Scope of the survey

At the outset, the research team identified literature associated with the subject matter based on searches in PubMed, the Institute of Electrical and Electronics Engineers (IEEE), Science Direct and Google Scholar databases using various combinations of the following keywords: upper extremity, arm, hand, rehabilitation, therapy, training, movement, motion, assistance, support, robot, robotized, robotic, mechatronic, and motorized. Additionally, referenced literature from the selected publications was included in the survey as well. The information obtained from this literature compendium is supplemented with the data acquired from professional caregivers and manufacturers’ catalogs and websites, as well as direct communications with rehabilitation professionals, manufacturers and patients. Over 120 systems are summarized and compared in Table 1; this tabulated summary constitutes the reference for information provided in subsequent sections. As previously mentioned, the scope of this report is generally limited to the devices that support or retrain movement or manipulation abilities of disabled individuals. This survey excludes systems developed for movement assessment, occupational purposes or boosting physical abilities of healthy people. We however considered some academic, not yet specialized systems, supporting upper-extremity movements, especially if they have potential to be used for rehabilitation purposes (e.g. CADEN-7[97]). This survey also excludes devices that substitute movements of the disabled extremity but do not replace the movement itself (e.g. wheelchair mounted manipulators or autonomous robots). Although these devices improve the patient’s quality of life, they differ significantly from systems described in this survey and constitute a separate category of devices. Some companies (e.g. CSMi Computer Sports Medicine, Inc.; Biodex Medical Systems, Inc.; BTE Technologies, Inc.) manufacture sensorized equipment for rehabilitation of various joints and muscles and whose principle of operation often resembles that of exercising devices found at fitness centers. Those devices are used mainly to strengthen muscles and joints and provide some predefined resistance (e.g. isotonic, isometric or isokinetic exercises) or active force (e.g. continuous passive motion exercises). These devices also constitute a different category from the systems included in this survey because their functions are performed along a predefined operation pathway. Although difficult to clearly identify, the aforementioned were also excluded from this review.

Throughout this report, the term “number of degrees of freedom (DOF)” describes the sum of all independent movements (i.e. displacements or rotations) that can be performed in all the joints of the device. The number of DOF is defined in order to determine the exact position and orientation of all segments of the device. Also, some sections in this report are supplemented by an explanation of the most important terminology for readers who are not familiar with the technical vocabulary.

Application field and target group

A description of the specific field of application for upper limb rehabilitation devices often determines solutions for which the device itself may be applied. Upper-extremity rehabilitation involves actions that stimulate patients’ independence and quality of life. Two main application fields of robotic devices stand out: support to perform some ADLs (e.g. by power assistance or tremor suppression) and providing physical training (therapy). Although there is a significant need for powered devices supporting basic ADL at home, there are only a few of such devices proposed so far (see sixth column in Table 1). This is mainly due to technical and economical restrictions. Such devices should significantly improve the lives of their users, otherwise patients become dissatisfied and discontinue their use shortly after. They should be also safe, easily to handle and inexpensive. Portability is also often expected from devices assisting patients to perform basic ADL; in such cases the amount of available energy is limited by the capacity to store energy. Furthermore, if the device is supposed to support movements of multiple joints, the number of needed actuators increases as well as the weight of the device. Therefore, the number of portable actuated devices supporting upper extremity movements is typically low. Instead, purely mechanical solutions are used for that purpose. A few examples of portable powered devices for upper extremity assistance used in daily living are PowerGrip system (Broaden Horizons, Inc., USA) and a system proposed by Hasegawa, et al. [98] (both for grasp assistance), as well as WOTAS orthosis [99] and a system proposed by Loureiro, et al. [100] (both for tremor suppression). However, portability is not always necessary. Often, especially after a stroke or a spinal cord injury, disorders of the upper extremity are accompanied by lower extremity disabilities. These scenarios are typically characterized by immobilized conditions and require a wheelchair. Therefore, many systems assisting upper limb movements are installed close to the patient (e.g. modular wheelchair-mounted system MUNDUS[101]).

