Mechanical Layout

Today’s wearable robotic devices feature a diverse variety of mechanical designs depending on the specific task they are intended to perform. Although examples of hybrid designs exist, most design approaches for wearable robots can be classified into two main categories: stiff devices, which consist of rigid, mostly metallic, structural elements, and so-called “exosuits,” which consist of flexible or even textile elements (Sanchez-Villamañan et al., Reference Sanchez-Villamañan, Gonzalez-Vargas, Torricelli, Moreno and Pons2019). The main challenge that engineers face in the development of wearable robot mechanics is to realize compact and light structures, combined with suitable kinematics so that the robot does not hinder the human’s musculoskeletal kinematics. In other words, good exoskeleton mechanical design must be able to effectively replicate or support the complex movements of the human limbs while being as unobtrusive as possible in their utilization context and in their management.

Stiff Exoskeleton Designs

Exoskeletons with stiff structure and active joint actuation are the most widely known wearable devices and will be addressed in this section by analyzing applications for both the upper and lower limb. Depending on the amount of net active torque they can generate, they can be classified as “high-torque” or “low-torque” solutions. Specifically, “high-torque” stiff exoskeletons can generate sufficient torque to realize absolute control over the limb to which the exoskeleton is attached. This subcategory is ideal either when the wearer needs substantial support by the machine or when high levels of controllability are needed on the limb. High-torque use cases include lower limb assistance for severely impaired people, (Esquenazi et al., Reference Esquenazi, Talaty, Packel and Saulino2012; Wang et al., Reference Wang, Wang, Meijneke, van Asseldonk, Hoellinger, Cheron, Ivanenko, La Scaleia, Sylos-Labini, Molinari, Tamburella, Pisotta, Thorsteinsson, Ilzkovitz, Gancet, Nevatia, Hauffe, Zanow and van der Kooij2015; Milia et al., Reference Milia, De Salvo, Caserio, Cope, Weber, Santella, Fiorini, Baldoni, Bruschi, Bigazzi, Cencetti, Da Campo, Bigazzi and Bigazzi2016; Vouga et al., Reference Vouga, Baud, Fasola, Bouri and Bleuler2017; van der Kooij et al., Reference van der Kooij, van Asseldonk, van Oort, Sluiter, Emmens, Witteveen, Tagliamonte, Tamburella, Pisotta, Masciullo, Arquilla, Molinari, Wu, Ijspeert, Dzeladini, Thorsteinsson, Arami, Burdet, Huang, Gregoor, Meijneke, Carrozza, Micera and Pons2019; Jyräkoski et al., Reference Jyräkoski, Merilampi, Puustinen and Kärki2020; Figure 2c; www.indego.com) or power augmentation applications in which also upper limbs might be involved (Kazerooni, Reference Kazerooni2005; Marcheschi et al., Reference Marcheschi, Salsedo, Fontana and Bergamasco2011; www.sarcos.com; www.raytheon.com; Figure 2a,b). While kinematics that enables high levels of controllability of the limb is frequently used in rehabilitation applications since it allows the movement pattern to be precisely tuned to the patient’s needs (Bai and Christensen, Reference Bai and Christensen2017; Molteni et al., Reference Molteni, Gasperini, Cannaviello and Guanziroli2018), power augmentation applications typically do not demand total control of the biological joints. Indeed, within the “high-torque” class of wearable robots, some devices do not constrain the full limb of the wearer and deliberately leave freedom of movement to some biological joints. A similar approach can be found in underconstrained exoskeletons, where dedicated kinematic solutions and redundancy are exploited to maximize the wearability and usability of the device (Lo and Xie, Reference Lo and Xie2012; Tiseni et al., Reference Tiseni2019). In fact, nonfully constrained architectures may also be employed in other applications where absolute control of limb joints is not needed, such as assistive/industrial scenarios (Guizzo and Goldstein, Reference Guizzo and Goldstein2005; Bogue, Reference Bogue2018). In these applications, biological joints might not exactly match those of the robot, resolving possible joint alignment issues.

Figure 2. Stiff exoskeletons: (a) the power augmentation exoskeleton Guardian XO from Sarcos (www.sarcos.com), (b) the Raytheon XOS 2 (www.raytheon.com), (c) the lower limb rehabilitation exoskeleton Indego (www.indego.com; Farris et al., Reference Farris, Quintero and Goldfarb2011; Jyräkoski et al., Reference Jyräkoski, Merilampi, Puustinen and Kärki2020), (d) assistive lower limb exoskeleton GEMS (Lim et al., Reference Lim, Lee, Jang, Kim, Park, Seo and Shim2019), (e) the rehabilitation upper limb Harmony Exoskeleton by ReNeu Lab (www.reneu.robotics.utexas.edu/harmony-exoskeleton; Kim and Deshpande, Reference Kim and Deshpande2017), (f) rehabilitation upper limb exoskeleton ANYexo from ETH (www.sms.hest.ethz.ch/research/current-research-projects/armin-robot/ANYexo.html; Zimmermann et al., Reference Zimmermann, Forino, Riener and Hutter2019), (g) the occupation active trunk exoskeleton Robo.Mate (Huysamen et al., Reference Huysamen, Bosch, de Looze, Stadler, Graf and O’Sullivan2018), and (h) the occupational passive upper limb exoskeleton EksoVest (www.eksobionics.com/ekso-evo/; Kim et al., Reference Kim, Kim, Seok, So, Lee and Choi2018a,Reference Kim, Nussbaum, Mokhlespour Esfahani, Alemi, Alabdulkarim and Rashedib).

In all these high-torque lower/full-body exoskeletons, the mechanical structure needs to be in direct physical contact with the ground (i.e., through the feet) to support the load of the machine and the user.Footnote ¹ Although significant engineering efforts have been made to lighten the structures of these devices (Farris et al., Reference Farris, Quintero and Goldfarb2011; Vouga et al., Reference Vouga, Baud, Fasola, Bouri and Bleuler2017), their mechanical structure still tends to be cumbersome due to the considerable load they need to bear as well as the high torques that have to be generated and transferred through the structure to the ground, the load, and possibly to human limbs.

On the other side of the spectrum, “low-torque” stiff exoskeletons are not capable of generating sufficient torque to realize total control over either the upper or lower limb, and thus provide only partial assistance to the user. Application domains of this class include gait assistance devices (Lim et al., Reference Lim, Lee, Jang, Kim, Park, Seo and Shim2019; Zhang et al., Reference Zhang, Xiong, Zhang, Liu, Lai, Rong and Fu2019; Figure 2d) and occupational exoskeletons, either active (Figure 2g) or passive (Figure 2h; Bogue, Reference Bogue2018; Huysamen et al., Reference Huysamen, Bosch, de Looze, Stadler, Graf and O’Sullivan2018; Theurel and Desbrosses, Reference Theurel and Desbrosses2019; www.eksobionics.com/ekso-evo/). The main function of this “low-torque” category is to assist the wearer in performing a specific motor task. As a result of the smaller force/torque exchange required by these applications, their structure is typically slimmer. These “low-torque” exoskeletons are considerably lighter than those in the “high torque” group, and to a certain extent their weight can thus be borne by the user.

Concerning the upper limb, most actuated exoskeletons for this limb are used for robot-assisted rehabilitation (Nef et al., Reference Nef, Mihelj, Kiefer, Perndl, Muller and Riener2007; Gopura et al., Reference Gopura, Bandara, Kiguchi and Mann2016; Kim and Deshpande, Reference Kim and Deshpande2017; Rehmat et al., Reference Rehmat, Zuo, Meng, Liu, Xie and Liang2018). In upper-body exoskeletons, the device can be fixed on an external platform (Caldwell et al., Reference Caldwell, Tsagarakis, Kousidou, Costa and Sarakoglou2007; Noda et al., Reference Noda, Teramae, Ugurlu and Morimoto2014), which severely limits its mobility; conversely, it can be carried entirely by the human (Balasubramanian et al., Reference Balasubramanian, Wei, Perez, Shepard, Koeneman, Koeneman and He2008), which requires sufficient fitness to support the weight.

