Multi-Robot Assembly Strategies and Metrics

3.1 Peg-in-Hole Assembly

Peg-in-hole assembly is the quintessential problem domain for assembly research. It consists of inserting a rigid object (i.e., a peg) into an appropriately shaped receptacle (i.e., a hole). Generally speaking, it is assumed that the peg is in constant contact with the surface of the assembly fixture until the hole is found such that the insertion point is located and the peg slips in. However, if constant contact with the assembly cannot be maintained, such searches may also be conducted by gently probing the surface until the insertion point is located. Given sufficiently large hole tolerances and minimal pose uncertainty of both the peg and the hole, peg-in-hole assemblies can be completed using position control with the robot simply moving directly toward a known goal state. However, the large tolerances and a priori knowledge required for success are typically not available in real-world applications.

This assembly process is challenging because of the physical contact required between the workpiece rigidly held by a robot’s end effector and some other object. Errors in sensing or control may result in excessive application of force or in the binding of partially engaged parts, thus damaging parts, tooling, and/or robots. Mechanisms to compensate for such errors include the remote center of compliance (RCC), which is passive mechanical tooling that accommodates slight orientation errors of the peg relative to the orientation of the hole. Machine vision can also be applied to identify, localize, and—in the case of moving assembly lines—track insertion locations. With increases in computational and sensing capabilities paired with decreasing costs, force-based control is becoming more commonplace in assembly operations. A number of force-guided peg-in-hole insertion strategies have been implemented in the literature, and robot vendors now provide off-the-shelf software options for their hardware platforms. Most of these solutions reduce the problem to a two-dimensional (2D) problem space, though some researchers have extended the problem into the full three-dimensional (3D) search space (e.g., Fei and Zhao (2003)). In the following subsections, we present these search strategies as they could be implemented on a stock industrial robot controller. Some of these strategies, namely stochastic search, spiral search, and raster search, are illustrated in Figure 1. Algorithms are presented as piecewise adjustments to the current position of the TCP as a function of iteration.

3.2 Gear Meshing

The generic “gear meshing” assembly seeks to align complex physical features of freely moving components, typically cogs or chains. With such assembly processes, the risks of damaging the parts, tooling, or robot stem from the potential for binding. When the components cease moving freely, the effort to mate the parts results in high torsional values, which can warp, crack, or shear component features. As such, gear meshing operations are often complex processes that rely on tight control loops with force sensing (e.g., Gravel and Newman (2001)) or dedicated tooling (e.g., Schimmels (2002)). For robotic assembly applications, a common approach is to pair a back-and-forth rotational action with a planar positional offset (here, a circular motion) to alternatingly increase and decrease contact forces through partial cog engagements.

3.3 Snap and Press Fits: Linear Search

In some cases, an assembly action consists simply of pushing two components together with sufficient force to engage snap-in tabs or tight, press-in connections. Such assemblies are common in consumer electronics and have been successfully implemented for decades using selective compliance articulated (or assembly) robot arm (SCARA) systems.

In the generalized case, snap and press fits are completed using simple linear searches under force control. The linear search introduces a back-and-forth linear motion along a defined vector from the robot’s initial position. Termination conditions include defined insertion depths, specific force profiles, or audible “snaps” (Genc et al. 1998).

For a linear search, let x be the user-specified X-axis offset, y be the Y-axis offset, z be the Z-axis offset, and s be the desired speed of the tool motion. Depending on the force control loop, s may be defined or limited by the desired applied force. Given the initial starting point p₀, search call iteration counter, c, and update frequency f, the linear distance change per time step, l_step, is defined as

and the total length of the search motion, l_total, is calculated as

The magnitude of the change in position as a function of search speed is calculated as

and the individual axis offsets are computed by

To ensure predictable behavior, a local search call iterator is maintained to determine where along the linear path the workpiece should be located. This local iterator, c_local, is defined as

This value is then used to determine the step-iteration X-, Y-, and Z-axis offsets as

3.4 General Assembly

There exist a number of motion strategies that are not specifically targeted toward completing particular assembly types. However, these general strategies are typically combined with other searches to either avoid or compensate for pose and localization errors and excessive forces and torques.

