Control strategies for active lower extremity prosthetics and orthotics: a review

Supraspinal neural activity

Recalling that motor intentions originate at the cortical level, several groups have investigated methods for triggering the device to provide assistance through Brain-Computer Interfaces (BCI) [45]. Recording of activity at this level has the potential to allow for a wide-variety of volitional movements, however, these may be difficult to decipher given that the brain is concurrently responsible for a multitude of tasks, including the control of the other limbs. In addition, many of the control loops responsible for physiological locomotion take place at the spinal level via reflex arcs (Figure 2), which may fundamentally preclude the use of neural activity to directly control the legs while maintaining balance during a dynamic task. However, there may still be utility in using brain activity to provide high-level commands to the device, which it will then execute (as in the shared control context promoted in [45–47] and demonstrated in [48, 49]).

Functional near-infrared spectroscopy (fNIRS) uses optical light emitters and receivers placed on the scalp to sense the haemodynamic response of the brain, which correlates with brain activity. This modality is subject to non-specific brain activity, motion artifacts, significant haemodynamic delay, and requires that optodes be worn on the head. Even so, a recent pilot study investigated the use of an fNIRS-BCI to detect the preparation for movement of the hip in seated stroke subjects, which may indicate its suitability in shared control with severely impaired subjects [50].

Electroencephalography (EEG) uses an array of surface electrodes to non-invasively record the electrical activity of the brain as evident on the scalp [45]. The EEG electrode arrays typically used in research are built into a snug-fitting skull cap that can be extremely difficult and time-consuming to put on by oneself, especially for the patient groups whose injuries would necessitate direct cortical input to the P/O controller. This supposedly could be countered with advancements in self-contained EEG headsets designed for consumer use. The electrodes can be either dry or wet, depending on whether an electrically conductive gel is required. Signals recorded via EEG can encode a wide variety of movements with high temporal resolution.

In practice, the use of EEG signals demands a high level of focus and concentration from the user and is susceptible to movement artifacts, autonomic neural activity and electrical noise. Use in real-world environments is further complicated by the presence of distractions and the performance of tasks that are unrelated to locomotion. However, EEG signals could be combined with other sensory inputs in the framework of the so-called hybrid BCIs [51, 52] in order to decode user’s high-level commands more reliably.

Environmental sensing (see section below) can add an additional layer of safety in the context of shared control with BCIs, as the controller may prevent certain movements due to the presence of obstacles [47]. For example, prior to executing a high-level command (e.g. go forward, turn left), the controller would check first whether there are any terrain features in the way. Similarly, the execution of the high-level command “sit down” would not require the user to align perfectly with the chair, but would rely on the controller’s ability to compensate for the misalignment. As these examples illustrate, shared control reduces cognitive workload, as the user does not need to care about the mid-to-low-level execution over long periods of time or during critical operations.

Implanted electrode arrays within the motor cortex enable measurements which may encode a wide variety of movements, with the noted downside of requiring a highly invasive (and still experimental) surgical procedure [15, 53, 54]. Such an interface may also be used to provide sensory feedback to the user, thus closing the sensory-motor control loop [54]. Intracortical electrode arrays have been successfully demonstrated to allow control of multi-degree-of-freedom reach and grasp movements with robotic arms in tetraplegic subjects [55, 56], though to date there are no known examples of cortically-implanted electrodes being used to control a lower limb device in humans. Similar experiments have been done, however, in rhesus macaques to demonstrate the prediction of leg movements to control of bipedal gait in a humanoid robot [57]. It remains to be demonstrated how well this technique would translate to the control of a wearable P/O device.

Peripheral neural activity

The closer that neural activity can be recorded to the innervated muscle, the more specific the motor commands become. Also interesting is the electromechanical delay between the motor commands and the generation of force in the muscle on the order of 10s of milliseconds [58], which would provide a significant head-start to a controller based on muscle activity over one based on mechanical feedback alone [59]. This delay, however, may also be a source of instability when a device with a faster control loop is coupled to the user to provide high levels of assistance [60].

