Abstract-- In this paper, we
present a state of the art regarding lower limb assistance systems. That is,
exoskeletons, exosuits, actuated orthoses and smart walkers intended to assist
users with or without walking difficulties in such task. We emphasize on those
systems headed towards motion intention prediction, especially those that
implement myosignals in order to control the system. However, there are some using
cortical or encephaloelectric signals. We make clear the importance of
developing this technology for the sake of the user’s health, when regarding people
with reduced motor capacity. After the reviewed systems, we bring forward our
own system intended to assist hemiplegic/hemiparetic patients walking by
predicting motion intention. The main differentiator of our design shall be
energy saving as it is being designed to actuate the hip only during swing
phase. Thus, hypothetically allowing the system to save 
60% energy,
compared to the same system actuated during the whole walking cycle.
Keywords-- Lower limb exoskeleton, motion intention prediction, myosignal, walking assistance, energy saving.
I. INTRODUCTION
Each year the number of people suffering from a lower limb (LL) mobility impairing disease increases. It is know from various medical reports that for people to keep good health it is very important that they to walk correctly (Organización Mundial de la Salud and Banco Mundial, 2011; Westerterp, 2013).
Mobility loss affects the patient’s economy, social status and yields negative effects that strike both the patient’s physical and mental health. Such negative effects can derive not from the disease itself but from a poor medical care. Accessing assistance devices that would allow patients going through a proper rehabilitation process and/or doing activities of daily living (ADL) independently would avoid such effects and, in some cases, recover in some grade. Nevertheless, in developing nations, only between 5 to 15% of the world population has access to the required attention (World Health Organization, 2013; Chen et al., 2016, 2017).
According to the WHO’s 2011 world report on disability (Organización Mundial de la Salud and Banco Mundial, 2011), jointly with The World Bank, disabilities have a social and economic negative impact, not only to those suffering from some disability but also their families and even their nations. On the same report, and also claimed by the UN, it is estimated that in 2050 22% of the world population will be over 60 years old and 2 billion would be adults with mobility issues affecting their ADL. Furthermore, this rate increases exponentially.
There are many reasons leading to lower limb mobility loss: an accident, disease or surgery aftermath. These people, as said before, face not only motor difficulties but a series of social and health problems. Namely, exclusion and discrimination, which, together with motor disability and realizing they need someone else’s assistance, leads them to depression. Moreover, they are also prone to suffering back disorders, obesity, cardiovascular diseases, muscle damage, to name a few, due to lack of motion or to unnatural motion (World Health Organization, 2013; Chen et al., 2016, 2017).
People suffering from motor disability require rehabilitation. Such process represents a great burden to the physiotherapist since it takes a large number of repetitions. Besides, it is not as effective as desired. That is one reason why experts are proposing mechatronic rehabilitation devices. This would relieve physiotherapist from the burden and provide the patient with a reliable and efficient rehabilitation (Ai et al., 2018).
Even though there are LL orthoses, they still have shortcomes that do not allow the user a natural motion. Moreover, some patients do not have enough strength to seize an orthosis, so it has become necessary to develop more advanced devices that would allow them performing ADL: exoskeletons (World Health Organization, 2013; Chen et al., 2016, 2017).
An exoskeleton is a type of assistance/rehabilitation robot that can be worn by a human, as in Fig. 1. It is used to assist motion, whether in rehabilitation process or through ADL or to augment human force (Marcheschi et al., 2011; Lee et al., 2014).
Lower limb exoskeletons (LLE) have been developed
in late years. Even though there has been great progress, and the use of
assistance and rehabilitation robots has been increasing, there are still
challenges to overcome: many do not allow performing many tasks, or allow just
a few, they use control strategies based on motion intention prediction (MIP)
via myosignals, to offer the patient a more natural motion and control it
implicitly with his bare intention, which implies a hard work on calibrating
the biosensors and factors that easily induce interference on the myosignals.
