Development of a Robotic Companion to Provide Haptic Force Interaction for Overground Gait Rehabilitation

The aim of gait rehabilitation is to achieve independent ambulation. Somatosensory augmentation with external haptic sources can improve the subject’s ability to walk or stand. This paper presents the development and evaluation of a robotic system prototype that delivers haptic forces to aid overground gait rehabilitation. This portable system is based on a compact, mobile robot that is equipped with force and LIDAR sensors. The robot is flexibly linked to the user, which allows the force interaction between the user and machine to be halted when desired. During operation, the system can dynamically transition between modes in which force is applied or distance is maintained to emulate the experience of a human walking a dog on a leash. The haptic feedback from our system was evaluated in a pilot study that involved six young, healthy subjects and one individual recovering from a hemiparetic stroke. The study comprised independent and device-assisted walking trials. When using the device, the subjects walked continuously as it transitioned between distance and force modes. Gait speed and step length increased when force was applied and decreased as the force was removed. The improvements exhibited by an individual suffering from stroke were similar to those exhibited by healthy subjects. The application of haptic forces has a high potential for improving the efficiency of overground gait training with simple interactions.


I. INTRODUCTION
Individuals suffering from stroke typically regain some locomotor function but continue to exhibit significant gait deficits [1], [2]. Therefore, stroke survivors experience a loss of independence in terms of daily activities and restricted community ambulation, which limits their ability to engage in work, leisure, and social activities. Regaining the ability to walk independently is an important factor in the reintegration of stroke survivors [1], [3], [4]. Gait rehabilitation protocols are used to minimize the risks involved in increasing a patient's mobility; these involve exercises that The associate editor coordinating the review of this manuscript and approving it for publication was Jenny Mahoney.
instigate functional gait adaptations [5], [6] by subjecting the patient to varying physical demands. Gait adaptations that occur during rehabilitation extend into unassisted walking behaviors [7], [8]. The continuation of these gait adaptations is believed to be caused by the retention of adapted motor patterns by the central nervous system (CNS) after training has ceased [9].
Following a stroke, overground gait adaptation training is complemented by assistive devices and other therapies [10]. Individuals with a diminished sense of balance and weak sensorimotor systems encounter problems related to gait failure, which include an increased chance of falls, and a loss of independence. To overcome these problems, individuals must improve their sense of balance and gait ability through the use of assistive devices [11], [12]. However, the fluidity of an individual's motion is restricted when common walking aids are used. Subjects have been observed to bend forward, put a considerable load on the assistive device, or reduce their gait speed [13]- [18]. Therefore, a number of researchers do not recommend the use of such devices for gait training [13]- [15].
In general, rehabilitation devices use multisensory stimuli [19]- [21] for task-oriented neuromotor rehabilitation following trauma [22]. Sensorimotor coupling is useful for rehabilitation as sensory inputs can be coupled with motor adaptation to enhance the spatiotemporal organization of movements [23]. Evidence shows that light, non-supportive contact with a fixed object can limit postural deviations during locomotion [24]- [27]. Furthermore, Afzal et al. observed that haptic cues delivered by a wearable device can emulate such contact, and effectively reduce the body sway of healthy subjects performing tasks while standing [28].
Specially trained rehabilitation dogs can be used as an alternative source of haptic cues for patients with chronic stroke symptoms. This concept originated from clinical observations [29] as subjects walking with trained service dogs exhibited better fluidity of gait and took longer steps than those supported by conventional walking aids. The use of rehabilitation dogs was evaluated in a pilot study of four individuals with stroke. All subjects exhibited an increased gait speed and improved gait pattern after three weeks of training [29]. To extend this concept, a recent study has measured the effects of haptic inputs delivered in the form of forward-leading tensile forces applied to the hand while patients were walking on a self-paced treadmill. It was found that the application of a leading force increased instantaneous gait velocity and step length [22]. Providing haptic inputs to a walking subject may alter trunk and whole-body motion and spatiotemporal gait parameters related to stability, muscle activity, and reflex modulation [30]- [32]. If haptic inputs have a positive effect on walking performance, they may also be effective tools for locomotor rehabilitation [33]. However, the kinematic and temporal gait parameters for a standard treadmill differ significantly from overground gait parameters [34]; and a standard treadmill does not provide the same voluntary walking experience. More complex adaptive treadmill systems can emulate a voluntary gait experience [7] and simulate an environment that is close to real-life walking when combined with immersive technologies [35]. Furthermore, users with limited balance function can be protected using suspended safety harnesses. However, overground gait training delivers a more realistic walking experience for patients with better balance function as obstacles and turns can be more easily introduced to stimulate multiple senses. Recovering stroke patients consume less metabolic energy [36] and improve their gait endurance [37] when walking on real ground rather than on a treadmill. Therefore, individuals with stroke that are capable of unassisted walking can benefit from the use of a system that enables overground gait training. Furthermore, a portable overground gait rehabilitation system is more accessible and therefore, may increase participation in long-term rehabilitation activities. Pyo et al. reported that the gait of a stroke survivor improved when the patient was walking with an instrumented cane that provided haptic cues [38]. In this context, haptic cues may be a useful therapy for intuitive interaction during overground gait rehabilitation; their suitability for post-stroke ambulation recovery can be explored further.
Inspired by these previous works, in this study, we present the design and evaluation of a companion robot that can provide haptic interaction in the form of a tensile force applied to the user's hand, for the purpose of overground gait training. The presented companion robot is based on the hardware layout developed by Pyo et al. [38], which was modified to simulate the haptic interaction between a human and dog walking together. The robot is equipped with force and LIDAR sensors that are used to determine haptic interaction forces and the distance between the robot and the user. The companion interacts with the user via a flexible fixed-length leash that is held by the user. The system can operate in a force mode in which the leash applies a force in the direction of motion, or in a distance mode in which a fixed separation is maintained between the user and companion without applying any substantial forces to the leash. The robotic companion is designed to combine the benefits of a robotic system and animal-assisted training while overcoming the limitations of existing works. Animal-assisted training is sometimes complicated by the unpredictable behavior of dogs as they are affected by environmental stimuli such as scents, sounds, and the presence of other animals or people. Some existing mechanical solutions require the use of a treadmill which may not be available or provide a suitably realistic experience. The performance of our system was evaluated by investigating the experience of healthy users and an individual recovering from stroke when haptic forces are delivered through the leash and following the removal of these forces. The design of the system is discussed in Section II, and the details of the pilot study are presented in Section III. The experimental results are provided and assessed in Section IV, and our conclusions are drawn in Section V.