These peripheral nerve signals can be sensed through the use of electromyography (EMG). Surface EMG is the least invasive technique, where electrodes are placed on the skin over the muscle belly of interest. Assuming that the musculature remains somewhat constant and that the device can be fastened to the body in a consistent manner, it may be possible to embed the electrodes within the human-robot physical interface, thus significantly reducing the amount of time required to don and doff the device [61, 62]. Surface EMG activity is susceptible to changes in electrode-skin conductivity, motion artifacts, misalignment of the electrodes, fatigue, and cross-talk between nearby muscles [60, 61, 63]. Myoelectric signals are also non-stationary in nature during a dynamic activity, which necessitates the use of pattern recognition techniques [64]. In practical use, a calibration routine is typically necessary each time the device is put on [60, 65].

In the event that a limb has been amputated, the residual neuromusculoskeletal stucture must be surgically stabilized. Depending on the location of the injury, the muscles responsible for the actuation of the amputated joints may still be present and natively innervated, albeit relocated and fixed to the bones in a non-physiological manner. In this case, it may be possible to record the EMG signals in the residual leg for the control of a particular joint (e.g. using muscles in the lower leg to control the ankle [66]). If the amputation is more proximally located (e.g. above the knee), the muscles to control the distal joint (e.g. the ankle) are altogether missing, and thus can not be used directly. However, given that the nerves that would normally control these muscles are still present, a technique called “targeted muscle reinnervation” (TMR) can be used [64, 67]. For TMR, the severed nerves are surgically reattached and allowed to reinnervate a foreign muscle, which can then be used as an EMG recording site for the amputated muscle. The reinnervated muscle acts as a “biological amplifier” for the severed nerve and provides a means to record its activity noninvasively via surface electrodes.

Joint torques and positions

Mechanomyography (MMG) can be used to estimate the force production in muscle by measuring the sound or vibrations evident on the surface of the skin using microphones or accelerometers [68]. A potential advantage of MMG over EMG is that the muscle force estimated through MMG is less sensitive to fatigue [69]. Force production can also be estimated via changes in muscle hardness [70, 71] and the volume of the muscle [72, 73]. A substantial downside to all of these approaches is their high sensitivity to motion artifacts, which may be significant given the nature of the physical coupling at the user-device interface.

Joint torques can be estimated via inverse dynamics provided measurements of the joint positions and external forces being applied to the limbs. Wearable sensors for estimating joint positions or limb segment orientations are summarized in [74] and include goniometers, inclinometers, accelerometers, gyroscopes, magnetometers, and inertial measurement units (IMUs). Ground reaction forces can be sensed using instrumented insoles worn under the foot (reviewed in [75]) or e.g. by measuring the load in the shank of a prosthesis. A variety of foot switches can also be used to deliver binary ground contact information, for example using force-sensitive resistors, sensed air pressure in a sealed tube under the foot, or a physical switch.

Furthermore, interaction forces can be measured at the physical interface between the user and the device. Useful sensors may include load cells, strain gages, pressure sensors, and force-sensitive resistors.

Alternative input modalities

Simple manual inputs (e.g. keypads, buttons or joysticks) may be effective even though the used signals are completely artificial [76, 77]. Voice commands or eye movements sequences have also been demonstrated as possible ways to interact with P/O devices [78–80]. Here again, the seamlessness and intuitiveness of these input methods are suboptimal, but they can represent viable alternatives when no input other methods are possible.

Activity mode recognition

Activity mode recognition is what enables the high-level controller to switch between mid-level controllers that are appropriate for different locomotive tasks, such as level walking, stair ascent, and standing. The cyclic nature and long-term repeatability of various modes of gait lend themselves to automated pattern recognition techniques for classification. The inputs to the classifier include the sensed states of the user, the environment, and of the device. Important considerations for choosing a classifier include the number of activities from which to choose, the procedure required for training, its error rate in real-world conditions, signals that are required as an input, and the classification latency i.e. the time required by the classifier to reach a decision.

