The greatest functional differences of reaction moments occur on the shank.

Although both actuator paradigms were capable of stabilising the model after tripping using the selected recovery strategy, this was not the case for both actuator locations. The presence of a reaction moment produced substantial functional differences while actuating the shank. While FM_Sk was able to successfully assist recovery from perturbations P1-P3, JM_Sk-Th was unable to prevent falling in any of the cases. In contrast, reaction moments appeared to only modestly affect actuating the thigh, wherein FM_Th, JM_Th-Tk, and JM_Th-Th yielded similar n and anterior MoS in perturbations P1-P3 and deviated only in P4. While exerting a reaction moment on the trunk appears to have a negative effect on balance recovery (FM_Th was able to stabilise P4 but JM_Th-Tk was not), exerting it instead on the contralateral thigh (JM_Th-Th) had a less obvious detriment that was discernible only by a lower MoS (Table 2 and Fig 5) and the smaller set of feasible parameters (Fig 4) in P4. However, the existence of feasible solutions for both actuator paradigms for hip flexion assistance suggests that the presence of reaction moments on the body is neither a fundamental requirement nor barrier to the assistance of balance recovery.

Assistance parameters reveal differences in model sensitivity.

In instances where actuators had a similar functional result (i.e., recovered gait), they can still be discriminated by subtle differences in the sensitivity of the model to the assistance parameters. The location and size of the contour regions of n in Fig 4 (and to a lesser extent Fig 5) indicate the sensitivity and robustness of successful stabilization to the assistance magnitude and onset time. With respect to the parameter ranges explored, there appears to be considerable tolerance to the assistance onset time—most early-mid values yielded similar results within and between perturbations and amongst all actuator types and locations (with the exceptions of P5 and JM_Sk-Th, which did not yield feasible results). The feasible moment magnitude, on the other hand, is confined to a smaller range with both lower and upper bounds (associated with insufficient and excessive anterior MoS, respectively, Fig 5), and shifts appreciably with perturbation magnitude. Compared to a free moment applied to the thigh (FM_Th), the minimum moment required to stabilise the model was 25–40% larger when a reaction moment was present on the trunk (JM_Th-Tk, Table 3) but similar when the reaction moment was instead exerted on the contralateral thigh (JM_Th-Th).

Location of reaction moments can determine whether balance recovery is feasible.

While both actuator paradigms can be useful in the appropriate context, these simulations suggest that, for the application and control scheme presented here, internal reaction moments can either catastrophically inhibit performance (shank) or decrease sensitivity to the assisting moment (thigh). However, the simulations of the thigh assistance also show that it is possible to mitigate this interference through judicial placement of the reaction moment (i.e., moving the reaction from the trunk to the contralateral thigh). Hence, care must be taken when designing an internal-reaction actuator. Prior to considering other (practical) factors pertaining to the design objective, external-reaction actuators may offer a conceptually simpler alternative for controlling body segment motion for high-level tasks like balance control.

Knee extension reaction moment induces early foot contact.

Both actuators with an action moment on the shank (FM_Sk and JM_Sk-Th) extend the knee joint (Fig 6) with the intention of bringing the foot forward with respect to the knee and increasing anterior foot placement; however when a reaction moment is simultaneously applied to the thigh (JM_Sk-Th), hip flexion is inhibited during the mid-swing phase and/or hip extension occurs prematurely during the late-swing, depending on when the assistance is initiated. This decreased flexion or increased extension lowers the knee with respect to the pelvis and, consequently, lowers also the foot and causes early ground contact. During the swing phase, the pelvis forward velocity is not substantially affected by the assistance, so early swing termination decreases step length (S1 Table) and creates a smaller anterior moment arm between the CoP and CoM, lessening the centroidal retarding moment needed to counteract the effect of the perturbation. This is evidenced by a decreased anterior MoS and number of successful steps when compared to FM_Sk in several of the perturbation conditions (Table 2, Figs 4 and 5) and the finding that JM_Sk-Th is often most effective when the assistance is too low or late to influence foot placement (Fig 5).