Another group of the robotic devices used for rehabilitation purposes, much bigger than the group of devices supporting basic ADLs, constitute devices providing physical therapy. These may be designed for either specialized therapeutic institutes or home-based conditions. A vast majority of these devices may be used only at therapeutic institutes since they require supervised assistance from qualified personnel. Their price is often prohibitive for personal use due to their complexity. The patient demand for home-based therapy is expected to increase. Along this context, the concept of the Gloreha system (Idrogenet srl) is provided in two versions: (1) a more complex and more adaptable professional version intended for use at hospitals and rehabilitation centers and (2) a simplified low-cost version intended for patient use at home. However, according to Dijkers, et al. [102], many therapists may stop using devices if set-up takes more than 5 minutes. Thus new developed devices for physical training should be intuitive, easy and fast to set-up and have a reasonable price.

Stroke is the most common cause among diseases and injuries for upper limb movement disorders. It is estimated that by 2030, stroke will be the fourth leading cause of reduced disability-adjusted life-years (DALY) in western countries. DALY takes into account years of life lost due to premature death as wells as years of life lived in less than full health [103]. Other causes include traumatic brain injury, spinal cord injury and injuries to motoneurons, as well as certain neurological diseases such as multiple sclerosis, cerebral palsy, Guillain-Barre syndrome, essential tremor and Parkinson’s disease. Currently proposed robotic systems for upper limb rehabilitation are typically tested on stroke patients. Only a fraction of these systems are investigated on subjects suffering from other diseases (see last column of Table 1).

Type of assistance

Haptic devices constitute another group of systems interacting with the user through the sense of touch. Haptic devices are similarly classified as either active or passive, depending on their type of actuator. In this report, haptic devices are independently categorized because their main function is not to cause or resist movement but rather to provide tactile sensation to the user. Other non-actuated devices for upper limb rehabilitation do not generate any forces but provide different feedback. These systems are labeled coaching devices throughout this report. Because coaching devices are sensorized, they serve as input interface for interaction with therapeutic games in virtual reality (VR) (e.g. T-WREX[106], ArmeoSpring from Hocoma AG) or for telerehabilitation (i.e. remotely supervised therapy). Coaching systems using video-based motion recognition (e.g. Microsoft Kinect) would also belong to this category if it were not for their lack of any mechanical part in contact with the patient. Therefore, these systems will not be further discussed in this survey.

Passive and non-actuated systems are less complex, safer and cheaper than their active counterparts. However, they are often modified in the development process with more active characteristics. Still, the main characteristic that identifies a non-actuated or passive device is the lack of the ability to perform movement; they may be an option for continuation of the rehabilitation process, rather than for training of people with significant movement disorders at an early stage of rehabilitation.

Mechanical design

The advantage of the end-effector-based systems is their simpler structure and thus less complicated control algorithms. However, it is difficult to isolate specific movements of a particular joint because these systems produce complex movements. The manipulator allows up to six unique movements (i.e. 3 rotations and 3 translations). Control of the movements of the patients upper limb is possible only if the sum of possible anatomical movements of patient arm in all assisted joints is limited to six. Increasing the number of defined movements for the same position of the end-point of the manipulator results in redundant configurations of the patient’s arm, thus inducing risk of injuries and complicated control algorithms.

The typical end-effector-based systems include serial manipulators (e.g. MIT Manus[107] - Figure 1B, ACRE[108]), parallel (e.g. CRAMER[109] and a system proposed by Takaiwa and Noritsugu [110], both for wrist rehabilitation), and cable-driven robots (e.g. NeReBot[111] - Figure 1C, MACARM[112]). The mechanical structure of HandCARE[113] may be also recognized as the series of end-effector-based cable-driven robots, each of which induce movement of one finger. In this system a clutch system allows independent movement of each finger using only one actuator.

Application of the exoskeleton-based approach allows for independent and concurrent control of particular movement of patient’s arm in many joints, even if the overall number of assisted movements is higher than six. However, in order to avoid patient injury, it is necessary to adjust lengths of particular segments of the manipulator to the lengths of the segments of the patient arm. Therefore setting-up such device for a particular patient, especially if the device has many segments, may take a significant amount of time. Furthermore, the position of the center of rotation of many joints of human body, especially of the shoulder complex [114], may change significantly during movement. Special mechanisms are necessary to ensure patient safety and comfort when an exoskeleton-based robot assists the movements of these joints [114]. For this reason, the mechanical and control algorithm complexity of such devices is usually significantly higher than of the end-effector-based devices. The complexity escalates as the number of DOF increases.