As useful definitions, Huang et al. coined the term critical time to describe the time by which a classification decision must be reached to ensure proper kinematic and kinetic transitioning between modes [59]. Thus, the classification latency must be shorter than the critical time to execute a proper transition. The critical time is an especially important constraint when transitioning between activity modes with substantially different characteristics, for example level walking to stair ascent, where excessive latency may cause a loss-of-balance. In subsequent work, Zhang et al. use the term critical error to describe any error that results in the subjective feeling of unstable balance [121]. This definition emphasizes not only that a loss-of-balance is to be avoided, but that the user must also feel secure with the performance of the device.

First, different types of classifiers that have been used for activity mode recognition will be discussed, then the sources of information that have been used as inputs to these classifiers will be presented. For additional information related to these topics, see the review of Novak and Riener [122] on sensor fusion methods in wearable robotics.

Direct volitional control

Volitional control grants the user the ability to voluntarily modulate the device’s state. Such functionality is especially important in scenarios where the locomotive activity is irregular or noncyclic (e.g. walking in a crowd or standing and shuffling), in situations where foot placement is critical (e.g. stair descent, walking on rough terrain), and during nonlocomotive activities (e.g. repositioning legs while sitting, bouncing a child on one’s knee). It is emphasized for consistency that, while the volitional intent is determined at the high level, the conversion to a desired device state occurs at the mid level.

Myoelectric signals are an intuitive approach to volitional control since they are already present during voluntary movement of the user’s own limbs. Sensing of peripheral neural activity for control does come with limitations, as were highlighted in the section on sensor modalities for human motion intentions. Surface EMG has been demonstrated for this purpose in transfemoral prostheses [61, 120, 130, 132, 149, 150], virtual above- and below-knee prostheses [151], a hip and knee orthosis [152], knee orthoses [60, 153, 154], a transtibial prosthesis [66] and an ankle-foot orthosis [155].

EMG-based control approaches differ in the way that the myoelectric signals recorded from the various muscle groups are mapped to the desired device state. The simplest approach is to directly modulate the actuator’s torque based on EMG activity [63, 152, 155]. A more complex approach uses a neuromusculoskeletal model to calculate net joint torques from the EMG signals of joint flexor and extensor muscles [60, 149, 154, 156]. One can also map processed EMG signals to a desired joint position, velocity, or acceleration by using a model of the coupled user-device system [153] or to the set-point angle or stiffness of an impedance control law [61, 66, 130, 132, 151].

It is also possible to use the EMG signals to contribute an additional flexor or extensor torque to the nominal torque output by a mid-level controller. This was demonstrated to allow stair ascent in a transfemoral amputee with a powered knee prosthesis [150]. This approach combines the inherent stability of the underlying controller (e.g. in the absence of any myoelectric input) while providing moderate levels of volitional control to the user.

As the user acclimates to and is able to predict the output behavior of the powered assistive device, it may be possible for him to volitionally manipulate the device by providing the appropriate set of inputs, possibly involving contrived or compensatory movements. This is likely true for mid-level controllers based on correlated postures [157] and invariant trajectories [158, 159] as discussed in the following section, though long-term studies would be required to show that users can learn to control the device in this manner.

Phase-based controllers

In time-based control, a set of actions is performed based on a programmed time delay following a clearly identifiable gait event, for example heel strike [160, 161] or toe-off [162]. This technique is simple to realize and relies heavily on the regularity of the steady-state stepping period. As such, the weakness inherent to time-based control is its inflexibility to accommodate irregular or unprogrammed gait patterns (e.g. walking in a crowd or over rocky ground), unexpected events such as tripping, or within-cycle switching of activity modes [163].

Invariant trajectories represent user states which vary with respect to the gait phase, and do not change substantially with respect to gait speed, between different users, or with minor intra-gait-mode variations [164, 165]. By projecting a set of invariant trajectories onto judiciously chosen axes, one can derive an invertible relationship between the user’s state and the gait phase that is ideally independent of time, gait speed, or subject. While this technique does not represent a controller by itself, it can be used to input phase information to a mid-level controller without relying on the elapsed time between gait events, as with time-based methods. As such, walking backwards is also possible with this technique without altering the controller[158, 159]. It remains to be shown, however, whether these trajectories remain invariant when the user dons an actuated assistive device, across different modes of gait, or during pathological gait.