3.5 Complex Assembly Strategies

With the exception of the tilt strategy, the general search methods described previously may be logically combined and chained to generate complex assembly actions. Because each search results in positional offsets from the initial pose, p₀, simply adding together all j p_Δ vectors for a given c results in offsets that reflect multiple transformations:

For instance, assume that an assembly consists of inserting a central sun gear into a planetary gear configuration. The search strategy may consist of a rotational search combined with a circular search and a hopping motion. Summing the offsets from these three search strategies will result in the rigidly held gear simultaneously oscillating back and forth along the Z axis, moving in circular motions along the X-Y plane, and moving up and down on along the Z axis.

Similarly, multi-stage assemblies are accommodated by daisy chaining assembly search strategies predicated on the successful achievement of positive termination conditions (e.g., Marvel et al. (2008)). For instance, the insertion of a sun gear fitted with a shaft into a hole will typically require a two-stage assembly. Stage 1 consists of a spiral search paired with a downward linear search to insert the axle into the bearing, and stage 2 will consist of a rotational search paired with a hopping motion. The circular search is omitted from stage 2, because the axle is engaged with the bearing and adding lateral offsets will impose undue stress on both the axle and the bearing.

4.1 Time Profiles

A given assembly subtask is typically allocated a certain time limit in which the robot must complete its task to maintain production throughput. Tasks that are completed within that timeframe are considered in tolerance and any time spent beyond that limit will impact system performance. When a timer event occurs, a signal is sent to the robot to perform some conditional next step. Such next steps may be to discard the components for a human operator to finish offline or to reassess and restart the assembly operation from a previously known state. Although typically associated with signals to abort an assembly in the event of failures, timers may also be used to indicate task success. For instance, if parts must be held together for a given duration of time to allow adhesives to set, a timer event marks the successful completion of that particular task.

4.2 Force Profiles

In-line forces and torques are often measured and used to transition from the approach stage to subsequent stages of an assembly task. Appropriate deviations from some nominal force, torque, or motor current profile indicate successful interactions with the workpieces. Typically, a search profile will continue until a certain force/torque threshold has been met or exceeded. For instance, force spikes may indicate an object has been found in a “move until touch” motion primitive or completion of a snap-fit task in the presence of steadily increasing force values. Alternatively, force profiles may be used to ensure that components are not damaged during an assembly. For instance, if a nominal force limit has been met, an assembly search may be aborted to prevent excessive wear on the robot, tooling, or workpiece.

4.3 Motion Profiles

Like the force profiles, motion and position profiles may also be used to transition between stages or mark the need to abort a given assembly. If predefined depth thresholds are exceeded, for instance, then a robot may conclude that the part that it is holding has been successfully inserted into its respective receptacle. Similarly, if the lateral distance from p₀ exceeds some nominal maximum expected value prior to observing some other signal, the system may conclude that the search has failed and needs to be restarted.

4.4 External Signals

In some instances, a robot system may not know the state of an assembly without the assistance of an external observer system. In such cases, a search may need to be terminated based on external signals prior to the robot system’s own termination conditions being met. Support for such early terminators allow for observer systems to preemptively abort a robot system’s assembly searches with defined “success” or “failure” without requiring complex logic on the robot controller.

5.1 Multi-Robot Coordination

An entire field of research has been dedicated to the coordination of multiple robot systems for applications ranging from manufacturing to military surveillance. Such applications employ technologies and algorithms ranging from rudimentary collision avoidance to goal-based optimized swarming of robotic agents. This field is necessarily vast, and extends well beyond the scope of this article. For more detailed coverage of multi-robot coordination, it is recommended the reader begin with the surveys by Yan et al. (2013) and Doriya et al. (2015) and the book by Balch and Parker (Balch and Parker 2002). The literature on multi-robot coordination is largely centralized on mobile robotic platforms. Nevertheless, many important lessons can be drawn and extended to other robotic design paradigms used in manufacturing. As such, it is worth providing some basic—albeit brief—coverage of the topic here.