In case of systems for the rehabilitation of the whole limb the number of DOF reaches nine (ESTEC exoskeleton[115]) or ten (IntelliArm[116]). Some systems for fingers or hand rehabilitation have an even higher number of DOF. Examples include the system proposed by Hasegawa, et al. with eleven DOF [98] and the hand exoskeleton developed at the Technical University (TU) of Berlin with twenty DOF [117]. Even at such a high number of DOF some of these devices still remain wearable (i.e. the user is able to walk within a limited area due to connections to power source and control unit, e.g. ESTEC and hand exoskeleton developed at the TU Berlin) or portable (i.e. the area within which the user may walk is not limited, e.g. the system proposed by Hasegawa).

Apart from purely exoskeleton- or end-effector-based devices, there are many systems combining a few approaches. In the ArmeoSpring system (Hocoma AG) for example, only the distal part – comprising the elbow, forearm and wrist – is designed as an exoskeleton. Therefore the limb posture is statically fully determined (as in exoskeleton-based systems) and the shoulder joint is not constrained, allowing easy individual system adaptation to different patients. A similar concept was applied in Biomimetic Orthosis for the Neurorehabilitation of the Elbow and Shoulder – BONES[118]. In that case, a parallel robot consisting of passive sliding rods pivoting with respect to a fixed frame provides shoulder movements. Such application of sliding rods allows internal/external rotation of the arm without any circular bearing element. The distal part allowing for flexing/extending the elbow resembles the exoskeleton structure. In the MIME-RiceWrist rehabilitation system [119] the end-effector-based MIME[120] system for shoulder and elbow rehabilitation is integrated with the parallel wrist mechanism used in the MAHI exoskeleton (known as RiceWrist[119]).

Another example is the 6 DOF Gentle/S[121] system allowing for relatively large reaching movements (three actuated DOF of the end-effector-based commercial haptic interface HapticMaster, Moog in the Netherlands BV [122]) and arbitrary positioning of the hand (connection mechanism with three passive DOF). The Gentle/S system was further supplemented with a three-active-DOF hand exoskeleton to allow grasp and release movements. This new nine DOF system is known as Gentle/G[123].

The HEnRiE[124] is a similar system based on the Gentle/S system. In addition to the three active DOF of HapticMaster, HEnRiE includes a connection mechanism with two passive DOF for positioning of the hand and grasping device (two parallelogram mechanisms allowing parallel opening and closing of fingers attachments) with only one active DOF.

Some systems combine more than one robot at the same time. This approach may be considered as the combination of end-effector-approach, where only the most distal parts of robots are attached to the patient’s upper limb, with the exoskeleton-based approach, where movements of few segments are directly controlled at the same time. Use of two robots to control the movements of the limb may allow for mimicking the operations performed by therapist using two hands. Examples of systems using two-robot-concept include REHAROB[125] (using two six-DOF manipulators), iPAM[126] and UMH[127] (both having six DOF in total). Researchers at the University of Twente, in Enschede, Netherlands, made an attempt to use two HapticMaster systems to provide coordinated bilateral arm training, but limitations in hardware and software caused the virtual exercise to behave differently to the real-life [128]. In some cases industrial robots have also been used. The REHAROB uses IRB 140 and IRB 1400H from ABB Ltd., while MIME[120] uses PUMA 560 robot. In general, industrial robots reduce costs; however, such robots have significantly higher impedance than the human upper limb and, according to Krebs, et al. [129], should not be in close physical contact with patients. Therefore most of the robots used for the rehabilitation of the upper limb are designed with a low intrinsic impedance. Some of those devices are also back drivable (e.g. HWARD[130] and RehabExos[131]), meaning that the patient’s force is able to cause movement of those devices when they are in passive state. Back-drivability further increases safety of the patient because the device does not constrain patient movements. It also allows for using the device as an assessment tool to measure patient’s range of motion.

