<?xml-model href='http://www.tei-c.org/release/xml/tei/custom/schema/relaxng/tei_all.rng' schematypens='http://relaxng.org/ns/structure/1.0'?><TEI xmlns="http://www.tei-c.org/ns/1.0">
	<teiHeader>
		<fileDesc>
			<titleStmt><title level='a'>Virtual Reality-Based Robot-Assisted Method for Gait Training Showing Retention of Anticipatory Motor Responses</title></titleStmt>
			<publicationStmt>
				<publisher>IEEE</publisher>
				<date>09/01/2024</date>
			</publicationStmt>
			<sourceDesc>
				<bibl> 
					<idno type="par_id">10553297</idno>
					<idno type="doi">10.1109/BioRob60516.2024.10719832</idno>
					
					<author>Bradley Hobbs</author><author>Panagiotis Artemiadis</author>
				</bibl>
			</sourceDesc>
		</fileDesc>
		<profileDesc>
			<abstract><ab><![CDATA[The recent increase in attention toward robotic assistance in gait therapy has prompted a clear need for useful paradigms that can interface with such devices. This is especially true for post-stroke rehabilitation, for which impairments in walking are particularly debilitating and notoriously difficult to overcome. Until recently, robot-assisted gait training methods tended to target only a single aspect of the human sensorimotor system, usually proprioception. To address this limitation, a virtual reality (VR) system is united with the unique robotic rehabilitation platform, the Variable Stiffness Treadmill (VST), in order to understand the complex interactions between visual and proprioceptive feedback in gait. This work proposes a new type of intervention that directly results in significant anticipatory responses to stiffness perturbations, even when those anticipated perturbations may not occur, showing that using a VR system can lead to the retention of training provided. The results show relevant and repeatable responses, which can lay the foundation for more effective protocols that can be used for a wider array of robotic systems and sensory modalities.]]></ab></abstract>
		</profileDesc>
	</teiHeader>
	<text><body xmlns="http://www.tei-c.org/ns/1.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:xlink="http://www.w3.org/1999/xlink">
<div xmlns="http://www.tei-c.org/ns/1.0"><head>I. INTRODUCTION</head><p>The majority of the human brain is devoted to vision and movement <ref type="bibr">[1]</ref>. Following traumatic brain injuries such as stroke, brain lesions can interrupt these neural circuits <ref type="bibr">[2]</ref>, which is why many rehabilitation protocols target both modalities <ref type="bibr">[3]</ref>. Numerous studies have demonstrated the importance of combining visual and proprioceptive feedback in rehabilitation <ref type="bibr">[4]</ref>  <ref type="bibr">[5]</ref>, especially for lower limb gait therapy <ref type="bibr">[6]</ref>. This highlights a gap for standardized protocols that can effectively interface with these modalities <ref type="bibr">[7]</ref>.</p><p>Early visual systems in rehabilitation included subjects looking at screens with kinematic-based feedback <ref type="bibr">[8]</ref>. This led to the use of gaming devices, such as virtual reality (VR) headsets <ref type="bibr">[9]</ref>, which immerse the wearer in complex environments for therapy goals <ref type="bibr">[10]</ref>. VR has shown promise as a valuable tool to interface with the brain's visual system and combine with other physical modalities of treatment <ref type="bibr">[9]</ref>. Given that motor learning requires full attention and intent <ref type="bibr">[11]</ref>, VR systems play a unique role in robotic gait rehabilitation.</p><p>In stroke rehabilitation, particularly VR-based approaches, a lack of motivation is a predictor for poor treatment adherence <ref type="bibr">[12]</ref>. Since patient success is dose-dependent on protocol intensity, finding protocols that decrease intensity *This material is based upon work supported by the National Science Foundation under Grants No. #2020009, #2015786, #2025797, and #2018905, and the National Institutes of Health Grant No. 1R01HD111071-01.</p><p>Bradley Hobbs and Panagiotis Artemiadis are with the Mechanical Engineering Department, at the University of Delaware, Newark, DE 19716, USA. bwh@udel.edu, partem@udel.edu * Corrresponding author: partem@udel.edu while maintaining outcomes could increase adherence <ref type="bibr">[13]</ref>. Overground walking requires both reactive and anticipatory control <ref type="bibr">[14]</ref>, with anticipatory control being crucial <ref type="bibr">[15]</ref> and affected by aging <ref type="bibr">[16]</ref>  <ref type="bibr">[17]</ref>. Training for anticipation of visual perturbations has shown improvements in variables relevant to stroke therapy, such as step length, walking speed, and balance <ref type="bibr">[18]</ref> [19], since the brain controls motor functions in anticipation of ground changes, rather than reflexively responding <ref type="bibr">[20]</ref>.</p><p>Previous studies are limited by inconsistent visual feed- Fig. 2. Experiment design showing a sample of Rigid, VO, and VP gait cycle sections where gait cycle is denoted as GC. The number of VO, VP, and PO patches in this figure is not to scale. The left side of the VST experiences physical perturbations, while the right side remains rigid for this study. The first phase, lasting for 500 gait cycles, contains VP conditioned gait cycles every 7 &#177; 2 steps. The second phase lasting 300 GC contains VO conditioned gait cycles every 7 &#177; 2 steps.