While there are many forms of multi-robot coordination, we want to draw particular attention to synchronization of independently operating robotic systems. Some robot manufacturers provide software solutions that enable several of their robots to be temporally and spatially synchronized using a single controller. Such solutions often provide some handling of process or environment uncertainty, but are otherwise limited in their ability to adapt to large errors given the reliance on strict constraints on the arrangement and control of the system. Independently operating robot systems, however, tend to be more flexible in handling uncertainties, but must be coordinated through communications to achieve any form of synchrony. While there are many approaches to facilitating this coordination, they are ultimately employing some variation of either centralized or distributed control. Similarly, communications may either be explicit in that messages of intent are directly conveyed between robots with or without goals defined a priori (e.g., Rubenstein et al. (2014)), or implicit in that the robots must infer intent by means of sensor-based observations given a shared goal (e.g., Tambe (1997)).

While centralized coordination of multiple robot systems may be used effectively for most assembly tasks, accommodating large uncertainties often requires the distributed coordination of independently intelligent robotic agents. Such coordination requires extensive communication, both implicit and explicit, to accommodate changes in the environment, task, and behaviors of other robotic agents (e.g., Kaminka et al. (2010)). However, such extreme uncertainties are unlikely for the majority of manufacturing tasks, and efforts to facilitate the communication and control complexity necessary for enabling this level of independent coordination in manufacturing robots are likely to experience diminishing returns for both assembly performance and system reliability.

5.2 Coordination Performance and Uncertainty

For any sufficiently advanced mechanical system, some means for evaluating the system and process performance are required to ensure continued operation, optimize production, assess root cause analyses of errors, and predict pending failures. Modern trends in smart manufacturing promote frequent turnover and customization of manufacturing processes to accommodate ever-shifting market demands, which makes developing long-term performance models for prognostics and health monitoring (PHM) particularly challenging (Choo et al. 2016).

Currently, there are no standardized test methods or metrics for the evaluation of multi-robot coordination and control. The basic robot performance standard from the International Organization of Standardization (ISO) (ISO 9283 1998), measures only for the spatial path and end-point accuracy and repeatability of a single robotic manipulator. It does not capture the ability of a given robot platform to coordinate its motions with those of another robot, human, or mechanical system. Moreover, this action-side test methodology does not account for the accuracy of the trajectories or motion plans, themselves, nor is it intended for real-time evaluations. Both of these further confounds PHM efforts, particularly with regards to process monitoring. The test methods in ISO 9283 may be extended and adapted, however, to account for multiple robots’ motions relative to one another as opposed to absolute measures. However, because one cannot expect robots to provide accurate real-time updates of its pose and joint configurations (Marvel and Norcross 2017), means of real-time monitoring by accurate, external observer systems will be required to enable such measures to be used in PHM.

The real-time position and configuration of multiple robots accounts only for a snap-shot of the effects of coordination and individual control of the independent platforms. These physical configuration errors are frequently a resulting symptom of some preceding series of root errors that, if identified earlier, could have been addressed to avoid potential damage to parts, robots, equipment, or human operators. Specifically, physical conflicts such as a lack of temporal or spatial synchronization, collisions, or the excessive build-up of internal stresses or forces of shared workpieces are likely the result of errors or uncertainty in the robots’ software coordination. At the team level, performance metrics have been proposed to evaluate system-wide errors in multi-robot planning (e.g., Sellner et al. (2006), Korsah et al. (2013), and Micalizio and Torasso (2014)), communications (Kalech 2012), and conflict negotiations (Kalech and Kaminka 2007). Alternatively, at the individual robot level, performance errors may result from shortcomings in software agility, possibly inferred based on the robot’s measured efficiency and effectiveness (Downs et al. 2016).

6.1 A Basis for Multiple Arms

A case study showing the benefits of using multiple collaborating robots to reduce the number of fixtures needed for welding in a Body-in-White application is presented in Papakostas et al. (2011). The authors compared two scenarios. In the first, two robots were used for welding components that were held in fixtures. This scenario required a total of six fixtures. In the second situation, three robots were used and the number of fixtures could be reduced to three. They showed a 49% reduction in cycle time for the case in which robots were used to hold the components instead of fixtures. In the first scenario, significant time (nearly 30s) was taken by having to frequently open and close the fixtures and grippers, which was substantially reduced in the second scenario. The use of robots also made the work cell more flexible in comparison to the fixture-based scenario.