With patients with impaired motion, it is often complicated, if not impossible
in some cases, to obtain a reliable signal. Nonetheless, the use of myosignals
to control an LLE is considered one of the most powerful techniques (Harvard, 2017; Farris et al., 2011; Westerterp, 2013; World
Health Organiza-
Figure 1[1]: Assistance lower limb exoskeleton (Mummolo et al., 2018).
tion, 2013; Talaty et al., 2013; Taslim Reza et al., 2013; Murray et al., 2014; Pransky, 2014; Wang et al., 2015; Tsukahara et al., 2015; Chen et al., 2016, 2017; Ai et al., 2.018)
LLE are usually mechatronic devices. The interaction between a mechatronic and a biological system should be double: both cognitive and physical. The cognitive one between the exoskeleton and the user sets a frame that allows an effective control by the user and the sensorial and motion feedback to the user. The aim of physical interaction in LLEs is to provide stability while walking on patients with muscular weakness. In such case, users often cannot stabilize their knee/s, therefore, the exoskeleton must provide torques in the right times to maintain stability during the different walking phases (Moreno et al., 2009).
Moreover, MIP is vital so that there can be a synergy between the exoskeleton and the user. There should be safe control algorithms to acquire the user’s intention and instruct the system to respond quickly (Sanz-Merodio et al., 2014). That way, the user would only have to concentrate in doing the desired activity and not in controlling the exoskeleton (Fleischer and Hommel, 2008).
However, designing an exoskeleton with a purely mechanical control system, based solely in the user’s motion, may be significantly advantageous. This would improve the device’s weight, required maintenance and components simplicity, all in comparison with electromechanical systems (Yap et al., 2019; Zhao et al., 2019; Zhou et al., 2020).
Additionally, one of the goals of developing an exoskeleton is energy saving by taking advantage of human biomechanics to minimize energy consumption and allow the user to perform ADL naturally, besides maintaining stability, as well as reducing the system’s weight and the metabolic cost (Griffin et al., 2003).
As it seems, there are several issues still being tackled in order to improve and develop assistance technologies. We made a thorough investigation to gather information about the latest in lower limb assistance systems, presented in this paper. We also propose our own design, intended to save a considerable amount of energy by adding a passive ratchet mechanism and a clutch at the hip joint that would allow the system to disconnect the hip motor during the stance phase of the walking cycle without losing stability.
Figure 2[2]: Anatomical planes.
II. LOWER LIMB DEGREES OF FREEDOM
The human hip is considered to have three degrees of freedom (dof): one in each anatomical plane, sagittal, transverse, and coronal, depicted in Fig. 2. The sagittal plane divides the body in a right side and a left side, it is a vertical plane located along a vertical line that passes through the body’s center of mass; the transverse plane divides the body in an upper side and a lower side, located at the body’s center of mass and it is horizontal; the coronal plane divides the body in a frontal side and a back side. Like the sagittal plane, it is a vertical plane that passes through the body’s center of mass, but perpendicular to the sagittal plane.
When the hip rotates in the sagittal plane, such that the thigh moves to the front, the movement is called hip flexion, and when the hip rotates in the same plane, with the thigh moving to the backside of the body, is called hip extension; when it rotates in the transverse plane is called internal or external rotation, whether it rotates towards the inside or the outside of the body, respectively; and when it rotates in the coronal plane is called abduction, when the thigh moves towards the outside of the body, and adduction, when it moves in the opposite direction (Marrero and Cunillera, 1998; Wang et al., 2013; Peterson and Brozino, 2014; Herman, 2016).
As for the human knee, it can be considered to have four dof, three rotations, one in each anatomical plane, and one translation between the thigh and the calf, although due to soft tissue and boney constraints is commonly consider that the joint has only one rotation. Such rotation would be in the sagittal plane and is called extension when the calf moves towards the front, and flexion when it moves towards the back. Regarding the translational dof, it derives in a moving center of rotation and ensures stability of the joint while flexing (Marrero and Cunillera, 1998; Wang et al., 2013; Peterson and Brozino, 2014; Herman, 2016).
Figure 3[3]: The human walking cycle divided in 8 subphases (Kibushi et al., 2018).
Figure 4: Angular trajectories described by the hip, knee and ankle joints during the walking cycle of a healthy person in the sagittal plane.