</p><p>back <ref type="bibr">[21]</ref> or walking speed <ref type="bibr">[18]</ref>, disturbances to proprioceptive feedback <ref type="bibr">[19] [22]</ref> and exclusion of physical perturbations <ref type="bibr">[19]</ref> or virtual environment <ref type="bibr">[23]</ref>. To our knowledge, no study has investigated anticipatory accommodation strategies with consistent gait speed and ground contact while controlling for visual and physical perturbations. Observing the ability to anticipate walking surface changes can improve VR-based protocols. Repetition drives rehabilitation <ref type="bibr">[24]</ref>, and current robotic methods in proprioceptive-based therapy allow for multiple ways of interfacing with the brain's control of movement. Instrumented robotic treadmill systems provide a way to implement gait training in a controlled environment <ref type="bibr">[25]</ref>. Errorinducing methods that force adaptation to force changes, such as perturbations, have advantages in requiring user focus and adaptation <ref type="bibr">[26]</ref>. Stiffness changes are useful because they mimic real-world ground changes. Previous studies have demonstrated this with stroke-affected subjects <ref type="bibr">[27]</ref><ref type="bibr">[28]</ref><ref type="bibr">[29]</ref>, but more research is needed to implement such modalities broadly, especially in robotic treadmill systems.</p><p>There is a need for visual feedback in motor learning and usefulness for proprioception-based protocols in robotassisted treadmill training. Preliminary research on combining the two modalities in perturbation-based training has shown promise in increasing muscle activity and gait parameters related to anticipation <ref type="bibr">[30]</ref>  <ref type="bibr">[31,</ref><ref type="bibr">32]</ref>, because it is well known that reduced muscle activity on the affected side of the body is associated with motor learning deficiencies, and anticipatory responses are an avenue for targeting these deficiencies <ref type="bibr">[33]</ref>. We aim to improve this by understanding how adaptations made during training can be retained, potentially reducing the need for constant physical perturbations, expanding the amount of eligible participants. We hypothesize that these anticipatory responses can be retained without the physical stimulus. In this study, subjects were convinced the next step would be on a compliant surface based on visual feedback alone after training with both visual and physical stimuli. These findings support and build upon previous research and allow for a new type of protocol that can increase retention, decrease training intensity, and be as effective as previous training modalities, offering the next natural step towards robot-assisted gait rehabilitation to a broader patient population.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>II. METHODS</head></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>A. Experimental Setup</head><p>An instrumented robotic split-belt variable-stiffness treadmill (VST) was used for the unique capabilities (shown in Fig. <ref type="figure">3</ref>). This device allows finely controlled changes in walking surface stiffness, influencing the subject's proprioception, through a lever-attached spring underneath the walking surface that is connected to a controlled fulcrum, achieves stiffness resolutions of less than 0.0001 kN/m <ref type="bibr">[34]</ref>. The effective stiffness can change from 0.1 kN/m to over 1,000 kN/m within an average human swing phase time <ref type="bibr">[34]</ref>.</p><p>Visual feedback was controlled using an Oculus Rift (Oculus Inc.) headset projecting a virtual reality (VR) en-  vironment. The custom-made environment in Unreal Engine 4 (UE4) provided a straight-line walking path resembling a hard, grassy surface, with the softer, sandy surface texture resembling a surface of lowered stiffness (see Fig. <ref type="figure">3</ref>). The wearer sees real-time leg motion through their avatar (see Fig. <ref type="figure">3</ref>, on the right).</p><p>Real-time subject motion is captured with a 3D motion capture system using standard reflective markers on the lower body, with eight cameras transmitting data to Vicon Nexus software (Vicon Motion Systems Ltd.). The marker positions are converted to joint angles and sent to UE4, moving the subject's avatar limbs at 200Hz. Delsys Trigno surface electromyography (EMG) electrodes (Delsys Inc.) captured bilateral muscle activity for biceps femoris, vastus medialis, tibialis anterior and gastrocnemius at 2000Hz, representing knee and ankle flexors and extensors, relevant for post-stroke gait analysis and therapy <ref type="bibr">[35]</ref>.