The versatility and complexity of arm and hand movements with unique functions such as unimanual reaching, grasping and manipulation, as well as bimanual separate and cooperative movements, differ fundamentally from stepping movements with a more automatic movement control. Skilled hand and finger movements reflect cultural achievements in the evolution [
59] that are associated with a specific cortico-motoneuronal control [
60], i.e. direct projections from the cortex to motoneurons in the spinal cord which innervate arm/hand muscles. As a result, arm, and especially distal hand function are often severely impaired following CNS damage, greatly limiting patients in their ability to perform ADL [
61]. The severity of impairment and, consequently, any recovery of function is related to the extent of damage of the corticospinal system [
62,
63]. Functional training approaches and, consequently, devices supporting unilateral arm and hand movements [
64,
65] should thus be directed towards the abilities patients require for ADL, i.e., most importantly unimanual and bimanual reach and grasp tasks [
66]. Furthermore compensatory approaches and assistive devices have to be considered for more severely impaired patients.
Neurophysiological factors influencing the recovery of upper limb function
In general, the recovery of arm/hand function following CNS damage is limited when compared to gait in post-stroke [
41] and cervical spinal cord injured [
67] subjects, even if intensive therapy is applied. In patients with a cervical SCI, arm function depends on the level of the lesion. An injury level at C5/C6 results in a ‘passive’ hand function (supination movement at the elbow joint for hand opening) or, frequently, at C6/7, in a tenodesis grasp. This grasp is defined as a hand function when some forearm extensor muscle activation is preserved [
68]. It allows to close the hand by wrist extension movements with the fingers in a slightly flexed contracture position. Some spastic muscle tone is required to perform such simple grasp movements [
24].
In post-stroke subjects, outcome of upper limb function critically depends on the integrity of the corticospinal tract (CST) [
63,
69]. A stroke with damage to the CST results in lasting impairment of hand and finger function and an unbalanced muscle tone with forearm flexor hypertonia and extensor weakness that contributes to the inability to perform finger extension and hand opening movements [
60]. These patients also suffer from difficulties in the grasping and manipulation of objects, while some proximal arm function is usually preserved. Most reports show that in patients with damage to the CST, even with intensive rehabilitation measures, little recovery [
28,
30], particularly of hand and finger function [
70], can be expected.
In contrast, the recovery in patients with an intact CST is proportional to the initial impairment, with patients recovering approximately 70–80% of the initial impairment (proportional recovery rule) [
28‐
30]. Some studies indicate that training effects in these patients are small or absent [
46], i.e. only a minor dose-response effects occur [
44]. However, there is also evidence that a higher dose of practice, especially when applied early after a stroke, leads to a better outcome of motor function of the paretic arm [
41,
43,
71].
Early after stroke flaccid arm muscle paresis prevails, i.e. the limbs are weak and do not resist passive displacement. With the development of some spastic muscle tone, needed to perform rudimentary grips, the training of residual muscle function can be initiated [
24]. In this stage, the focus of therapy/training should be directed to enable the execution of simple reach and grasp movements. In the weeks following stroke, spastic muscle tone usually becomes more pronounced in the forearm flexor than in the extensor muscles, as the antigravity muscles have more muscle mass [
39,
72]. This can again impair the execution of functional reach and grasp movements. However, some spastic muscle tone in the forearm muscles allows the performance of a tenodesis grasp, which is important for the execution of ADL, not only in SCI but also in post-stroke subjects.
Patients typically compensate for their sensorimotor deficits through the involvement of the non-paretic arm/hand, leading to learned non-use of the paretic arm [
64,
73]. Therefore, one important approach to rehabilitate hand function after stroke was presented in the form of constraint-induced movement therapy (CIMT). This was based on the idea of enhancing recovery of function by reducing interhemispheric inhibition of the stroke hemisphere [
74]. By immobilizing the non-affected hand the patient is forced to use the paretic hand/arm for the performance of ADL [
64]. However so far, a superior effect of CIMT compared to other therapy approaches was not reported [
75].
During the course of upper limb rehabilitation, the support provided should always be kept to a minimum in order to make the training challenging with a maximum of individual effort and contribution to movement performance by the patient (for review [
24]). However, the optimal level of assistance also depends on the severity of impairment [
70]. Most stroke patients will benefit from gravity support, allowing them to perform functional movements by their own effort [
76]. Without such support, shoulder abduction, which is important for object manipulation, may limit elbow extension and result in concurrent elbow, wrist and finger flexion, i.e. so-called flexion synergies after stroke [
77]. This can affect the execution of functional hand movements.
Many upper limb movements involve the use of both hands. However, only a few studies provide a neurophysiological basis for the training of bimanual movements [
78]. Bimanual training of reaching and grasping tasks in stroke patients has been suggested to be more effective in improving unilateral execution of these tasks with the affected arm than unilateral training alone [
79]. This might be a result of stronger recruitment of the contralesional hemisphere through bilateral compared to unilateral training [
80]. However, there is currently no clear evidence that bimanual training is superior to CIMT [
65,
81,
82], or unconstrained unimanual training [
83].
