt mobile manipulators in a sustainable and economically feasible manner.

We envision mobile manipulators operating alongside humans within an environment in the future. To achieve that, absolute safety needs to be guaranteed. While there have been lots of works on human-safe industrial robotic manipulators [430–432] and automatic guided vehicles (AGVs) or mobile robots [433–435], there are not as many works for mobile manipulators. A mobile manipulator comes with its own challenges for the integration of robotic arm(s) onto a mobile base. Marvel and Bostelman [436] describe safety standards specifically for mobile manipulators in modern manufacturing and discuss the need for a new class of test artifacts for mobile manipulators in collaborative environments. Ensuring safety requires development on three major fronts. First, we need superior and reliable sensing capability, which has zero to little uncertainty in perceiving its environment dynamically. Second, the system needs to find, plan, and execute appropriate corrective actions based on the sensor feeds. Finally, in the case of an unavoidable collision, the system needs to minimize human injury. Reliability in all these features will ensure safety for the surrounding humans, which would escalate the real-world deployment of mobile manipulators in the future.

Mobile manipulator systems often get very complex, and their complexity is only going to increase owing to the rapid deployment of robots for several advanced tasks in industrial settings. As the complexity of the systems increase, often the modeling of these systems using first principles become extremely difficult, rendering model-based control inept. There has been a steady increase in the deployment of data-driven control methodologies, as many modern industrial processes produce huge amounts of process data containing all the valuable state information about its own dynamics and state-flows. Designing controllers or estimators directly using these data, online and/or offline, is extremely appealing especially given the lack of accurate process models. However, in applications such as mobile manipulators that operate amongst humans, safety is of great concern, hence requiring rigorous verification and safety guarantees before deployment. Here is where model-based control holds a greater advantage due to the fact they are verifiable, repeatable, and explainable. It is clear that both these control methodologies hold significant strengths that make a case for each; however, we envision hybrid methods that combine the best of both worlds can greatly benefit the automation of mobile manipulators. Although explainable artificial intelligence and hybrid control strategies are popular research topics in robotics, there are significant gaps particular to mobile manipulators, and hence, research in hybrid control theory applied to mobile manipulators is a necessary future direction.

There are many open challenges toward next-generation human–robot collaboration with mobile manipulators. We have more works on HRIs for robotic manipulators and AGVs separately than we have for mobile manipulators. We need improved communication modalities to ensure natural communication between robots and humans. The most common kinds of communications would be humans providing instructions to the robot and the robot relaying new information to humans. Voice-based or natural language-based communication can be extremely natural for humans, but gestures, AR or virtual reality systems, or even joystick and computer keyboard-mouse and screen displays can also provide useful functionalities for HRI. Another aspect includes mobile manipulators performing physically collaborative tasks with humans. Most of the features we mentioned for safe operation in the presence of humans will be useful for such collaborative tasks.

Research interest in utilizing the mobility of mobile robots and the manipulability of manipulators to provide enlarged workspaces and adaptability in skills has had a significant presence in robotics in the past few decades. Despite its theoretical merit, the implementation of such systems in actual applications has largely been limited due to significant challenges related to online planning and control. However, in recent times, the triple convergence of developments in miniaturized computing, sensing, and advanced algorithms have provided a lot of capabilities that enable robots to graduate from mere automated systems to autonomous ones. These capabilities have now brought about a renewed interest in mobile manipulators, which are being considered in new applications at a rapid rate. Autonomous deployment for mobile manipulators is emerging as new fields of research fostering advances in industry 4.0. In recent times, research in all aspects of mobile manipulators, from design optimization and architecture to perception and decision-making, has been on the rise. Our focus has been limited to the decision-making methodologies of wheeled mobile manipulators in this article. Several research challenges and future directions are outlined in this study that needs to be addressed and pursued to further expand the applicability of mobile manipulators in challenging applications. The scope of necessary research is not only limited to particularly robotics but, as outlined in this study, also expands to several fundamental theoretical areas that need advances. Further advances in these areas will enable not only mobile manipulators to be more effectively deployed in existing applications but also enable new applications.

The authors would like to acknowledge support of their colleagues from RROS Lab at the University of Southern California and ARMLab at the Clemson University International Center for Automotive Research. We also acknowledge partial support for this work from the National Science Foundation National Robotics Initiative grants 1925084 and 1924721, the Computing Community Research Infrastructure grant CNS-1925500, and the ROSEHUB-IUCRC grant CNS-1939058. Opinions expressed are those of the authors and do not necessarily reflect opinions of the sponsors.

There are no conflicts of interest.

No data, models, or code were generated or used for this paper.