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>B. Protocol</head><p>All subjects gave informed consent of the protocol approved by the University of Delaware Institutional Review Board (IRB ID: 1544521-2). Nine healthy subjects participated in this study (6 male, 3 female; age = 25 &#177; 0.7 years; body mass = 71 &#177; 4 kg; height = 169 &#177; 7 cm) and were absent of gait abnormalities and regular VR use. Subjects wore a safety harness (Litegait Inc.) and walked on the VST at 0.9m/s for 800 gait cycles (see Fig. <ref type="figure">1</ref>), divided into two phases (see Fig. <ref type="figure">2</ref>). The first phase (500 GC's) included Rigid (R) and visually and physically perturbed (VP) gait cycle conditions. Gait cycles during the Rigid condition show the hard, grassy path while walking on 1,000 kN/m surface stiffness, with VP condition showing the soft, sandy texture while walking on 60 kN/m stiffness <ref type="bibr">[36]</ref> on the left belt, returning to rigid before the next left heel-strike (see Fig. <ref type="figure">4</ref>). The VP condition appeared after 5-9 gait cycles with an average R:VP ratio of 7:1, with the aim being to learn association between visual and physical stimuli.</p><p>The second phase (300 GC's) included Rigid (R) and visually only perturbed (VO) conditions. The condition of physically only perturbed has been studied and is outside the scope of this experiment <ref type="bibr">[31,</ref><ref type="bibr">32]</ref>.</p><p>Gait cycles during Rigid condition match the first phase, with VO condition involving the soft, sandy texture while walking on 1,000 kN/m stiffness. The VO condition appeared after every 5-9 gait cycles with an average R:VO ratio of 7:1, assessing if subjects showed similar anticipatory responses as the first phase, but without physical stimuli. Intermittent VP condition gait cycles ensured subject confidence in the association, appearing after every 3-5 VO condition with an average VO ratio of 4:1, and are omitted from analysis.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>C. Data Processing</head><p>The Vicon Plug-in-Gait (PiG) model was used for marker placement, real-time heel-strike event detection <ref type="bibr">[37]</ref>, and extraction of hip, knee, and ankle joint angles and velocities, and hip-ankle span. Each gait cycle was normalized to start at the left heel strike (LHS) for kinematic and EMG data, with EMG data normalized as a percentage of the experiment maximum (EM) to prevent excess fatigue during maximum voluntary isometric contraction. Since this work is focused on anticipatory changes, all gait cycles for each individual type of data discussed hereafter are grouped into three classifications based on the condition that will follow that gait cycle.</p><p>Gait cycles are classified based on the next step's condition, excluding data from that next step. Outlier gait cycles were removed based on criteria from <ref type="bibr">[38]</ref>, and a 2-sample t-test determined if peak values of VP condition were significantly different from Rigid condition (&#945; = 0.05), with the same test conducted for VO and Rigid conditions. If both tests reached statistical significance, the anticipatory responses for the first phase VP, were not only statistically different from rigid gait cycles, but also sustained even in the VO phase, regardless of physical perturbation.</p><p>Figures denote statistical significance of peak values with an asterisk, and discussed significance range with a horizontal line segment. Because statistical significance testing is not useful for data with steep slope, only the peak values within the range were checked statistically, however it is highly probable that the region within the line segment is also significant to the discussion and impact of the results from this work, which are detailed below. Fig. <ref type="figure">5</ref>. Left hip flexion mean from 0% to 100% of the gait cycle, beginning at left heel-strike. The profile for VP (blue) and VO (orange) are compared against the rigid (grey) to show unified anticipation response immediately before the perturbation. Peak magnitude reaches statistical significance and is markedly changed after the peak. Fig. <ref type="figure">6</ref>. Right hip adduction mean from 0% to 100% of the gait cycle, beginning at left heel-strike. The profile for VP (blue) and VO (orange) are compared against the rigid (grey) to show unified anticipation response for over half of the gait cycle before the perturbation. Peak magnitude reaches statistical significance during the left swing phase, but is clearly changed throughout the gait cycle as a whole.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>III. RESULTS</head><p>The results of the current study show that the subjects were convinced they were going to step on a surface of lowered stiffness, and retained those anticipatory responses even when the physical perturbations were removed (VO steps) for the duration of the trial. The subjects did this in a variety of ways that not only corroborate with previous works <ref type="bibr">[32,</ref><ref type="bibr">35,</ref><ref type="bibr">39]</ref>, but also are important to post-stroke robot-assisted gait rehabilitation <ref type="bibr">[29]</ref>. To demonstrate the key findings clearly, all figures shown are a snapshot from one subject, representing the majority of changes across subjects, and because all specific patterns of muscle activity and kinematic data are unique to each subject.</p><p>Focusing on the kinematics, within the hip joint, the subject shows significant increases in left hip flexion (see Fig.  <ref type="figure">7</ref>. Left knee velocity mean from 0% to 100% of the gait cycle, beginning and ending at left heel-strike. The VP (blue) and VO (orange) profiles are compared against Rigid (grey) to show unified anticipation response in 60% to 90% of the gait cycle before perturbation, with the peak reaching statistical significance and is markedly changed after the peak. 5) and right hip adduction (see Fig. <ref type="figure">6</ref>) during the left swing phase, just before stepping on the compliant surface. This particular finding is important because it shows not only that subjects were preparing to experience the perturbation even at the hip joint level, but also that left hip flexion was seen, as in previous works <ref type="bibr">[32]</ref>. It is important to note that the left hip was observed to increase and project more anterior, further increasing the significance of the left hip flexion response; it would be expected that since the left hip joint was higher, further left hip flexion would be more difficult to reach with a higher hip joint. In addition, this increase in left hip position explains the increase in right hip adduction. Since the right leg is in the stance phase at this time, right hip adduction is needed to raise the left side of the hip. Upon further investigation, left hip abduction relative to the pelvis was found to not be statistically different, implying that subjects also increased activity of the left abductors, since the left hip was at a higher position, but the left femur-pelvis angle was unchanged. Since abductor muscle activity was not recorded, more research is needed to further elucidate this finding.</p><p>At the knee level, statistically significant findings can be seen in the left swing phase immediately before the perturbation in both the left knee velocity (see Fig. <ref type="figure">7</ref>) and left knee flexion (see Fig. <ref type="figure">8</ref>). Further, this increase was also sustained for both occurrences through the majority of the swing phase of the subject, showing a lasting effect of this response. This finding is significant in its own right because it shows that subjects may be anticipating the upcoming stiffness change by recoiling their leg longer to give more time to process the correct foot placement, since the VR is showing a clear finite sandpatch length, that should be fully stepped on, with the foot close to the center, and not partially on the sandpatch. Higher and more sustained knee flexion could be functioning as a way to give just enough time for this decision. Further analysis of the biceps femoris muscle activation showed that it was not consistently higher Fig. <ref type="figure">8</ref>. Left knee angle mean from 0% to 100% of the gait cycle, beginning and ending at left heel-strike. The VP (blue) and VO (orange) profiles are compared against Rigid (grey) to show unified anticipation response in the middle third portion of the left swing phase before perturbation, with the peak reaching statistical significance and is markedly sustained.</p><p>(p = 0.4) during this phase for all subjects, which could be due to other non-recorded hamstring or even the bi-articular gastrocnemius activation. However, this is still a relevant finding that should be explored further.</p><p>At the ankle joint level, there is a markedly higher peak left gastrocnemius (LGA) activity during the left stance phase and a decrease in left hip-ankle span during the left swing phase (see Fig. <ref type="figure">9</ref>). The increase in LGA activity was initiated during the end of the terminal stance phase and was even sustained until the left toe-off occurred. This finding suggests that additional propulsion force was needed by the subject to prepare for the perturbation. It is expected that the placement of the right foot is critical to properly provide stability to the subject immediately before and during the perturbation. Since the right foot can only be planted once, and not shifted again before the perturbation begins, it can be hypothesized that the subject desired the position of the body (supported by the right foot just before the perturbation at the left heel strike) to be closer to the sand patch so that more time and less energy can be put into placing the left foot. This hypothesis is founded on findings of both the hip and the knee. Hip-ankle span is shown to be a marker of particular importance in stroke rehabilitation, as it combines and relates to other key outcome measures <ref type="bibr">[35]</ref>. Here, a decrease in left hip-ankle span is seen during the exact period of increased knee flexion (see Fig. <ref type="figure">10</ref>). Additionally, this hip-ankle span stayed markedly lower until the left heel strike occurred, showing that compared to stepping on a rigid surface, a time delay occurred prior to making the step on the compliant surface, regardless of whether that step was revealed to have a stiffness change or not.