In this paper we have presented the development, characterization and wearability evaluation of a fully portable, powered one DoF wrist exoskeleton designed for independent and unsupervised training. The results of the characterization showed that the current prototype fulfils the technical requirements of output torque (up to 3.7 Nm), angular velocity (up to 530 deg/s) and RoM (154∘ or up to 215∘ if required), distal weight (238 g for forearm module) and autonomy (125 min) as previously specified in the literature. Furthermore, the wearability evaluation revealed that all participants (healthy and stroke) embraced the device and were able to don and doff it independently and quickly after a few practice trials.
Our approach of directly integrating the actuator and its drive locally at the wrist has the advantage of a rather simple implementation and good control of the wrist joint (PD steady-state error <0.12
∘). However, the motor alone (69 g) accounts for about 29% of the forearm module weight (238 g) and is therefore a major contributor of the weight placed distally on the arm. Fixing the second module on the upper arm reduces the weight distally and facilitates donning with the other hand, but still impacts arm motion in patients. This could be avoided by moving it to the back or less affected body side [
72,
96,
97], however, the further the exoskeleton is removed from accompanying modules, the more difficulties arise for donning and doffing independently. Thus, our solution is a compromise between good usability for donning/doffing and reducing the weight attached distally to the affected arm. The weight of the forearm module is comparable or lower than for other similar devices [
32,
98,
99].
The dynamics assessment has demonstrated that the angular velocities and accelerations achievable with the
eWrist in the restrained RoM are comparable to those observed in healthy skilled workers which perform typical manual activities [
100]. High achievable velocities and accelerations are necessary to render transparency. Despite a rather low position bandwidth of 1.7 Hz our impedance planes show that the implemented admittance controller can stably (cf. high R
2 and low residuals) provide transparent or resistive dynamic behaviour, which is important for accommodating different rehabilitation training settings [
101]. The capacity to provide all of these training modalities is important for haptic rehabilitation devices [
102,
103] for (i) training a wide range of impairments (i.e. from plegic to moderately impaired function), and (ii) quantitatively assessing the patient’s ability to perform movements without being disturbed by the device dynamics [
82].
PD controllers were implemented in both steady-state error and position bandwidth assessments and were tuned for maximum performance in each case. Although proportional Kp and derivative Kd tuning parameters were set to different values for each assessment, their ratio was kept the same to preserve stability (Kp/ Kd=10 in both cases). Moreover, Kp and Kd were 45% larger for the position bandwidth assessment compared to the steady-state error assessment in order to exhibit a more dynamic behavior. The two PD controllers were implemented for the sole purpose of performing these assessments and tuned independently to demonstrate the best capability of the device in each experimental context. Only the admittance controller is used during human-robot interaction, and dictates the experience of the user with the device.
Our autonomy assessment is comparable to other studies [
32,
39,
53] and would provide an extensive training dose for the user. However, the obtained autonomy must be interpreted with caution because it depends on the movement regime, for example, higher interaction velocities or higher interaction forces might arise if the hand is not completely passive. Surprisingly, we observed during the autonomy assessment that for a given time and with a substantially larger angular velocity (+42%), the energy consumption was reduced (-5%), revealing a non-linear effect which decreases with increasing velocity. This observation could partially explain the large disparity (-42%) between the desired virtual damping
Bvirt and the measured apparent damping
Bapp seen in Fig.
3a but not in Fig.
3b. With larger velocities this non-linear effect is lower, leading to a lower apparent damping felt by the user.
In the same vein, the large discrepancies (about +50%) observed in both renderings (i.e. transparent and resistive) between the virtual inertia
Mvirt and the apparent inertia
Mapp can be mainly explained by the intrinsic mechanical feature of our design, which requires that an interaction force needs to be applied first in order to illicit a motion. In the time delay (due to processing) between the force measurement and the handle motion, the force increases. And stronger forces will cause stronger friction between the gears and eventually resistance to the movement, thus leading to a larger apparent inertia experienced by the user compared to the one initially set in the controller. The steel-bronze combination for the worm drive is a fair compromise between low friction coefficient and high strength [
104]. Nevertheless, special attention must be given to optimising the manufacturing of these parts to keep their weight low. Moreover, the first-order characteristics of Eq.
1 also introduces a time lag in the control command which is directly linked to the inertia term. It would thus be tempting to minimize or even suppress this term to decrease time lag, however, we observed empirically that both terms (inertia and damping) are required to stabilize the exoskeleton during human-robot interaction. More specifically, stability was enhanced when the ratio between inertia and damping remained constant, as shown in other studies [
105‐
107]. Finally, since the worm screw can shift up relative to the worm wheel due to the oblong fixation points, our experience showed that the safety of the user’s wrist and the mechanics are preserved in case of unexpected high torque.
