Robotics for Training After Injury

Jose A. Galvez and David J. Reinkensmeyer Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021

Locomotor training can help people with to regain some mobility. However, the amount of hands-on therapy that patients can receive is limited, as economic pressures are inherent in the health care system. Therefore, worldwide efforts are being made to automate locomotor training. Automation has the potential to make therapy more affordable and thus more available for more patients and for more time. This article reviews efforts to automate locomotor training. We suggest that developing robotic devices that assist only as needed during step training is an important direction for future research. Key words: biomedical technology, gait, locomotion, lower extremity, rehabilitation, robotics, therapy

obots are devices that can move and vous system diseases, including traumatic exert mechanical forces in a con- injury, multiple sclerosis, Parkinson’s Rtrolled way. They are usually de- disease, and spinal cord injury (SCI).2 signed to substitute for human labor in tasks Studies of movement therapy after stroke that are dangerous or difficult for humans to and SCI have shown that recovery is better perform. Robotic capabilities differ from when more therapy is given.6–9 More therapy those of humans; robots excel at accuracy, can be given if robots are built that can repeatability, and speed but are typically not replicate the exercise provided by rehabilita- as adaptable or intelligent as humans in the tion therapists. Robots do not get tired, and ways they perform their tasks. they have the potential of delivering training Robots have been studied and used as an more consistently and of quantifying and assistive technology, such as an aid to daily controlling the variables associated with the living of impaired people, since the 1960s (see, e.g., Hillman1 for a review). Their use for therapy, such as facilitating exercise and Jose A. Galvez, PhD, is Postdoctoral Researcher, Department of Mechanical and Aerospace Engineer- subsequent recovery of motor function, is ing, University of California, Irvine. more recent, growing during the 1990s (see, e.g., Reinkensmeyer2 for a review). A major David J. Reinkensmeyer, PhD, is Associate Professor, Department of Mechanical and Aerospace Engineer- research focus has been the development of ing and Department of Biomedical Engineering, Uni- devices that can provide movement therapy versity of California, Irvine. for the arm after stroke.3–5 It may also be possible to use therapeutic robotics to treat Top Spinal Cord Inj Rehabil 2005;11(2):18–33 © 2005 Thomas Land Publishers, Inc. impairments resulting from other central ner- www.thomasland.com

18 Robotics for Gait Training 19 Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021

Figure 1. Locomotor training with body-weight support on a treadmill. assistance more accurately than human partial body-weight support.16,17 therapists. Robots also have the potential of Partial body-weight support treadmill implementing dynamic regimes of training training (PBWSTT) is a rehabilitation tech- that are beyond the capacities of human nique based on the animal studies of spinal therapists. cord plasticity. Three therapists assist the Increasing scientific evidence has accu- legs and hips of the patient on a mulated that training improves recovery of treadmill while part of the patient’s body mobility in patients with SCI. Studies with weight is supported by an overhead harness cats decades ago demonstrated that the mam- (Figure 1). PBWSTT is based on the prin- malian spinal cord can learn after injury.10,11 ciple of generating normative, locomotor- Cats and rodents with complete spinal cord like sensory input to promote the functional transection regain hindlimb stepping ability reorganization and recovery of the injured after training with partial support of the body spinal cord circuitry.18 In automated weight and assistance of leg movements on a PBWSTT, the three therapists shown in Fig- treadmill.10–13 In humans, patients with both ure 1 are replaced by a robotic system; thus acute and chronic incomplete SCIs improve only one therapist is required for initiating functional mobility both after a period of and monitoring the activity. Automated conventional mobility training tech- PBWSTT is arguably less strenuous for the niques14,15 and after treadmill training with therapist than any other manual locomotor 20 TOPICS IN SPINAL CORD INJURY REHABILITATION/FALL 2005