The majority of the devices presented in Table 1 allow movements in three dimensions; however there are also planar robots, i.e., systems allowing movements only on a specified plane (e.g. MEMOS[132] and PLEMO[105]). Also the MIT Manus system initially allowed movements only on one plane [107]. Subsequently, an anti-gravity module added possibility to perform vertical movements [133] (Figure 1B). Designing the device as a planar robot reduces the range of movements that can be exercised for particular joint. It also reduces the cost of the device. Furthermore, when the working plane is well selected, the range of training motion may suffice in most of therapeutic scenarios. Some of such planar devices allow changes in the working space between horizontal and vertical (Braccio di Ferro[134]) or even almost freely selecting the working plane (e.g. PLEMO and Hybrid-PLEMO[135]). It further increases the range of possible exercise scenarios while keeping the cost of the device at a minimum.

In the ARM Guide[136] (Figure 1A) and ARC-MIME[137] systems, with which patients practice reaching movements, the working space is limited to linear movements because the forearm typically follows a straight-line trajectory. However the orientation of the slide that assists forearm movements can be adjusted to reach multiple workspace regions and fit different scenarios.

Modularity and reconfigurability are concepts that could reduce therapy costs by adopting therapeutic devices for various disabilities or stages in patient recovery. However there are still only a few systems using these concepts. For example, InMotion ARM robot (the commercial version of MIT Manus, previously called InMotion 2.0) from Interactive Motion Technologies, Inc., may be extended by InMotion WRIST robot (previously InMotion 3.0), developed at MIT [138] as a stand-alone system, and InMotion HAND add-on module (previously InMotion 5.0) for grasp and release training. Another example of modular system is MUNDUS[101], consisting of various modules that may be included depending on the patient condition, starting from muscle weakness to complete lost of residual muscle function. For example as command input residual voluntary muscular activation, head/eye motion, or brain signals may be used. However, this system’s complexity might make commercialization of the device very difficult.

A very interesting solution was implemented in the Universal Haptic Drive (UHD)[139]. It has only two DOF and, depending on the chosen configuration, it can train either shoulder and elbow during reaching tasks or forearm (flexion/extension) and wrist. For the latter setting option, it is also possible to select a flexion/extension or pronation/supination training for the wrist. See Figure 2 for an explanation of anatomical terms used for description of upper limb motion.

Actuation and power transmission

A few systems use pneumatic actuators. Pneumatic actuators are lighter and have lower inherent impedance than the electric counterparts. Because such actuators require pneumatic pressure, the system is generally either stationary (e.g. Pneu-WREX[145]), its service area is limited (e.g. ASSIST[146]) or the compressor is installed on the patient’s wheelchair (e.g. system proposed by Lucas, et al. [147]). Special type of pneumatic actuators, called Pneumatic Artificial Muscles (PAMs), Pneumatic Muscle Actuators or McKibben type actuators are often used in rehabilitation robotics (e.g. Salford Arm Rehabilitation Exoskeleton[148] or system proposed by Kobayashi and Nozaki [149]). Such actuators consist of an internal bladder surrounded by braided mesh shell with flexible, but non-extensible, threads. When the bladder is pressurized, the actuator increases its diameter and shortens according to its volume, thus providing tension at its ends [150]. Due to such physical configuration, PAMs’ weight is generally light compared to other actuators, but also have slow and non-linear dynamic response (especially large PAMs), in consequence they are not practical for clinical rehabilitation scenarios [131, 151]. In addition, at least two actuators are necessary in order to provide antagonistic movements due to the unidirectional contracting. The ASSIST system has a special type of PAM with rotary pneumatic actuators that allows bending movements [146].

A total of four systems using hydraulic actuators were identified in this survey. All four of them are not standard and use actuators developed specially for that purpose. Reasons to evade industrial hydraulic actuators include weight, impedance, fluid leakages and difficulties to provide fluid. Large, noisy systems are usually necessary for that purpose. Mono-and bi-articular types of Hydraulic Bilateral Servo Actuators (HBSAs) are used in the wheelchair-mounted exoskeleton proposed by Umemura, et al. [152]. Miniaturized and flexible fluidic actuators (FFA) were applied in the elbow orthosis proposed by Pylatiuk, et al. [153]. Hydraulic SEAs are used in two other systems: the Dampace system [154] is equipped with powered hydraulic disk brakes; the Limpact system [155], developed by the same group, uses an active rotational Hydro-Elastic Actuator (rHEA).

