We are witnessing an exciting era in which robotic technology for gait training is becoming a common tool in rehabilitation hospitals around the world. Advances in computing, materials, sensors, interfaces and manufacturing processes, along with the incorporation of basic knowledge of the neuro-physiological principles involved in motor recovery into intelligent controllers, are enabling better robotic rehabilitation services. Nevertheless, there is still no consensus on whether or not robot-assisted gait training benefits patients more than conventional therapy [1
]. Moreover, the effectiveness of locomotor therapy is limited regardless of the training approach [3
]. The use of ambulatory robotic exoskeletons, compared to robotic static gait trainers, may provide the patient with a more realistic and physiological gait condition, thereby increasing active participation in the therapy while providing task-consistent sensory and visual feedback.
However, the outcomes attained with ambulatory exoskeletons are still controversial. Published studies and reviews show considerable differences among protocols, targeted populations and variables analyzed, in addition to the specific differences among exoskeletons (number of joints, type of actuators and controllers, among others) [3
]. Recent research claims that robot-assisted walking arises from the interaction between the human body, driven by the central neural system (CNS) through the muscles, neural loops, reflex mechanisms and the mechanical structure of each exoskeleton, driven by the controller through the joint actuators and sensors [5
], there is still no consensus on the specific adaptation mechanisms, as well as what is the role of user preference on the performance of the human–exoskeleton system [8
Further investigation of some of these physiological mechanisms, as well as involving user preferences within the control loop, would allow better identification of the patients who can benefit the most from robotic therapy, how to shape and customize the exoskeleton structure and control the functional needs of the patient, as well as designing a personalized exercise program to conduct with the exoskeleton, in order to maximize motor learning and, ultimately, recovery [9
Along with the number and configuration of the joints, the main exoskeleton characteristic that affects human–robot interaction is how the robotic joints deliver torque to the human ones through the physical interface. Movement reference and joint actuator control are the two major areas of research. Several control algorithms for joint trajectory tracking have been proposed in the literature: from the simplest proportional-integrative-derivative (PID) control family, [11
] to more sophisticated algorithms such as fuzzy control [12
], robust variable structure control [13
] and sliding mode variable structure control [14
] have been used for lower limb exoskeleton robots [15
]. A common feature of these algorithms is that they do not consider the wearer in the system besides limb inertia in some cases [16
]. Exoskeleton–limb contact stiffness and user movement (either voluntary or reflex) greatly affect human–robot performance [17
]. Therefore, a smooth and efficient movement might be achieved by focusing on the human–robot interaction instead of on the accuracy of the trajectory tracking, to achieve smooth and efficient movement [18
]. Several control strategies have also been proposed that focus on the human–robot physical interaction: computed torque control [12
] and different versions of impedance/admittance controllers [16
]. However, despite the variety of control approaches investigated in the literature, the learning mechanisms of the user in response to robotic assistance are yet to be clearly established [19
In order to improve human–robot interaction, muscle electromyography (EMG) has also been investigated as a predictor of human intention during walking [20
], applied as a trigger for controlling prostheses and exoskeletons [22
]. EMG signals show characteristic patterns of activation associated with each activated muscle in terms of onset timings, burst duration and levels of activation [27
In addition, an analysis of combined muscle activation in terms of the number and characteristics of synergies has been proposed to provide a reliable representation of a person’s motor deficits and the degree of adaptability of their motor patterns [29
]. It has been shown that the use of ambulatory exoskeletons does not alter muscle coordination, independently of the level of assistance [10
EMG-based control algorithms may offer improved performance in terms of: (1) accuracy of movement selection, (2) intuitiveness and (3) response time of the control system [32
]. Movement accuracy is relevant for achieving precise execution of a user’s intended task; an intuitive interface relieves the struggle of the user on the use of the control system; and, finally, the response time is important for avoiding any possible delay perceived by the user, which may hinder proprioceptive mechanisms. The patterns differ between healthy and pathological gait conditions and therefore can be used to assess improvements in muscle function, motor control and neuromuscular adaptations following rehabilitation interventions [33
The main goal of this study was to compare the effects on human–exoskeleton interaction as well as user perception of the three main types of control strategies described above: joint trajectory tracking, joint mechanical admittance control and EMG-triggered control. We hypothesized that the EMG-Onset-triggered controller would improve human–exoskeleton interaction in terms of muscle coordination, physiological effort and walking experience compared to the admittance and trajectory controllers.