training technique. Automation has the po- tients, 92% of those initially wheelchair- tential to make locomotor training available bound became independent walkers after to more people, more intensively, for more PBWSTT, in contrast to only 50% after time and also can provide training in chronic conventional therapy. Among the chronic stages of SCI and even months or years after patients, 76% of those initially wheelchair- injury. bound learned to walk independently after Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021 PBWSTT compared to only 1/14 after con- Clinical Studies of Manually Assisted ventional training. Locomotor Training In a follow-up study,22 it was found that this functional recovery was maintained or The first studies that reported improve- improved in most patients 6 months to 6 ments in ambulatory capacity in chronic in- years after locomotor training without fur- complete SCI patients who underwent ther training. A prerequisite to maintaining PBWSTT date back to 1989–199216,19,20 from these long-term beneficial effects appears to two independently working groups. be a minimum of regular activity, such as In 1994, a third research group21 included walking daily at least short distances.23 This acute complete paraplegic patients in their is consistent with observations made in study study and showed that coordinated stepping of cats after SCI that indicated that hindlimb movements could be induced on the tread- stepping ability on a treadmill was main- mill in complete SCI patients. Leg extensor tained 6 weeks after cessation of training but electromyographic (EMG) activity in- decreased significantly after 12 weeks.24 creased during training, while the inappro- When subsequently retrained, the spinal cats priate, non–locomotor-like EMG activity regained stepping ability more rapidly than decreased. However, the group of complete when trained initially. SCI patients did not improve their ability to More recently, a large multicenter ran- walk over ground; their improvement was domized clinical trial with acute incomplete limited to stepping on the treadmill with SCI patients compared PBWSTT at high partial body-weight support. Patients with treadmill speeds with conventional mobility acute incomplete paraplegia improved their training.25 Preliminary reports indicate that overground walking. both groups improved their outcome mea- In 1995, Wernig and colleagues17 re- sures relevant to walking performance, but ported the first study that included a com- no significant differences were found be- parison with a control group that underwent tween the PBWSTT and the conventionally conventional training. Eighty-nine patients trained groups.26,27 Interventions started from with incomplete SCIs (44 chronic and 45 2 to 8 weeks after onset of injury; the total acute) underwent daily PBWSTT and were number of patients was 140. The convention- compared with 64 patients (24 chronic and ally trained group received an equal duration 40 acute) treated conventionally. The of therapy consisting of standing and step- therapy duration varied between patients in ping using a tilt table, standing frame, paral- a range from 1 to 5 months. The PBWSTT lel bars, or over ground. An article reporting group improved their mobility more than the results has not been published yet, and the control group. Among the acute pa- thus it is difficult to determine why the results Robotics for Gait Training 21

of this study may differ from those of Wernig two footplates that attach to the patient’s and colleagues.17 feet. They are driven by a singly actuated Even if PBWSTT is only as effective as mechanism that moves the along a conventional training, it is of particular inter- fixed ellipsoid-like trajectory with a est because one can argue that it is more doubled crank and rocker system.28 The amenable to automation than conventional sagittal movements of the torso are driven in Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021 locomotor training techniques. PBWSTT is a phase-dependent manner by ropes at- done on a stationary set-up with the patient tached to the harness and connected by held in a fixed reference position and orienta- another crank to the foot crank. The body tion. This makes PBWSTT easier to equip weight is unloaded as needed by an over- with the appropriate tools and robotic systems head harness. The GT I is installed in tens of than conventional training techniques, be- locations, mainly across Europe. Clinical cause the robotic devices do not have to move trials have been carried out with the GT I, themselves over ground along with the pa- and a more advanced version has been de- tient. Conventional training techniques signed and built, as described below. change when the patient progresses, from tilt table or standing frame to parallel bars and Lokomat overground walking, which adds to the com- The Lokomat (Hocoma Medical Engi- plexity of its automation. An automated neering, Inc., Zurich, Switzerland; Figure PBWSTT system addresses the problem of 2B) is a motorized exoskeleton worn by training (both traditional or PBWSTT) being patients during treadmill walking.29 Four an arduous and labor-intensive job, particu- rotary move hip and knee flexion/ larly with severely impaired patients. The extension for each leg. The joints are driven subsequent dissemination and intensification by precision ball screws connected to di- of therapy might have a significant impact on rect-current motors that are programmed to rehabilitation for individuals with SCI. In the drive the legs in a gait-like pattern along a next sections, we describe the worldwide ef- fixed position-controlled trajectory. The forts toward automating PBWSTT. device attaches to the thighs and shanks through padded straps. A passive parallelo- Gait-Training Robots Currently in Use gram mechanism allows vertical translation in Clinics of the patient’s torso, restricting lateral translation. The weight of the leg exoskel- Two European research groups and a US eton is counterbalanced through a gas rehabilitation provider have developed de- spring attached to the parallelogram. The vices for automating PBWSTT that are al- patient’s body weight is unloaded as needed ready being used in several clinics world- through an overhead harness connected to a wide: the Mechanized Gait Trainer,28 the stack of counterbalancing weights. The Lokomat®,29 and the AutoAmbulator™.30 Lokomat has been installed in tens of re- search labs and clinics worldwide. Reports Gait Trainer GT I of using the device with SCI patients and The Gait Trainer (GT I; Reha-Stim, Ber- attempts at refining the robot’s control strat- lin, Germany; Figure 2A) is comprised of egies are detailed below. 22 TOPICS IN SPINAL CORD INJURY REHABILITATION/FALL 2005 Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021

, and (C)

®

Gait-training robotic systems currently in use in clinics: (A) Gait Trainer GT I, (B) Lokomat

AB C

Figure 2. AutoAmbulator™. (A) Reproduced with permission from Reha-Stim, Berlin, Germany. www.reha-stim.de/ eprogt.htm. (B) Reproduced with permission from Hocoma AG, Volketswil, Switzerland. www.hocoma.ch. (C) Reproduced with permission from Healthsouth Corporation, Birmingham, Alabama. Robotics for Gait Training 23