A. THE PORTABLE HAPTIC FORCE SYSTEM CONCEPT
The system presented in this paper has been designed for use by individuals who are in the chronic stage of recovery after stroke and are capable of walking without assistance. Figure 1 shows the two simplified interaction modes between a human and dog when walking overground. The dog may walk at pace with the human so that there are no interaction forces between them. In this case, the dog maintains its distance from the human by matching its velocity with that of its human companion. Alternatively, the dog may try to increase its speed and pull on the leash so that the human experiences a tensile force acting on their hand. Although the leash is always in the human's hand, a force is only applied in the second case.   We propose the use of a robotic companion that has a flexible haptic interface with the user and simulates these two states. Figure 1(b) shows the concept of our prototype robotic companion system. Walking without force interaction is simulated by the distance mode (as shown in Figure 2(a)), and walking with a constant force interaction is simulated by the force mode (as shown in Figure 2(b)). The target of this system is to deliver a force interaction with a controlled magnitude and duration at a specified time. The distance and force modes were implemented so that the system could controllably transition between the two modes.
In order for the robotic companion to operate in the two modes mentioned above, it requires a means of measuring the horizontal distance between itself and the user and a means to establish a flexible linkage with the user at hand level such that the force applied through it can be measured. In a previous work [38] we had developed a wheeled haptic cane. In order to provide removable haptic input at the hand during walking, we built upon this experience and designed a free-standing mobile robotic platform based on the haptic cane hardware layout. As shown in Figure 1(b), the hardware layout was modified by adding a caster wheel at the back so that there is ample support in the fore-aft direction to handle the forces generated by the flexible linkage. The caster wheel also allows passive turning. A special mounting assembly for the LIDAR sensor was also added that allowed for the easy placement of the sensor on the right or left side of the robot to ensure ease of adaptability for different users. The flexible linkage was attached in such a way that the force sensor built into the structure can be used to determine the horizontal force carried by the linkage. This particular design was developed because it allows for the easy mounting of the LIDAR sensor to measure the distance between the robot and the user's center of mass and also allows tensile force to be applied at the height of the user's hand. Although the system has a narrow wheelbase, all of the heavy components such as batteries, motor, etc. are placed very close to the ground resulting in a very low center of gravity (COG). The low COG and the added caster wheel ensure sufficient amount of stability according to the system's target ground conditions, i.e. flat, hard surfaces such as tiled floors, concrete pavement, tarmac, etc. Furthermore, a completely new software was developed to control the companion robot that implemented distance and constant force running modes and incorporated special algorithms for safe transitions between the modes. The details are given in the following parts of this section.
The human-robot interaction must transition between continuous and constant force, and no force. This may be achieved using a rigid linkage between the human and robot, but it is challenging to produce a zero-force sensation with such an arrangement. Therefore, the system presented here uses a flexible linkage similar to the leash used to walk a dog. A detailed description of the system hardware is given in Section II-D.