The human ankle, is a very complex joint that is considered by some authors as two joints, but in exoskeleton research is often considered as a single joint with three dof, one in each anatomical plane and the names of the movements are analog to those of the hip (Marrero and Cunillera, 1998; Wang et al., 2013; Peterson and Brozino, 2014; Herman, 2016).
III. THE WALKING CYCLE
Human walking cycle can be divided in two phases: stance, comprehending 60-65% of the walking cycle, and swing phase, in which it alternates between single and double support, i. e., when one foot is touching the ground and when both feet are touching it, respectively. These two phases can be subdivided in eight sub-phases, as shown in Fig. 3.: 1) Initial contact, or heel contact. When the heel hits the floor. That is when the walking cycle, of one lower limb, begins; 2) Weight acceptance, loading response, or initial stance. Right after the initial contact until the foot is completely touching the ground; 3) Mid stance. It ends when the hip reaches its highest point; 4) Final, or late, stance. It ends when the hip reaches its lowest point and the opposite heel hits the floor; 5) Pre swing. From opposite heel contact until the foot takes off; 6) Initial swing. From foot takeoff until the hip reaches its highest point again; 7) Mid swing. Finishes instants before heel contact; 8) Terminal, or late, swing. Ends at heel contact (Kapandji, 2010; Herman, 2016; Prieto Villalba, 2016; Choi et al., 2018).
During the walking cycle, double support takes place during subphases initial contact, weight acceptance and mid stance, ergo, to transfer the body weight from one foot to another.
According to (Herman, 2016), during the walking cycle, the hip, knee and ankle joints describe an angular trajectory, in the sagittal plane, as shown in Fig. 4.
IV. LOWER LIMB ASSISTANCE SYSTEMS
Exoskeleton researching began in the 1960s. It was the US department of defense interest that yielded the development of arming and human force augmentation suits. The first exoskeletons were too big and heavy. Due to technology limitations, researching stopped. However, back in the 80s, it was again with military interest that researching restarted. Since then, there has been human force augmentation exoskeletons developed in universities such as the University of California at Berkeley and the Massachusetts Institute of Technology. This researching led to assistance and rehabilitation exoskeletons (Fleischer and Hommel, 2008; Taslim Reza et al., 2013).
In recent years, exoskeleton researching has been heading towards biosignalled control. On one hand, there are reports about exoskeletons controlled via corticoelectric (Fifer et al., 2012) and encephaloelectric (Villa-Parra et al., 2015; Delisle-Rodriguez et al., 2018; Long et al., 2019; Vinoj et al., 2019; Gordleeva et al., 2020) signals. On the other hand, development of exoskeletons using myosignals started over a decade ago and has become common for LLE researchers to implement myosignalled control strategies. Nonetheless, the pure myosignal is not enough to control an LLE. They are rather combined with force, inertia or gyroscopic sensors (Chen et al., 2014).
After reviewing some commercial LLEs it was noted that their walking speed is much lower than the standard human mean walking speed, ranging from 0.05 up to 0.72 m/s, their weight varies from 11.8 to 38 kg, while their cost reaches 160, 000 USD. Also, many of them are complemented by crutches (Rebiotex, 2020; ReWalk, 2020; Rex Bionics, 2020; suitX, 2020; Neuhaus et al., 2011; Kim et al., 2014; Wang et al., 2015; Herman, 2016; Ren et al., 2018; Indego, 2020).
Naik et al. (2018) developed a four dof passive LLE, two dof at the hip, one at the knee and one at the ankle. With one unrestricted dof at the hip, while the remaining is restricted by a ratchet mechanism controlled by a servomotor to restrict them in both ways. The mechanism is not designed to assist the user’s motion but to help him lift loads without transferring them to the user. The system weights 6 kg and is capable of bearing 30 kg without the user’s effort.
Kundu et al. (2014), Mazumder et al. (2014) merged myosignals from six muscles and used the merged signal in a PID controller. They report excellent results on ankle trajectory tracking with the PID controller during the walking cycle. They calculated the moving windows RMS and traced the knee angular trajectory to identify the walking phases with hopes to apply the detection system to an LLE.
Figure 5[4]: Walking assistance system developed by (Yap et al., 2019).