AutoAmbulator therapy in the opposite order (B-A-B). The patients were not significantly different be- Healthsouth, a US provider of health care fore therapy in their overground walking services, has developed the AutoAmbu- ability, but after therapy the robot group lator,30 which is a body-weight supported performed better. Walking ability improved treadmill robot system (Figure 2C). Robotic more rapidly for patients in both groups dur- Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021 arms are strapped to the patient’s legs at the ing the robot-assisted training. Differences thigh and , driving them through a step- were not significant at a 6-month follow-up. ping pattern. A touch screen computer moni- These results suggest that robot-assisted gait tors knee torque, thigh torque, and speed, training can be as effective as conventional among other session data. An automatic shut training, with the added benefit of a reduction down feature is triggered through severe in manual labor and in the strenuous nature of spasms or an improper foot strike on the the job. When asked about their therapy treadmill by the patient. The AutoAmbulator preference, 29 out of the 36 stroke patients is currently installed in 26 rehabilitation cen- preferred the gait-training robot and 7 pre- ters, all of them in the United States. ferred the nonrobotic treadmill therapy. The participants who preferred robotic training Clinical Studies of Robot-Assisted found it less demanding (18 participants) and Locomotor Training more comfortable because it required less manipulation during therapy (11 partici- Clinical studies of gait-training robotic pants). The seven patients who preferred the systems have been carried out for stroke treadmill therapy said that swinging the patients with the GT I28,31 and for SCI patients paretic on the treadmill seemed more with the Lokomat.9,32,33 natural (with or without help) and thus more In 2000, the GT I was reported to have effective to them, compared with the ongo- trained two patients who were 2 months post ing footplate support of the “swinging” limb stroke.28 The patients received 4 weeks of on the gait trainer.31 gait training with the device, consisting of In 2001, the Lokomat was reported to have five 20-minute sessions per week. The pa- trained two patients with SCI, one complete tients improved significantly in their and one incomplete.32 The elicited EMG pat- overground walking ability. terns were compared with those elicited dur- One randomized, controlled study of ro- ing manually assisted PBWSTT, and no sig- botic therapy for gait training after stroke nificant difference was found. The patients was published in 2002.31 The GT I was used reported that the automated training was more in a crossover design to train subacute stroke comfortable than the manually assisted one. patients to walk. A robot training group (n = Also, longer training sessions became pos- 15) received 2 weeks of robotic gait training sible with the automated gait trainer. While a (A), followed by 2 weeks of human-assisted manually assisted training usually lasted up to treadmill therapy (B), followed by 2 more 20 minutes, the automated training could be weeks of robotic gait training (A). A control performed for up to 60 minutes.32 group (n = 15) received the robotic gait In 2005, Hornby and coworkers33 reported training and human-assisted treadmill the use of the Lokomat to train three incom- 24 TOPICS IN SPINAL CORD INJURY REHABILITATION/FALL 2005

plete SCI patients, two of them acute and the moved by the patient if he or she can generate other chronic. The two acute patients recov- the required motor output. A backdrivable ered independent overground walking; the gait-training robot could be made to grade chronic patient, who was ambulatory before assistance, that is, to assist as needed. locomotor training, improved his gait speed Also in 2005, Hidler and Wall34 reported and endurance, although only slightly. the use of the Lokomat with seven unim- Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021 Wirz and colleagues9 recently reported paired participants to compare EMG patterns preliminary results of the effects of robot- in various leg muscles with those exhibited assisted PBWSTT with the Lokomat. Train- during treadmill walking without assistance. ing was for 8 weeks, three to five sessions per Significant differences were found. In the week, 45 minutes per session, in 20 chronic swing phase, EMG activity of the quadriceps incomplete SCI individuals more than 2 and hamstrings was significantly higher dur- years after injury. The improvements in ing Lokomat walking than during unassisted overground walking speed and endurance treadmill walking, whereas activity in the were statistically and functionally signifi- ankle flexor and extensor muscles was re- cant: approximately 50% on average in the duced with the Lokomat. Hidler and Wall 16 patients who were ambulatory prior to suggest that the changes in muscle activity entry into the study. Endurance and func- are caused by the restrictions of leg move- tional performance were also improved. ments imposed by the robot. Adding more There were no significant changes in the degrees of freedom, adaptability, and requirement of walking aids, orthoses, or backdrivability might result in more locomo- external physical assistance, as assessed by tor-like, normative muscle activity. the Walking Index for Spinal Cord Injury In another study,35 six people with SCI (WISCI-II) test.9 The four patients who were stepped with a newer version of the Lokomat nonambulatory before training did not regain that is not yet commercialized. This proto- locomotor function. The improvements ap- type permits gait-pattern adaptation by pear to be comparable to those achieved by changing five parameters per leg that param- similar SCI patients who received manually eterize the reference kinematic trajectories. assisted PBWSTT. Gait speed and endur- The gait pattern is adapted in a way that ance tests, which are more sensitive to reduces the interaction between the Lokomat changes than the WISCI-II test, showed that and the patient. The rate of change of the improvements among ambulatory patients trajectory parameters was 2 seconds. In a were significantly greater for the most im- questionnaire, the patients said they pre- paired patients. Wirz et al. suggest that ferred the training modality in which they manually assisted PBWSTT might result in could change the gait pattern to a more de- greater improvements for SCI people who sired one rather than training with a fixed gait already ambulate at relatively high speeds, as pattern. compared to robot-assisted PBWSTT. As we argue in the next section, a backdrivable Assisting As Needed and Backdrivability robotic device might elicit more significant motor gains in less impaired patients. The assistance provided to SCI patients Backdrivable means that the robot can be during gait training on a treadmill can be Robotics for Gait Training 25