Another study of multi-robot assembly was conducted by Rojas and II (2012). The authors used two heterogeneous robots, one of which was an industrial manipulator and the other an anthropomorphic, dual-armed robot. The task was to insert a pipe into a fitting. The “push-hold” and “push-push” strategies were both implemented. The authors found that the “push-push” strategy was quicker at completing the insertion than the “push-hold” scheme and reduced the likelihood of stiction, jamming, and wedging. Another conclusion was made for the case of “push-hold” schemes, where industrial robots are smoother and faster pushers, while compliant robots are better holders.

Shi and Menassa (2010) presented some challenges to assembly in automotive applications. They described a system for loading wheels and tires onto vehicles that used two robots and an automated guided vehicle. Computer vision was used to locate the parts, insert the wheels onto the lugs, and attach the bolts. The two robots worked independently, each on one wheel, but had to synchronize their actions with the motion of the AGV. All of the programming to achieve this was manually developed.

6.2 Multi-Arm Designs

Several multi-arm robotic systems have been developed that are aimed specifically at human-scale manufacturing tasks. For example, in the mid 1980s, Hörmann, Lueth, Nassal, and Rembold developed the mobile, dual-armed Karlsruhe Autonomous Mobile Robot (KAMRO) platform (Hörmann and Rembold 1991; Lueth et al. 1995). The robot consisted of two Puma 260 arms rigidly mounted on a mobile platform. The mobile base moved the dual arms to predefined, dockable locations within a work cell for on-table assembly. The assembly component consisted of several action primitives, including pick, place, and transfer (specifically, for handing parts between arms) (Hörmann et al. 1989). The KAMRO was used principally for test and evaluation of multi-robot control architectures and was applied in research testbeds toward assembly-related material handling problems.

In the 1990s, Yamada et al. (1995) presented their design for a multi-arm robot configuration for performing preprogrammed differential gearboxes assemblies (Figure 3). The robots and their attached parallel grippers were designed to hold a variety of parts while simultaneously maximizing the robots’ combined working volume. Parts were fed to the robots in a fixed kit of parts, and the assembly consisted of (1) the insertion of two sets of vertical pinion gears and spherical washers, (2) the insertion of two sets of horizontal pinion gears and spherical washers, and (3) the insertion of a pinion shaft through the differential case, gears, and washers. The two robot arms used position control and were largely independent of one another. However, each could potentially impact the other’s performance during the phase 2 insertion due to the engagement and rotation of the gears as they were inserted.

In the early 2000s, the HRP2 robot, a dual-armed, humanoid robot, was developed by Kaneko et al. (2002) for generalized task-driven application development of humanoid robotics. Like its predecessors, the HRP2 was targeted for assembly operations due to its dexterous nature. Dauchez et al. (2005), for instance, used a simulation of the HRP2 to perform peg-in-hole assemblies based on the assumption that both the peg and the receptacle were rigid extensions of the robot end effectors. However, no a priori assumptions were made regarding how the parts were held or where they would first make contact when brought together. During the approach phase, the arms would slowly bring the two assembly components together. Once the parts made contact, however, the goal configuration was assumed known, and the robot controller generated a trajectory using a force feedback law to move directly to the goal state while maintaining constant contact between the parts. This represents a simplified application of the tilt method described in Section 2, where the force control law is used to maintain contact and minimize the potential for parts to bind.

More recently, the ABB YuMi robot was developed for assembly tasks for the electronics industry. It focuses on safe operation and easy reconfiguration for new tasks in a human-occupied environment (Kock et al. 2011). Each arm has 7 degrees of freedom (DOF) to ensure that the entire workspace is reachable without bumping into obstacles. The robot uses an integrated industrial controller and requires manual programming. The robot was developed as part of the European RObot control for Skilled ExecuTion of Tasks (ROSETTA) project (Patel et al. 2012), which focused on robot safety, easy programming, and the use of machine learning. It is assumed ABB’s force-based assembly library (ABB) is leveraged to enable and control assembly operations.