Incidentally, if cadence remains similar, a decreased swing time and step length due to knee joint assistance may also explain the empirical finding that unimpaired users of powered knee orthoses do not walk faster than when not wearing a device, contrary to the expectation that this would augment physical performance (although it is unclear to what extent this may also be due to the added inertia or impedance of the device) [49].

From the simulated scenarios, a conclusion cannot be drawn on whether JM_Sk-Th is an invalid mode of assistance for tripping recovery, or whether success was limited by the control scheme or moment constraints. An alternative control scheme might first flex the knee to increase toe clearance before extending the knee to increase step length.

Hip flexion reaction moment destabilises trunk posture.

All three actuators with an action moment on the thigh (FM_Th, JM_Th-Tk and JM_Th-Th) flex the hip; however, when the reaction moment is applied on the trunk, as with a traditional hip exoskeleton (JM_Th-Tk), there is also an increase seen in the forward trunk lean angle. This effect on the trunk lean angle increases as the magnitude of the assistance increases for larger perturbations. This is an important trend, since kinematics elements such as the trunk lean angle can be associated with imbalance after perturbation [33]. The fact that balance was only partially recovered after P4 by JM_Th-Tk, whereas FM_Th resulted in a full recovery, also points at the destabilising effect of the trunk reaction moment. Even if the effect on forward trunk lean does not directly cause instability, it is important to consider how the user would perceive such an effect. Among other factors, control of the trunk angular velocity is one of the critical factors in balance recovery [33], especially in elderly people [48]. Adding moments to trunk in the direction of fall is likely to hinder in a person’s own active control of their trunk angular velocity.

Hip flexion reaction moment on contralateral thigh is a good alternative.

The actuators with action moment on the thigh and external reaction moment (FM_Th) and internal reaction placed on the contralateral thigh (JM_Th-Th) show very similar performance and kinematics. The reaction moment on the contralateral thigh could promote trailing limb hip extension, which could also promote the forward walking progression. This is seen here as a small effect seen in the trailing limb kinematics (S4 Fig), namely a small extension of the hip and knee, which likely also slightly increases push-off. We therefore propose that the reaction moment placement on the contralateral thigh is a superior alternative to placing it on the trunk.

Assisting thigh or shank affects gait kinematics.

When unassisted, the perturbation causes the swing foot to contact the ground prematurely. Contacting the ground early does not allow the knee time to fully extend, which substantially shortens the step length. Actuating the thigh flexes the hip and lifts the knee, increasing foot clearance and extending the swing time with respect to the unperturbed case. This lengthened swing phase gives the knee sufficient time to fully extend, increasing anterior foot placement. On the other hand, assisting knee extension directly does not increase the swing time with respect to the unperturbed case but instead increases knee extension velocity, bringing the knee to full extension quicker (Fig 6). In the cases where both strategies successfully assist recovery, the step lengths are similar—substantially larger than the unassisted case but smaller than the unperturbed case.

In the cases where the FM action on the shank successfully assists recovery, hip extension is also indirectly affected, even though there are no moments exerted on the thigh as in JM_Sk-Th. The hip briefly extends as the knee reaches full extension, which may be the result of the reflex model reacting to increased swing leg velocity by increasing hip extension near the end of swing phase. Similar behavior is seen in the unperturbed case, as the hip extends to retract the swing leg before making foot contact. Moreover, such increase in hip extension could also be facilitated indirectly by knee joint reaction forces. When the knee is extending, a reaction force at the knee joint would extend the hip if the knee is not stiff.

Assisting hip flexion out-performs assisting knee extension.

For both actuation paradigms, actuating the thigh was more likely to result in successful tripping recovery and require a smaller magnitude of applied moment to do so. In the case of internal-reaction actuators, actuating the shank (JM_Sk-Th) was not capable of stabilising the model for the given perturbation and assistance magnitudes, while both realizations of thigh actuators (JM_Th-Tk and JM_Th-Th) performed relatively well. Similarly, external-reaction actuators at the thigh (FM_Th) required 20–30% less assistance than an identical actuator placed on the shank (FM_Sk) for perturbations P1-P3 (Table 3).

Assisting subfunctions for tripping recovery.