Normalized-trajectory control takes a prototypical joint trajectory from a set of previously recorded gait data and scales it to match the pace and physical size of the user as a function of gait phase. So-called “dynamic pace control” is one such example that uses the fast Fourier transform (FFT) to represent the prototypical trajectory by a set of Fourier coefficients. These can be used to scale and generate a desired device trajectory by taking the inverse transform [158]. One of the biggest challenges with this approach is identifying trajectories that can be appropriately scaled to both the speed and weight of the user. The output of the control law is typically a desired position.

Echo control is a combination of time-based and normalized-trajectory control, in which the position trajectories of the unassisted limb are recorded and replayed on the assisted limb with some time delay and scaling during steady-state reciprocal gait [114, 166, 167]. During certain activities (e.g. sit-to-stand transitions), no phase shift is required and the movement of the sound limb can be mapped directly to the impaired one. Such an approach assumes symmetry of movement between the two sides, and as such is inappropriate in cases where bilateral assistance is required or where the required stepping pattern is inherently asymmetric. Activities involving an odd number of steps can also be problematic. Note also that any undesired/compensatory movements that are recorded will also be replayed, which may result in instability or loss of balance. Gait mode switching can only be achieved at the beginning of a stride, and must be initiated with the unassisted limb.

Virtual constraint control is a strategy that has been used to control locomotion in bipedal robots [168], which has since been implemented and demonstrated with a powered knee and ankle prosthesis [159, 169, 170]. For this method, the anterior-posterior location of the CoP on the prosthetic foot is used as a phase variable, which is possible given that this trajectory monotonically increases throughout the stance phase of level walking. The so-called “effective shape” of the ankle-foot and knee-ankle-foot complex resembles a circular rocker when plotted against the CoP phase variable that is invariant with respect to speed, heel height, and body weight [165]. These effective shapes constitute virtual constraints that can be enforced through actuation of the device during stance, while a finite-state impedance controller (described below) was used during swing. The choice of a different phase variable may enable swing-phase virtual constraint control as well. As implemented, this technique enables walking patterns that are qualitatively similar to nominal ones using only generic, normalized shape parameters from literature (i.e. subject-specific tuning is not required).

Finite-state controllers (FSCs) decompose gait as a periodic activity that is described by a series of distinct phases, typically further delineated on the basis of foot contact events or joint velocities, as illustrated in Figure 3 and elaborated in the caption. A FSC implements a discrete set of parametric control laws that will cycle through as each new phase of gait is entered. These control laws differ in the way that the desired state of the device is computed, with popular choices being position and impedance control, as will be discussed below. A different FSC is required for each activity mode included in the high-level controller.

FSC is far-and-away the most popular mid-level control approach. Many groups have successfully implemented FSCs using different numbers of activities and states with a wide variety of devices [2, 3, 14, 24, 41, 44, 107–109, 114, 117, 118, 121, 123, 124, 134, 137–139, 143, 150, 163, 171–186].

The most of the referenced FSCs use a set of static parameters that are hard-coded into the controller, usually requiring a heuristic tuning routine involving the end-user. Given the inherent variability in the gait patterns between individuals, such tuning is likely necessary and indeed desired to optimize the user’s comfort and efficiency. However, this approach quickly becomes unwieldy, as the number of tunable variables rapidly increases with the number of parameters per control law, the number of states per activity, the number of activity modes, the number of joints to be actuated, and the number of limbs to be controlled [176].

To reduce parameter tuning time, Simon et al. documented the “modified intrinsic control strategies” to reduce the number of tunable parameters in a knee and ankle prosthesis by using impedance control laws based on joint position or load as opposed to a set of static parameters [187]. Aghasadeghi et al. demonstrated a model-based method to predict initial parameter values for a particular user based on invariant trajectories and his particular anthropomorphic characteristics [188]. Wang et al. describe the use of an expert system to automatically tune the impedance parameters to match that of healthy human gait based on fuzzy logic inference [189].