6.3 Programming and Control

The use of human-scale robot systems allows for direct teaching of assembly strategies by demonstration. In the system presented by Park et al. (2009), an operator recorded motion trajectories by directly physically interacting with a custom dual-arm robot to complete a peg-in-hole assembly. These trajectories were then played back verbatim to complete subsequent assemblies without further human intervention. In the approach presented by Takamatsu et al. (2007), teaching the robot was done by means of a vision system that recognized and tracked rigid polyhedral structures held by human operators. The assembly task was broken down into defined motion primitives to define the assembly process, which was then replayed by the custom, dual-armed robot after having acquired the parts in its grippers.

A master-slave control strategy was used by Su et al. (2009) to enable two robots to screw a nut onto a bolt. A three-layer architecture was used with the top layer responsible for motion planning of the master robot, computing the kinematics cooperation relation, and computing the motion of the slave robot. The second, servo layer, handled joint control and execution of the motions of each robot. The third layer was responsible for adjusting the poses of the robots based on a sensor that determined the locations of the bolt and nut using a resolved motion rate control method. The master robot held and translated the bolt while the slave robot held and turned the screw. The speeds of the robots were controlled to ensure that the nut meshed with the threads of the bolt during the insertion.

A master-less control strategy developed by Šurdilovic et al. (2001) used two lightweight industrial robot arms modified to enable impedance control to assemble truss structures. Assemblies consisted of rods with locking mechanism integrated to snap fit into the vertex nodes (Figure 4). The structure was systematically assembled using different “push-hold” strategies that were dependent on the assembly operation. In operations where connection nodes were attached to a rod on the assembly structure, one robot would hold the rod by grasping it tightly while the other robot inserted the node over the rod. In contrast, when rods were attached to nodes on the assembly structure, one robot would hold the node by pushing it to keep it in place while the other robot inserted the rod into the node. In both instances, the hold operations were performed by robots operating in position control mode; while in position control, the robot would monitor force magnitudes but would not respond to them. Due to the lack of a defined master, the two robots coordinated through message exchanges to ensure proper timing for correct assembly completion. All of the assembly experiments were evaluated in conditions in which the node and rod alignment errors were kept at a minimum to eliminate the need for positional search. A similar approach was taken by Marvel (2010), in which two industrial robot arms were coordinated via message passing to complete a force-guided snap-fit of cellular telephone case parts.

Much of the academic multi-robot research makes use of agent-based control approaches or behavior-based methods. An approach described as agent based, but that looked like a state-based method, was Zone Logic by van Dyke Parunak (1999). In Zone Logic, each agent was a state-based machine with no knowledge of other agents, although agents could coordinate their actions by propagating constraints or by using an auction mechanism. Zone Logic was developed as a commercial product and has been used to control large transfer lines in factories. Jarvis et al. (2001) developed a system based on Zone Logic but also using a belief-desire-intention (BDI) agent structure. There were two kinds of agents: manufacturing agents controlled devices and machines, while part/product agents planned, scheduled, and executed tasks. A Programmable Logic Controller (PLC) implemented primitive actions for a physical device, such as opening or closing a clamp. The manufacturing agents controlled these actions by setting flags.

In many cases involving multiple arms simultaneously working on an assembly, the arms are used to hold different parts and work either independently or together to complete the assembly task. In contrast, systems like the one presented by Yuan (2006) use two robots to simultaneously hold and control a shared workpiece for assembly. An admittance control law was developed that compliantly moves a shared object across a surface, and then performs a peg-in-hole assembly using the tilt method described in Section 2. In this approach, one robot minimized the incurred forces and torques imposed by part misalignments while the other provided force tracking to maintain part orientation and gravity compensation.

