A soft robotic exosuit designed by Harvard University was placed over the paretic limb to enhance forward propulsion and ground clearance, contributing to more normal walking post-stroke (Awad et al., 2017b). A bio-inspired soft wearable robotic device has been proposed by Carnegie Mellon University for ankle-foot rehabilitation (Park et al., 2014).

An active AFO has been developed by the Massachusetts Institute of Technology (MIT) to assist foot-drop gait (Blaya and Herr, 2004). The powered exoskeleton has been proposed at North Carolina State University (Takahashi et al., 2015). An ankle rehabilitation robot was designed by the Chinese University of Hong Kong for robot-assisted gait training of stroke survivors (Yeung et al., 2018).

An overview of recent wearable ankle rehabilitation robots is listed in Table 1. Comparative analysis of the actuator, gait event detection, control strategies, and performance evaluation of the wearable ankle robots will be detailed in the next section in terms of their merits and demerits.

The PMA consists of an inner layer made from butyl rubber tubing with two end-caps forming the terminal connectors to seal the muscle cylinder (Klute and Hannaford, 2000; Klute et al., 2002; Davis et al., 2003). The main merits of PMA are its high torque-to-weight ratio and natural compliance. PMA was applied to these studies (Bharadwaj et al., 2005; Ferris et al., 2006; Gordon et al., 2006; Sawicki et al., 2006; Cain et al., 2007; Ward et al., 2007; Kao et al., 2010; Park et al., 2014; Takahashi et al., 2015; Jacobs et al., 2018). Powered ankle exoskeleton based PMA are listed in Table 2.

A large proportion of developed ankle robots include unidirectional devices that provide only plantar flexion assistance (Sawicki et al., 2005, 2006; Gordon et al., 2006; Gordon and Ferris, 2007; Kao et al., 2010; Koller et al., 2015). In Kao et al. (2010), two PMAs were connected to the posterior of the orthosis to provide plantar flexion assistance. The maximum torque created by the ankle exoskeleton was 50.09 ± 12.05 Nm. Moreover, PMA was attached to the anterior of the orthosis to provide dorsiflexion assistance (Bharadwaj et al., 2005; Bharadwaj and Sugar, 2006; Ward et al., 2007). Specifically, a robotic gait trainer was proposed by Arizona State University to assist stroke survivors during walking (Ward et al., 2007). The RGT employs a pneumatic spring over muscle (SOM) to generate bidirectional forcing, which improves the shortage of PMA acted only in contraction forcing. The SOM actuator consists of a standard compression spring linked in parallel with a traditional pneumatic muscle and has bi-directional, compliant, and lightweight characteristics. When the two SOM actuators move in the simultaneous direction, the ankle is driven in plantar flexion/dorsiflexion in the sagittal plane. Similarly, the ankle is moved by the two SOM actuators in the opposing motion to generate inversion/eversion in the frontal plane. Two pneumatic muscles actuators were attached to the anterior and posterior shank sections to assist with plantar flexion and dorsiflexion (Ferris et al., 2005, 2006).

In Park et al. (2014), a bio-inspired soft wearable robotic device powered by four PMAs was used to assist dorsiflexion, plantar flexion, inversion, and eversion. The design mimics the morphology and functionality of the biological muscle-tendon-ligament structure to provide active assistance. This bio-inspired design is lightweight, providing multi-DOFs assistance, and natural degrees of freedom are not limited. However, the disadvantage of the device is that it is complex and not portable. Moreover, the details regarding the assistance torque were not reported in the study.

The series elastic actuator (SEA) is composed of a traditional brushless direct current (DC) motor in series with a standard spring (Veneman et al., 2016). To control the impedance of the orthotic ankle joint for sagittal plane rotations, Robinson et al. (1999) previously developed the SEA, which is used for lower-limb robots. Advantages of the SEA include: (1) it has low impedance; (2) it has high control precision and storing energy; (3) the effects of backlash, torque ripple, and friction are filtered by the spring; (4) the shock load is isolated from the motor (Pratt and Williamson, 1995). SEA is limited by its large volume, heavy mass, and complicated structure (Zhang et al., 2017).

