Limitations to mobility impact millions of individuals worldwide. These challenges can make simple activities like walking or climbing a flight of stairs difficult due to lack of muscle strength, coordination, or balance. Recently, powered ankle and hip exoskeletons have demonstrated the potential to improve mobility in clinical populations by generating sagittal plane torques that can support the human joint function. For example, ankle exoskeletons assisting the user’s plantar/dorsiflexion can increase level-ground self-selected walking speed in individuals poststroke [1], [2] and in children with cerebral palsy [3]. Assisting the user’s ankle plantarflexion with an ankle exoskeleton can also decrease the metabolic cost of transport during ramp and stair ascent in individuals with cerebral palsy [3], [4]. Similarly, hip exoskeletons assisting hip flexion/extension can increase level-ground self-selected walking speed in stroke survivors [5], [6], decrease metabolic cost of transport in above-knee amputees and elderly individuals [7], [8], and improve recovery after anterior and posterior slip events in elderly individuals and individuals with above knee amputation [9], [10]. Thus, walking speed and metabolic cost of transport in clinical populations can be improved through exoskeleton assistance to the sagittal plane. However, current exoskeleton assistance strategies have had limited success improving medio-lateral balance of clinical populations [9].

Medio-lateral balance is an important factor in gait. Clinical populations often struggle to maintain medio-lateral balance, putting them at high risk of falling [11], [12], [13]. To maintain mediolateral balance, healthy individuals use their hip abductor and adductor muscles to regulate hip torque in the frontal plane [14], [15]. By controlling frontal plane hip torques, healthy individuals can shift their center of mass with respect to their center of pressure when the foot is in contact with the ground. Moreover, individuals can use frontal plane hip torque to reposition their feet when the foot is off the ground, increasing their base of support in the subsequent step, which also improves balance [14], [16]. The hip abductors and adductors play a large role in both these balancing strategies, generating 0.7 to 1.0 N·m/kg of frontal plane torque during walking and climbing stairs [14], [17], [18]. Unfortunately, most people affected by mobility impairments have weakened lower limbs [19], [20], [21]. This lower limb weakness makes balance more challenging and increases the prevalence of falls [22], [23]. Therefore, powered exoskeletons could also improve balance and mobility in clinical populations by generating torques in the frontal plane.

To the best of our knowledge, only two autonomous exoskeletons have attempted to improve balance by providing assistive hip abduction and adduction torques in addition to hip flexion and extension. The first is the MindWalker, a 28 kg hip–knee–ankle exoskeleton developed for people with spinal cord injury [24]. This device provided the first demonstration of an exoskeleton that actively modifies the lateral foot placement in response to a perturbation. A 9.2 kg powered hip-only exoskeleton from North Carolina State University extended this control scheme to include admittance-based control [25]. Unfortunately, both these devices are quite heavy and bulky, which substantially reduced their usability in the real world. To address this problem, researchers developed an exoskeleton that assists the hip abduction/adduction only [26]. This device is lighter but cannot assist with body propulsion and step length symmetry, which require hip flexion/extension assistance. Thus, there is an unmet need for a lightweight and compact autonomous hip exoskeleton that can assist both hip flexion/extension and abduction/adduction.

Existing hip exoskeletons that power both sagittal and frontal plane motion are also limited by their series kinematic configuration. In this series configuration, the powered frontal plane joint meant to generate hip abduction/adduction is placed proximal to the powered sagittal plane joints meant to generate hip flexion/extension. With this configuration, at high degrees of hip flexion (e.g., during terminal swing of walking or ascending a step), exoskeleton frontal plane torques cause anatomical hip eversion/inversion torques instead of hip abduction/adduction. Therefore, powered hip exoskeletons using a series kinematic configuration cannot provide hip abduction/adduction torques during terminal swing in level ground walking or when climbing stairs. Moreover, in a series kinematic configuration, the frontal plane actuators are located in the back, preventing the exoskeleton users from sitting comfortably in a chair.

