Individuals with complete paraplegia due to spinal cord injury (SCI) have impaired motor control and sensory feedback that limits their ability to walk (Shepherd Center). While wheelchairs provide alternative mobility to individuals with paraplegia, they are not a complete substitute for natural ambulation. Current research has addressed this issue with wearable lower extremity exoskeletons (Dollar and Herr, 2008; Contreras-Vidal et al., 2016). The past decade has witnessed a dramatic growth in the study and implementation of such technology, not only for those with paraplegia due to SCI, but also for individuals with cerebral palsy, stroke, traumatic brain injury and multiple sclerosis (Canela et al., 2013; Murray et al., 2015; Kozlowski et al., 2017; Lerner et al., 2017; Patané et al., 2017; Androwis et al., 2019; Karunakaran et al., 2019). These exoskeletons are mechanically similar, consisting of a set of linkages that parallel the wearer's thighs, calves, and feet, and augmented with actuators to provide alternatives to muscle torque at the joints. Unlike devices developed for military and industrial tasks, most rehabilitation exoskeletons rely on the subject's use of crutches or canes to provide balance, as the devices lack sensory feedback and balance compensation (Contreras-Vidal et al., 2016).

Where current rehabilitation exoskeletons differ significantly is in their detection of user initiation of gait patterns (Dellon and Matsuoka, 2007; Strickland, 2012)^1,2. The Ekso (Ekso Bionics) has motors at the hip and knee, with passive springs at the ankles to provide gait only in the sagittal plane (Strickland, 2012)¹. The Ekso has two options to initiate the gait cycle. (1) The first allows a clinician to control gait by means of an external switch pad for training and therapy. (2) For independent control, sensors embedded in the device detect changes in the hip position¹. A step can be initiated by the user moving the hip forward and laterally or by changing tilt angle and making ground contact with sensors on the crutches¹. Goldfarb et al. developed what has been commercialized as the Indego (Parker Hannifin) (Farris, 2012; Farris et al., 2012; Quintero et al., 2012). It also has powered degrees of freedom (DOF) at the hip and knee, and passive ankle support in the sagittal plane. It uses Hall effect sensors, potentiometers, and accelerometers to detect the user's center of pressure (COP) (Farris, 2012; Quintero et al., 2012). When the user leans forward with both crutches touching the ground, the COP shifts in the direction of movement, and the exoskeleton initiates swing of the most rearward leg (Farris, 2012; Quintero et al., 2012). The Rewalk (Rewalk Robotics) also has two active DOFs with the ankle consisting of a simple orthotic joint with limited motion and spring assisted dorsiflexion (Esquenazi et al., 2012). The control system includes a tilt sensor to determine changes in trunk motion and center of gravity. Shifts in the center of gravity initiate the preprogrammed hip and knee displacement in the appropriate leg (Esquenazi et al., 2012). The HAL (Cyberdyne) employs a combination of EMG gait initiation detection with an accelerometer and gyroscope to sense body posture (Lee and Sankai, 2002, 2003; Hayashi et al., 2005). In contrast to the above exoskeletons that have two active degrees of freedom, the REX (REX Bionics) has 5 motors per leg and is the only available assistive exoskeleton to provide movement in the coronal as well as sagittal planes, and to be self-balancing (i.e., no crutches²). It is controlled by a joystick that signals one of 8 discrete directions of ambulation, and has button selection for sitting and rising².

Electrophysiological signals have also been employed to initiate exoskeleton gait (Kilicarslan et al., 2013; Kwak et al., 2015; Lebedev and Nicolelis, 2017). Contreras-Vidal et al. have demonstrated the use of EEG triggers to select various REX exoskeleton's discrete control commands (Farris et al., 2012). Similarly, other investigators have focused on detecting gait initiation and termination events using EEG, EOG, evoked potentials, and other bioelectric signals (Nicolelis, 2003; Kilicarslan et al., 2013; Kwak et al., 2015).