given in different ways. Currently, it is un- et al.38 reported that active training of repeti- clear what the optimal assistance techniques tive thumb movements resulted in encoding are. Is it better to assist only if needed and of motor memories in the primary motor only as much as needed, letting the patients cortex, accompanied by characteristic do the effort if they have the motor output to changes in corticomotor excitability, do so, or is it equally good to assist always, whereas passive training did not. It remains Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021 driving the legs through a fixed rigid trajec- to be experimentally studied if these neural tory? Well-controlled, clinical tests are plasticity findings also apply to gait training needed to answer these questions. We hy- of impaired people. pothesize that such tests will favor assis- Third, assistance-as-needed is also sup- tance-as-needed for the following reasons. ported by a computational, theoretical model First, common sense suggests that a key to of sensorimotor control that models locomo- motor learning is experiencing the real dy- tor adaptation as a process of learning an namics of the task to the most complete altered sensory-motor transformation.39 This extent possible. The real-world dynamics of model is based on three assumptions. The the task under consideration here are those of first assumption is that the central nervous unassisted walking. Therefore, less assis- system probabilistically interprets sensory tance results in experienced dynamics that information in real time to generate motor are more similar to the real dynamics. Pa- output. Consider, for example, that changes tients will likely learn better if they experi- in load-related afferent information during ence the consequences of their motor drive gait training cause functionally appropriate than if they learn to rely on the trainer’s levels of limb extensor muscle activity to assistance, be it robotic or human. support the body40 or that afferent input from Second, basic motor training research for hip joints plays a role in triggering swing other tasks has shown that rigidly assisting (see, e.g., Dietz41 for a review); however, movement, with the participant remaining inherent variability exists in this response. relaxed, produces less motor learning when This first assumption is captured abstractly compared to active voluntary practice in un- using a Markov model. The second assump- impaired participants.36–38 For instance, tion is that sensorimotor pathways become Lotze and colleagues37 compared perfor- more reliable with repetitive activation in a mance gains after a training period of either sort of Hebbian learning. In other words, subject-driven (i.e., active) or robot-driven each time a sensorimotor transition is (i.e., passive) wrist movements. Motor per- crossed, the probability associated with it is formance, measured as the number of move- incremented and competing pathways are ments that hit a target window duration, was decremented. The third assumption is that significantly better after active training than normal sensory input sometimes elicits ab- after passive training. Passive training did normal motor output after neurologic injury not lead to significant behavioral gains. In owing to disrupted neural organization. addition, the magnitude of cortical reorgani- Given these assumptions, the adaptive zation and the size of the engaged brain areas Markov model predicts that the best recovery were larger with active than with passively occurs when an external trainer intervenes to elicited movements. Recently, Kaelin-Lang correct errant movements on an as-needed 26 TOPICS IN SPINAL CORD INJURY REHABILITATION/FALL 2005

basis, compared to no assistance or continual and locomotor-like EMG activity, this would assistance.39 provide support for the principle of assis- As already mentioned, this rationale for tance-as-needed. assistance-as-needed still remains to be vali- dated experimentally for gait rehabilitation Other Considerations for the of individuals with SCI. The best tool to test Facilitation of Locomotor-like Afferent Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021 the assist-as-needed concept is a robotic sys- Feedback tem that can be moved by the patient with the smallest possible resistance opposed by the In addition to the hypothesis that assis- robot. In other words, a highly backdrivable, tance-as-needed is better than assisting al- compliant robotic system is needed as a re- ways, we propose the following working search tool. Depending on the results of the hypotheses for the mechanical and control research, backdrivable robots may have an design of a gait-training robotic system: important impact as a therapy tool. 1. A design is good if it maximizes the The backdrivable machine should be able sensory inputs associated with natural to assist the motion if the patient lacks walking (e.g., normative locomotor- enough motor output and should reduce the like impact forces on the feet, norma- amount of assistance provided as the patient tive locomotor-like evolution and dis- recovers. It should also be able to deviate tribution of the foot forces, normative from the controlled path when the patient locomotor-like kinematics of the pel- exerts uncoordinated forces, providing di- vis and legs). Afferent input from re- rect and natural kinematic feedback of move- ceptors is essential for the activation of ment control errors. The patient should al- spinal locomotor centers (see, e.g., ways be able to influence the trajectory. The Dietz41 for a review), and therefore it is quality of the backdrivability is given by the probably also important for stimulat- lowest endpoint mechanical impedance ing training effects. achievable, which should ideally be zero; the 2. The design will be even better if it robot should ideally be able to fade to nothing minimizes sensory inputs that do not if it turns out that the patient does all the work belong to normal locomotion, that is, it required to move. is important that the robot is not no- In collaboration with colleagues at UCLA ticed by the patient except when it is and Rancho Los Amigos National Rehabili- strictly necessary for completing a step- tation Center, we are working to test the ping motion. assist-as-needed concept on manually as- In light of the above hypotheses, a possible sisted PBWSTT. We have developed a sen- improvement to current commercial gait- sor system that measures the forces applied training robots would be a more natural mo- by the therapists during PBWSTT, as well as tion of the pelvis, because current designs the patient’s shank kinematics.42 We have restrict this motion to one degree of freedom found that different therapists assist very (dof). This is important not only for experi- differently in terms of forces and elicited encing natural kinematics but also for shap- kinematics. If the therapists providing less ing the distribution and timing of foot forces assistance elicit better stepping kinematics and, in particular, for appropriate weight- Robotics for Gait Training 27