In passive systems, it is often desired to modify the amount of resistance during the exercise. This modification increases the resistance when the patient departs from the desired trajectory or to provide haptic feedback for VR interactions. In existing systems, different solutions for provision of adjustable resistive force have been implemented. Powered hydraulic brakes, for example, controlled by electromotors in a SEA are used in Dampace system [154]. Magnetic particle brakes are used in ARM Guide[136] (Figure 1A), in its successor ARC-MIME[137] to resist other than longitudinal movements of the forearm, and in the device for training of multi-finger twist motion proposed by Scherer, et al. [156]. A few groups have also investigated the application of brakes incorporating magnetorheological (MRF brakes) and electrorheological fluids (ERF brakes). These fluids change their rheological properties (i.e. viscosity) depending on the applied magnetic or electric field, respectively. Thanks to those properties it is possible to achieve brakes with high-performance (with rapid and repeatable brake torque) [105]. MRF brakes are used in MRAGES[157] and MEM-MRB[104] systems. ERF brakes are used in PLEMO[105] and MR_CHIROD v.2[158] systems. The same group that developed the PLEMO also proposed ERF clutches to control the force provided by an electric motor in active systems. Such an actuation system was implemented in EMUL[159], Robotherapist[160] and Hybrid-PLEMO[135] devices.

The natural actuators of body muscles can be used instead of external actuators. For this purpose, an electrical stimulation of the muscles leading to their contraction can be applied. This specific electrical stimulation is known as Functional Electrical Stimulation (FES). FES significantly reduces the weight of the device. From a therapeutic point of view, FES allows patients to exercise muscles, improving muscle bulk and strength and preventing muscular atrophy [161]. It has been also shown that FES, complemented by conventional physiotherapy, may enhance the rehabilitation outcome [162]. However, FES may cause strong involuntary muscle contractions and can be painful for patient. Furthermore, it is difficult to control movements using FES because of the non-linear force characteristic of contracting muscles, muscles fatigue and dependency of the achieved contraction on the quality of the contact between stimulating electrodes and the body tissue. There are two commercial systems using FES for upper limb rehabilitation: Ness H200 (Bioness, Inc., US) and NeuroMove (Zynex Medical, Inc., US). Two other systems combining FES with assistive force were proposed by Freeman, et al. [163] and Li, et al. [164].

When selecting actuators, it is also important to consider their location, especially with exoskeleton-based mechanical structures. The actuators can be placed distally, close to the joints on which they actuate (e.g. ArmeoPower system, Figure 1D). This specification simplifies the power transmission by using direct drives. However, it increases the weight of the distal part of the device and inertia and makes it more difficult to control the system. On the other hand, locating the actuators in the proximal part of the device, often in the part that remains constrained, reduces the weight and inertia of the distal part. However, a power transmission mechanism complicates the mechanical structure and may lead to difficulties with control due to friction. For example, the same group who developed InMotion HAND system proposed an earlier prototype of the hand module with eight active DOF and cable-driven mechanism for power transmission. The friction in that mechanism and its level of complexity was too high for clinical scenarios [165]. Nevertheless, there are systems, in which power transmission using cables and gear drives was successfully applied, like for example CADEN-7[97] and SUEFUL-7[166].