In this paper, we present a lightweight and compact powered hip exoskeleton with an alternative kinematic design that can provide hip abduction/adduction independently of the hip flexion/extension angle. Specifically, we propose a parallel kinematic actuation system that concurrently assists hip flexion/extension and abduction/adduction. Leveraging the proposed parallel actuation system, we developed a powered hip exoskeleton that is lightweight (5.3 kg), has a slim profile (adding only 3 cm posteriorly, and 8 cm laterally at the hip), and can provide high torque during gait (up to 30 N·m). This article presents the kinetostatic model and the simulation framework used to design the proposed parallel hip actuator. Benchtop testing shows the accuracy of the transmission model including feedforward friction and inertia compensations. Treadmill-based walking studies with five subjects demonstrate the ability of the proposed exoskeleton hardware and controller to provide assistance both in the sagittal and frontal planes during walking. Thus, the main contributions of this paper are the kinematic model of the actuator, the application of adaptive oscillators and phase-based torque planning to control assistive torques in the frontal plane, and the verification of the actuator and controller performance on the bench top and with human subjects. Due to its compact size, small mass, and unique kinematic configuration, we believe that the proposed powered hip exoskeleton has the potential to improve balance and reduce metabolic cost in individuals with lower-limb impairments, improving their mobility and quality of life.

The exoskeleton uses a hierarchical control structure similar to previous work [30] (see Fig. 6). The high-level controller generates a real-time gait phase estimate. The gait phase estimate is the product of a gait timer and a gait cadence estimate generated by an adaptive oscillator [44], [45]. The gait timer is reset every cycle by a state machine that identifies peak hip flexion. The parameters of the state machine can be selected to be robust across ambulation tasks [30]. The phase estimate generated by the high-level controller is utilized by a mid-level torque planner. The torque planner generates desired assistive torque profiles for both the sagittal and frontal plane exoskeleton joints. Each assistive torque profile is the sum of two Gaussian functions in the gait-phase domain. In the sagittal plane, this assistive strategy has successfully reduced the metabolic cost of walking of individuals with lower limb amputation [7]. A similar approach generating abduction assistance during stance has reduced the metabolic cost of walking for healthy individuals [46].

The duration, magnitude, and phase-timing of the Gaussian functions is manually tuned to achieve the desired torque profiles. The desired joint torques are converted by the low-level controller into desired motor currents for the linear actuator modules. Corrections are made for inefficiencies in the hip exoskeleton actuation. Additional feed-forward friction and inertia terms are added to the desired current to compensate for the linear actuator dynamics. The desired motor currents are sent to motor drivers which perform closed-loop current control and torque vectoring.

The high and low-level controllers (see Fig. 6) run on the microcontroller on the low-level electronics board at 1000 Hz [see Fig. 5(a) and (c)], while the mid-level torque planner (see Fig. 6) runs on the single-board computer on the high-level electronics board at 500 Hz [see Fig. 5(a) and (b)].

We estimated the hip actuator steady-state error, actuator bandwidth, and impedance by performing benchtop tests. Similar to our prior work [30], we connected the hip actuator crank to a grounded six-axis force-torque sensor (Sunrise Instruments M3713D). The hip actuator joints were both set to zero degrees and the distal portion of the exoskeleton frame was also grounded. Five N·m of preload was applied, and we commanded torque steps of 5 or 15 N·m in three separate conditions: frontal plane torque, sagittal plane torque, and combined torque. Forces and torques measured at the load cell were reflected to the hip actuator joint centers. Each step was conducted 5 times, and the results are reported in (see Fig. 7). The rise time, percent overshoot, and steady-state error, were extracted from each trial and averaged (see Table III). Across the 6 experimental conditions, the risetime, percent overshoot, and rms steady state error are between 9.4 and 17.2 ms, 6.9% and 39.3%, and 0.2 and 1.0 N·m.