Upper limb rehabilitation devices need to deliver torque to the wearer during physical exercise (Lo and Xie, Reference Lo and Xie2012; Mehrholz et al., Reference Mehrholz, Hädrich, Platz, Kugler, Pohl and Mehrholz2012; Wendong et al., Reference Wendong, Hanhao, Menghan, Yang, Xiaoqing, Xing and Bing2020), and typically involve designs that allow high levels of controllability of the limb (Nef et al., Reference Nef, Mihelj, Kiefer, Perndl, Muller and Riener2007; Gopura et al., Reference Gopura, Bandara, Kiguchi and Mann2016; Bai and Christensen, Reference Bai and Christensen2017). The robot is usually used in one of three modes: (a) passive mobilization, where the user relaxes their arm and the robot moves it to improve its range of motion (Brahmi et al., Reference Brahmi, Saad, Rahman and Ochoa-Luna2019); (b) assist as needed, where the robot assists voluntary movements of the patient (Reinkensmeyer et al., Reference Reinkensmeyer, Wolbrecht, Chan, Chou, Cramer and Bobrow2012); or (c) resistive or error augmentation mode, where the robot acts in opposition to the wearer’s movements (Marchal-Crespo and Reinkensmeyer, Reference Marchal-Crespo and Reinkensmeyer2009). Please refer to Section “Control” for further details. In all three modes, the torques applied to the user are relatively low compared to lower limb or power augmentation scenarios. This is also true in the case of upper limb exoskeletons for teleoperative haptic applications, which need to generate feedback torques toward the user (Mallwitz et al., Reference Mallwitz, Will, Teiwes and Kirchner2015; Buongiorno et al., Reference Buongiorno, Chiaradia, Marcheschi, Solazzi and Frisoli2019). It is worth mentioning that, although the assistive torques that need to be generated are much smaller compared to those in lower limb or power augmentation applications, the structure of these exoskeletons may need to support significant load due to the “cantilever” effect generated when the arm is extended—for example, during reaching. Therefore, despite the low required net torque, the structures of these exoskeletons may become bulky due to the need to withstand considerable gravitational load. In addition to this, upper limb exoskeleton kinematics are generally far from trivial due to the high number of degrees of freedom of the human and the complexity of the shoulder joint (Figure 2e,f; www.reneu.robotics.utexas.edu/harmony-exoskeleton; www.sms.hest.ethz.ch/research/current-research-projects/armin-robot/ANYexo.html; Kim and Deshpande, Reference Kim and Deshpande2017; Zimmermann et al., Reference Zimmermann, Forino, Riener and Hutter2019). Therefore, this exoskeleton type may need dedicated designs to allow correct replication of biological kinematics and consequently facilitate their control (Perry et al., Reference Perry, Powell and Rosen2009). In contrast to the fully actuated and typically heavy upper limb machines addressed above, there is a growing interest in exoskeletons for arm support or gravity compensation in industrial applications (Spada et al., Reference Spada, Ghibaudo, Gilotta, Gastaldi and Cavatorta2017; Kim et al., Reference Kim, Kim, Seok, So, Lee and Choi2018a,Reference Kim, Nussbaum, Mokhlespour Esfahani, Alemi, Alabdulkarim and Rashedib; Van Engelhoven et al., Reference Van Engelhoven, Poon, Kazerooni, Barr, Rempel and Harris-Adamson2018; Pacifico et al., Reference Pacifico, Scano, Guanziroli, Moise, Morelli, Chiavenna, Romo, Spada, Colombina, Molteni, Giovacchini, Vitiello and Crea2020; Goršič et al. Reference Goršic, Regmi, Johnson, Dai and Novak2020). These devices are typically passive and use elastic elements to provide the needed compensation torque.

Regardless of the application domain or body area, stiff exoskeleton designs may present wearability issues since mechanical stress is generated on the user’s skin due to the generation of shear forces at the interface point. These shear forces may occur when there is misalignment between the exoskeleton rotation axes and the corresponding human joints. This represents a nonnegligible aspect of stiff exoskeletons, where the joints’ centers of instantaneous rotation are often fixed while the human joints’ (e.g., knee or shoulder) centers of instantaneous rotation change their relative position during the motion (Veeger, Reference Veeger2000; Hollman et al., Reference Hollman, Deusinger, Van Dillen and Matava2002). As a consequence, aligning a stiff exoskeleton to a human limb might lead to undesired stresses. The amount of these stresses depends on the quality of the alignment and the capability to maintain it during motion. To deal with this, researchers are increasingly adopting original solutions that allow the exoskeleton structure to self-align to the anatomical kinematics of the wearer using additional passive degrees of freedom (Ball et al., Reference Ball, Brown and Scott2007; Toxiri et al., Reference Toxiri, Ortiz, Masood, Fernandez, Mateos and Caldwell2015; Näf et al., Reference Näf, Junius, Rossini, Rodriguez-Guerrero, Vanderborght and Lefeber2018), compliant actuation systems (Vitiello et al., Reference Vitiello, Lenzi, Roccella, De Rossi, Cattin, Giovacchini, Vecchi and Carrozza2013), or specific mechanisms (Stienen et al., Reference Stienen, Hekman, van der Helm and van der Kooij2009; Ergin and Patoglu, Reference Ergin and Patoglu2012). However, these solutions require additional space and volume and are thus not always feasible in portable devices.

Soft Exoskeleton Designs

In contrast to stiff devices, soft exoskeletons (also called exosuits) deliver mechanical power to the user using anchors that are typically made of clothing textiles and powered by cable-driven actuation (Awad et al., Reference Awad, Bae, O’Donnell, De Rossi, Hendron, Sloot, Kudzia, Allen, Holt, Ellis and Walsh2017; Schmidt et al., Reference Schmidt, Duarte, Grimmer, Sancho-Puchades, Wei, Easthope and Riener2017; Li et al., Reference Li, Yang, Yu, Chang, Zhao, Zhao, Elhajj, Xi and Liu2018; Di Natali et al., Reference Di Natali, Poliero, Sposito, Graf, Bauer, Pauli, Bottenberg, De Eyto, O’Sullivan, Hidalgo, Scherly, Stadler, Caldwell and Ortiz2019; Kim et al., Reference Kim, Kwon, Han, Park and Jo2019a,Reference Kim, Lee, Heimgartner, Arumukhom Revi, Karavas, Nathanson, Galiana, Eckert-Erdheim, Murphy, Perry, Menard, Choe, Malcolm and Walshb; Xiloyannis et al., Reference Xiloyannis, Chiaradia, Frisoli and Masia2019; Yang et al., Reference Yang, Huang, Hu, Yu, Zhang, Zhou, Carriero, Yue and Su2019; Goršič et al., Reference Goršič, Song, Dai and Novak2021; Figure 3). Furthermore, soft exoskeletons are typically ergonomic, lightweight, and often designed following bioinspired techniques to better reproduce human physiological movements (Lessard et al., Reference Lessard, Pansodtee, Robbins, Trombadore, Kurniawan and Teodorescu2018; Figure 3b). They can be employed in several domains, from medical/rehabilitation scenarios to improving the physical strength of healthy individuals (Kim et al., Reference Kim, Kwon, Han, Park and Jo2019a,Reference Kim, Lee, Heimgartner, Arumukhom Revi, Karavas, Nathanson, Galiana, Eckert-Erdheim, Murphy, Perry, Menard, Choe, Malcolm and Walshb; Figure 3c). The same work also demonstrated how a lower limb soft exosuit can reduce the metabolic cost of walking and running. Another important advantage is that, compared to stiff designs, exosuits largely maintain the natural freedom of movement of the wearer (Wehner et al., Reference Wehner, Quinlivan, Aubin, Martinez-Villalpando, Baumann, Stirling, Holt, Wood and Walsh2013; Figure 3d). However, the preservation of the freedom of movement offered by soft exosuits, as well as their comfort, depends on ergonomically appropriate design. Indeed, several recent works have focused on the design of ergonomic soft exoskeletons to minimize user discomfort (Andrikopoulos et al., Reference Andrikopoulos, Nikolakopoulos and Manesis2015; Varghese et al., Reference Varghese, Mukherjee, King, Keller and Deshpande2018; Choi et al., Reference Choi, Kang, Jung and Cho2019; Chiaradia et al., Reference Chiaradia, Tiseni, Xiloyannis, Solazzi, Masia and Frisoli2021). These studies showed how the combination of textile fibers, 3D printed plastic materials and sometimes even metal parts, should be carefully designed to offer appropriate ergonomics and avoid possible limitations of the device’s functionality.