These simulations show that multiple strategies are feasible for assisting balance recovery following tripping—protectively increasing step length can be achieved by either increasing the rate of knee extension or prolonging the swing phase by increasing hip flexion. However, despite yielding similar step length, they produce different trajectories of the swing foot. Previous simulations indicate that augmenting knee extension before mid-swing decreases the ground clearance of the foot when it is at its highest velocity [29], which can increase the risk of tripping over obstacles or uneven terrain or result in foot scuffing over even smooth surfaces. While we have thus far assumed a single discrete tripping obstacle, it is conceivable that prematurely assisting knee extension could even impair tripping recovery if presented with further variations in surface height. Either care must be taken to ensure that knee-extending assistance is initiated only after mid-swing, or the assistance should bimodally first flex the knee to increase foot clearance before switching to extension. In either case, accurately estimating the gait phase and synchronising the assistance may be critical. In contrast, the same previous simulations showed that assisting instead hip flexion in the early swing phase both lifts and flexes the knee and increases ground clearance while simultaneously increasing step length [29]. This may constitute the more robust and safer alternative.

While these simulations show that it is possible to influence step length from either the shank (knee) or thigh (hip) independently, it may also be possible to combine these strategies. Placing multiple smaller actuators on the legs could allow the knee and hip joints to be independently manipulated but, in the context of providing only emergency assistance, the obtrusiveness of the larger number of points of contact and greater overall mass may detract from any functional benefit. Alternatively, the hip and knee degrees of freedom might be coupled to mimic the functions of the biarticular leg muscles [50], which has been investigated for facilitating steady locomotion and sitting transfers (e.g., [19, 51]) but not yet for balance control.

Actuator placement for tripping recovery.

Our simulations suggest that actuators on the hip or thigh would likely not need to exert moments as large as those placed on the knee or shank for a similar degree of functional support (n or MoS), which could potentially entail smaller and lighter actuators. For a wearable aid, it is generally advisable to not only minimise the amount of mass added to the body, but also minimise its distance from the center of mass—placing mass distally on the lower extremities can undesirably increase both the metabolic cost of walking [52] and the rotational inertia of the leg about the hip joint, which may reduce the efficacy of assistance or require more powerful (and potentially heavier) actuators. These factors visibly influence the design of many existing minimally-obtrusive wearable robots, for example, recent tendon-driven JM systems in which the actuators and power supplies are placed near the body’s center of mass [18, 19].

Placing an actuator distally on the leg could instead be preferable if it could be integrated with an existing device worn by the user, for example, incorporating an FM actuator into an ankle orthosis or a transtibial prosthesis, or either a JM or an FM actuator into a transfemoral prosthesis.

Internal-reaction wearable actuators.

The vast majority of existing wearable robots use internal-reaction actuators. Monoarticular joint actuators, in which each biological degree of freedom is controlled by a separate dedicated actuator, have been conventionally preferred in the design of exoskeletons. Perhaps because these systems have traditionally been designed to primarily oppose gravity and facilitate movement rather than specifically regulate balance, relatively little attention has been paid to how reaction moments affect stability during dynamic activities. The simulations presented here suggest that such architectures can undesirably influence passive mechanics during gait subfunctions and be disadvantageous for balance control. Contrary to most state of the art, if it is assumed that the user is independently ambulatory and the robotic orthosis is exclusively for emergency balance correction, then there is also a potential usability drawback of such a design.

Recent trends in tendon-driven exosuits have drawn inspiration from biological multiarticular musculature to offer greater flexibility in addressing these concerns. Multiarticular muscle-tendon complexes span and actuate multiple joints simultaneously, simplifying neurological control and power transmission from powerful proximal muscle groups to distal joints [50]. Similarly, minimalistic wearable robots can employ a small number of favourably-placed artificial tendon actuators to support a core set of subfunctions necessary for gait without actuating or constraining all leg joints [50, 53, 54]. Significantly, this yields greater freedom in specifying the locations of the action and reaction forces on the body, and enables heavier components to be placed close to the CoM to minimise energetic cost and disturbance to natural gait patterns. The actuator JM_Th-Th simulated here represents an uncommon parallel biarticular construct without a direct biological counterpart. However, a similar robotic hip orthosis using instead two coordinated monoarticular actuators has previously shown great potential for aiding balance recovery following slipping [45].