An active ankle-foot orthosis was designed by the Massachusetts Institute of Technology (MIT) to provide motions in plantar flexion/dorsiflexion and to mainly prevent foot-drop and foot slap during walking (Blaya and Herr, 2004). Its total weight is 2.6 kg. It does however, have some disadvantages such as being tethered and being heavy. An active ankle-foot orthosis (AAFO) has also been developed at the Yonsei University (maximum torque 97.2 Nm plantar flexion/dorsiflexion) to help avoid foot slap during weight acceptance and foot-drop in the swing phase (Hwang et al., 2006; Kim et al., 2007, 2011). The active ankle-foot orthosis was powered by the SEA to provide for plantar flexion/dorsiflexion. The MIT's AFO and AAFO designed by Yonsei University have similar designs, with a SEA actuator attached posteriorly to a conventional AFO. A portable robotic device has been developed by Arizona State University to improve gait kinematics and to enhance gait speed and walking duration for stroke survivors (Ward et al., 2010, 2011). The above discussed powered ankle-foot orthosis is powered by the SEA.

A motor is often used as a common actuator in ankle rehabilitation robots. For ambulatory individuals, soft robots have contributed to a more natural man-machine interaction, and have minimized the disruption of stroke survivors during the natural dynamics of walking. The most recent innovative design is a soft exosuit (Bae et al., 2015, 2018a,b; Awad et al., 2017b). Awad et al. (2017b) reported that a soft robotic exosuit designed by Harvard University was placed over the paretic limb to enhance forward propulsion and ground clearance, contributing to more normal walking post-stroke. An exosuit transmits mechanical power generated by the actuator to the wearer through the interaction of garment-like, functional textile anchors and cable-based transmissions. The overall mass of the exosuit is only 0.90 kg. The soft exosuit was developed to assist paretic ankle dorsiflexion and plantar flexion.

The Anklebot which was actuated by two linear actuators, has been used for gait rehabilitation post stroke (Roy et al., 2007, 2009, 2011, 2013, 2014; Forrester et al., 2016). The dorsiflexor/plantar flexor torque is generated when both dc-motor pull/push in simultaneous directions, while the actuators move in opposing directions, and creates the inversion/eversion rotational torque. The design is similar to the robotic gait trainer (RGT). However, MIT's Anklebot has some disadvantages such as being heavy, its bulky size, and being tethered. Recently, a lightweight and autonomous ankle robot has been developed for gait training of chronic stroke survivors (Yeung et al., 2017, 2018). The ankle robot is compact, lightweight, and portable.

Additionally, an Electro-hydraulic ankle-foot orthosis (EHO) has been developed by the Laval University (Noel et al., 2009; Blanchette et al., 2014). The custom-design uses a hybrid drive system which includes pneumatic, hydraulic and electric systems. More specifically, the EHO is characterized by its high power and light weight. Moreover, Tanida et al. (2009), Kikuchi et al. (2010) proposed the intelligently controlled ankle-foot orthosis (I-AFO) powered by compact magneto-rheological fluid brakes (MRB). Additionally, the portable powered ankle-foot orthosis, which was designed by the University of Illinois, provides plantar flexor/dorsiflexor torque assistance through a bidirectional pneumatic rotary actuator.

Ankle robots can also be divided into untethered and tethered devices. A few of the presented ankle rehabilitation robots are portable devices (Shorter et al., 2011a; Awad et al., 2017b; Choi et al., 2018; Yeung et al., 2018). Portable devices can be used to assist impaired users in daily life activities. It has been found that SEA and motors are common actuators in portable ankle robots. Moreover, the small size and light weight of the systemic component is a key requirement in the portable robot. In contrast, some tethered devices which are suitable for the gait rehabilitation of stroke survivors in hospitals or rehabilitation centers were designed in studies (Blanchette et al., 2014; Bae et al., 2015; Awad et al., 2017b). The advantage of tethered robots is that they do not add heavy components to the body which results in a lower negative impact for users.