B. DESIGN OF THE CONSTANT-FORCE MODE
The low-level controller shown in Figure 3 controls the system speed while operating in both system modes. The low-level motor controller determines the system velocity through proportional-integral (PI) control. In the block diagram, v d is the desired command input, v c is the system speed feedback obtained from the motor encoder, ω d is the desired wheel angular velocity, T m is the torque input from the PI controller, θ m is the actual angular velocity, T D is the torque disturbance, and k e is the gain used to convert the encoder rotation angle into system velocity. As shown in Figure 2(b), the system works to equalize the interaction and target forces (F interaction and F desired , respectively) when operating in the force mode by generating the v c that delivers a constant force to the user. Figure 4 shows  a block diagram of the proportional-integral-derivative (PID) force mode controller that is used to apply F desired to the user independent of gait speed. The target of this closed-loop controller is to minimize the interaction force error (F desired -F interaction ). The system velocity generated from the force error delivers a tensile force to the user (see Figure 2(b)). The F interaction between the user and system is measured using a force sensor. When the system interacts with the user, F interaction is the difference between the force generated by the system (F system ) and the force applied by the user (F user ). If the user applies an arbitrary force to the leash, the PID controller compensates for the change in load by increasing or decreasing the system's target velocity to maintain F desired . Therefore, the system controls the relevant forces without measuring the user or system velocity. The saturation block (shown with a dotted line) shown in Figure 4 is only applied when the system is operated in the distance mode, and is excluded in the transition to force mode. This will be discussed further in Section II-F.

C. DESIGN OF THE DISTANCE MODE
The distance mode was designed to use a robust control scheme. In this mode, the distance between the robot and the user is maintained with no interaction forces between them; i.e., the robot should reach the desired distance from the user, x d , and then match the user's velocity, v w , as shown in Figure 2(a). The implemented control scheme is similar to that of a self-paced treadmill in which the velocity of the belt is matched to the user's velocity [39]. The user's initial gait speed was estimated with a robust integral of the sign of the error (RISE) controller and a low-gain observer. The RISE controller is a feedback control system; thus, the positional error was reduced by estimation of the time-varying uncertainty of the closed-loop system [40]. The low-gain observer was used as the feedforward term to estimate the user's gait speed at low acceleration. Through implementation of this method, the system can be made more comfortable by smoothing the rapid changes in velocity.
According to Figure 2(b), the user-system interaction state space model can be represented as follows [41]: where, x e is the user position error, v w represents the intentional velocity of the user, and v c and a c represent system velocity and acceleration, respectively. According to equation (1), v c is the direct current (DC) motor driver control input from the low-level controller. However, as described in Ref. [41], equation (1) can be modified with equation (2) such that a c is used as the input to prevent a large acceleration during system movement. The control command v c can be obtained by integrating a c . The controller's objective is to reduce x e to zero. Here, a positional error variable and its dynamics must be derived.
The desired controller output is calculated as follows: where v d c is the desired velocity command, andv w is a feedforward term that is an estimate of v w that can be obtained from the linear observer [41]. The robust term µ is a continuous input, which includes the RISE controller, defined in Ref. [40] as follows: (4) where k s , α 1 , α 2 and β are adjustable positive control gains, and 'sgn' represents the standard sign function. The RISE controller learns the slowly time-varying uncertainties of the close-loop dynamic system, which compensates for the estimation error of the feedforward controller.
The error signal is defined as q = v c − v d c . We assume that the low-level control can accurately follow target commands. Therefore, the dynamic model can be written as follows: With a c as its input, a controller for (5) can be described by This controller includes the terms v w andv w , which cannot be directly measured. However, the velocity is the system input command, so it can be obtained from v c (t) = t 0 a c t dt. Thus, the final control command is approximated as follows:

D. SYSTEM HARDWARE
The haptic companion robot shown in Figure 1(b) was controlled by an onboard controller and a motor driver was used to control the DC motor that powered the wheels; the system was supplied by a battery-powered switched-mode power supply (SMPS). A LIDAR sensor (TF Mini) was used to measure the distance between the robot and the user, and a load cell was used to measure the interaction force. The trailing caster wheel provided stability and enabled the robot to be turned freely in any direction. Detailed hardware specifications are provided in Table 1. High-level control was executed using a laptop computer, which was also used for data logging. The laptop and mobile robot communicated via a WiFi (TCP/IP) link with a 1-kHz data rate (see Figure 6). For safety reasons, the software interface included an emergency stop button for use by the operator.

E. OPERATIONAL VERIFICATION OF THE DISTANCE AND FORCE MODES
Operational verification tests of the system were carried out outdoors with a single healthy user. First, the force mode was tested by applying a force between 2.5 and 10 N while, from a standing start, the user walked for 30 s. As shown in Figure 7, the maximum force fluctuation was not more than 2.5 N, which shows that the system can apply force in 2.5 N increments.  The distance mode was tested by asking the user to walk with the system at a slow, normal, and fast pace. As shown in Figure 8, the average velocities for these three trials were 0.63, 0.81, and 1.33 m/s, respectively. The target distance for this test was set as 70 cm. The root mean square (RMS) errors of the distance between the device and user were 6.68, 5.76, and 5.77 cm, respectively, and the maximum error was 18.07 cm. The minimum safe distance between the user's COM and the device is 40 cm. Thus, the maximum acceptable amount of error in distance maintenance is 70 -40 = 30 cm. Hence, these results demonstrate that the distance error was acceptably small and did not show any relationship with walking speed. Thus, the system was shown to operate with acceptable accuracy in both force and distance modes. In the force mode, up to 10 N could be applied to the hand-held leash in the direction of motion. In the distance mode, no force was applied through the leash, but the specified distance was maintained with a maximum user velocity of 1.5 m/s.

F. MODE TRANSITION ALGORITHMS
The system was programmed so that it was possible to apply a force for a programmed duration at any time during operation. This functionality required that the system could make automatic transitions between the distance and force modes. Safe operation required that these transitions did not pose any risk of an abrupt halt of the robot, or collision with the user. Furthermore, these transitions needed to be gradual so that the user did not feel compelled to make sudden gait adjustments as the system switched modes. Figures 9(a) and 9(b) illustrate the mode transition concepts that were implemented. The transitions were performed smoothly and safely using custom-designed algorithms, as described in Figure 9(c).
To transition from the distance to force mode, the robot must accelerate to reduce the slack in the leash, and then apply a leading force. The system is driven by the RISE and PID controllers when in the distance and force modes, respectively. Both controllers are kept in operation at all times to ensure smooth operation of the robot. In the distance mode, to ensure that only the RISE controller controls the system, the PID controller output is nullified by feeding the current system velocity back as the output saturation value (see Figure 9(c)). This output saturation is removed when a transition to the PID controller is initiated, which allows the PID output to increase and accelerate the robot. As the system accelerates, the distance between the user and robot increases to above the reference position value, which drives the RISE controller output to zero so that the PID controller output is the only non-zero motor control signal. This process is described pictorially in Figure 9(c).
During the transition from the force to distance mode, as shown in Figure 9(b), the device must decelerate to slacken the leash and remove the interaction force. During this transition, the user-robot separation is greater than that required in the distance mode, x ID (set initially as the desired distance). Therefore, the robot stops abruptly to wait for the user to reach the reference position. Because the system stops, the user may also have to stop and wait for the robot to move again. The device may also become unstable due to the large x e and user intentional velocity h w .
The system makes this transition more gradual by measuring the average distance from the user during the first half of the last second of the force mode; this is the device's measured distance value, x m . During the following 5 s (i.e., the first 5 s of the distance mode), the device linearly reduces the distance between itself and the user from the measured distance to the initial distance (x m → x ID ). The initial distance is the separation that was maintained between the device and user during the previous distance mode operation. This gradual reduction in distance removes the tension from the leash and ensures continuous motion of the device. The transition process is described in Figure 9(c), and the effects of these transition algorithms are shown in Figures 11 and 12 and discussed further in Section IV.