External-reaction wearable actuators.

From a theoretical standpoint, our simulation results appear to favour the use of external-reaction actuators for this particular use-case. Although the hypothetical actuators have thus far been idealised as a mass and moment source(s), their physical realisation can take various forms with different practical consequences. We define here two subclasses of wearable external-reaction actuators: contact-manipulating actuators, such as active ankle orthoses [55, 56] and supernumerary limbs [57] that exert forces directly against the environment, and inertia-manipulating actuators, such as artificial tails [58], resonating pendula [59], reaction wheels [28], control-moment gyroscopes [30], and cold-gas thrusters [31] that induce moments or forces by manipulating the motion of a mass contained within (or, in the latter case, expelled from) the actuator itself.

Contact-manipulating actuators are capable of generating relatively large, sustained forces to support balance and gait. However, a constraint is that this is realised through a (temporarily) stable mechanical connection with the environment and is therefore effective only when the actuated limb is in stance and on surfaces that are preferably neither compliant nor slippery. The design and manner of achieving contact with the surroundings can also influence compatibility with different terrains (e.g., stairs) and the likelihood of tripping over components or inhibiting natural stepping patterns.

Inertia-manipulating actuators, in contrast, do not require ground contact and can assist body movement even if contact with the surroundings is limited or unreliable. Nor do these actuators necessarily need to be placed around the feet to influence balance (in fact, in most examples they are not even on the lower body [28, 30, 31]), thus preventing issues with tripping or terrain compatibility. There is, in principle, considerable freedom in where and how these actuators attach to segments of the body, since they do not need to span the biological joints to exert reaction moments. The relatively few points of contact with the body simplifies the donning and doffing of the device and reconfiguration of the interface for other users. However, the principle of operation also introduces several limitations. The presence of a passive mass for energy storage, and sometimes multiple motors (e.g., control moment gyroscopes), can entail that inertial-manipulating actuators are often larger and heavier than alternatives with a similar power output. While the detrimental effects of added mass on gait is similar for all wearable devices, actuators with rotational energy storage (e.g., reaction wheels and control moment gyroscopes) also have the potential to disturb gait through unintended gyroscopic moments induced through rotational movement of the actuated body segment—however, these induced moments can be partly or wholly cancelled internally via mechanical constraints [60] or control [61] or even be exploited to produce intentional changes in gait patterns [62]. Finally, although all actuator types are limited in their maximum output moment and power, inertia-manipulating actuators also have impulse constraints that limit the duration of moment application and prevent them from providing continuous support in static directions.

Given the relatively recent appearance of these technologies in wearable applications, there are few empirical analyses and it is yet unclear which actuation architecture is best (or even feasible) for practical implementation. For example, most empirical studies involving control moment gyroscopes have thus far sought to influence balance through control of the motion of the CoM [30, 63, 64], whereas control of the lower extremities has been investigated only in simulations [65]. However, parallel research has explored the uses of these actuators on the limbs for other purposes, such as tremor-suppression (GyroGlove by GyroGear; London, UK), upper-extremity prostheses [66, 67], and emulation of virtual environments [68]. Ongoing research continues to investigate the extent to which optimisation of mechanical designs can mitigate technical limitations and exploit the proposed advantages [69]. While much development and validation of these emerging technologies remains, they may soon fulfill a complementary roll in the technological landscape and offer versatile solutions for facilitating limb movement and balance.

Potential beneficiaries.

The target group is one of the essential factors to consider when designing an assistive device. We believe that elderly, people suffering from stroke and neuromuscular issues could benefit from forward foot placement assistance. Although in this study we only investigated the forward foot placement assistance, a FM device placed on either the thigh or shank could also assist in ab/adduction and endo/exorotation of the hip. For instance, it is commonly seen that people suffering from stroke circumduct the swing leg as a compensatory movement to avoid foot scuffing. Moreover, ad/abduction assistance can compensate for weaker muscles which otherwise results in scissored gait in patients with cerebral palsy.