Of those discussed above, the three main types of actuators all have benefits and limitations. PMA has the advantage of a high torque-to-weight ratio and compliance, but in portable systems, the compressed air generator restricts their potential use. SEA is a popular actuator choice in exoskeletons. This is due to advantages including high control precision, low impedance, storing energy, and high back-drivability. However, SEA is limited by the large volume, heavy mass, and complicated structure. Moreover, since the mass, size, and output torque are suitable for most applications, the motor plays the largest role in exoskeletons. Nevertheless, a long power supply provides to the series elastic actuator (SEA), and the motor has traditionally been stated as one issue in the development of portable ankle exoskeletons.

Although the rigid ankle exoskeletons discussed above improve mobility and independence of stroke survivors, it has some drawbacks. Since the rigid ankle exoskeleton adds a burden to the survivor's lower extremity, this will inevitably limit the gait kinematics, resulting in a slow and inadequate gait. Moreover, if the exoskeleton's rigid joints misaligns with the user's biological joints, it can give rise to uncomfortable stress on the soft tissue and bones (Bae et al., 2015). In contrast, soft exoskeletons hold some tremendous advantages such as compliance, natural interaction, adaptation, light weight, small size, less energy used, and is easier to wear. However, it has some demerits. Cable length and routing are not reconfigurable for different survivors with various body types and different paretic sides because the Bowden cable is fixed in series with the actuation system. The soft ankle exoskeleton also has difficulties in transferring power from the area of the body to the ground and motors and sensors are also difficult to mount.

In the works presented in Kim et al. (2007), Tanida et al. (2009), Shorter et al. (2011a), Park et al. (2014), foot switches or foot pressure insoles were utilized to identify and judge gait events. Once sensor magnitudes located in the heel and the metatarsal exceeded the predefined thresholds, gait events are detected. It is noted that force sensor thresholds are adjusted for each user to determine event boundaries during the gait cycle. Moreover, different ankle exoskeletons use a different number of sensors. Most commonly, two force sensors are embedded in the forefoot and heel, respectively, to identify toe off (TO), and heel strike (HS).

Awad et al. (2017b) reported that kinematic information measured by gyroscope sensors identified gait events to control the assistance time of the exosuit. Combining the detection ankle kinematic and kinetic information of all kinds of sensors can enhance the accuracy of the gait phase detection to make sure that the exoskeleton robot will work effectively and reliably (Blaya and Herr, 2004; Yeung et al., 2017; Choi et al., 2018). It was found that an inertial measurement unit and two force sensors could be used to identify the swing phase and stance phase and three walking modes, namely level walking and stair ascent and descent (Yeung et al., 2017).

Ferris et al. (2006) reports that the soleus electromyography (EMG) signal activates the PMA to generate plantar flexor assistance and the tibialis anterior EMG signal controls the PMA to produce dorsiflexor assistance. More specifically, when the soleus EMG signal magnitudes exceeds the predefined threshold, the control system completely inhibits the activation of the PMA applied to the dorsiflexor motion. Additionally, Joshi et al. (2013) separated eight different phases of gait by utilizing the lower limb EMG signal, which is based on Bayesian Information Criteria (BIC), standard feature extraction methods, and a Linear Discriminant Analysis (LDA) classification algorithm. However, Taborri et al. (2016) suggest that a method based on threshold rules applied on electromyography (EMG) signals performs better than machine-learning algorithms.

In short, mixed sensing techniques which combine kinematic and kinetic information could detect motion intention in a multimodal condition, namely level walking, stair ascent and descent, and ramp ascent and descent. As such, this technique embedded in a portable powered ankle exoskeleton can contribute to the daily life of stroke survivors.