Figure 3. Examples of soft exoskeletons: (a) the research upper limb soft exosuit from Aries Lab (Xiloyannis et al., Reference Xiloyannis, Chiaradia, Frisoli and Masia2019), (b) the CRUX soft upper limb exosuit (Lessard et al., Reference Lessard, Pansodtee, Robbins, Trombadore, Kurniawan and Teodorescu2018), (c) the research lower limb soft exosuit from the Wyss Institute (Kim et al., Reference Kim, Kwon, Han, Park and Jo2019a,Reference Kim, Lee, Heimgartner, Arumukhom Revi, Karavas, Nathanson, Galiana, Eckert-Erdheim, Murphy, Perry, Menard, Choe, Malcolm and Walshb), and (d) a lower limb pneumatic soft exoskeleton (Wehner et al., Reference Wehner, Quinlivan, Aubin, Martinez-Villalpando, Baumann, Stirling, Holt, Wood and Walsh2013).

The main disadvantage of soft exoskeletons is that the mechanical power they can deliver is still limited as compared to stiff designs (Schmidt et al., Reference Schmidt, Duarte, Grimmer, Sancho-Puchades, Wei, Easthope and Riener2017). This is mostly because the forces produced by soft exosuits inherently increase the shear stress on the human body. Indeed, although these devices may potentially offer higher comfort than stiff systems due to their compliant nature, soft exoskeletons often present problems of slipping and chafing at the anchor points due to the generation of shear forces at the interface point with the wearer (Asbeck et al., Reference Asbeck, Dyer, Larusson and Walsh2013; Yandell et al., Reference Yandell, Quinlivan, Popov, Walsh and Zelik2017). On the other hand, the effects of the mentioned shear forces on joint loading are usually negligible: the radius over which they apply forces is significantly greater than that exploited by the biological tendon, and the joint compression force decreases as a consequence of this (Lamers et al., Reference Lamers, Yang and Zelik2018). The shear force problem is instead minimized in stiff designs, where forces are applied orthogonally to the human segment (Yandell et al., Reference Yandell, Quinlivan, Popov, Walsh and Zelik2017). Another common issue of soft design approaches is the friction introduced by the employed cable-driven actuation systems (e.g., Bowden-cable transmission), which typically results in low efficiency and bad control performance of the overall system if not compensated (Jeong et al., Reference Jeong, Woo and Kong2020). Furthermore, given their soft nature, the interface points where the human limb is anchored introduce a low-stiffness series compliance that would need high actuation power and appropriate bandwidth to be overcome (Chiaradia et al., Reference Chiaradia, Xiloyannis, Solazzi, Masia and Frisoli2019). Figure 3 shows examples of upper and lower limb soft exoskeletons.

Actuation

Actuators are a major bottleneck for the development of active wearable robots. The demanding requirements imposed upon them include high torque density (torque per unit mass), backdrivability, compactness and, for untethered devices, high energy efficiency. The actuator’s torque density is especially important for actuators at distal joints due to the large metabolic cost associated with carrying mass in these locations (Royer and Martin, Reference Royer and Martin2005; Browning et al., Reference Browning, Jesse R., Rodger and Ambarish2007). Highly backdrivable actuators allow the wearer to move unimpeded when the robot is not providing assistance, for example, during the swing phase of walking or during donning and doffing. Finally, wearable robots should be as unobtrusive as possible. Actuators, which tend to protrude from the device, therefore need to be designed in small and compact form factors.

The majority of actuators for wearable robotics incorporate electric motors as prime movers (Windrich et al. Reference Windrich, Grimmer, Christ, Rinderknecht and Beckerle2016; Gopura et al., Reference Gopura, Bandara, Kiguchi and Mann2016; Veale and Xie, Reference Veale and Xie2016). The main competing technologies, hydraulic and pneumatic actuators, are typically less efficient and reliable, tend to be noisy, and require pressurized fluid supplies that are difficult to make portable in an acceptable form factor (Madden, Reference Madden2007; Veale and Xie, Reference Veale and Xie2016). Hydraulic power units driven by combustion engines do have very high power densities, which makes them interesting when high power output is required, for example, for exoskeletons intended for load carrying and human performance augmentation (National Research Council, 2004). However, the high power density of hydraulic actuators comes at the cost of high impedance (Hollerbach et al., Reference Hollerbach, Hunter and Ballantyne1992; Zinn et al., Reference Zinn, Khatib and Roth2004), which is undesirable for wearable robots. Overall, for most wearable robots that are not used for human performance augmentation, electric motors are considered to present the best trade-off between the different requirements.

Interestingly, in terms of power-to-mass ratio and energy efficiency, electric motors perform better than human muscles (Madden, Reference Madden2007). However, in wearable robots, motors need to operate dynamically at varying torques and speeds. In these conditions, their efficiency is far below the nominal value (Verstraten et al., Reference Verstraten, Flynn, Geeroms, Vanderborght and Lefeber2018). Moreover, the torque output of electric motors is low relative to the demands of wearable robots, which is why they are typically paired with high-reduction gearboxes. The dominant gearbox technologies in wearable robotics, harmonic gears and planetary gearboxes, offer high reduction ratios in compact sizes. However, their low efficiency and poor backdrivability are detrimental to the overall performance of the actuators and the wearable robot as a whole (Kazerooni, Reference Kazerooni1995; Nef and Lum, Reference Nef and Lum2009; Lv et al., Reference Lv, Zhu and Gregg2018; García et al., Reference García, Crispel, Saerens, Verstraten and Lefeber2020).

Two paths are currently followed by the research community to tackle this problem. On one hand, novel gearbox technologies with improved characteristics are being explored, as summarized by García et al. (Reference García, Crispel, Saerens, Verstraten and Lefeber2020). Recent examples include the Gear Bearing Drive (Brassitos and Jalili, Reference Brassitos and Jalili2019), and the Bilateral Drive Gear (Matsuki et al. Reference Matsuki, Nagano and Fujimoto2019). These gearboxes achieve high gear reductions at high efficiencies and in compact sizes through the use of the Wolfrom gearbox topology, combined with performance-enhancing modifications such as a carrierless design (Gear Bearing Drive) and profile shift modification (Bilateral Drive). On the other hand, quasi-direct drive has emerged as a new actuation paradigm. It combines high torque-motors with low-reduction gearboxes, resulting in energy-efficient actuators with high backdrivability (Sensinger et al., Reference Sensinger, Clark and Schorsch2011). Ideally, these gearboxes are single-stage (gear ratio below 10) since efficiency is mostly related to the number of stages. Following their success in legged robots (Seok et al., Reference Seok, Wang, Chuah, Hyun, Lee, Otten, Lang and Kim2015; Ding and Park, Reference Ding and Park2017), quasi-direct drive actuators are currently being implemented in wearable robots (Wang et al., Reference Wang, Li, Huang, Yu, Li, Chen, Carriero, Oh-Park and Su2018; Elery et al., Reference Elery, Rezazadeh, Nesler and Gregg2020; Yu et al., Reference Yu, Huang, Yang, Jiao, Yang, Chen, Yi and Su2020; Zhu et al., Reference Zhu, Nesler, Divekar, Peddinti and Gregg2021).