III. EXPERIMENT
A pilot study was carried out in which the gait of six healthy, young subjects, and one individual who was recovering from a hemiparetic stroke was studied while they used our system. The primary focus of the study was to observe the VOLUME 8, 2020 changes in user gait velocity when a haptic force was applied through the leash and then removed. Furthermore, we aimed to determine whether the introduction of a force changed user gait velocity, and how much time the user would require to return to their pre-force velocity. Our prototype device and the system developed by Sorrento et al. [22] operate on the same principles, so the experimental protocol used in Ref. [22] was adapted to include overground walking instead of treadmill walking, and different timings and force magnitudes. The six young, healthy participants (gender: male, mean age: 28.2 ± 4.7 years, mean height: 172.7 ± 3.0 cm, mean weight: 68.6 ± 16.6 kg) had no prior history of musculoskeletal or neurological injuries. The system is designed for people capable of unassisted locomotion, so the participant who was recovering from a hemiparetic stroke was selected accordingly (gender: female, age: 60 years, height: 148 cm, weight: 48 kg, right-sided hemiplegia, 22 days since onset, cause of stroke: infarction, Brunnstrom stages of stroke recovery: 4/4/5, modified Barthel index score: 44, mini-mental state exam (MMSE) score: above 24). All participants gave written informed consent prior to participation in the study.

A. PROTOCOL
All subjects followed the same walking trial protocol. First, the participants were introduced to the system and its functionality. The testing protocol was explained verbally, and the users were instructed to release the system if they felt unstable or uncomfortable. The users then underwent familiarity training which included walking with the system in both distance and force modes, and the transitions between them.
Once familiar with the device, each participant was subjected to two iterations of normal-walking and with-devicewalking trials. In the normal-walking trial, the user walked in a straight line at their preferred gait speed for 14 m; each user's walk during the middle 10 m was timed using a stopwatch, and observations made during this period were used for data analysis. As shown in Figure 10, the with-device-walking trial required the subject to walk with the system for 70 s. During the first 15 s of the trial, the system led the user at their desired speed (distance mode), and the leash remained slack. Then the system automatically transitioned (distance to force mode transition) and for the next 25 s (including transition time), the system applied a force to the user's hand (force mode). At 40 s from the start of the trial, the system automatically transitioned (force to distance mode transition) and during the last 30 s (including transition time) it again led the user at their preferred speed (distance mode), such that the leash was slack. A 5-N haptic force was maintained throughout the force mode during all trials. The healthy subjects completed the with-device-walking trials in a 100-m-long hallway, whereas the individual with stroke performed the test in a 60-m-long hallway. All subjects were allowed to sit for 5 min between trials.

B. DATA COLLECTION AND ANALYSIS
The device velocity was calculated in real-time using the output of the optical encoder (which was used for low-level device control), and data were stored on the laptop for postexperimental analysis.
The interaction force between the user and the device was measured using the installed force sensor, and the haptic force felt by the user in the direction of motion is the horizontal component of this measured force. The time-dependent interaction force was also logged using the laptop.
During the with-device-walking trials, an experimental supervisor walked two paces behind the subject and placed markers on the ground 15, 40, and 70 s after the start of the trial to mark the distance travelled by the subject. After each trial, these distances were recorded. The distance and time information were used to calculate the gait speed of the subjects during each of the periods that corresponded to the three with-device-walking conditions (pre-force distance mode, force mode, and post-force distance mode). Data from the first 2 m of each trial were not considered during analysis.
All subjects wore shoes with instrumented insoles (Moticon, ReGo AG) that were used to measure the number of steps taken during each trial. This value and the distance travelled in each phase were used to estimate the step length of the subject for each device control mode.
A one-way repeated-measures analysis of variance (RMANOVA) was used to identify the effect of different walking modes on the gait speed and step length of the young, healthy subjects. Furthermore, Mauchly's test of Sphericity was used to confirm the validity of the RMANOVA results and post-hoc tests were carried out. All statistical analyses were performed using SPSS V20.0 (IBM Corp., Armonk, NY, USA).