Beyond the triggering of preplanned gait patterns, extensive research has been pursued on control methods that allow users to plan and execute novel gait patterns similar to those of individuals with no disability. Gancet et al. have tried to interpret EEG signals from the motor cortex to calculate the kinematics of the gait cycle. A dynamic recurrent neural network was used to train the network to detect the gait patterns in the EEG signal (Gancet et al., 2012). Lebedev and Nicolelis (2017) has also attempted to use BCI to communicate the user's desired gait cycle kinematics. Unfortunately, both groups have reported numerous challenges such as difficulty in identifying the user intention of each joint, removal of mechanical artifacts caused by relative movement of the EEG cap, and physiological artifacts due to muscle activity in the vicinity of the cap. Even with extensive signal processing, they were not able to completely isolate the relevant signals at all time periods. López-Larraz et al. (2016) have concluded that the current state of the art of non-invasive BCI knowledge is insufficient for precise decoding of neutrally intended leg kinematics.

Similar to our intention of redirecting controllable neuromuscular activities to define real-time novel gait patterns, Durandau et al. (2019) have explored the use of residual muscle force activity by using detecting EMG signals. These investigators explain that no other exoskeleton has the ability to amplify weak user muscle strength. Another group (Ferris and Lewis, 2009) have considered the use of proportional EMG signals to activate pneumatic muscles that power a lower extremity exoskeleton. Yet another group has recognized the contralateral synchronization of the arms and legs in unimpaired walking and has used the shoulder angles to define unique sets of hip, knee, and ankle angles that are provide to the exoskeleton (Fang et al., 2017).

Similar to those efforts, the long-term objective of our work is to allow users to intuitively express their desired gait kinematics and dynamics using their arms and hands as sensory and motor alternatives to their legs and feet. Our prior lower extremity research (Karunakaran et al., 2014, 2017) evaluated the feedback conditions required by the hand to produce gait kinematics. That study included 18 subjects controlling virtual feet using hand movements to produce gait trajectories in a virtual environment. Our results indicated that users, provided with haptic through a physical link, and visual feedback (both sensations felt by the hands and visual observation), produced hand and virtual foot trajectories similar to biological gait trajectories (Karunakaran et al., 2014, 2017). We showed that for this to be a viable exoskeleton control method, the hands must be haptically connected either contralaterally or ipsilaterally to the feet. This ensures that the hands and feet move in precise time synchrony, and that hands sense the movement of the feet in order to provide the central nervous system with both position and force feedback from the feet (Karunakaran et al., 2014, 2017). This study as well as our other prior work has shown that neural control of arm and finger walking-like movements generates kinematics very similar to biologically intact neurally determined leg movements and foot placement (DeMarco and Foulds, 2002; Birmingham et al., 2003; Karunakaran et al., 2014, 2017).

The goal of this work was to develop and user test an intuitive control mechanism that independently controls both legs, while producing symmetrical gait kinematics using trajectories generated by hand movements in real time. Balance and co-ordination are being addressed in the later phase of the work, and will be presented in a subsequent paper.

We have chosen to allow the user to control movement of each foot with movements of the ipsilateral hand and arm. This may appear counter intuitive since the arms swing contralaterally with respect to leg movements during normal walking. However, we believe that ipsilateral control presents little if any physiological impediment to a successful user/exoskeleton interface, while its advantages are significant.

Although contralateral arm swing is commonly observed, its function in human gait is not entirely understood. Recent studies (Meesen et al., 2006; Meyns et al., 2013) agree there may be an enhancement in balance and stability, however, they also note that walkers can carry objects and make purposeful arm movements without compromising balance.