To further compensate for the inherent limitations of commercial motors and gearboxes, researchers have resorted to specialized actuator designs. Our musculoskeletal system is built in such a way that elastic tissues in our body (tendons, aponeuroses, elastic tissues between muscle fibers, etc.) can perform a large amount of work by passive stretch and recoil (Alexander, Reference Alexander1990). This is possible because the joint torques applied during human movements typically exhibit a linear relationship with respect to the joint angle during consistent parts of the movement (Feldman, Reference Feldman1980; Hasan, Reference Hasan1986; Frigo et al., Reference Frigo, Crenna and Jensen1996; Toxiri et al., Reference Toxiri, Ortiz, Masood, Fernandez, Mateos and Caldwell2015). Engineered actuators for wearable robots, which are designed to provide assistive torques with a similar profile, can therefore also benefit from incorporating elastic elements with well-chosen stiffness. These specialized actuators are called elastic actuators.

Springs with well-designed stiffness allow actuators to deliver peak powers well beyond the capabilities of the motor itself (Paluska and Herr, Reference Paluska and Herr2006). Elastic actuators can therefore incorporate a smaller electric motor and gearbox, resulting in an overall reduction of weight and size compared to a traditional, stiff actuator. Moreover, springs offer the possibility to mechanically store energy when the actuator performs negative work and release it when positive work is required. As such, the average power flow through the actuator is reduced, resulting in an improvement in energy efficiency of the actuator (Hollander et al., Reference Hollander, Ilg, Sugar and Herring2006; Verstraten et al., Reference Verstraten, Geeroms, Mathijssen, Convens, Vanderborght and Lefeber2017; Bolivar Nieto et al., Reference Bolivar Nieto, Rezazadeh and Gregg2019).

Elastic actuators can be subdivided into two main categories based on the placement of the spring element. By adding a (unidirectional) parallel spring to a stiff motor, one obtains a parallel elastic actuator (PEA). In this case, the parallel spring takes over a part of the load, leading to a reduction of the motor torque. The required motor speed remains unchanged since the motor is directly coupled to the output. Another option is to place the spring in between the motor and load, that is, in a series configuration. This type of actuator is called a series elastic actuator (SEA). In contrast to the PEA, a well-tuned spring lowers the speed required from the motor while the motor torque is unaffected. Although SEA and PEA provide similar levels of peak power amplification and increased energy efficiency during locomotion tasks (Verstraten et al., Reference Verstraten, Geeroms, Mathijssen, Convens, Vanderborght and Lefeber2017), the PEA has the specific ability to compensate for static torques by tuning the equilibrium angle of the spring (Beckerle et al., Reference Beckerle, Verstraten, Mathijssen, Furnemont, Vanderborght and Lefeber2016). This makes the PEA particularly suitable for upper-body exoskeletons that assist with overhead work or back-support exoskeletons for lifting tasks. Here, the parallel spring can provide the torque required to support the weight of the arms or torso as well as the load.

The performance of elastic actuators can be further enhanced by adding locking mechanisms and clutches (Plooij et al., Reference Plooij, Wolfslag and Wisse2017). Passive occupational exoskeletons often allow the user to disengage the spring providing the assistance so that it does not impede the user by generating a torque when it is not needed (Zhang and Huang, Reference Zhang and Huang2018). In a similar fashion, since little torque is required during the swing phase of gait, lower limb exoskeletons and prostheses can benefit from clutched PEAs (Haeufle et al., Reference Haeufle, Taylor, Schmitt and Geyer2012; Geeroms et al., Reference Geeroms, Flynn, Jimenez-Fabian, Vanderborght and Lefeber2017) and SEAs (Rouse et al., Reference Rouse, Mooney and Herr2014; Convens et al., Reference Convens2019) that only engage the spring during stance. In robotic ankles, a catapult-like mechanism consisting of a spring and locking mechanism could be used to deliver the explosive burst of power required during push-off in an energy-efficient way (Collins and Kuo, Reference Collins and Kuo2010; Cherelle et al., Reference Cherelle, Grosu, Matthys, Vanderborght and Lefeber2014). Clutches have also been implemented in new actuator designs with multiple springs, which can be recruited progressively to contribute to the output torque (Mathijssen et al., Reference Mathijssen, Lefeber and Vanderborght2015; Krimsky and Collins, Reference Krimsky and Collins2020).

Most springs are manufactured to exhibit a linear relationship between torque and deflection. However, in wearable robotics, the required torques are rarely perfectly linear with respect to the output angle over the entire operating range of motion. The joint stiffness profiles required for walking, for example, exhibit low stiffness at low joint flexion and high stiffness at high joint flexion (Seyfarth et al., Reference Seyfarth, Geyer, Blickhan, Lipfert, Rummel, Minekawa and Iida2006). A closer match between the spring characteristic and the required torque can be achieved by means of nonlinear transmissions. The simplest way to generate a nonlinear transmission is by providing torque to the joint through a lever (Sugar and Holgate, Reference Sugar and Holgate2013). The linear actuators driving the lever often consist of brushless DC motors combined with ball screw drives, a solution which offers relatively high reductions at high efficiencies. Other common nonlinear transmissions are cams (Realmuto et al., Reference Realmuto, Klute and Devasia2015; Gao et al., Reference Gao, Liu and Liao2018) and four-bar mechanisms (Bergelin and Voglewede, Reference Bergelin and Voglewede2012; Hyun et al., Reference Hyun, Park, Ha, Park and Jung2017).

Specialized actuator designs achieve impressive performance by relying on springs and nonlinear transmissions optimized for a specific task. The downside of this approach is significantly worse performance in off-design conditions. This issue can be mitigated by variable stiffness springs (Vanderborght et al., Reference Vanderborght, Albu-Schaeffer, Bicchi, Burdet, Caldwell, Carloni, Catalano, Eiberger, Friedl, Ganesh, Garabini, Grebenstein, Grioli, Haddadin, Hoppner, Jafari, Laffranchi, Lefeber, Petit, Stramigioli, Tsagarakis, Van Damme, Van Ham, Visser and Wolf2013; Shepherd and Rouse, Reference Shepherd and Rouse2017), variable transmissions (Mooney and Herr, Reference Mooney and Herr2013; Tran et al., Reference Tran, Gabert, Cempini and Lenzi2019), multispeed gearboxes (Sugar and Holgate, Reference Sugar and Holgate2013) or multimotor solutions (Verstraten et al., Reference Verstraten, Furnémont, López-García, Crispel, Vanderborght and Lefeber2019), which allow the relationship between the required output torque and the spring characteristic to be shaped depending on the task. However, these solutions require additional motors or mechanisms, which increase the weight of the actuator. The metabolic penalty resulting from the additional weight can be avoided by remote actuation through, for example, cable- or wire-based transmissions, as implemented in several rehabilitation exoskeletons (Veneman et al., Reference Veneman, Ekkelenkamp, Kruidhof, van der Helm and van der Kooij2006; Mao and Agrawal, Reference Mao and Agrawal2012). In portable devices, these transmissions can decrease metabolic cost by enabling a more proximal placement of the actuators, near the human’s center of mass (Asbeck et al., Reference Asbeck, De Rossi, Holt and Walsh2015).