Wearability evaluation and general considerations
The effort directed towards the development of adjustable attachment systems which ease the donning and doffing procedure of the eWrist was positively received by the participants according to the scores obtained in our questionnaires. Although not standardized, the customized SUS questionnaire allowed us to get a better understanding of which specific aspects were favored and which were disliked. Encouragingly, the majority of participants quickly endorsed the mechanisms and found them efficient in terms of gripping force and adjustability. Generally, the doffing was found more straightforward than the donning. Stroke survivors judged wearability similar to healthy participants in the customized questionnaire. However, one clear limitation of our study is that we tested only two patients with moderate to minor impairment. In order to generalize our results to stroke patients, it would be valuable to also test wearability in more severely impaired patients. One important difference between the two cohorts was that healthy participants, but not stroke survivors, found their movements to be hindered by the device, most likely reflecting a difference in the perceived benefit of motor assistance via the exoskeleton.
As mentioned in the design review and also clearly expressed in the feedback, a critical phase during the donning is the correct placement of the forearm module to match the biological joint and mechanical axis of the eWrist. Most of the participants struggled with this aspect during the first four trials. During this phase, the forearm module must be balanced on the forearm and the hooks of the attachment system locked. However, the combined weight of the actuator, the gear drive and the load cell, all located on the same side of the module, tends to tip the device over. Nonetheless, our experience suggests that with slightly more practice both of these phases can easily be mastered.
According to a survey of 22 studies scoring mechanical tasks with the TLX [
108], the obtained score of 22.3 in the RTLX questionnaire (average workload score without considering
Performance) is below the 25th percentile of the scores (i.e. better than 75% of all scores). Nevertheless, despite this encouraging result, the scores comparison in Fig.
4b reveals that stroke survivors perceived mental and physical demands of donning/doffing much differently from healthy participants. This disparity, and more generally the wearability evaluation, should be further assessed by testing the device with more stroke survivors of different impairment levels and over several sessions. Nonetheless, it has been shown that the most critical usability problems are likely to be detected in the first few subjects, and that the likelihood of uncovering new problems decreases as more and more subjects participate [
109]. In our usability study, we consistently observed that difficulties encountered by healthy participants affected stroke subjects in a similar manner.
The weight of the exoskeleton was found to be acceptable. The rating was sometimes biased when participants would hold their whole arm over the table during the donning instead of laying it down, thus increasing their weight perception. Unfortunately, some participants felt discomfort mainly due to size mismatch. This can be addressed by tailoring the device to the individual user. For this study, two
eWrist of different dimensions were built, one for the right arm and one for the left. Based on anthropometric measurements (width, length and circumference) of the forearm, the wrist and the hand, an individualized exoskeleton can be printed. Tailor-made manufacturing with 3D printing techniques has already been adopted in community settings to offer simple prosthetics for impaired children [
110] and could potentially be applied for powered and more complex robots [
111]. Nevertheless, although the structure and 3D printed parts can be adapted, the electronics, load cell and actuator remain the same and would not properly suit small patients (i.e. <1m60 tall).
There were a number of general limitations to the wearability assessment. First, introducing the concept of the device before its assessment might have biased the participants towards higher ratings regarding functionality. Second, certain participants might have evaluated their own performance rather than the actual wearability of the device. Third, the wearability assessment was also limited in its design since participants were only evaluating the device during a single session. For instance, it would have been worthwhile to evaluate whether participants had memorized the procedure by retesting them after a week. Finally, the use of the
Think Aloud Method conjointly with the observations of the two experimenters allowed identification of where participants were experimenting difficulties in the task. However, even though participants were given preparation in verbalizing their thoughts, the use of this method with naive users had a tendency to slow down the execution time, especially with S2. Additionally, one has to keep in mind that we only evaluated the donning and doffing of the device but did not yet test its usability within a rehabilitation setting. Even though wrist extension/flexion function is highly relevant for post stroke recovery [
49], only supporting this movement in such a setting might limit some activities.
In its current form, the
eWrist is an important preliminary step towards a rehabilitation technology that could be donned, used and doffed independently by the patient in unsupervised settings, and which would complement a conventional therapy. Target patients would ideally start training with this device in the acute or sub-acute phase post stroke. The main inclusion criterion is low spasticity (i.e. MAS <3). However, patients who will likely benefit the most are those that have some remaining EMG activity in the forearm muscles and suffer from impaired hand and/or wrist function. In the initial phase, patients would use the device in a supervised manner, but as rehabilitation progresses and their impairment decreases, they would use the device more independently in daily life settings. As currently envisioned, rehabilitation training with the
eWrist will be in the form of a visuomotor task where the wrist angle of the exoskeleton is visualized as a cursor on a computer display and the patient performs wrist extension and flexion movements to move the cursor to different targets [
112], with an adaptive level of mechanical support from the exoskeleton based on sEMG amplitude. The control of robotic devices with sEMG signals have been extensively studied and one of the most preferred approach is to proportionally match sEMG to position [
75] or force [
34,
113]. We believe that a visual feedback combined with the mechanical support can not only reinforce sensorimotor loops and enhance the recovery process, but perhaps more importantly, boost motivation. Moreover, the wearable aspect of the device gives more freedom to the user and could easily be combined with a smartphone or a tablet.