Brogårdh (2009) discussed the evolution of control in industrial robots. He showed how model-based control has greatly improved the speed and accuracy of robots and described the need for expanding the models to include sensor-based control, especially for new and more demanding applications, such as assembly and safety in human-robot collaboration. He stated that the main challenges for multi-robot control are in the architecture of the motion control software. It has to generate servo commands to the robots at exact times and be able to transition smoothly between coordinated and independent robot motions. It also needs to recover from failures without collisions between the robots. Since collaborating robots have serially connected kinematics chains, errors in the servo loops and in the robot models will cause pose deviations between the tool and the work object that are larger than for single robots. This means that collaborative robots must have higher-accuracy servo loops, servo references, feed-forward calculations, dynamic models, and kinematics models.

The ongoing software-driven robotics (SoftRobot) research project by Hoffmann (2014) is principally aimed at reducing the intellectual overhead required for programming complex robotic tasks. As a demonstration of this capability, the researchers have developed a complex assembly task illustrating the performance capabilities of two industrial robot arms and an omnidirectional mobile robot (Hoffmann 2012). The assembly consisted of loose slip-fit metal components held together by machine screws. The parts were delivered on kit trays to the robot arms by the mobile base. The robot arms registered, acquired, and moved these kit trays to locations within the task environment. Parts were then taken from the kit trays in known configurations using specialized tooling, which reduced the complexity of the assembly problem. Once assembled, the completed work pieces were then put back onto the kit trays and transferred back to the mobile base. The assembly process was controlled entirely by the SoftRobot architecture, including (1) the force-based registration of the kit trays, (2) the force-based acquisition of parts, (3) the coordinated motion of the robots for assisted lifting, and (4) the force-based manipulation of parts.

11.1 Trade-offs with Multi-Robot Assembly

In Section 1, we briefly discussed the trade-offs associated with the transition from robotic assembly lines to robotic assembly cells. Moving from single- to multi-robot assembly processes introduce additional trade-offs, some of which may be directly measurable (see Section 11.3). When introducing multiple robots, for instance, throughput for a given assembly cell may be increased given the capacity to parallelize processes within a relatively limited footprint. In contrast, a single-robot cell would require serializing the assembly process, frequent tool changes, and, while only housing a single robot, careful planning of the locations and order of operations to accommodate the limited reach of a single robot manipulator. Similarly, by increasing the number of robots in the work cell, one increases agility of the cell. This enables not only a greater variety of assemblies being produced by a single cell, but also allows for multiple assemblies to be created in parallel per cell.

As a trade-off, however, by increasing the number and variety of robots in a given work cell, the challenge of controlling, coordinating, and debugging the robots increases commensurately. Parallelization, in particular, adds to the complexity of the code. A single-robot cell’s serial nature promotes easier tuning of parameters and motions, while a multi-robot system would necessitate optimizing a complex dance of moving manipulators and parts. This takes more time and programming effort to coordinate, which does not necessarily translate to increases in process or product quality. Similarly, with more robots simultaneously active within a confined workspace, the likelihood of collisions increases, thus necessitating increasingly complex observer systems to ensure the safety of human operators, robots, parts, and tools.

11.2 Trends in Multi-Robot Assembly

In general, increasing the number of robots actively servoing to complete an assembly task greatly increases the control complexity while offering some benefit in terms of assembly quality or timing. The principal positive impact of such applications to assembly tasks is measured in terms of process agility and a reduction on infrastructure investment. Specifically, by using a robot as a reconfigurable and dynamic fixture, the required work cell complexity is reduced due to the obsolescence of rigid, non-reconfigurable fixtures, conveyors, and careful planning of real estate. Similarly, a multi-robot work cell may be quickly re-tasked with a different assembly process of similar scale (e.g., with the introduction of a new model year product) with minimal work cell and task redesign.

In small-scale assembly, the active coordination of multiple robots appears to have minimal impact on the actual performance of the assembly in terms of quality or time necessary to complete certain assembly tasks. Many of the successful strategies rely on one robot acting as a modular fixture while the other performs the actual assembly action. Many coordinated applications of robot-robot collaborations for assembly provide proofs of concept but do not demonstrate actual manufacturing capabilities. Regardless, the potential for realizable performance improvements is illustrated, and it is expected that continued research in this domain will net positive results.