The gait event detection technique may not be robust enough in over-ground walking, as it depends on identifying heel strikes and the foot-flat phase during the gait cycle. Nevertheless, we also observed that some stroke survivors land with the mid-foot rather than walk with a heel strike. Survivors often utilize “vaulting” compensations to reduce foot-drop (Kerrigan et al., 1997), eliminating the foot-flat phase. Therefore, this technique may not be suitable for post-stroke populations. Bae et al. (2018b) reported that the new gait detection algorithm used foot angle and angular velocity measured by foot IMUs to recognize paretic and nonparetic toe-offs, and nonparetic mid-swing. The results showed that the new gait detection algorithm was implemented to improve gait event detection reliability, compared with previous gait event detection algorithms which detected heel strikes and foot flat phases (Bae et al., 2015).

The position control method is trajectory-tracking control, which is to drive the patient's foot to move about on the reference trajectories with the help of the ankle robot. These reference trajectories data were normally measured from the healthy limb or healthy subjects, using the motion sensors in the biomechanics labs (Andersen and Sinkjaer, 1995; Wheeler et al., 2004; Ward et al., 2007, 2010; Shorter et al., 2011a; Park et al., 2014; Awad et al., 2017b).

A force signal is produced by limb contraction and interactions with a mechanical structure. Compared with the EMG signal, the force signal has better determinacy, which can reflect the motion intention of the patient better, so the control based on force signal is feasible and relatively steady. Force control schemes were found in some studies (Tanida et al., 2009; Blanchette et al., 2014; Choi et al., 2018), however, position control and force control might not be the most appropriate control scheme for robots with a medical purpose, which mainly requires dynamic interaction. To solve this problem, the impedance/admittance control scheme was proposed in the robot control field (Hogan, 1985). Since then, impedance control has gained extensive application (Chiaverini and Sciavicco, 1993; Ziren and Goldenberg, 1995; Seraji and Colbaugh, 2002). Impedance control schemes were found in multiple studies (Blaya and Herr, 2004; Roy et al., 2009, 2011, 2013, 2014; Forrester et al., 2016). The impedance control scheme provides a natural, comfortable, and safe touch interface, effectively avoiding secondary damage. An additional advantage of impedance control is that the achievement of impedance control was independent of prior knowledge (Tsoi and Xie, 2009).

An EMG signal is the electrical activity produced by the skeletal muscle (De Luca, 1997; Robertson et al., 2013; Zhang et al., 2017; Li et al., 2018). To effectively extract motor control command from myoelectric signals, substantial work has been carried out (Nurhanim et al., 2014). Several wearable ankle rehabilitation robots have been controlled by EMG signals (Ferris et al., 2005, 2006; Sawicki et al., 2005, 2006; Cain et al., 2007; Takahashi et al., 2015; Koller et al., 2018). EMG signal control has been employed by the University of Michigan's AFO (Ferris et al., 2006; Cain et al., 2007). Ferris et al. (2006) proposed that the soleus electromyography (EMG) signal activated the PMA to generate plantar flexor assistance and the tibialis anterior EMG signal controlled the PMA to produce dorsiflexor assistance. Cain et al. (2007) demonstrated that proportional EMG control led to a larger decrease in muscle activation and gait kinematics closer to normal than footswitch control. Compared with the force signal, EMG has the following advantages: (1) the acquisition of EMG is simple; (2) use of the EMG signal can identify finer movements than force signal; (3) the interactive control based on EMG has more flexibility, which can realize the control of the diseased limb through the healthy limb according to the coordination of the body movement. Nevertheless, the myoelectric signals of stroke survivors may be weakened. Additionally, the ankle muscles post stroke become too weak or paralyzed to generate abnormal muscle activation (Wright et al., 2014). Furthermore, the myoelectric signals are affected by electrode-skin conductivity, improper electrodes alignment, fatigue, and the interaction between nearby muscles (Fleischer and Hommel, 2008; Tucker et al., 2015). The EMG signals are also non-stationary during dynamic activity, which requires the utilization of pattern recognition techniques (Souza et al., 2014). It is usually necessary to calibrate each time the device is installed in practical use (Dawley et al., 2013).