</p><p>In general, anticipatory responses were seen relatively quickly after initial VP training. Many subjects confirmed after the trial that they quickly learned to associate visual and physical feedback. From this, it is also likely that the amount of time spent training could even be decreased, and Fig. <ref type="figure">9</ref>. Left GA activation mean from 0% to 100% of the gait cycle, beginning at left heel-strike. The profile for VP (blue) and VO (orange) are compared against the rigid (grey) to show a unified anticipation response in the middle of the right swing phase of the gait cycle before perturbation. Peak magnitude reaches statistical significance and is markedly sustained after the peak, only noticeably decreasing when left toe-off occurs. Fig. <ref type="figure">10</ref>. Left hip-ankle span mean from 0% to 100% of the gait cycle, beginning at left heel-strike. The profile for VP (blue) and VO (orange) are compared against the rigid (grey) to show a unified anticipation response, reaching a peak at around 70% of the gait cycle before perturbation. Peak magnitude reaches statistical significance and is markedly sustained after the peak. still achieve results. Future studies will investigate further the training dose requirements. Taken to the extreme, a similar protocol could simply include a higher VO:VP2 ratio, or only VO gait cycles in the last phase, but it is likely that subjects may ignore the previous training and walk normally without regard to the virtual environment, since the premise of the current work requires an established association between the sand patches and the compliant surface. Future studies should test for how large this ratio can grow before the participants start to disconnect this association and tailor the protocol accordingly to avoid any negative effects caused by this cognitive conflict. Overall, anticipatory responses were similar after repeated VO inputs, without significant trends of the effects decreasing over time. While every subject expressed being convinced by the majority of VO gait cycles, it could be reasonably hypothesized that eventually subjects may become comfortable with not receiving physical perturbations, and the expectation could wear off, however, this trial was designed to mitigate those effects. It is also possible that some findings were caused by the subjects choosing to exhibit patterns related to stepping on a compliant surface, even though subjects had evidence that the surface might be rigid, because assuming that the surface will be compliant may be more comfortable for the subject due to the reduced joint stiffness and resulting impact forces <ref type="bibr">[40]</ref>. Although the scope of this work only includes anticipation changes to answer the primary hypothesis, it is also important to note that all subjects were able to recover from the lack of stiffness changes within the next gait cycle, and without significant gait abnormalities. Another limitation of the current work is the relatively short study duration. Because the current work investigates a novel experimental design, a single-day study duration was chosen. Further studies should investigate how long the responses are retained over a larger timescale, and what changes occur in that timescale in order to have even more applicability to robot-assisted rehabilitation. No motion sickness was reported in this study.</p><p>Each finding shows that subjects learned to associate virtual sand patches with stiffness changes in the first phase and were still convinced that the next step would be a compliant surface in both virtual and physical environments. Subjects retained this anticipatory response throughout the experiment, demonstrating coordinated effort to prepare for expected surface changes in VR. These results highlight the crucial role of visual and proprioceptive feedback interplay in the body's responses to environmental changes, influencing long-term expectations, even before those changes occur and even without being presented repeatedly. These anticipatory responses are directly related to targeted therapy goals, providing evidence for coupling visual feedback with expected perturbations to improve recovery in rehabilitation. Such findings support using similar protocols for other robot-assisted gait therapies, improving outcomes through additional modalities.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>IV. CONCLUSIONS</head><p>This paper unifies multiple modalities of sensorimotor learning during robot-assisted therapy for post-stroke gait rehabilitation. The Variable Stiffness Treadmill (VST) combined with a VR system enhances understanding of gait training mechanisms, specifically how the visual system can be used to train responses to visual and proprioceptive feedback. Stiffness changes evoke bilateral anticipatory responses in all three lower limb joints in coordinated ways, particularly when subjects expect the next step to be perturbed based on visual feedback. Increases in left hip flexion, right hip abduction, left swing phase knee flexion and velocity, and gastrocnemius activity are observed. Moreover, the combination of proprioceptive and visual feedback extends the training effect during trials with visual stimuli alone.</p><p>These significant findings provide substantial displays of how proper protocol design utilizing connected visual and proprioceptive modalities can drastically improve our understanding of gait training from a motor learning perspective. This study seeks to revolutionize robot-assisted gait rehabilitation, opening possibilities in the integration of neuroscience and robotics toward applications in intent recognition, fall evaluation, and fields outside of rehabilitation, such as dynamic control of robotic prostheses or exoskeletons.</p></div></body>
		</text>
</TEI>