This exploratory study was conducted within the benchmarking initiative of the EUROBENCH European project, which developed the first unified benchmarking framework for robotic systems in Europe [34
], comprised of a testing facility located at the Center for Clinical Neuroscience
of Hospital Los Madroños (Brunete, Madrid, Spain)
in Madrid (Spain), as well as a comprehensive set of testbeds with dedicated experimental protocols and performance indicators (PIs) [35
] (hereinafter only testbeds). These results will allow companies and/or researchers to test the performance of their robots at any stage of development. In this work, we are therefore users of EUROBENCH’s project results. As shown in detail in Section 2
, we designed our experimental protocol with the two testbeds most suitable for our objective, which are the EXPERIENCE [36
] and PEPATO [38
] testbeds: the EXPERIENCE testbed aims at evaluating the user’s physiological response and subjective experience during exoskeleton-assisted walking on a treadmill, while the PEPATO testbed aims at analyzing muscle coordination during exoskeleton-assisted gait. More details are provided in Section 2
Therefore, the objectives of this work are twofold: (1) to compare the effects on human–exoskeleton interaction as well as user perception of the three main types of control strategies: joint trajectory tracking, joint mechanical admittance control and EMG-triggered control and (2) to provide third-party experience concerning the use of the EUROBENCH facility and testbeds, for both the project consortium and the scientific community.
This study focused on comparing the effects on human–exoskeleton interaction as well as user perception of three widely used exoskeleton control strategies—joint trajectory tracking (TC), joint mechanical admittance control (AC) and EMG-triggered control (OC)—while providing a third-party user-experience of the EUROBENCH testbeds, protocols and benchmarks. Our hypothesis was that the OC controller would improve human–exoskeleton interaction, in terms of muscle activation and coordination, and also reduce physiological effort while providing a better walking experience compared to the TC and AC assistance controllers. To test this hypothesis, we selected amongst the EUROBENCH available testbeds and protocols, the EXPERIENCE and PEPATO ones, which were combined and adapted to our objectives.
Overall, the admittance-based controllers—AC and OC—allowed the users to modify their walking kinematics, reducing the joint ROM (Figure 7
and Table 1
), while also delaying the maximum and minimum of the curves. These results are consistent with other studies in which able volunteers walked at lower speeds using an exoskeleton when compared to self-paced slow walking speed (with no exoskeleton) [31
]. The OC control strategy, while allowing the user to trigger each step individually, did not result in relevant kinematic alterations when compared to AC.
Regarding the synergy analysis obtained from the PEPATO software, we obtained an EMG reconstruction quality above 90% for all walking conditions, which indicates that four synergies were enough to account for the EMG variability for the three walking conditions. However, no statistical differences were found across walking conditions in any of the muscles within the four synergies. We hypothesize that a bigger sample size could have reached statistically significant differences in some of the synergy parameters, based on the trends observed.
We used a prototype EMG recorder (Section 2.2.2
) designed initially for upper limb tremor assessment. This device uses pre-gelled single-use disposable EMG electrodes connected by cables to a connection board that is plugged into the device. This configuration might have a good performance for the target application—upper limb EMG monitoring in quasi-static configuration—but showed to be not adequate for EMG monitoring of walking with the exoskeleton. Firstly, walking is a dynamic task that produces constant movement and friction between cables, which might induce electronic noise coming from this constant cable movement and contact. Note that, although the electronic noise arising from that phenomenon can be relatively low, the cables translate a raw EMG signal which is also a very low-potential signal. In addition, the exoskeleton motors, drivers and power cables also produce considerable electronic noise within the cables. In order to minimize these effects, an EMG system that integrates measuring, filtering and analog-to-digital conversion close to the measuring point would be beneficial for obtaining high-quality and good signal-to-noise EMG signals.