A. HEALTHY SUBJECTS
The system velocity and interaction force during the withdevice-walking trial of a single healthy subject are shown in Figures 11 and 12.
As shown in Figure 11, the user comfortably reached their preferred gait speed during the first 15 s of the trial (pre-force distance mode) and maintained it while walking with the device. During the following 25 s (force mode), the user increased their gait speed when the haptic force was applied through the leash and maintained this greater speed.  The system decreased its velocity and transitioned from the force mode to the distance mode during the 5-s time period starting at 40 s into the trial. At this time, the device slackened the leash and reduced the haptic force by reducing its speed. This transition was carried out according to the algorithm described in Section II-F such that there were no abrupt changes in speed so severe that they may cause the user to trip over or collide with the system. Although at the start of the force to distance mode transition there is a drop in speed, it is very quickly recovered, and during trials none of the participants felt that this drop was a cause of disturbance for them. Upon removal of the force, the user's gait speed started to decrease slowly. At the end of the 70-s trial period, the user was nearing their preferred gait speed.
The interaction force experienced by a user during the with-device-walking trial is shown in Figure 12, which illustrates that the user walked with the system without any interaction force during the distance-mode periods, and only experienced the set 5-N interaction force during the force-mode phase.
The mean and standard deviation (SD) of the healthy subjects' velocities during the normal-walking and with-devicewalking trials are shown in Figure 13. The users' gait speeds during the normal-walking and pre-force distance modes were not significantly different, which indicates that the system did not hinder the user when operating in distance mode. The application of the haptic force delivered by the leash increased the mean gait speed, and a corresponding decrease in mean gait speed was observed after the force was removed.
One-way RMANOVA revealed that the differences in gait speeds of the young, healthy subjects among the different walking modes were statistically significant. (F (3, 15) = 22.101, P-value < 0.001). Post hoc tests showed that the gait speed was significantly greater during the force (P-value = 0.002) and post-force modes (P-value = 0.027) compared to the subjects' normal walking speeds. Likewise, the preforce distance-mode gait speed was significantly different from the force (P-value < 0.001) and post-force distancemode speeds (P-value = 0.001). The young, healthy subjects exhibited a statistically significant decrease in gait speed during the post-force distance mode compared to the force mode (P-value = 0.002).
The changes in gait speed observed during the with-devicewalking trials agree with the findings of Sorrento et al. [22]. Furthermore, we verified that the use of our system in pre-force distance mode does not interfere with the user's preferred overground gait speed. The maximum difference in gait speeds between normal walking and the pre-force distance mode was 0.18 m/s, and the minimum was -0.01 m/s. This signifies that the current system is an appropriate source of on-demand force interaction. The step lengths estimated during the normal-walking and with-device-walking trials are shown in Figure 14. The increase in step length following the application of a haptic VOLUME 8, 2020 force can be attributed to the immediate increase in user gait speed. Similarly, the decrease in step length following the removal of force can be attributed to the decrease in gait speed. One-way RMANOVA revealed that the step length of the young healthy subjects differed significantly among different walking modes (F (3, 15) = 4.476, P-value = 0.020). Post-hoc tests revealed that step length increased significantly during the force-mode phase relative to normal-walking (P = 0.046). Similarly, the pre-force distance-mode step length was significantly different from those measured during the force (P-value < 0.009) and post-force distance modes (P-value = 0.039). Furthermore, young, healthy subjects exhibited a statistically significant decrease in step length in the post-force phase compared to that during the force mode (P-value = 0.043). The changes in step length observed during the with-device-walking trials are also consistent with the findings of Sorrento et al. [22]. The mean system velocity of all healthy subjects during the with-device-walking trials, averaged over 1-s intervals, is shown in Figure 15. An increase in the mean system velocity represents an increase in the gait speed of the user following the application of force. These data show that upon removal of the force, the users' gait speeds tended to revert back to the pre-force distance mode gait speed over an extended period. We hypothesized that users would require a similar amount of time to revert to their preferred gait speed after experiencing an increase in gait speed due to the application of force for 25 s. The time-dependent velocity time-series revealed that user gait speed decreased between 45 and 70 s ( Figure 15). This trend is in agreement with our hypothesis, but the time required to return to the pre-force gait speed may be greater than we expected. Further study is required to find a correlation between the duration of force application and the time required for a user to return to their preferred gait velocity.