Control signals

Some systems use surface electromyography (sEMG) as an input signal, which provides information about intention of the person to perform particular movement. Therefore it is possible to detect and support it. Most of such systems support elbow movements, as sEMG signals from muscles controlling this joints (i.e. biceps brachii or triceps brachii) are relatively easily measured. Among proposed solutions are both stationary (e.g. systems proposed by Rosen, et al. [169] and Kiguchi, et al. [170]) and portable systems (e.g. systems proposed by Ögce and Özyalçin [171] and Pylatiuk, et al. [153]). So far the most successful of those systems is the one DOF portable orthosis developed at the Massachusetts Institute of Technology (MIT), Cambridge, US [172]. The system successfully sustained clinical trials, received FDA approval and was commercialized as Myomo e100 system (Myomo, Inc.) [173]. Examples of sEMG-controlled systems supporting movements of other joints include those proposed by Kiguchi, et al. [114] for shoulder rehabilitation, W-EXOS[174] for forearm and wrist rehabilitation, PolyJbot[175] for wrist rehabilitation, SUEFUL-7[166] exoskeleton for whole limb (excluding fingers) movement assistance, TU Berlin Hand Exoskeleton[117] for fingers rehabilitation, as well as 11-DOF portable orthosis for grasp assistance proposed by Hasegawa, et al. [98]. The sEMG signals from the contralateral healthy limb have been also used to control movements of the affected one (see system proposed by Li, et al. [176]). The concept of using movements of the not affected limb to control motion of the affected one has been also implemented in Bi-Manu-Track system (Reha-Stim, Germany), ARMOR exoskeleton [177] and device proposed by Kawasaki, et al. [178]. Using the other limb to control the affected one is especially useful during rehabilitation after stroke. In cases of hemiparesis (or hemiplegia), often only one side of the body is affected.

In some systems also contact-less movement detection methods have been used. For example, Ding, et al. [179] proposed a system to assist the load of arbitrary selected muscles using motion capture systems in order to calculate the actual muscle force.

Feedback to the user

Different types of feedback may be provided to the user, among them: visual, tactile, audio and in the form of electrical stimulation. Some systems, for example those proposed by Lam, et al. [180] and Nathan, et al. [167], use vibrational stimulation of the muscle tendons to support their contraction. It was also suggested that providing tactile feedback to flexor and extensor surfaces of the skin at the appropriate location could produce more naturalistic movements and improve clinical outcomes [3]. Some other systems combine other types of the feedback, for example system proposed by Casellato, et al. [181] combines visual and haptic feedback to improve motor performance of children with dystonia.

A significant number of systems provide training in virtual reality (VR) scenarios. VR provides a much more interesting training surrounding to the patient, compared to the typically available conditions presented in therapeutic units. Furthermore, it allows for fast modification of training scenarios, increasing patient attention and motivation to perform the exercise. Therefore it may also improve positive outcome of the therapy. It also adapts the system for various patients in a very short time frame and restarts the task if the object was dropped or misplaced. Haptic devices are especially well suited for provision of therapy in VR because they provide an impression of manipulating the virtual objects. Some groups developed own versions of haptic systems. For example Takahashi, et al. [182] proposed haptic device for arm rehabilitation, which can apply multiple types of force including resistance, assistance, elasticity, viscosity and friction. Other examples are: a two DOF Haptic Interface for Finger Exercise (HAFI), which provides rehabilitation of only one finger at a time [183]; a force reflecting glove, named MRAGES, using magneto-rheological fluid [157]; MR_CHIROD v.2, a one DOF grasp exercise device for functional magnetic resonance imaging [158] and force-feedback glove Rutgers Master II-ND[184], developed at the Rutgers University (Piscataway, US) and used in hand therapy scenarios (e.g. [185–187]).

Many groups have investigated application of a few of commercial haptic devices for rehabilitation of upper extremity. Among such haptic interfaces are: - HapticMaster incorporated for example in Gentle/S[121] (for other examples see Table 1),- in-parallel robots Phantom Omni and Premium (Geomagic, Inc., US) - used e.g. in experiments performed by Casellato, et al. [181], Brewer, et al. [188], and Xydas and Louca [189], - parallel robot Falcon (Novint Technology, Inc., US) - used in My Scrivener system for hand writing training (Obslap Research LLC, US) [190],- force-feedback glove CyberGrasp (CyberGlove Systems LLC,US) - used among others in therapeutic scenarios investigated by Adamovich, et al. [191, 192].

Because the entertainment industry have recently introduced many new devices to capture motion of the healthy people for interaction with VR-based games, it may be expected that soon some of those devices will be also adapted for rehabilitation purposes, providing so called “serious games”.

Control strategy

Clinical evidence

Clinical studies

As previously discussed, there has been a significant effort during last two decades to improve the design and control strategy of robotic rehabilitation devices. Yet, less has been done to prove the efficacy of such systems in rehabilitation settings. Although the results of clinical evaluation of therapy applying robots are still sparse, the problem is slowly being recognized. The focus in rehabilitation robotics is starting to move from technical laboratories to clinics. References to clinical trials in which robotic rehabilitation devices have been used are provided in the last column of Table 1. The classification of clinical trials used in this review is summarized in Table 7.