Furthermore, because these parameters are fixed in value, the P/O device may not optimally accommodate changes in the user’s gait, for example due to user fatigue, a sudden weight change, or variations within a particular gait mode. One approach to overcome this is to modulate a torque superimposed on the output of the control law through surface EMG [150]. Another approach implements a stance-phase control law based on a neuromuscular model, which was demonstrated to adapt to changes in both terrain [172] and gait speed [190] in a transtibial prosthesis.

Non-phase-based controllers

Complementary Limb Motion Estimation (CLME) infers the intended motion of affected limbs from the motion of the residual limbs, and maps this to a reference trajectory for robotic P/O joints to track [191]. This is possible due to the strong inter-joint coordination exhibited during physiological human motion [36]. The mapping is derived through regression of physiological gait recordings of healthy subjects.

CLME is sometimes confounded with echo control, as described above. By contrast, CLME complements residual body motion without delay. In addition, the recorded joints and their sensed states (e.g. position, velocity, acceleration) need not directly correspond to the controlled joint, as with echo control. For example, the motion of the arms and upper body are often highly correlated to the motion of the lower limbs, and thus represent potential recording sites for CLME-based control [192]. CLME does not depend on information of the gait phase, and no change of control is necessary between stance and swing. However, instrumentation of selected residual body parts must be provided.

CLME was first evaluated with the LOPES gait rehabilitation robot, showing that it can both replace (unilateral) leg function when needed [193], and be transparent when no assistance is needed [191]. Later, it was shown that an amputee subject can walk at different speeds and negotiate stairs [157] with a CLME-controlled actuated knee prosthesis. This study also demonstrated a qualitative reduction in the compensatory motion of the sound-side foot relative to a commercial controlled-damping prosthesis.

Recently, CLME has been extended to generate not only reference kinematics, but also stiffness. To obtain training data, active muscle stiffness during gait was quantified according to [194]. In a case study, an amputee subject successfully walked with CLME-controlled impedance [195].

In all of these studies, the training data used to generate the mapping came from a different subject than the controller was tested on. Even so, the controller was shown to be effective to restore functional gait patterns. This is an important consideration as it is often difficult to obtain nonpathological gait data from the user himself, which would contain the “ideal” joint trajectory mappings. CLME is only suitable for patients who can still control specific body parts, under the premise that these body parts are sufficiently correlated with the other limbs in physiological motion. Generalization of this method for different activities remains to be investigated.

Force-feedback control measures the interaction force between the user and the device and acts to reduce it according to an assistance ratio. Zero-force or transparency control, also given the colloquial name “get-out-of-the-way control”, is a special case of force-feedback control. It serves various practical purposes, especially with execution of non-locomotive tasks. The first purpose is for assisting the user with tasks that require force amplification while maintaining lower limb agility, as with exoskeletons for performance augmentation [11, 13, 196]. The second use would be to render the device to be transparent such that the user’s own movement is unrestricted. The renderable transparency of a device can also be used as a performance metric that is indicative of how well the physical dynamics of the device can be controlled [197–199].

Forms of control laws

The output from the mid-level controller is the desired state of the device. This state may consist of a combination of joint positions, velocities, and torques. The design of the device and the actuators it contains will have a strong influence on how well the desired state can be achieved, and thus must be taken into account during the development of the mid and low level controllers.

Commanding joint positions or velocities is a straightforward approach to robot control, for which an abundance of theory exists. In this case, the robot is tasked with the precise reproduction of a pre-defined trajectory. This type of controller works best when the output mechanical impedance (as defined below) of the actuator is high relative to the load, thus enabling the device to reject perturbations from the user or the environment. Controlling the position or velocity of the device is useful when the desired trajectory and the interaction forces are well-characterized; but both of these may be difficult to predict given the dynamic nature of locomotion. Interaction with stiff objects (e.g. the environment) can lead to instability and the generation of high forces under this type of control [200]. Presumably, it is because of these issues that very few examples of assistive devices for lower limbs were found that use position control.

The causal dual of angular velocity is torque. Torque control is possible when the output mechanical impedance of the actuator is low relative to the load, and is useful for providing assistive forces when the desired position is ill-defined or unimportant. Problems can arise, however, when torques are applied without regard to the position of the joint (e.g. uncontrolled motions, bumping into RoM limits).