B. INDIVIDUAL RECOVERING FROM A STROKE
Based on the observations made during the trials with healthy subjects, a pilot trial was carried out with an individual suffering from hemiparesis following a stroke to examine the effects of our prototype device. We used the same protocol, with the addition of a therapist walking a couple of paces behind the subject to ensure their safety. The gait speed and step length of the subject during normal-walking and with-device-walking trials exhibited trends similar to those observed with young, healthy subjects.  Figure 16 shows the gait speed of the individual recovering from a stroke during normal-walking and with-devicewalking trials. As observed with the healthy subjects, the gait speeds during normal-walking and the pre-force distance modes were not significantly different. Following the application of a 5-N haptic force during the force mode, an increase in gait speed was observed, and a decrease was seen upon removal of the force.  Figure 17 shows the subject's step length during the trials. The changes in step length may have resulted from the changes in gait speed associated with the introduction of the haptic force.
Changes in gait speed and step length indicate that the individual had adapted their gait to changes in system modes. Although only short-duration adaptations were tested in this work, these results indicate that long-term adaptations may occur through the implementation of extended training protocols.

V. CONCLUSION AND FUTURE WORK
We have presented the development of a companion robot system that provides haptic interaction, which may be a useful tool for overground gait rehabilitation of individuals with stroke who are in the chronic stage of recovery and are capable of walking stably without any assistance. This system emulates the interactions between a human and a dog on a leash that may be useful for facilitating gait recovery while eliminating the unpredictable behaviors associated with live animals. This system, unlike those in previous studies [22], does not require the use of a treadmill, which increases accessibility for outpatients who need to continue their overground rehabilitation. The trends in gait speed and step length observed during the trials involving healthy subjects are consistent with the observations of researchers that have explored the use of real and virtual companion animals for gait recovery [22], [29]. Furthermore, experimental observations demonstrated that the system presented here did not have a significant influence on the user's gait when no force was applied (distance mode). Overall, the findings presented here show that this system can be used as a robotic companion for gait rehabilitation.
The tilted-cane portable robot was designed to have a flexible linkage to the user, allowing for the total elimination of force interaction when desired. It can move at a wide range of speeds and generate a range of interaction forces, which makes it suitable for many users' needs. The patient found the system easy to use and exhibited gait speed and step length behaviors that were similar to those of healthy subjects.
The increase in gait speed observed after a haptic force was applied declined once the force was removed. This suggests that further study of the correlation between the force application duration and the time required for the gait speed to return to the user's preferred speed may be required. Furthermore, the effects of different haptic-force loads should be investigated in addition to the effects of long-term training as a means of instilling more permanent gait adaptations.
This study assessed the performance of six young, healthy subjects, and one individual recovering from a stroke. All participants demonstrated outcomes that were consistent with the observations of previous researchers. Further studies comprising comprehensive gait analysis, including kinematic and postural measures to evaluate the system's effects on individuals who have suffered a stroke need to be conducted to further explore the device's potential as a tool for gait rehabilitation.
Certain design changes are required to enable this system to be used in an outdoor setting. The system track width is narrow, which provides high maneuverability, but may make the system unstable on uneven surfaces such as those encountered outdoors. The system does not have powered direction control or obstacle detection and avoidance functions [42], which may aid outdoor operations. An update to the system design would improve the robustness of the system for outdoor use and may enable the implementation of outdoor training protocols that are closer to the users' real-life environments. From 2005 to 2017, he was a Professor with the School of Mechanical and Aerospace Engineering, Gyeongsang National University, Jinju, South Korea. In 2017, he joined the School of Integrated Technology, Gwangju Institute of Science and Technology, Gwangju, South Korea, where he is currently an Associate Professor. He is also a Technical Editor of the IEEE/ASME TRANSACTIONS ON MECHATRONICS and an Associate Editor of Frontiers in Robotics and AI. He has authored or coauthored more than 80 peer-reviewed journal articles and patents. His current research interests include bio-nano robot control, virtual reality haptic devices, and rehabilitation robots. VOLUME 8, 2020