Chen et al. (2014) designed a 10 dof LLE, two at the hip, one at the knee and two at the ankle, with pneumatic linear actuators. They used myosignals and an insole to measure the pressure at the heel to help identify the user’s intention. They observed that the leg’s anterior compartment muscles tense first than those in the posterior compartment during sit-to-stand and the opposite during stand-to-sit. In addition, myosignal amplitude from the thigh’s frontal muscles is bigger than the one from posterior muscles of the leg during such transitions. Authors suggest that, in order to avoid motion intention misprediction, an LLE should not depend merely on myosignals.
Choi et al. (2018) reported using myosignals from the soleus muscle to predict the walking speed, based on the fact that the max value of the soleus’s myosignal, right at the end of terminal stance, is proportional to the walking speed during the swing phase. Furthermore, they used eight force sensors to measure the pressure at the sole and identify the current phase and obtained a relation between the walking speed and the sole force. For walking speed characterization, they used the square root of the sum of max amplitude of the soleus during weight acceptance and terminal stance. By comparing real versus predicted values, they obtained a coefficient of determination of 0.75. They intend to apply this system to an LLE for hemiparetic patients.
Joshi et al. (2013) implemented the Bayesian information criterion, standard methods for feature extraction and linear discriminant analysis (LDA) as classifier in order to design a walking cycle phase detection system. They used myosignals from four muscles and obtained a precision that ranges from 46.67 to 93.83%. Features used were four coefficients of the fourth order auto-regressive model and four time domain features. They report that in spite of good results, there is still uncertainty that could be reduced, according to authors’ hypothesis, by combining the myosignals with the sole and inertia sensors signals.
Ma et al. (2016) designed a rehabilitation LLE with a smart perception system. It consists of surface myosensors, piezoelectric film sensors and photoelectric encoders. Myosignals from the vastus lateralis are used to identify user’s intention; piezoelectric sensors are placed on the straps used to adjust the LLE to the user in order to measure the user-exoskeleton interaction force; and encoders are used to measure the hip and knee angles. To identify the current phase, an adaptive neuro fuzzy inference system (ANFIS) was implemented. They report a correct walking phase identification over 92%.
At Universidad de Cuenca, in Ecuador (Mora Tola et al., 2020), an exoskeleton to assist knee motion has been in development. Motion intention is detected through myosignals and a backpropagating artificial neural network (ANN). Authors concluded that lower limb in stance phase is best for a better classification.
Yap et al. (2019) parted from a principle, the mutual entrainment principle, that states that the walking cycle is a motion controlled by four extremities, two lower ones and two upper ones, to develop a walking assistance exoskeleton, shown in Fig. 5, upon the hypothesis that “it is reasonable to consider intervening [stimulate/actuate] in upper extremities motion to assist lower extremities motion”. The system consists on the actuator module (two DC motors, adjustable harnesses and velcro straps), which is attached to a harness that adjusts to the upper limb via velcro straps; the control module (two DC motor controllers, two encoders and an I/O); and the power module (a rechargeable battery and an emergency switch). Its total weight is 5.8 kg. The system is adjusted to the user via velcro straps around the chest and waist. Experiments were performed with 12 subjects from 71 to 77 years-old that reported no neurological disorders. The results show that the angular displacement amplitude increases, while the period decreases, while using the exoskeleton. Nevertheless, the improvements are lesser than 8% for the amplitude and 3% for the period. Nonetheless, the results prove that the system could have a beneficial impact.
Auberger et al. (2018) developed a brace-like adjustable system to assist knee motion. It is controlled by a finite state machine (FSM). They added an elastic element to store energy during the stance phase and release it at the swing phase. The brace system contains a hydraulic actuator, a plank mechanism and two servo valves to provide high impedance while knee flexing and low while extending it.
Chung et al. (2015) presented a sit-to-stand two-channel surface myosignals (SM) classification method. It uses LDA as classifier and a multiple windows feature extraction approach, consecutive time-windowed feature extraction (CTFE). Additionally, they implemented a majority voting algorithm as a postprocessor to reduce false positives. Results show a prediction precision over 90%. Furthermore, they found no muscle activation pattern in sit-to-stand.
Figure 6[5]: Lower limb exosuit developed by (Schmidt et al., 2017).