shifting between the legs during walking. treadmill walking. In normal walking, the Are there better places to place the horizontal foot force component is directed handholds where the robots interact with the forward to propel body and limbs; whereas in patient? Studies in cats show that maintained treadmill walking, it is directed backward to load of the Achilles tendon (i.e., continued move the stance leg backward. Thus, the activation of extensor load-sensitive recep- sensory input associated with treadmill Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021 tors) prevents the initiation of flexor muscle walking differs from the sensory input asso- activity necessary for swing.43 Based on ex- ciated with normal walking. A possible solu- tensive clinical experience, Behrman and tion is to incorporate the concept of assis- Harkema44 provide some guidelines regard- tance-as-needed into a split-belt treadmill; ing which tendons to stimulate and which to the motor could provide the force needed to avoid stimulating in different phases of the move the treadmill band at a constant speed step cycle to facilitate more successful step- only if the patient is not able to generate that ping on the treadmill. A key challenge in the force. design of robotic gait training is to effec- Further research is needed to determine tively reproduce the human touch, if the how much the patterns of sensory input gen- specific pattern of this touch is essential for erated by physical interaction with a patient optimizing locomotor training. contribute to locomotor training outcomes. Because many patients with SCI experi- Such research will shape the evolution of ence foot drop during gait, it is difficult for a robotic gait trainers. gait-training robot to ensure safe toe clear- ance without generating conflicting, unnatu- Current Developments and Future ral sensory feedback that may inhibit step- Clinical Technologies ping, particularly the activity necessary for swinging the leg. The Lokomat has elastic Several groups are working toward de- straps attached to cuffs surrounding the signing and building improved gait-training metatarsal heads. The GT I attaches through robotic technologies. footplates, which has the possible drawback of concealing the impact forces at the transi- Haptic Walker tion from swing to stance in normal walking. The Haptic Walker is a major redesign of On the other hand, the footplate concept has the GT I that maintains the permanent foot/ the benefit of permitting the implementation machine contact but allows the footplates to of haptic emulation of ground conditions move along arbitrary 3 dof trajectories. The (e.g., stair climbing), as has been done with device incorporates force sensing, compli- the GT I’s follow-up, the Haptic Walker45 ance control, and haptic simulation of ground (see below). conditions (e.g., stair climbing). A prototype Another possible limiting factor of current machine has been built.45 approaches at automating PBWSTT is the use of standard treadmills. In normal walk- Adaptive Lokomat ing, the horizontal component of force ap- Jezernik and colleagues have improved plied by the ground to the foot is usually the Lokomat by adding algorithms for auto- directed in the opposite direction than in matic adaptation of the gait patterns with 28 TOPICS IN SPINAL CORD INJURY REHABILITATION/FALL 2005 Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021

(A) ARTHuR leg robot; (B) PAM and POGO gait training robot system.

AB

Figure 3. Robotics for Gait Training 29

help of force sensors.46 Five parameters per low them to be considered for applications leg that parameterize the reference trajectory for which previously only electric motors are changed every 2 seconds in a way that the were suitable. Pneumatic systems are inher- interaction effort between patient and ma- ently compliant, which enables a softer and chine is reduced, thus adapting to a more more natural interaction with the patient. desired motion. Three adaptation algorithms Moreover, pneumatic actuators offer a Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021 have been implemented: one is based on higher power-to-weight ratio than electric inverse dynamics for interaction torque motors. Their low price could potentially minimization, a second one is based on an facilitate clinical dissemination. estimation of the variation in acceleration ARTHuR, POGO, and PAM are capable desired by the patient, and a third one is based of providing assistance-as-needed, applying on an impedance-control scheme. The algo- a wide range of forces from fully assisting to rithms were tested on unimpaired and SCI “just going along for the ride.” participants who were able to influence the gait pattern with their remaining motor activ- String-Man ity.35,46 The String-Man is a wire robot for ma- nipulating the torso of the patients based on ARTHuR, POGO, and PAM the string-puppet principle (Figure 4).50 Our research group has developed robotic Seven wires connected to the patient’s trunk devices with the aim of providing natural gait achieve weight bearing and 6-dof manipula- movements, using backdrivability and com- tion. Force sensors between the wires and the pliance as the starting points for the robot actuators allow force or impedance control; design.47 The first gait-related robot proto- this permits adjustment of the interaction type was ARTHuR (Figure 3A), a 2-dof control from totally passive to completely robot that makes use of linear electric direct- active and allows the patients to test their drive (nongeared) motors to precisely apply balancing capabilities. The String-Man de- forces to the leg during stepping.48 It was velopers have designed a control scheme that designed primarily to conduct research on can control both the zero-moment-point lo- human motor control of gait and has been cation and the ground reaction force, using successfully used with unimpaired partici- foot force sensors in addition to the wire pants to create novel dynamic environments force sensors.50 that can enhance motor adaptation.49 Our research group has also developed Safety Issues PAM, a 5-dof compliant robot that accom- modates natural pelvic movement during Gait-training robots pose more risk than walking, and POGO, a leg robot aimed at the upper-limb robots because they need to be goal of being used in the clinic for daily stronger and more powerful. Work toward locomotor training rather than just for re- developing safety standards for rehabilita- search, but which also preserves good force tion robots has already started.51 This work control (Figure 3B).47 PAM and POGO’s includes the design of a certification system actuators are pneumatic. Recent develop- to guarantee that risk assessment is carried ments in pneumatic actuators and valves al- out. Manufacturers of gait-training robotic 30 TOPICS IN SPINAL CORD INJURY REHABILITATION/FALL 2005 Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021