Contralateral arm movement is the preferred pattern at walking speeds above 0.8 m/s, yet at speeds below 0.8 m/s, individuals will often adopt an ipsilateral pattern or not swing the arms at all (Ford et al., 2007). In studies of walking with constrained or impaired arm movements (Meesen et al., 2006; Ford et al., 2007) there were only minor reductions in walking speed, that could be voluntarily corrected by participants. Studies of ladder climbing (Armstrong et al., 2009) show a preference for ipsilateral hand and arm coordination. Most importantly, in Meesen et al. (2006) subjects were asked to make four types of simultaneous arm/leg movements: ipsilateral same direction movements (e.g., right arm and right leg raised up and down, or adducted/abducted), opposite direction ipsilateral movements (e.g., right arm up while right leg down, or right arm adducted and right leg abducted), contralateral same direction movements and contralateral opposite direction movements. When examining the quality to interlimb coordination, the investigators found that the ipsilateral, same direction movements were modestly more accurate in absolute position/angle and phase than both contralateral conditions (Meesen et al., 2006). Thus, we are confident that we could employ either ipsilateral or contralateral arm/hand control method.

For our purposes, ipsilateral control of foot movement has several significant advantages. Admittance control requires the user to apply forces to the foot of the exoskeleton. This is most easily accomplished via a rigid trekking pole on the same side as the foot. A pole on the same side of the body facilitates directing the foot to move vertically and horizontally in the sagittal plane as well as controlling ab/adduction in coronal plane. Also, the poles allow the user's to feel the impact of the ground reaction through the ipsilateral hands. This quality of sensory feedback is unavailable in any proposed exoskeleton control method, and has been shown in our earlier work (Karunakaran et al., 2017) to be of vital importance to controlling the movement of the feet.

Lower extremity exoskeletons show potential for restoring ambulation in individuals with paraplegia due to spinal cord injury. Currently, many lower extremity exoskeletons produce a preprogrammed gait that can be initiated by the user, but not completely controlled by the user. Thus, users have limited control over their gait (i.e., step length, width, and speed), and little feedback of foot placement and ground contact. In this paper, we proposed a novel control mechanism for lower extremity exoskeletons to address these shortcomings; allowing the users to control their gait in real-time, as well as receiving haptic feedback. The accuracy and efficacy of the control mechanism was evaluated using the following outcome measures: accuracy of the foot position and joint angles; time delay between user-desired and actual robot kinematics; temporal, spatial, and step height symmetry; duty cycle of stance and swing phases; and similarity of biped hip and knee angles to those of human gait.

Our results demonstrate the effectiveness of the control mechanism, which allows the user to interactively control both legs of a ½ scale robot to produce a gait trajectory closely resembling human gait. As required for smooth real-time human-robot interaction (Van Der Linde et al., 2002), the time delay of the control mechanism is below 10 ms; ensuring smooth and stable movement. The difference between user-desired and actual robot Cartesian position and joint angles were <1 cm and 1°, respectively; showing that real-time accurate trajectories can be obtained over time by our system.

Admittance control provides the user with the following advantages: one-to-one correspondence between hand and foot movement; force amplification; intuitive control (as the user applies desired force in the intended direction of movement); and accurate time-varying trajectory control (the leg moves in the direction of the force) (Glosser and Newman, 2002; Van Der Linde et al., 2002). Research has shown that admittance control is an effective strategy for human-robot interaction, and that it requires a minimum of 100 Hz for acceptable human interaction (Glosser and Newman, 2002; Van Der Linde et al., 2002); which we have achieved.

Human gait is a rhythmic movement that is symmetrical but out-of-phase between the two legs. Hence, an effective exoskeleton control strategy should be able to reproduce the same pattern of movement over a period of time, as well as coordinate inter-limb movement (Vaughan et al., 1999; Pearson, 2003). Inter-limb coordination or symmetry is the ability to maintain temporal and spatial symmetry, which results in healthy gait (Patterson et al., 2008). Temporal and spatial symmetry close to “1,” signifies that both limbs are performing a symmetrical movement (Patterson et al., 2008). Our analysis shows that the participants controlling the biped were able to maintain inter-limb temporal, spatial, and step height symmetry close to 1 across all trials; irrespective of speed variations, and similar to that observed in human gait. Studies have shown that there is a correlation between large deviations in temporal & spatial symmetry away from 1 and falls & reduced walking speed (Balasubramanian et al., 2007; Patterson et al., 2008). Thus, we can conclude that our biped control mechanism will promote stable and safe human-exoskeleton walking. Our robot gait cycle has an average swing phase of just over 40% and a stance phase of just under 60%, as shown in Figure 7, across all trials. The deviation from the normal human gait cycle is small and likely due to the biped performing a flat-footed gait, as the foot is constrained to remain parallel to the ground, resulting in a shorter stance period. In addition, the percentage of time spent in stance and swing phase was the same between both legs, further indicating consistent temporal characteristics.