Villa-Parra et al. (2015) designed a system that combines myosignals and encephaloelectric signals. It comprises an LLE to assist knee motion, a smart walker that guides the user, keeping him in a safe zone, besides helping him maintain a stable posture during walking. Motion intention is detected from lower limb myosignals via a support vector machine (SVM) and an ANN. Features extracted were mean absolute value (MAV), wave length (WL) and autoregressive coefficients (AC). Authors compared the SVM versus the ANN through total error rate (TER), sensibility (SS), specificity (SP) and positive predictive value (PPV) and determined that the ANN provides better results than the SVM.
Yu et al. (2013) developed an out-of-clinic rehabilitation system involving a pair of series elastic actuators (SEA), one at the knee and the other at the ankle. It was designed with carbon fiber and weighs less than 4 kg. It has myosensors, angular sensors, pressure sensors at the sole and inertial measurement units (IMU) to determine optimal assistive torque and rehabilitation progress.
Hua et al. (2019) designed a variable magnification ratio system. It has active joints in the sagittal plane and passive joints in the frontal plane at the knee and hip to allow lateral movement while walking. The system is designed to assist the user walking, climbing up and down stairs, walking up and down a slope and during sit-to-stand and stand-to-sit. They also implemented a feedforwarding PID controller with extreme machine learning techniques and tracking differentiator, to compensate for human-robot interaction force (HRIF). They report that HRIF was reduced by 70.6%. To identify each of the six modes a deep neural network (DNN) is used with a precision between 97.2 and 97.3%, and up to 99.2 to 99.7% with an optimization algorithm. For walking phase detection, an ANFIS was set. They report a 17.5 to 21.3% rise in beats per minute (bpm) while using the system, and a rise up to 40.3% while not using it. Concluding that the system helps diminish the burden of the user while using it.
Xie et al. (2019) developed a variable stiffness flexible exoskeleton. Variable stiffness is obtained by cable actuation. Exoskeleton joints are actuated by an air cylinder, which also varies joint stiffness, and does not constraint the joint’s dof, for it is flexible. Stiffness and energy consumption are controlled through air pressure. It requires only one air supply to actuate the joints and vary their stiffness, and that the system can lock the joints when certain pressure is applied. Nevertheless, joint actuation and stiffness variation require four solenoid valves and two actuators, and that the system provides a low assistive torque.
Schmidt et al. (2017) presented an exoskeleton/exosuit, Fig. 6, consisting of three main components: user-exoskeleton interface; “ligaments”, elastic elements for energy storage and passively joint assistance, and; power component, an actuator and an artificial tendon connecting hip and leg. Force in artificial tendons is measured, as well as length change. Legs and torso angular acceleration and velocity are also measured. Control of the system is explicit, meaning that the user needs to press a bottom, in this particular case, for the system to operate in certain way. The system is biarticular: hip and knee are actuated by the same motor.
Shamaei et al. (2014) report a quasi-passive exoskeleton with a pair of springs and a clutch. The last one allows the system to move freely during the swing phase, for knee assistance. One of the springs stores energy during load acceptance and releases it during the rest of the walking cycle to let the joint rotate freely. The other spring stores energy during knee flexion and releases it during knee extension. To accomplish the springs’ storing-releasing task, they are controlled with a latching mechanism and an FSM. The inputs of the FSM are signals from the sole of the foot. Authors report coefficient of determination values between 91 and 99, besides reducing joint torques.
Zhou et al. (2020) aim for an LLE with springs for gravity compensation. The torques at the LLE’s hip and knee are generated by the force in the springs connected to gears. To provide safety to the user and not to increase the system’s size, springs were set inside the LLE’s elements. Moreover, spring’s pretension is adjusted with a leadscrew in order to balance different weights. They determined, via simulations, that, for periods from 3 to 8 seconds of the walking cycle, torques decrease up to 84% at the hip and 69% at the knee. Nevertheless, as the design compensates only for gravity, reducing the walking cycle’s period diminishes the spring’s efficiency, as they do not compensate for dynamic effects. According to the authors, the five test subjects, healthy persons, advised sensing the assisting force provided by the system. These latter results were based on qualitative data.