Studies (Shorter et al., 2011a; Takahashi et al., 2015) evaluated the effect of ankle rehabilitation robots on the muscle activation of the lower limb of survivors. In Takahashi et al. (2015), a powered ankle exoskeleton was used for five stroke subjects to walk three sessions for 5 min each session. The results showed that the soleus activation of the paretic side of the three subjects during the propulsion phase was decreased with the powered assistance compared to unassisted walking. Meanwhile, in the powered assisted walking test, the exoskeleton increased the paretic plantar flexion torque by 16% relative to the unassisted walking condition. However, in Shorter et al. (2011a), it can be seen that the tibialis anterior activation in stance and swing phase was reduced during assisted walking trials for non-disabled subjects. Moreover, Kao and Ferris (2009) studied the effect of an active dorsiflexion assist orthosis that was proportionally controlled by tibialis anterior electromyography on the muscle activation of the neurologically intact subjects. It was shown that in the continuous group, the amplitude of tibialis anterior EMG during the swing-to-stance phase transition was reduced by 28%, in contrast, the amplitude of tibialis anterior EMG during the stance-to-swing phase in both groups was not decreased. Therefore, the results only partially supported their hypothesis that the subject's tibialis anterior muscle activity is reduced during walking with a powered dorsiflexion assist orthosis which is proportionally controlled by the tibialis anterior EMG.

Furthermore, Young and Ferris (2017) reported that exoskeletons were to reduce muscle recruitment of the lower limb during walking. The EHO (Blanchette et al., 2014) has been developed to assess the residual adaptive capacity of ankle dorsiflexor when the paretic limb of stroke survivors was added to a perturbation. An interesting result was reported that the mean amplitude of tibialis anterior (TA) in four of the six participants was significantly increased after the walking period with the perturbation. Moreover, modifications in tibialis anterior (TA) activation in three of the four participants persisted after perturbation removal.

From the results mentioned above, it can be found that for healthy subjects, reducing plantar flexion and dorsiflexion muscle recruitment walking with ankle exoskeletons reduces the user's metabolic cost of walking. In contrast, for stroke survivors, there is no consensus that the ankle exoskeletons can reduce dorsiflexion and plantar flexion muscle activity.

Post-stroke paresis of the ankle musculature causes a weak ability of the active dorsiflexion during the swing phase. The ankle rehabilitation robot can enhance the ability of ankle dorsiflexion of the paretic limb and ground clearance, which reduces the risk of falling. Dorsiflexion angular range during the swing needs to be measured to quantify the reduction of the second major complication of foot-drop. The comparison of peak dorsiflexion angle during the swing is listed in Table 5.

Studies (Roy et al., 2013; Forrester et al., 2016; Awad et al., 2017b) showed the improvement of the dorsiflexion angle on the paretic side during the swing after rehabilitation training. Forrester et al. (2016) reported that the peak dorsiflexion angle on the paretic side of stroke survivors during the swing in the treadmill robot training group was increased, with gains sustained at a follow-up test. A similar effect was obtained in Roy et al. (2013), achieving bigger gains in a follow-up test. Awad et al. (2017b) assessed the peak dorsiflexion angle on the paretic side during the swing in a treadmill group for unpowered vs. powered over-ground training. Compared with the over-ground group, the treadmill group achieved a greater improvement of the peak dorsiflexion angle on the paretic side. An increase of 5° (bold from Table 5) in the dorsiflexion movement was usually regarded as clinically significant (Rose et al., 2010). However, an increase of the dorsiflexion angle reported in Blanchette et al. (2014), Awad et al. (2017a), Awad et al. (2017b) was <5°. The results may be related to a single session of walking with powered ankle exoskeletons. In contrast, multiple sessions of training contributed to statistically significant increases in the dorsiflexion angle on the paretic side (Roy et al., 2013; Forrester et al., 2016). Therefore, gait rehabilitation post-stroke may benefit from multiple sessions of gait training using an ankle rehabilitation robot. By comparing immediate effects between unpowered and powered ankle rehabilitation robot, it would be more valuable to assess statistical significance before, and after the rehabilitation training of multiple sessions with an ankle rehabilitation robot. In particular, follow-up evaluation is warranted to validate the sustainability of gait recovery.