Motor and gearbox technologies continue to improve and these improvements will promote further adoption of wearable robots in the near future. However, commercial devices for specific and simple tasks (e.g., occupational exoskeletons) may very well continue to rely on passive solutions until actuators with superior torque densities and energy efficiencies become available. The major breakthrough may arrive through the development of radically new actuator technologies such as dielectric elastomer actuators (Veale and Xie, Reference Veale and Xie2016). Until these technologies reach a mature state, we expect that advances will most likely come from specialized actuator designs as those described in this section.

Sensing

Wearable robots require advanced sensors so that they can determine the user’s actual and intended motions, determine the state of the robot itself, and use this information as a basis for robot control. Any limitations in the sensing lead to suboptimal control, which limits the overall performance of the wearable robot. These limitations can occur at one of two levels: at the lower level of individual sensors and at the higher level of multisensor fusion.

Body Interface

Wearable robots cannot provide any benefits if humans do not use them. State-of-the-art wearable robots can effectively reduce workload; however, limitations related to user comfort and acceptance are a major remaining barrier to widespread use (Kermavnar et al., Reference Kermavnar, de Vries, de Looze and O’Sullivan2021), together with high cost. It was recognized in early exoskeleton research that a certain amount of discomfort was experienced by the users (Abdoli-E and Stevenson, Reference Abdoli-E and Stevenson2008; Wehner et al., Reference Wehner, Rempel and Kazerooni2009), while newer models have been reported to minimize this discomfort to a tolerable level (Sirawattanakul and Sanngoen, Reference Sirawattanakul and Sanngoen2020; Kermavnar et al., Reference Kermavnar, de Vries, de Looze and O’Sullivan2021). Some wearable robots are heavy and bulky (Stadler et al., Reference Stadler, Altenburger, Schmidhauser, Scherly, Ortiz, Toxiri, Mateos and Masood2016; Näf et al., Reference Näf, Koopman, Baltrusch, Rodriguez-Guerrero, Vanderborght and Lefeber2018; García et al., Reference García, Crispel, Saerens, Verstraten and Lefeber2020), which can interfere with the user’s freedom of movement. Even with lighter models, a certain level of discomfort and movement hindrance has been reported (Baltrusch et al., Reference Baltrusch, van Dieën, van Bennekom and Houdijk2018; Goršič et al Reference Goršic, Regmi, Johnson, Dai and Novak2020). Inevitably, the unloading of one part of the body results in an increased loading (i.e., force) on another. In spinal exoskeletons, the reduction of the load to the back is typically achieved via increased loading to the lower limbs, which can be sometimes reflected in increased leg muscle activity (Luger et al., Reference Luger, Bär, Seibt, Rieger and Steinhilber2021) and discomfort to the chest area (Hensel and Keil, Reference Hensel and Keil2019), although back exoskeletons may also reduce leg activity (Bosch et al., Reference Bosch, van Eck, Knitel and de Looze2016; Jeong et al., Reference Jeong, Woo and Kong2020). In comparison, upper limb exoskeleton interfaces can increase the loading to the spine (Weston et al., Reference Weston, Alizadeh, Knapik, Wang and Marras2018). This trade-off, as well as the effects of wearable robots on task performance, discomfort and perceived freedom of movement, must be taken into account when designing wearable robots and studying their effects. While these functional and subjective aspects of wearable robots are increasingly being addressed, research is predominantly done in laboratory environments with healthy young male participants (Kermavnar et al., Reference Kermavnar, de Vries, de Looze and O’Sullivan2021). Therefore, on-field research with different groups is needed to support the development of body interfaces, suitable for broader population. Interface design has also begun to address the issue of load transfer to other body parts. One possible way to address this is by extending the legs of the exoskeleton all the way to the ground. For instance, a recent prototype of a semi-active exoskeleton, designed in this fashion, has been reported to decrease the activity of both back and leg muscles (Wang et al., Reference Wang, Wu, Zhang, Chen, Liu, Liu, Peng and Ma2021). Novel exoskeletons such as the Guardian XO (CRI Writers, Reference Writers2018), are also being developed to support immense loads without stressing any human body parts. On the other hand, back exoskeletons with good design seem to decrease the load to the legs in addition to unloading the back (Bosch et al., Reference Bosch, van Eck, Knitel and de Looze2016; Jeong et al., Reference Jeong, Woo and Kong2020).

Exoskeleton interfaces should enable efficient transmission of mechanical power without compromising user comfort. Langlois et al. (Reference Langlois, Moltedo, Bacek, Rodriguez-Guerrero, Vanderborght and Lefeber2018) have proposed three crucial design requirements that need to be fulfilled to achieve this. First, the attachment points of the device on the human body segments should be as distant to the joint as possible to reduce interaction forces (e.g., attachment points of the hip exoskeleton on the thighs should be as close to the knees as possible). At the same time, contact areas should generally be maximized. However, caution is needed as some parts of the body may be more sensitive to pressure and increasing the surface area may not be desirable in such cases. Next, the exoskeleton should be designed in a way that prevents interface migration (e.g., sliding along the user’s body), as misalignments caused by sliding commonly reduce power transmission efficiency (Langlois et al., Reference Langlois, Moltedo, Bacek, Rodriguez-Guerrero, Vanderborght and Lefeber2018). Power transmission is also increased with larger contact point stiffness (Asbeck et al., Reference Asbeck, De Rossi, Holt and Walsh2015; Langlois et al., Reference Langlois, Rodriguez-Cianca, Serrien, De Winter, Verstraten, Rodriguez-Guerrero, Vanderborght and Lefeber2020). This can be achieved by placing the contact points over the bony prominences, such as tibial surface in case of lower limb exoskeletons (Mooney and Herr, Reference Mooney and Herr2016) or pelvis bones in case of trunk exoskeletons (Näf et al., Reference Näf, Junius, Rossini, Rodriguez-Guerrero, Vanderborght and Lefeber2018). Alternatively, the pads can be tightly attached to the user, but this can cause discomfort. Keeping in mind the importance of contact point placement, prevention of interface migration and contact stiffness, the compactness of the design should not be overlooked. For instance, the additional inertia due to the device’s weight can increase actuation demands (Langlois et al., Reference Langlois, Moltedo, Bacek, Rodriguez-Guerrero, Vanderborght and Lefeber2018).