In contrast, with large-scale assembly tasks, mobility and component support are key factors in defining the robot system design and effective functionality. Large-scale robots with heavy lift capacities are useful for macro positioning, while dexterous manipulation is a function of smaller, cooperative robotic platforms. Like the small-scale assembly tasks, successfully building large-scale assemblies does not necessitate the coordinated, active servoing of multiple robots. However, the synchronized control of support- and manipulation-role robots will reduce the need for stop-and-go corrections of workpiece positioning and thus improve system cycle time as a whole.

11.3 Multi-Robot Assembly Metrics

As mentioned briefly in Section 2, the most common performance metrics associated with single-robot assembly tasks include profiles of the timing, motions, and forces of the assembly process. Outside of the obvious question of whether the assembly was successful, some of the performance metrics found in and inspired by the literature landscape include:

These metrics hold true also for multi-robot assembly. With additional robots included in the assembly task, however, the motions of a single robot are expected to impact the motions and performances of the others. This results in increased positional errors and incurred forces on the robots’ joints, the tooling, and the assembly workpieces. Assembly strategies that minimize these impacts are expected to improve the quality of the assembly but are also likely to result in increased assembly times and computational complexity of the control laws. For multi-robot configurations, additional metrics include:

• Efficiency:

  • How much time is required to configure and tune the multi-robot assembly solution as compared with the time for a single robot to perform the same task?

  • What is the cost (material) difference between the multi-robot assembly configuration versus the single-robot setup for the same assembly task?

  • How much additional resources (e.g., space, equipment, or energy) are required to support a multi-robot configuration versus a single-robot solution?

• Timing:

  • How does the timing change when adding or changing robots to the assembly application?

• Forces:

  • What is the force transfer from one robot to others in the multi-robot configuration?

A brief summary of some common metrics applicable to multi-robot assemblies is given in Table 3. Ultimately, it is worth remembering that robotic assembly tasks in manufacturing are guided by the same tenets that guide any adoption of automation technology: (1) process quality, (2) process efficiency, and (3) return on investment. For any given application of automation technology, there must be a measurable positive impact on the manufacturing process for such solutions to be considered viable. A given solution’s ease of use is anticipated as being one of the key factors when determining broad acceptance of robotic assembly. If the time and effort necessary to accommodate even simple assembly tasks eclipses the expected return on investment, then solutions in robot assembly are unlikely to be accepted. Online and adaptive process optimizations have demonstrated realizable potential for automatically improving the performance of robots performing assembly tasks, but the initial, pre-optimized integration and implementation remain as hurdles to broad adoption.

11.4 Open Challenges in Multi-Robot Assembly

Despite the technological advances toward robotic assembly, there remain a number of challenges that need addressing. Many of these issues are not limited to the multi-robot assembly problem, however. For instance, ease-of-programming and effective user interfaces are an open challenge. The industry is responding to such calls for improving the user experience, yet progress is slow, and is largely evolving separate from the communities of end users. Similarly, sensor-based, adaptive control of individual manipulators remains difficult, despite increases in computing and sensing capabilities while the cost of entry decreases. Continued research will only improve research, though broad improvements are unlikely to be realized in the near future.

A particular challenge facing multi-robot assembly, however, is centered on the lack of support for heterogeneous robot configurations. Single-manufacturer solutions are already challenging, but the difficulty of integrating, registering, configuring, and coordinating robots increases significantly when one intends to use systems from different manufacturers. System-agnostic solutions will be required to enable the existence of such configurations, let alone effective functionality. While individual research efforts are focused on enabling certain capabilities (e.g., improved calibration and registration (Marvel et al. 2015; Van Wyk and Marvel 2017) and common command languages (Edwards and Lewis 2012)), realizable improvements will require the active cooperation of the robot manufacturers. As is the case with the robot-on-mobile-platform paradigms, garnering such cooperation requires a clear, marketable industry need to exist. Although the needs and technologies exist today, they have yet to grow to the level that warrants concerted effort.