Zhao et al. (2019) developed an exosuit for knee assistance on the elderly while climbing stairs. It has two twisted string actuators (TSA) weighing in 390 g and steel cord artificial muscles. The system is fastened to the user with flexible straps at the waist and foot, and between these there are two nylon straps, set in series with Bowden cable, firmly fitted, one on the thigh and one on the calf muscle. According to the authors, combining the Bowden cable, the flexible straps and the nylon straps results in an artificial muscle acting similar to the quadriceps. Suit’s modular structure allows LL’s free motion, quick dressing/undressing and easily carrying it. With the suit there are kneepads included for making its use more comfortable. Due to the suit’s flexibility, it does not constraint LL’s dof, so it does not modify the user’s movements. To measure the joint angles, they used IMUs, one at each thigh and calf muscle and to detect gait phases during stair climbing, angular position and velocity were used, with which a fast and precise detection is reached and it allows the users to climb faster than their normal speed. Exosuit’s total weight is 3.5 kg and according to the results, a 29.8% mean assistance efficiency is provided.
Table 1 lists all the exoskeletons reviewed along with their main characteristics. From the investigation carried out, and from Table 1, we can conclude that it is possible to:
1) Implement a modern control strategy for an acceptable trajectory tracking.
2) Detect phases and sub phases of the walking cycle by measuring the pressure at the sole of the foot and/or lower limb joint angles.
3) Implement an FSM to obtain a fine precision in motion prediction.
4) Include elastic elements for energy storage during stance phase and release it during swing phase. This would cooperate for a lighter system.
5) Consider a quasi-passive exoskeleton that yields energy saving without diminishing LLE functionality.
It can be concluded that the three main subjects approached, as to lower limb assistance systems (LLAS), are: mechanical design, myosignal classification and control strategies. Regarding the last two, myosignal classification methods, even though they offer good results, require high energy consumption. Also, many researchers try and combine myosignals with joint angle measures. As to mechanical design, there is a trend to develop systems that would allow minimizing energetic consumption, most, mainly assistance LLE, with elastic elements, others with variable mechanical impedance.
V. ENERGY SAVING APPROACH FOR A HEMIPLEGIC/HEMIPARETIC PATIENTS ASSISTANCE LLE
Upon researching the subject, we hypothesized that it would be possible to design an LLE that considerably saves energy, when compared to another LLE under the same operating conditions with no means to save energy.
To calculate the energy consumed during the walking
cycle, first we mathematically modeled the lower limb based on diagram from
Fig. 7, where 
is the mass
of the upper body. If the center of masses of the thigh, leg and feet are given
by
Figure 7: Representation of the human lower limb and the upper body mass as a concentrated mass.
, (1)
, (2)
, (3)
, (4)
(5)
, (6)
where 
represents the
horizontal location of the 
-th member’s
center of mass (with 
for the
thigh, leg and foot, respectively), 
represents
the vertical location of the -th member’s
center of mass, 
represents
the center of mass location along the -th member
with respect to the -th joint (with
for the hip,
knee and ankle, respectively), 
represents
the length of the -th member, 
and 
represents
the sine and cosine of 
, 
and 
, and the
center of mass of the upper body is given by
, (7)
where 
is the
vertical distance from the hip to the center of mass of the upper body which,
for simplicity, and since its variation is small and the user of the
exoskeleton is desirable to keep a straight shape while walking, we considered
constant.
Then, we developed the dynamic model based on the
Lagrangian dynamic formulation (Craig, 2006). Parameters and , the masses
of the thigh, leg and foot where obtained from table 2, based on data from (Herman, 2016). Angular trajectories used
where the ones shown in Fig. 3.
The model was simulated for unitary height and mass.
Figure 8 shows the moments, in 
, for hip,
knee and ankle; Fig. 9 shows the angular velocities, in 
; and Fig. 10
shows the power, in 
, required for
each joint, defined as
, (8)
where 
is the moment
and 
is the
angular velocity.
Energy consumed was calculated by numerically
integrating power curves from Fig. 10. Trapezoid rule (Chapra and Canale, 2007) was chosen as
method, for an approximation. Energy consumed by the hip, knee and ankle where 
, 
and 
,
respectively. Therefore, the total energy is 
.