Figure 4. The String-Man wire robot for torso manipulation.50 Reprinted with permission of Springer Science and Business Media from Surdilovic D, Bernhardt R, Schmidt T, Zhang J. String- man: A novel wire robot for gait rehabilitation. In: Bien ZZ, Stefanov D, eds. Advances in Rehabilitation Robotics. Lecture Notes in Control and Information Sciences. Berlin: Springer- Verlag; 2004:418. Copyright © 2004 by Springer-Verlag GmbH. systems must spend considerable effort on implements foot safety release bindings and safety concerns. For instance, the Lokomat adjustable ankle range-of-motion limit developers have designed an optical light switches.52 The staff that operates each in- sensor system that monitors the patient’s stalled system must be properly trained on feet; if the patient trips, the Lokomat and safety issues. treadmill are automatically shut off. The AutoAmbulator is equipped with safety in- Conclusion terlocks, redundant travel limits, variable torque limits, and a safety light array to detect Locomotor training can help people with improper foot drop. The Haptic Walker SCI recover some mobility. Traditional loco- Robotics for Gait Training 31

motor rehabilitation approaches are labor training for recovery of locomotor capacity. intensive and strenuous for the therapists. In parallel, several groups are working to- Therefore, researchers, clinicians, and engi- ward improving gait-training robotic tech- neers are working to develop alternatives. A nologies. These improvements include en- gait-training technique that is amenable to hancing the facilitation of normative, automation is treadmill training with partial locomotor-like sensory feedback that gait- Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021 body-weight support. This technique has training robots provide. They also include been shown to restore and maintain some the incorporation of adaptation and the prin- degree of ambulatory capacity in incomplete ciple of assistance-as-needed, that is, the SCI individuals. It also has been shown to ability of the robots to grade the forces that elicit locomotor patterns on the treadmill in they apply to the patient. This encourages the individuals with complete SCI, although patient to contribute to the stepping motion these patterns do not transfer to improve- as much as possible, which is likely essential ments in overground walking capacity. for maximizing locomotor plasticity. An- An emerging theme in motor rehabilitation other important contribution of robotic gait research is that more therapy is better. Recent training systems will be to provide a con- studies show that gait training may be benefi- trolled and quantified test bed for under- cial at any time after SCI, even years after the standing the physiological principles of loco- injury. Robotic gait training may help to make motor recovery. locomotor training more affordable and thus more available for a longer time and also Acknowledgments available in the chronic stages of SCI. A number of gait-training robotic systems This work was supported by the National are already in use in clinics worldwide. Clini- Institute on Disability and Rehabilitation cal trials have shown that robotic gait train- Research and the Spanish Ministry of Educa- ing can be as effective as manually assisted tion and Science.

REFERENCES

1. Hillman M. Rehabilitation robotics from past 4. Krebs HI, Volpe BT, Aisen ML, Hogan N. to present - a historical perspective. In: Bien Increasing productivity and quality of care: ZZ, Stefanov D, eds. Advances in Rehabilita- robot-aided neuro-rehabilitation. J Rehabil tion Robotics. Lecture Notes in Control and Res Develop. 2000;37(6):639–652. Information Sciences. Berlin: Springer-Verlag; 5. Reinkensmeyer DJ, Kahn LE, Averbach M, 2004:25–44. McKenna A, Schmit BD, Rymer WZ. Under- 2. Reinkensmeyer D. Rehabilitators. In: Kutz M, standing and promoting arm movement recov- ed. Standard Handbook of Biomedical Engi- ery after chronic brain injury: progress with the neering & Design. New York: McGraw-Hill; arm guide. J Rehabil Res Develop. 2003:35.1–35.17. 2000;37(6):653–662. 3. Burgar CG, Lum PS, Shor PC, Van der Loos 6. Smith DS, Goldenberg E, Ashburn A, et al. HFM. Development of robots for rehabilita- Remedial therapy after stroke: a randomised tion therapy: The Palo Alto VA/Stanford expe- controlled trial. Br Med J. 1981;282:517–520. rience. J Rehabil Res Develop. 2000;37: 663– 7. Dam M, Tonin P, Casson S, Ermani M, 673. Pizzolato G, Iaia V, Battistin L. The effects of 32 TOPICS IN SPINAL CORD INJURY REHABILITATION/FALL 2005