Wearable robots for industrial environments are usually designed primarily with a focus on load reduction during strenuous physical tasks such as load lifting, overhead work and trunk bending (de Looze et al., Reference de Looze, Bosch, Krause, Stadler and O’Sullivan2016; Bao et al., Reference Bao, Pan, Fang, Wu, Yu, Cai, Yu and Wan2019). However, to allow efficient work and freedom of movement, the interface should not interfere heavily with other movements such as walking, stair climbing and sitting down or standing up. Indeed, wearable robots often decrease performance in tasks that require quickness and range of motion (Kermavnar et al., 2020). While these aspects were partially addressed in earlier research (Wehner et al., Reference Wehner, Rempel and Kazerooni2009; Bosch et al., Reference Bosch, van Eck, Knitel and de Looze2016), assessments involving subjective reports have only recently been introduced (Kozinc et al., Reference Kozinc, Baltrusch, Houdijk and Šarabon2020a). Such subjective reports allow potential drawbacks of specific wearable robots, especially those related to the interface and discomfort, to be identified. For instance, specific contact points that exert pressure on the user or limit their range of motion can be identified and considered in subsequent design (Abdoli-Eramaki et al., Reference Abdoli-Eramaki, Stevenson, Reid and Bryant2007; Baltrusch et al., Reference Baltrusch, van Dieën, van Bennekom and Houdijk2018; Baltrusch et al., Reference Baltrusch, Houdijk, van Dieën and Kruif2021; Kozinc et al., Reference Kozinc, Baltrusch, Houdijk and Šarabon2020b). Identifying the drawbacks of state-of-the-art devices should help developers optimize subsequent models. One recent example of approaches to improve the exoskeleton interface is found in the SPEXOR spinal exoskeleton. Specifically, a self-aligning mechanism was integrated into the hip joint, which prevents the sliding of the interface (Näf et al., Reference Näf, Koopman, Baltrusch, Rodriguez-Guerrero, Vanderborght and Lefeber2018; Babič et al., Reference Babič, Petrič, Mombaur, Kingma, Bornmann, González-Vargas, Baltrusch, Šarabon and Houdijk2019). With such improvements, exoskeleton interfaces could become more user-friendly. Another possible consideration for interface improvement is the option to adjust and switch off the exoskeleton’s support (Babič et al., Reference Babič, Petrič, Mombaur, Kingma, Bornmann, González-Vargas, Baltrusch, Šarabon and Houdijk2019; Wang et al., Reference Wang, Wu, Zhang, Chen, Liu, Liu, Peng and Ma2021), which allows the user to freely and efficiently perform the tasks that do not require support (e.g., walking).

As mentioned, the unloading of one part of the body inevitably results in an increased loading somewhere else. For instance, the mechanism by which spinal exoskeletons contribute to reduced workload is generation of torque around the spine and/or the hips (Bosch et al., Reference Bosch, van Eck, Knitel and de Looze2016; Stadler et al., Reference Stadler, Altenburger, Schmidhauser, Scherly, Ortiz, Toxiri, Mateos and Masood2016; de Looze et al., Reference de Looze, Krause and O’Sullivan2017, Näf et al., Reference Näf, Junius, Rossini, Rodriguez-Guerrero, Vanderborght and Lefeber2018; Babič et al., Reference Babič, Petrič, Mombaur, Kingma, Bornmann, González-Vargas, Baltrusch, Šarabon and Houdijk2019; Kermavnar et al., 2020). Even though spinal exoskeletons are designed to compensate for relatively small amounts of required spinal torque (Näf et al., Reference Näf, Koopman, Baltrusch, Rodriguez-Guerrero, Vanderborght and Lefeber2018), this unloading of the spine results in increased levels of force (and consequently pressure) on the exoskeleton attachment sites on the body (i.e., distal thighs and shoulder girdle). Similar observations have been reported for upper limb (Huysamen et al., Reference Huysamen, Bosch, de Looze, Stadler, Graf and O’Sullivan2018) and lower limb (Luger et al., Reference Luger, Seibt, Cobb, Rieger and Steinhilber2019) exoskeletons. Therefore, it is crucial to consider discomfort and pain tolerance thresholds at typical attachment sites for wearable robotic devices, and to design the devices in a way that minimize the chances of these thresholds being reached. While many studies have investigated pain thresholds with algometry (Sterling, Reference Sterling2011), their relevance to wearable robotics is limited, as the surface of force application in case of standard algometry is typically set at 1 cm² (Graven-Nielsen and Arendt-Nielsen, Reference Graven-Nielsen and Arendt-Nielsen2010). Potentially more useful evidence is found in studies using inflatable cuffs to induce pressure at different sites, typical for attachments of wearable robots to the human body (Kermavnar et al., Reference Kermavnar, Power, de Eyto and O’Sullivan2018). There are only two studies to date that explore this issue directly. Yandell et al. (Reference Yandell, Ziemnicki, McDonald and Zelik2020) applied the pressure to the human body with fabricated exosuit interfaces for the shoulders, thigh, and shank, and thereby provided normative data for discomfort thresholds. Similarly, Kozinc et al. (Reference Kozinc, Babič and Šarabon2021) applied a pressure using exoskeleton pads and straps to the thigh, pelvic and chest/shoulder areas. Exoskeleton developers are encouraged to consider these values during the design of wearable robots. One of the most important findings of both studies is the large (up to 10-fold) interindividual variability in discomfort thresholds. Thus, further studies should critically include sufficient sample sizes to encompass this variability. Additionally, an issue that remains largely unexplored is discomfort during longer-term use of exoskeletons (i.e., over a workday), where higher discomfort or pain levels might be seen due to temporal summation (Nie et al., Reference Nie, Arendt-Nielsen, Andersen and Graven-Nielsen2005).

It is unquestionable that wearable robots have the potential to increase work performance and reduce physical loads to the body. Some exoskeletons with inferior design and poor kinematic structures and human interfaces have been shown to reduce freedom of movement and increase discomfort, which hinders widespread use. On the other hand, several models have achieved high levels of support without compromising overall user experience. Recent advances in methods to assess functional performance and discomfort, as well as the availability of normative data on discomfort tolerance, should help the developers to design better interfaces in the future.

Control

When synthesizing controllers for advanced robotic systems, it is common to create a hierarchical controller scheme where high-, mid-, and low-level control actions are locally isolated from each other, thus providing a broader understanding of the control problem in question. Wearable robots and exoskeleton systems are no exceptions to that; it is important to separately plan these controllers on different levels so that interrelations between high-level decision-making, mid-level planning and low-level actuator control can be clearly identified (Jimenez-Fabian and Verlinden, Reference Jiménez-Fabián and Verlinden2012; Tucker et al., Reference Tucker, Olivier, Pagel, Bleuler, Bouri, Lambercy, Millán, Riener, Vallery and Gassert2015). In this section, we succinctly cover low- and mid-level controllers, while high-level control is discussed in the next section.

Human–Robot Interfacing and Coadaptation

Low-level control of wearable hardware is an essential basis for practically applicable exoskeletons. Moreover, exoskeletons can only successfully perform various tasks in synergy with their human users if they are able to adapt to the user while considering the fact that the user simultaneously adapts to the exoskeleton. To do this, a high-level control system is required to infer human intention and create and adapt reference actions to make sure that the desired task is executed as intended by the user and assisted as intended by the exoskeleton. On the other hand, for the human to understand the exoskeleton’s behavior, it is also important that it behaves consistently and predictably; thus, the high-level control system needs to be designed with human-in-the-loop aspects in mind.

If a wearable device, such as an exoskeleton, is to perform actions to accommodate and assist humans in performing their tasks, the high-level control system should be able to extract the human user’s intention. Several signals can be employed to infer the intention, and the high-level control system thus tightly connects to the sensing hardware and methods. While the Sensing section focused on how to obtain the signals, this section focuses on how to use the signals in the high-level controller in order to achieve effective interfacing and coadaptation between the user and the exoskeleton.

Since the human and wearable device are physically coupled at least to some degree, haptic sensing is a common choice to detect the intended actions of the user. Such signals can be obtained by force/torque sensors or pressure sensors that are mounted on or inside the exoskeleton and can measure the interaction forces between its structure and the user’s body (Kazerooni and Guo, Reference Kazerooni and Guo1993; Nagai et al., Reference Nagai, Nakanishi, Hanafusa, Kawamura, Makikawa and Tejima1998; Kong and Jeon, Reference Kong and Jeon2006; Lee et al., Reference Lee, Lee, Kim, Han, Shin and Han2014). When the exoskeleton contains elastic elements like springs, the interaction force and thus control intention can also be estimated indirectly through the position of the spring (Pratt et al., Reference Pratt, Krupp, Morse and Collins2004). An alternative is to simply use the measured position or motion induced by the user (Kazerooni et al., Reference Kazerooni, Steger and Huang2006). Position and motion states can also be used to detect the human intention and then play out a specific predefined feedforward behavior of an exoskeleton (Kawamoto et al., Reference Kawamoto, Kanbe and Sankai2003a; Varol et al., Reference Varol, Sup and Goldfarb2010; Li and Hsiao-Wecksler, Reference Li and Hsiao-Wecksler2013). Information about the Center or Pressure or Center of Gravity can be used to a similar end for walking tasks (Farris et al., Reference Farris, Quintero, Murray, Ha, Hartigan and Goldfarb2014).