Furthermore, the active range of motion in the dorsiflexion of a seated ankle rehabilitation robot was also reported in Forrester et al. (2016), showing that the treadmill robotics (TMR) group made significant progress and continued to improve during the 6 w period after rehabilitation training. We believe that an active range of motion in dorsiflexion of standing posture without any assisted device might be a key and valuable evaluation indicator.

Ankle rehabilitation robots designed for assisting propulsion on the paretic side could reduce propulsion asymmetry and facilitate more normal walking in patients after stroke. Moreover, the improvement of propulsion on the paretic side plays a decisive role in the improvement of walking speed.

The results reported in Table 6 show the effect of powered exoskeletons on the ankle paretic propulsion, propulsion asymmetry and propulsion impulse of stroke survivors. All the reported studies found an improvement in propulsion on the paretic side, propulsion asymmetry and propulsion impulse in Table 6. However, peak paretic propulsion during assisted push-off with powered exoskeleton was reduced in the sham group vs. the robotic group after 20 training sessions in Yeung et al. (2018). One of the possible reasons that might explain this result is traditional rigid ankle-foot orthoses (AFOs) in the sham group, which limits normal push-off and reduces gait adaptability (Vistamehr et al., 2014). Furthermore, it can be seen that compared to unassisted walking, the exoskeleton did not affect the percentage of paretic propulsions in stroke survivors walking with the powered assistance (Takahashi et al., 2015). One possible reasons is the low sample size. Other factors may include suboptimal timing of exoskeleton driving and insufficient adaptation when using the exoskeleton. Only a few studies that have been performed with powered exoskeletons assessed the effect on propulsion asymmetry (Bae et al., 2015; Awad et al., 2017b) in Table 6. It has been shown that walking assisted push-off with powered ankle exoskeletons can reduce propulsion asymmetry.

Forrester et al. (2016) verified that the ankle robot combined with a treadmill was more significantly effective in increasing paretic push-off impulse than the seated training after 6 w of rehabilitation training. Therefore, for chronic stroke survivors, a treadmill combined with ankle robot training improves gait function more effectively than robots that focus on ankle training.

Stroke survivors are characterized by hemiplegia gait, causing a slow, metabolically inefficient gait and an increased risk of falling. In Young and Ferris (2017), it is suggested that in these post-stroke populations, the clinical measurement method is the walking speed, usually using the 6 min walk test (SMWT). Higher priority walking speeds indicate better clinical outcomes because walking speed is closely related to social mobility. Yeung et al. (2018) reported that after the 20-session gait training, the SMWT of robotic and sham groups showed a certain extent of improvements in walking capacity and endurance, which improved both + 5.7 m (post-pre), and + 22.2 m (follow-up-pre) in sham groups, and increased + 16.9 m (post-pre) as well as + 41.5 m (follow-up-pre) in robotic groups. The 17% increased walking capacity was greater than the minimal clinically important difference (MCID) of SMWT for walking endurance, which is a 11.5% proportional change in walking distance for stroke survivors (Tang et al., 2012).

Studies (Ward et al., 2007, 2010) assessed the effect of an ankle rehabilitation robot on walking capacity of a stroke survivor. In Ward et al. (2007), it is reported that a stroke survivor obtained higher gains from RGT (Robotic Gait Trainer) across the 16-sessions of gait training, contributing to an increase in SMWT. Moreover, the results from SMWT for subject 1 and 3 between the pre-and post-intervention test showed improvements (Ward et al., 2010). Both ankle rehabilitation robots have been assessed on stroke survivors for at least three consecutive weeks in gait training sessions. Although the two clinical studies did not have control trials and the sample size was small, the researchers found that the longer the walking distance of the SMWT, the better the clinical outcomes.