The involvement of the unaffected hemisphere in movement control of the paretic hand might be even stronger in a special type of bimanual movement, where one hand supports the action of the other one by generating equal but opposed forces/torques, e.g. when opening a bottle or cutting bread. Such cooperative hand movements are based on a task-specific control: a ‘neural coupling’ of the hemispheres, i.e. both ipsi- and contralateral hemispheres become involved in the control of each of the two hands during cooperative hand movements [
84]. Consequently, in post-stroke patients during the training of cooperative hand tasks, the unaffected hemisphere supports movements of the paretic hand and arm [
85]. However, the effect of a cooperative training on the outcome of hand function remains to be determined.
Finally, while the recovery of finger function is limited, basic functions such as opening and closing the hand should also be trained, as most of the interaction with the environment during ADL involves grasping and releasing objects. Besides motor function, somatosensory function is also of importance during object grasping: shaping and maintaining a stable grasp during the manipulation of an object relies on the processing of somatosensory input, determined by the mechanical properties of the manipulated object [
86]. Somatosensory function is often impaired after CNS damage, leading to a visual compensation of movement control. However, in some patients it can recover spontaneously or through dedicated training [
87].
Implications for robot-assisted therapy of upper limb function
The combination of kinematic complexity and functional impairment makes the design of robotic devices to train arm, hand and finger function after CNS damage particularly challenging. Following the initial developments based on stiff industrial manipulators, end-effector-based devices for planar (MIT-MANUS; [
88]) and 3D (Gentle/S, [
89]) reaching movements were introduced to allow more active contribution of the patient while limiting the apparent impedance of the robot. Subsequent developments focused on incorporating additional degrees of freedom (DOF) related to wrist [
90] and hand opening/closing function (Gentle/G [
91]). For the functional training of three-dimensional arm movements with guidance at the three proximal joints, ARMin, a grounded, powered exoskeleton was developed, which also integrates grasp and release function [
92,
93] (Fig.
2, upper panel).
Independent of their kinematic configuration, all of these systems can partially or fully unload the arm against gravity. This approach reduces the effect of flexor synergies, and allows the performance of hand movements within a larger workspace. However, the complex structure and geared actuators of such devices with their reflected inertia limit the interaction quality and the ability to adapt the level of support [
16]. The large output impedance may render the active initiation of movements more difficult, and potentially alter natural movement dynamics. Therefore, a trade-off between the number of DOF and the quality of the physical interaction exists, limiting the application of these devices to specific stages of recovery. For example, training with a powered whole-arm exoskeleton is mainly indicated for stroke subjects with severe arm paresis early after the incident. Similar effects can also be achieved by using passive devices for gravity support to the upper limb, to enable self-initiated movements [
94].
Robot-assisted approaches should also consider the training of bimanual and cooperative movement tasks as they are important during ADL (Fig.
2, lower panel). Bimanual training was a focus of some early studies [
10], but its potential has not been sufficiently explored and deserves further investigation. Many upper extremity systems developed and clinically evaluated so far could also be used for bimanual training, by combining two devices in a mirrored configuration. The training of cooperative hand movements (e.g. opening a bottle) has been proposed using a dedicated device [
84], and can also be achieved by virtually coupling two unimanual devices through control.
Due to the biomechanical and neural complexity of hand and finger movements, robot-assisted rehabilitation of hand and finger function became a focus only recently. Most rehabilitation robots for hand function have been based on end-effector designs, used either independently or in combination with grounded exoskeletons or end-effector type arm devices (Fig.
2). Several groups have also made attempts to develop exoskeleton systems for the hand, some of which assist independent finger motion, generally resulting in highly complex devices that underwent none to little clinical evaluation. A review [
19] found that only 25% of 30 hand rehabilitation robots had been clinically tested, and many devices had been considered too complex for clinical use. However, such complexity might not be necessary when the focus is directed to the basic function of opening and closing the hand [
95]. This might be sufficient given the limited potential for the recovery of finger function following CNS damage, while remaining highly relevant for ADL. Finally, hand opening/closing can also be supported through wearable assistive technology, such as soft robotic gloves [
96,
97], which could be worn during the performance of ADL.
Interaction with the environment occurs mainly through the hands and generates somatosensory feedback. However, somatosensory function is often impaired after CNS damage. Therefore, neurorehabilitation devices for the upper extremity should train hand and, as far as possible, finger function, providing both visual and haptic feedback [
53]. Training should include tasks which are functionally relevant for ADL, such as grasping and releasing objects with rendered virtual dynamics to also train somatosensory function and sensorimotor integration [
98]. Finally, most upper limb training devices are embedded in computer games to reflect the cognitive nature of these tasks and motivate patients. In a meta-analysis, the application of virtual reality (VR) games was found to be potentially useful for the improvement of arm function after stroke [
34].
In conclusion, a good, mainly spontaneous, recovery of upper limb function after a stroke can be expected when the integrity of the CST is preserved. There is some evidence that higher dose of practice leads to improved function, especially early after stroke. Nevertheless, in cases with damaged CST the recovery is limited and neither depends on the approach nor on the dose of training. Unimanual robot-assisted therapy approaches should be complemented by bimanual (cooperative) approaches. These should also incorporate the training of basic hand function and interaction with virtual object dynamics that generate somatosensory feedback. In the future, it will be possible to at least partially compensate for remaining deficits with wearable assistive robotics.