Another common type of signals used for high-level control involves measuring the human body’s electrical activity: for example, muscle activity by EMG (Rosen et al., Reference Rosen, Brand, Fuchs and Arcan2001; Kawamoto et al., Reference Kawamoto, Lee, Kanbe and Sankai2003b; Fleischer and Hommel, Reference Fleischer and Hommel2008; Huang et al., Reference Huang, Kuiken and Lipschutz2009; Kiguchi and Hayashi, Reference Kiguchi and Hayashi2012), brain activity by EEG (Soekadar et al., Reference Soekadar, Witkowski, Vitiello and Birbaumer2015; AL-Quraishi et al., Reference AL-Quraishi, Elamvazuthi, Daud, Parasuraman and Borboni2018) and eye activity by electrooculography (EOG) (Soekadar et al., Reference Soekadar, Witkowski, Vitiello and Birbaumer2015). Such electrical signals are advantageous since they can provide the information slightly prior to the actual human movement and potentially even if the human cannot make a physical movement (Fleischer and Hommel, Reference Fleischer and Hommel2008). On the other hand, their main disadvantage compared to haptic or position signals is that they require probes to be attached to the user, which usually comes with complex equipment and time-consuming calibration procedures.

Based on the detected human intention, the exoskeleton can devise appropriate physical actions to adapt and assist the users in their tasks. The simplest method is to amplify a signal that represents the user’s effort. For example, the measured human interaction force in a given direction can be amplified by the exoskeleton to assist the motion in that direction (Kazerooni et al., Reference Kazerooni, Steger and Huang2006; Kong and Jeon, Reference Kong and Jeon2006). If the user induces motion in the desired direction, then the exoskeleton can derive the assistive movement through a dynamical model (Kazerooni et al., Reference Kazerooni, Steger and Huang2006). In case of EMG, the intention signal (i.e., muscle activation) first has to be transformed into force by biomechanical models of muscles before it can be amplified (Rosen et al., Reference Rosen, Brand, Fuchs and Arcan2001; Fleischer and Hommel, Reference Fleischer and Hommel2008). However, biomechanical models are user-dependent and can be hard to obtain. In some cases, a proportional mapping between EMG-derived muscle activation and the resulting exoskeleton assistive force can be sufficient for practical applications (Kawamoto et al., Reference Kawamoto, Lee, Kanbe and Sankai2003b; Ferris et al., Reference Ferris, Gordon, Sawicki and Peethambaran2006; Lenzi et al., Reference Lenzi, De Rossi, Vitiello and Carrozza2012). In complex tasks like locomotion, effort amplification alone might not be sufficient and dynamical models might be required; furthermore, different phases of the task may require distinct control modes (e.g., separate mode for a stance phase and a swing phase) (Kazerooni et al., Reference Kazerooni, Steger and Huang2006).

If the high-level control system of an exoskeleton has an insight into more complex human biomechanical models, it can exploit them and consequently optimize specific features during assistance (i.e., besides generally reducing the user’s effort), thus making task execution more ergonomic. For example, limb and muscle manipulability properties change the ability of humans to produce endpoint forces and motions in different directions of Cartesian space. For example, a lower limb exoskeleton can use geometric limb manipulability to optimize for energy efficiency in walking (Kim et al., Reference Kim, Lee, Kang, Han and Han2010). In another example, the muscle manipulability model can be used by an arm exoskeleton to keep the force and motion production capacity equal in all directions irrespectively of human arm configuration (Petrič et al., Reference Petrič, Peternel, Morimoto and Babič2019).

More complex movements can be encoded by trajectories; however, they typically have to be preplanned or be of repetitive nature, like in rehabilitation scenarios (Tsagarakis and Caldwell, Reference Tsagarakis and Caldwell2003; Riener et al., Reference Riener, Lunenburger, Jezernik, Anderschitz, Colombo and Dietz2005). There are numerous ways to encode the trajectories, ranging from simple spline methods to more sophisticated methods involving nonlinear oscillators (Ronsse et al., Reference Ronsse, Vitiello, Lenzi, van den Kieboom, Carrozza and Ijspeert2011; Aguirre-Ollinger, Reference Aguirre-Ollinger2013) or dynamical movement primitives (DMPs) (Huang et al., Reference Huang, Cheng, Guo, Chen and Lin2016; Peternel et al., Reference Peternel, Noda, Petrič, Ude, Morimoto and Babič2016; Lauretti et al., Reference Lauretti, Cordella, Ciancio, Trigili, Catalan, Badesa, Crea, Pagliara, Sterzi, Vitiello, Garcia Aracil and Zollo2018). These trajectories can be learned from EMG feedback (Peternel et al., Reference Peternel, Noda, Petrič, Ude, Morimoto and Babič2016) or from recorded motion directly (Huang et al., Reference Huang, Cheng, Guo, Chen and Lin2016; Lauretti et al., Reference Lauretti, Cordella, Ciancio, Trigili, Catalan, Badesa, Crea, Pagliara, Sterzi, Vitiello, Garcia Aracil and Zollo2018). However, unlike effort-amplification methods, methods that merely involve repetition of learned trajectories do not enable the desired online and real-time adaptation to the user.

It has been shown that the human neuromechanical system can adapt to exoskeletons (Galle et al., Reference Galle, Malcolm, Derave and De Clercq2013; Gordon et al., Reference Gordon, Kinnaird and Ferris2013). Exoskeletons that to adapt to the human user can adaptively update their assistive behavior (e.g., joint torques or motion) to minimize the user’s effort (e.g., measured by muscle activity or interaction force) (Kawamoto et al., Reference Kawamoto, Lee, Kanbe and Sankai2003b; Krebs et al., Reference Krebs, Palazzolo, Dipietro, Ferraro, Krol, Rannekleiv, Volpe and Hogan2003; Riener et al., Reference Riener, Lunenburger, Jezernik, Anderschitz, Colombo and Dietz2005; Balasubramanian et al., Reference Balasubramanian, Wei, Perez, Shepard, Koeneman, Koeneman and He2008; Huang et al., Reference Huang, Cheng, Guo, Chen and Lin2016; Peternel et al., Reference Peternel, Noda, Petrič, Ude, Morimoto and Babič2016; Teramae et al., Reference Teramae, Noda and Morimoto2018). For example, in Peternel et al. (Reference Peternel, Noda, Petrič, Ude, Morimoto and Babič2016), torque trajectories are increased as long as the user has to produce some effort to perform the given task (e.g., movement); therefore, the learned trajectories in a steady state compensate for all underlying dynamics (i.e., human, robot, and task). If the dynamics change, the user has to compensate for the change to perform the given task, and the changed human effort signal again reshapes the trajectories in real-time, thus achieving coadaptation between the user and the exoskeleton (Peternel et al., Reference Peternel, Noda, Petrič, Ude, Morimoto and Babič2016). Alternatively, reinforcement learning can be applied to optimize learned trajectories based on some specific metrics while the user performs the task (Huang et al., Reference Huang, Cheng, Guo, Chen and Lin2016). The main advantages of adaptive learning methods are that they can reshape complex trajectories in a real-time manner and, unlike direct effort amplification, do not require a constant user effort when the task movement or force is repetitive (Peternel et al., Reference Peternel, Noda, Petrič, Ude, Morimoto and Babič2016).