Impedance is defined as the transfer function between an input flow and an output effort [201]. In the domain of mechanical rotation, controlling the impedance of a joint is to control the relationship between the accepted angular velocity and the yielded torque. Since the output of an impedance controlled joint is ultimately a torque, the output mechanical impedance of the actuator must be relatively low. A wide variety of virtual dynamics can be rendered under impedance control, with common intuitive examples being linear stiffness, damping and inertia. By applying passivity constraints to the rendered dynamics, the stability of the impedance-controlled system can be guaranteed [200, 202].

A leading theory regarding physiological movement is that the CNS controls the limbs through impedance control [203, 204]. It is partly due to property that human movements can be robust to perturbations despite the delays inherent to transmission of efferent and afferent signals via neural pathways, as explained in the section on nominal control of human locomotion. Thus, it is possible to realize a bio-inspired approach to P/O control through impedance modulation. An improved understanding of how humans control the impedance of the lower limbs is necessary to optimize the mechanical design and control of P/O devices and is the topic of ongoing biomechanics and neuroscience research [156, 194, 205–209].

Admittance is defined as the inverse of impedance, and thus also defines a relationship between effort and flow. Admittance control has been proposed as an approach to masking the undesired dynamics (e.g. added friction and inertia) imposed by an exoskeletal device [210]. This type of control has been used with treadmill-based gait orthoses [211], but no examples of admittance control implementations with a portable device were found.

It is impossible to say whether it is best to control a device’s position, velocity, torque, impedance, or admittance, though the choice may be constrained by the physical dynamics of the device and how it is actuated. There are distinct advantages and disadvantages to each type of control, and there may be different conditions and gait phases that lend themselves to switching between control modes [134, 185, 212].

Coordinated motion control and load sharing

Prosthetic and orthotic devices are part of an overall system involving physical interactions, and thus cannot operate as if in isolation. A coordinated mid-level control scheme is necessary whenever there are multiple actuated degrees of freedom, whether contained within one device [117, 170, 181, 213, 214] or distributed across multiple devices [176]. Uncoordinated motion between limbs and joints can result in loss of balance or falls, and so the states of each should be communicated and taken into account to prevent such an occurrence. Monitoring of the state of the user (e.g. ground contact state, joint positions) also provides relevant information to the controller.

It should also be noted that the joints in the human body are subject to biarticular coupling on both a mechanical and neural level [215]. As a result, the realizable voluntary torque output and range-of-motion (RoM) of the user’s joints are functions of posture, which potentially involves joints that are not specifically actuated by the device. Considering this, the mid-level controller should take into account the configuration of the user to ensure that the device does not over-power the joint nor exceed his RoM.

Recalling that an orthotic device typically acts in parallel to the user’s limb, it is possible (and indeed desirable, as per the ”Slacking Hypothesis" [20]) to make the user responsible for sharing a portion of the load carried at the joint. The fraction of the load borne by the device is termed the support ratio [60] or the assistance ratio [216], and is normalized to either the net torque required of the joint for a particular movement or to the maximum torque of the device. The load carried by the physiological joint can be estimated from muscle activity [60, 70, 154, 216] or through direct force measurement [217].

Care must also be taken that the device does not provoke unphysiological muscle activity, for example excessive cocontraction or off-nominal timing of muscle activation. Note that the net joint torque may look “normal”, even though the underlying muscular activity is not natural. This is an especially important consideration in the context of retraining gait patterns following neurological injury, and may be monitored e.g. via EMG.

In the wake of certain impairments (e.g. spinal cord injury), however, the transmission of voluntary locomotor commands is interrupted. Short of physiological recovery or bridging the communication gap with neuroprostheses [21], functional electrical stimulation (FES) can be used to actuate the user’s muscles in cooperation with an orthotic device [218, 219]. Such stimulation must be provided in concert with the user’s intentions and locomotive state and provides a means for the user to actively participate in the locomotive task with the support of the orthosis. One obstacle to every-day use of FES is the need to precisely place electrodes over the muscles, which may be difficult for the user to achieve without assistance. A mid-level controller, such as a FSC, must underly the FES to manage the movement of the device itself.