Term	Description
Category 0	Initial feasibility studies: Trials performed with low number of healthy volunteers, often using the prototype of a device, in order to evaluate its safety and clinical feasibility.
Category I	Pilot Consideration-of-Concept studies: Clinical trials aimed at testing device safety, clinical feasibility and potential benefit. They are performed in a small population of subjects suffering from the target disease. There is either no control group in the trial, or healthy subjects are used as control.
Category II	Development-of-Concept studies: Clinical studies aiming at verification of device efficacy. Include a standardized description of the intervention, a control group, randomization and blinded outcome assessment.
Category III/IV	Demonstration-of-Concept-Studies/Proof-of-Concept studies: Further evaluation of the device efficacy. Similar to category II, however, usually these are multicentered studies with high number of participants.

This classification is based on guidelines provided by Lo [198] and supplemented by Category 0.

From the developer and manufacturer’s point of view, there may be at least three objectives in performing clinical trials. The first one is to address regulatory requirements. The devices described in this review are considered medical devices in most countries and as such the studies proving device efficacy and safety may be required before they are authorized for distribution. Although in some cases the exemption from providing the clinical data may be granted, e.g. if the device is recognized as low risk (Class 1 device in the European Union and the USA) or if equivalent device has been already approved for commercialization, the clinical data may be required by health insurance authorities in order to provide reimbursement. In this case the objective of the trial is to obtain a proof of clinical or financial benefit of the use of the device as compared to the existing modes of therapy. The third objective of clinical trials is to provide the professional community with the clinical evidence of device’s efficacy. Although, the three objectives may seem to be similar, the requirements are not the same, therefore when designing a clinical trial it should be considered if the obtained results will allow to satisfy requirements of those three objectives. For the study design requirements to satisfy the marketing and reimbursement objectives, we refer the readers to the legal regulations in the country of interest. Whereas, for a review on the process to design a clinical train with sound scientific results we refer to Lo [198].

From the clinical point of view, the objective of the clinical study may be different than to validate a particular device. For therapists the robotic device is a tool that provides a therapy protocol rather than an end product [198], thus they are rather interested in responses to questions concerning optimal training intensity, disorders for which such form of training may be beneficial, and whether robotic therapy should substitute or complement other forms of therapy.

This survey includes a search into the US Clinical Trials database (http://clinicaltrials.gov/) from October 2013 using a combination of keywords: robotic, robot, therapy and rehabilitation. This is an web-based database existing since 1997 and maintained by the US National Library of Medicine at the National Institutes of Health. Under the American Food and Drug Administration Amendments Act (FDAAA) of 2007 all the applicable clinical trials (what concerns category II and III/IV studies in our survey) performed in the USA and starting after 2007 have to be registered in this database. However, it includes also some category I studies and many other studies performed in other countries. Results of this search identified 197 clinical trials out of which 62 are relevant to this survey. The selected trials are divided into two categories. The main objective of the first category is to proof the efficacy or safety of the device, therefore there was either no control group or a control group was undergoing the standard form of the therapy. The main objective of the second category is to determine a more efficient form of the therapy. In the latter category, the participants were assigned to groups undergoing similar forms of therapy, but at different intensities, using various devices or undergoing various forms of therapy in different order. A total of 31 studies aim at device safety or efficacy validation while 27 address better forms of therapy. A total of four trials were excluded. The objective of these trials was to validate other forms of therapy; devices described in this review have only been used as a reference. As indicated in Figure 3, the number of participants enrolled in studies for therapy improvement significantly increased during last three years compared to the number of participants in the device safety/efficacy validation studies. This suggests that the objective of the studies changes from validating forms of therapy to finding optimal applications methods. This survey identified a total of 21 devices out of the 58 clinical studies. However, it was not possible to determine the robotic device in 11 studies. Surprisingly, almost only stroke survivors (54 studies) were enrolled. In the remaining four studies subjects with cerebral palsy, spinal cord injury, traumatic brain injury and rotator cuff tear were involved.