As outlined above, several major fundamental principles for high-level control of wearable devices have been proposed in past decades. We predict that future research will mainly focus on improving upon existing fundamental principles, as has been already done in the recent years. Most of the fundamental human–robot coadaptation control methods still require various improvements in terms of processing speed and accuracy, control intelligence and adaptability, and design practicality in order to make them effective in realistic real-world scenarios. Nevertheless, some methods are also inherently reliant on research in complementary fields, especially sensing-related research to obtain meaningful and information-rich intention signals that can be used by the control system. For example, decoding brain activity (either through EEG or some other method) heavily relies on developments in the field of neuroscience, while EMG-based and model-based methods rely on developments in the fields of human motor control and biomechanical modeling. We believe that the most powerful and convenient way for the high-level control system to govern user-exoskeleton coadaptation would be to have a complete and detailed understanding of brain functions and a more practical means to measure it. However, as of now, we are still quite far from this point.

Benchmarking

Demonstrating that a device is working properly for a given application and communicating this information to end-users and stakeholders is an essential task when developing a new product. Very often, the performance of an exoskeleton prototype is considered “demonstrated” after the execution of a few experiments in a laboratory using ad hoc sets of metrics, which are usually established by the same researchers involved in the development process. This common practice should be seriously reconsidered if the wearable robot community wants to promote rigorous comparison across existing systems as a basis of optimized and efficient product development (Torricelli et al., Reference Torricelli, Rodriguez-Guerrero, Veneman, Crea, Briem, Lenggenhager and Beckerle2020), and benchmarking may help in this process.

Benchmarking is a comparative evaluation process often used by large companies to measure and improve the quality of products, services or processes. By means of comparisons with a standard, or benchmark, it is possible to identify the gaps in the state of the art and find strategies to improve competitiveness. But how can we translate this concept to wearable robotics? What features should be the object of comparison? Which devices or behaviors should be considered as standard or reference? The main problem is that exoskeleton research is still at an early stage, with hundreds of solutions at low technology readiness levels and only a few commercial solutions with unclear performance levels. One of the main mistakes in comparing robotic systems is to think in terms of “this robot is better (or worse) than that robot.” This viewpoint, common in competitions, does not reflect the multifactorial nature of performance. For instance, a faster system is not necessarily better than a slower one, and similar principles apply to many variables and features that may be used to describe the performance of a system. Nonetheless, comparing a characteristic of a system with the same characteristic of another system is still a very valuable process in order to understand why that system is behaving in a certain way. And this is what benchmarking is about: comparing to understand.

Recent efforts in the wearable robotics community made apparent the difficulty of defining the performance of a system on simple scales, triggering interesting discussions on the terminology to be used to define a given aspect of performance (Torricelli et al., Reference Torricelli, Rodriguez-Guerrero, Veneman, Crea, Briem, Lenggenhager and Beckerle2020). What we understood from these discussions is that the evaluation of performance is strongly dependent on the robotic technology under examination. For instance, “interaction” may assume very different meanings if applied to an industrial manipulator rather than to a wearable robot. This also applies to other system abilities, for example, “adaptability” and “decisional autonomy.”. One of the first steps in benchmarking is to agree on the definition of these concepts with respect to the specific technology considered. The motor function and the application domain are two additional factors that influence the definition of performance. For instance, would “stability” mean the same for rehabilitation and industrial robots? Would a unique definition apply to walking and reaching skills equally? Without clear definitions, it becomes impossible or even worse, incorrect, to compare two systems.

Another crucial aspect of benchmarking is categorizing the multiple facets of a given system into meaningful subgroups. Some years ago, Torricelli et al. (Reference Torricelli, Gonzalez-Vargas, Veneman, Mombaur, Tsagarakis, Del-Ama, Gil-Agudo, Moreno and Pons2015) proposed to categorize the “motion abilities” of a bipedal system into four classes: efficiency, stability, kinematics, and dynamics. Each of them was composed of four to five quantitative indicators obtainable by direct measurements. Whether this classification is appropriate or not is yet to be proven. However, the methodological approach of decomposing the performance of a system into measurable items seems promising. Future research efforts aiming at identifying quantitative indicators for all the missing system abilities would be strongly valuable for the wearable robotics community. In this respect, one largely overlooked aspect is human–robot interaction, which is particularly difficult to characterize since it is strongly dependent on the subjective perception and experience of the user (Beckerle et al., Reference Beckerle, Salvietti, Unal, Prattichizzo, Rossi, Castellini, Hirche, Endo, Amor, Ciocarlie, Mastrogiovanni, Argall and Bianchi2017). Another crucial research direction in robotic benchmarking is establishing causal relations between the specific actuator typology and the overall performance in a given application. This would greatly help developers in identifying how improvements in components would contribute to better system behavior.

The discourse on system abilities and their definitions should be taken to a practical level to bring a real impact to the community. It is useless to design a meticulous protocol able to test a specific performance aspect very rigorously if such a protocol cannot be easily reproduced in different labs and systems. If benchmarking is about comparison, the reality of experimental settings should be considered. For this reason, the concept of “reproducibility” as opposed to “replicability” is worth clarifying. As well explained by Plesser (Reference Plesser2018), reproducibility allows comparability of results from experiments conducted using different equipment (e.g., different sensors or environment) and even different procedural steps (e.g., a different number of trials). Replicability, instead, requires two tests to be executed exactly in the same way, using identical experimental setups, to produce comparable results. For benchmarking to be extensively adopted, we support the direction of reproducibility, even when this implies renouncing accuracy to a certain level. Very recently, several researchers have started to identify reproducible metrics, protocols and testbeds to quantify performance of exoskeletons in a wide range of motor skills and application domains (Torricelli and Pons, Reference Torricelli and Pons2019; Baltrush et al. Reference Baltrusch, Houdijk, van Dieën and Kruif2021; Kozinc et al., Reference Kozinc, Baltrusch, Houdijk and Šarabon2020a; Madinei et al., Reference Madinei, Alemi, Kim, Srinivasan and Nussbaum2020; Pinto-Fernandez et al., Reference Pinto-Fernandez, Torricelli, Sanchez-Villamanan, Aller, Mombaur, Conti, Vitiello, Moreno and Pons2020; Pesenti et al., Reference Pesenti, Antonietti, Gandolla and Pedrocchi2021). This notable increase in scientific interest is being accompanied by the intense work of several international committees for standards (e.g., ASTM F48, ISO TC299, and CEN/CENELEC), which contribute to fill two important gaps: the availability of standardized testing methods and rigorous quantitative data on exoskeleton performance.

Nevertheless, even in the presence of agreed terminology, reproducible protocols and testbeds, and data-driven evidence, a missing element to ensure wide adoption of any potential standard testing methodology is the data format. Anyone with a clear and reproducible protocol and a proper experimental setup can run an experiment, obtain data and calculate the desired metric. But, in order to benchmark such system, a comparison with a reference should be done. If results cannot be easily shared and put together, for example, using unified databases, individual efforts in producing datasets are likely to remain limited to local (single lab-based) comparisons, or relegated to simplistic comparisons using easy-to-understand but poor metrics (e.g., speed and power consumption). To exploit all the potential that a unified database may provide, it is imperative to first find a common format for data, and second, make the existing computational algorithms to be compliant with such standardized data format. This way, anyone can compare their own system with another one, or with normative data obtained from all similar previously tested systems. To reach this situation, some bottlenecks should be overcome, the most important one being the regulatory constraints on sharing data given that tests with wearable robots intrinsically include a human in the loop.