long-term rehabilitation therapy on poststroke Wernig A, ed. Plasticity of Motoneuronal Con- hemiplegic patients. Stroke. 1993; 24:1186– nections. Amsterdam: Elsevier; 1991:461– 1191. 474. 8. Kwakkel G, Wagenaar RC, Twisk JWR, 20. Wernig A, Mueller S. Improvement of walking Lankhorst GJ, Koetsier JC. Intensity of leg and in spinal cord injured persons after treadmill arm training after primary middle-cerebral- training. In: Wernig A, ed. Plasticity of Moto-

artery stroke: a randomised trial. Lancet. neuronal Connections. Amsterdam: Elsevier; Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021 1999;354:191–196. 1991:475–485. 9. Wirz M, Zemon DH, Rupp R, Scheel A, Co- 21. Dietz V, Colombo G, Jensen L. Locomotor- lombo G, Dietz V, Hornby TG. Effectiveness of activity in spinal man. Lancet. 1994;344: automated locomotor training in patients with 1260–1263. chronic incomplete spinal cord injury: a 22. Wernig A, Nanassy A, Muller S. Maintenance multicenter trial. Arch Phys Med Rehabil. of locomotor abilities following laufband 2005;86:672–680. (treadmill) therapy in para- and tetraplegic 10. Lovely RG, Gregor RJ, Roy RR, Edgerton VR. persons: follow-up studies. Spinal Cord. Effects of training on the recovery of full 1998;36:744–749. weight-bearing stepping in the adult spinal 23. Wirz M, Colombo G, Dietz V. Long term cat. Exp Neurol. 1986;92:421–435 effects of locomotor training in spinal humans. 11. Barbeau H, Rossignol S. Recovery of loco- J Neurol Neurosurg Psychiatry. 2001;71:93– motion after chronic spinalization in the adult 96. cat. Brain Res. 1987;412:84–95. 24. de Leon RD, Hodgson JA, Roy RR, Edgerton 12. Timoszyk W, de Leon R, London N, Roy R, VR. The retention of hindlimb stepping ability Edgerton V, Reinkensmeyer D. The rat lum- in adult spinal cats after the cessation of step bosacral spinal cord adapts to robotic loading training. J Neurophysiol. 1999;81:85–94. applied during stance. J Neurophysiol. 2002; 25. Dobkin BH, Apple D, Barbeau H, et al. Meth- 88:3108–3117. ods for a randomized trial of weight-supported 13. Cai LL, Fong AJ, Otoshi CK, et al. Effects of treadmill training versus conventional training consistency versus variability in robotically for walking during inpatient rehabilitation af- controlled training of stepping in adult spinal ter incomplete traumatic spinal cord injury. mice. Soc Neurosci Abstr. 2004;418.3. Neurorehabil Neural Repair. 2003;17:153– 14. Klose KJ, Schmidt DL, Needham BM, Brucker 167. BS, Green BA, Ayyar DR. Rehabilitation 26. Dobkin B. From bench to bedside, a perspec- therapy for patients with long-term spinal-cord tive on spinal cord injury research. Presented injuries. Arch Phys Med Rehabil. at the Spinal Cord Conference; June 3–5, 1990;71:659–662. 2004, Long Beach, CA. 15. Dobkin B. The Clinical Science of Neurologic 27. Dobkin BH, Havton LA. Basic advances and Rehabilitation. 2nd ed. New York: Oxford new avenues in therapy of spinal cord injury. University Press; 2003. Annu Rev Med. 2004;55:255–282. 16. Visintin M, Barbeau H. The effects of body- 28. Hesse S, Uhlenbrock D. A mechanized gait weight support on the locomotor pattern of trainer for restoration of gait. J Rehabil Res spastic paretic patients. Can J Neurol Sci. Develop. 2000;37(6):701–708. 1989;16:315–325. 29. Colombo G, Joerg M, Schreier R, Dietz V. 17. Wernig A, Muller S, Nanassy A, Cagol E. Treadmill training of paraplegic patients with Laufband (treadmill) therapy based on “rules a robotic orthosis. J Rehabil Res Develop. of ” is effective in spinal 2000;37(6):693–700. cord injured persons. Eur J Neurosci. 1995; 30. Healthsouth, Inc. Autoambulator. Accessed 7:823–829. 2005. Available at: http://www. 18. Barbeau H. Locomotor training in autoambulator.com neurorehabilitation: emerging rehabilitation 31. Werner C, Von Frankenberg S, Treig T, Konrad concepts. Neurorehabil Neural Repair. M, Hesse S. Treadmill training with partial 2003;17:3–11. body weight support and an electromechani- 19. Barbeau H, Blunt R. A novel interactive loco- cal gait trainer for restoration of gait in sub- motor approach using body weight support to acute stroke patients: a randomized crossover retrain gait in spastic paretic subjects. In: study. Stroke. 2002;33:2895–2901. Robotics for Gait Training 33