The results show that the joint angles produced by the biped (Figures 5D–F) are within the range of healthy biological gait in the sagittal plane (Winter, 2009), and that the joint angles of the reference participant and the biped are similar.

The relative horizontal and vertical excursions of biped walking were consistent with our earlier work on the benefits of haptic feedback to the hands for controlling the position of virtual feet (Karunakaran et al., 2017). More importantly, they have the same appearance as similar measures captured from unimpaired biological gait. Figure 6D shows that the right foot data from the biped (plotted as X and Y positions) bears a strong resemblance to unimpaired human foot data (Meesen et al., 2006). The user feedback questionnaire indicated that biped control was easy, intuitive, and required only minimal force.

Current commercial exoskeletons are generally slow, with small step lengths and step heights (Kozlowski et al., 2015). They usually rely on control mechanisms that initiate each step separately, which results in a delay between steps (Dellon and Matsuoka, 2007; Strickland, 2012)^1,2 and hence a prolonged stance phase (Esquenazi et al., 2012)¹. Our control method, on the other hand, provides continuous gait (smooth transition from swing to stance) with no delay between steps. It also provides the user with the ability to vary step length and height, based on the required speed. Thus, it produces a more natural gait in terms of temporal & spatial characteristics.

This preliminary work shows the feasibility of our algorithm to control exoskeleton kinematics and produce symmetrical gait patterns, using hands as a controller. However, there are several limitations that will need to be addressed in future work. This paper does not address balance, which will be addressed in a separate future publication. We also acknowledge that the benefits of our control mechanism must be thoroughly evaluated with participants walking independently in a full-scale exoskeleton. A wearable exoskeleton is under development for this purpose (Al Rashdan, 2016; Androwis et al., 2017).

While our work has focused on a specific SCI population of users, we are excited about expansion of the technique to individuals with other disabling conditions, such as people with diplegic cerebral palsy who have sufficient arm control to operate our trekking poles. We also see potential for the application of variations of our control mechanism in individuals with stroke and traumatic brain injury (both for assistance with ambulation, as well as for gait therapy), and multiple sclerosis.

We have developed and tested an admittance-control-based user-robot control strategy that allows the user to control foot trajectories with hand-generated forces and hand-sensed foot kinematics, and thus gait in real-time. This new approach has the potential to be used with a wearable rehabilitation exoskeleton, as a control mechanism that provides users with lower leg disability complete control over gait. When implemented with wearable exoskeletons, our method has the potential to greatly improve community ambulation in individuals with lower extremity paralysis.

The datasets presented in this article are not available because New Jersey Institute of Technology Review Board approval does not include data sharing with external institutions.

The studies involving human participants were reviewed and approved by New Jersey Institute of Technology. The patients/participants provided their written informed consent to participate in this study.

KK designed the algorithm, evaluated the design, analyzed the data, and drafted and finalized the manuscript. KA assisted with the design. GA reviewed the manuscript. RF assisted with the design, revised, reviewed, and finalized the manuscript. All authors contributed to the article and approved the submitted version.

The research for this study was conducted when the authors were employed by NJIT. Currently, KA is now employed by the company Stryker Corporation, and RF consults at Really Useful Robots, LLC, in addition to being Professor Emeritus at NJIT. The companies had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.