The four synergies showed a similar composition for TC and AC walking conditions and showed a shift, not statistically significant, in the OC walking condition. With respect to TC and AC walking conditions, we found that synergies 1, 3 and 4 showed ankle plantarflexion activity and synergy 2 ankle dorsiflexion, whereas the knee extension was explained by the activity of synergies 1, 2 and 3. Synergy 4 showed knee flexion activity accompanied by some antagonist (extension) co-contraction from the ReFe
muscle. These results align with what has been described in the literature. For example, Barroso et al. [50
] also calculated four synergies, finding that ankle plantarflexion activity was described by synergies 1, 3 and 4, ankle dorsiflexion by synergy 3, knee extension by synergy 2 and flexion by synergy 4, with antagonist co-contraction of the VaLa
muscle. Furthermore, Zhang et al. [53
] investigated the effects on muscle synergies due to exoskeleton-assisted walking compared with free walking. Their results during free walking are slightly different, but also showed ankle plantarflexion activity in synergies 2 and 3, ankle dorsiflexion in synergy 1 and knee flexion in synergies 2 and 4 with considerable antagonist co-contraction in the latter (ReFe
muscles). Exoskeleton-assisted walking did not modify the synergies although some significant changes in the balance of the muscle groups were observed, i.e., changes in muscle activation while the overall effect of the synergy in the joint movement remained the same. However, synergy 4 showed an increment in the agonist–antagonist activation of the knee muscles. We also found this agonist–antagonist effect in the knee joint in synergy 4.
All four synergies decreased the FWMH duration (expressed as % of the gait cycle, Table 4
) compared to the other two walking conditions. These results are consistent with a greater effort made by the user on the ankle plantarflexion and dorsiflexion movements in order to activate the onset algorithm. Note that estimation of the walking intention relied on the EMG from the Sol
of the ipsilateral leg and the ReFe
of the contralateral (i.e., loading) leg. Given the noisy quality of the EMG signals, the user had to increase muscle activation beyond what was natural to him/her to initiate the step in order to increase the EMG amplitude and to be successfully detected by the threshold algorithms. After the step detection, the user would decrease the overall muscle activation toward the level needed to complete the step. Note that the FWHM is a measure of the duration of the peak activation (i.e., the smaller the FWHM value, the higher the ability of the user to contract the muscle).
Regarding onset detection, we can see in Figure 9
that the right step was triggered, on average, in around 12% of the gait cycle, and the left step was triggered with both right ReFe
and left Sol
in around 60% and 62% of the gait cycle, respectively. The percentage of gait cycle timings obtained for both steps are in accordance with the values found in the literature [54
Results also show that to start the left step, the algorithm detected, in all cases, muscle onset in the right Sol, not using the left ReFe as a trigger of the exoskeleton in any step. Despite the filtering stage, the signal from the ReFe muscle remained too noisy for the algorithm to work properly. Although the cables and electrodes were carefully revised, as well as the EMG readings, this problem could not be solved.
Differences observed in the kinematics and muscle activation were not reflected in differences in the overall physiological impact of walking (Figure 11
). Furthermore, the group-averaged subjective perception did not differ across walking conditions (Table 5
). Therefore, although the AC and OC walking conditions allowed the user to modulate their walking pattern and to adapt to the exoskeleton actions, these were not flexible enough to actually reflect a change either in the overall physiological cost of walking or in the perceived experience of use. Furthermore, the item that received the lowest score (Perceptibility
, average 2.8 out of 7) indicates that the users experienced a low embodiment, agency and emotions (as defined in Section 2.5.2
) during all the walking experiments.
Along with the already discussed limitations on the EMG recordings, another limitation of this study is its reduced sample size (n = 7 healthy subjects), which affects the likelihood of obtaining statistical differences and thus the generalizability of the results. However, the study allowed us to confirm the usefulness and limitations of our EMG-Onset control versus conventional AC and OC control.