32. Colombo G, Wirz M, Dietz V. Driven gait burst generation by loading ankle extensor orthosis for improvement of locomotor train- muscles in walking cats. Brain Res. ing in paraplegic patients. Spinal Cord. 2001; 1980;187:321–332. 39:252–255. 44. Behrman AL, Harkema SJ. Locomotor training 33. Hornby TG, Zemon DH, Campbell D. Ro- after human spinal cord injury: a series of case botic-assisted, body-weight supported tread- studies. Phys Ther. 2000;80:688–700.

mill training in individuals following motor 45. Schmidt H, Hesse S, Bernhardt R. Downloaded from http://meridian.allenpress.com/tscir/article-pdf/11/2/18/1982806/damj-g43a-16eh-1bdk.pdf by guest on 25 September 2021 incomplete spinal cord injury. Phys Ther. Hapticwalker - a robotic walking simulator for 2005;85:52–66. gait rehabilitation. In: Proceedings of the 14th 34. Hidler JM, Wall AE. Alterations in muscle International Conference on Mechanics in activation patterns during robotic-assisted Medicine and Biology; September 16-18, walking. Clin . 2005;20:184– 2004, Bologna, Italy. 193. 46. Jezernik S, Colombo G, Morari M. Automatic 35. Jezernik S, Scharer R, Colombo G, Morari M. gait-pattern adaptation algorithms for rehabili- Adaptive robotic rehabilitation of locomotion: tation with a 4-dof robotic orthosis. IEEE Trans a clinical study in spinally injured individuals. Robotics Automation. 2004;20:574–582. Spinal Cord. 2003;41:657–666. 47. Reinkensmeyer D, Aoyagi D, Emken J, et al. 36. Lippman LG, Rees R. Consequences of error Tools for understanding and optimizing ro- production in a perceptual-motor task. J Gen botic gait training. J Rehabil Res Develop. Psychol. 1997;124:133–142. Accepted for publication. 37. Lotze M, Braun C, Birbaumer N, Anders S, 48. Reinkensmeyer D, Wynne J, Harkema S. A Cohen L. Motor learning elicited by voluntary robotic tool for studying locomotor adaptation drive. Brain. 2003;126:866–872. and rehabilitation. In: Proceedings of the Sec- 38. Kaelin-Lang A, Sawaki L, Cohen LG. Role of ond Meeting of the IEEE Engineering in voluntary drive in encoding an elementary Medicine and Biology Society and the Bio- motor memory. J Neurophysiol. 2005;93: medical Engineering Society (pp 2353–2354); 1099–1103. October 23-26, 2002; Houston, TX. 39. Reinkensmeyer D. How to retrain movement 49. Emken JL, Reinkensmeyer DJ. Robot-en- after neurologic injury: a computational ratio- hanced motor learning: accelerating internal nale for incorporating robot (or therapist) assis- model formation during locomotion by tran- tance. In: Proceedings of the 25th Annual sient dynamic amplification. IEEE Trans Neu- International Conference of the IEEE Engineer- ral Systems Rehabil Eng. 2005;13:33–39. ing in Medicine and Biology Society (pp 1479– 50. Surdilovic D, Bernhardt R, Schmidt T, Zhang J. 1482); September 17–21, 2003; Cancun, String-man: A novel wire robot for gait reha- Mexico. bilitation. In: Bien ZZ, Stefanov D, eds. Ad- 40. Harkema SJ, Hurley SL, Patel UK, Requejo PS, vances in Rehabilitation Robotics. Lecture Dobkin BH, Edgerton VR. Human lumbosac- Notes in Control and Information Sciences. ral spinal cord interprets loading during step- Berlin: Springer-Verlag; 2004:413–424. ping. J Neurophysiol. 1997;77:797–811. 51. Nokata M, Tejima N. A safety strategy for 41. Dietz V. and locomotor disor- rehabilitation robots. In: Bien ZZ, Stefanov D, ders. Nat Rev Neurosci. 2002;3:781–790. eds. Advances in Rehabilitation Robotics. Lec- 42. Galvez JA, Kerdanyan G, Maneekobkunwong ture Notes in Control and Information Sci- S, Weber R, Scott M, Harkema SJ, ences. Berlin: Springer-Verlag; 2004:177– Reinkensmeyer DJ. Measuring human train- 185. ers’ skill for the design of better robot control 52. Schmidt H, Hesse S, Bernhardt R. Safety con- algorithms for gait training after spinal cord cept for robotic gait trainers. In: Proceedings of injury. In: Proceedings of the IEEE 9th Interna- the 26th Annual International Conference of tional Conference on Rehabilitation Robotics, the Engineering in Medicine and Biology Soci- June 28 - July 1, 2005, Chicago, IL. ety (pp. 2703–2706); September 1–5, 2004, 43. Duysens J, Pearson KG. Inhibition of flexor San Francisco, CA.