The feed-forward torque bandwidth of the system was estimated using a similar experimental setup. The crank of the hip actuator was connected to a grounded force-torque sensor (Sunrise Instruments M3713D). The thigh frame was also grounded. We set the actuator’s desired torque in the frontal or sagittal plane to a sinusoidal profile with an amplitude of 20 N·m and a frequency that increased from 0 to 30 Hz over a 30 s period. The forces and torques measured by the force-torque sensor were reflected to the joint center [see Fig. 8(a) and (c)]. A close-up shows the tracking performance between 0 and 4 s (approximately 0 to 4 Hz signal). In this section, the root mean square error between the desired and measured signal is 7.2 N·m in the frontal plane [see Fig. 8(a)] and 6.8 N·m in the sagittal plane [see Fig. 8(c)]. The desired and measured torques were also transformed into the frequency domain using a fast-Fourier transform (MathWorks MATLAB). We obtain the magnitude and phase of the ratio of the measured to desired torque [see Fig. 8(b) and (d)]. From this data, we estimate the −3 dB crossing as 21.2 and 19.4 Hz in the frontal and sagittal plane, respectively. The corresponding phase lag at the −3 dB crossing is −329° and −309° in the frontal and sagittal plane, respectively.

The output impedance of the hip actuator was estimated by mounting the crank of the exoskeleton to a grounded six-axis force-torque sensor and manually driving the frame through a sinusoidal motion in the sagittal plane to mimic walking at different frequencies. Forces and torques measured by the load cell were reflected to the joint center. We performed the test under two conditions, with and without feedforward compensations, the results of which are shown in Fig. 9. The output impedance was estimated in the frequency domain by fitting a 2 pole 1 zero transfer function between reflected joint torque and joint velocity. We estimate the reflected damping and inertia without compensations to be 1.13 N·m/rad/s and 0.53 N·m/rad/s². With motor compensations on, the reflected damping and inertia decrease to 0.37 N·m/rad/s and 0.35 N·m/rad/s². The feedforward compensations decrease the reflected damping by 68% and the reflected inertia by 35%. These results suggest that an exoskeleton user will experience limited resistance when walking or running.

Powered hip exoskeletons have the potential to improve gait balance and efficiency in clinical populations by concurrently generating torques in the frontal and sagittal planes. To the best of our knowledge, only two autonomous powered hip exoskeletons can provide assistive torques in both planes concurrently. Unfortunately, these devices are too heavy and bulky for unrestricted use in the real world, especially for clinical populations with limited strength and balance. To address this issue, we present the first powered hip exoskeleton using a parallel kinematic design with two linear actuators located along the user’s thigh. Because the two linear actuators concurrently provide assistance in abduction/adduction and flexion/extension, the proposed exoskeleton is much lighter than previous designs (e.g., 5.3 kg versus 9.2 kg) while being able to provide similar torque (sagittal plane torque 30 N·m versus 34 N·m, frontal plane torque 20 N·m versus 25 N·m) [25], [38] (see Table V). Moreover, due to its unique parallel actuation, the proposed exoskeleton is compact (adding only 3 cm posteriorly, and 8 cm laterally at the hip), which facilitates use in the real world. Human studies with healthy individuals show that the proposed exoskeleton can assist in the sagittal plane while providing frontal plane torques to alter step width, which is an important indicator of balance [49]. Moreover, comparisons to existing exoskeletons suggest that the proposed parallel actuation system satisfies the size, mass, and torque requirements necessary to potentially improve balance and reduce metabolic cost in individuals with lower limb impairments in the real world [7], [43], [46], [50], [51].

Virtually all powered hip exoskeletons that provide assistance in the frontal plane use a series kinematic design, and place the frontal-plane actuators posteriorly to the user. However, this posterior location prevents the user from comfortably sitting in a chair [24], [25], [26]. Furthermore, this posterior location means that at high degrees of hip flexion that occur during sit-to-stand transitions or stair ascent, the abduction and adduction torque generate hip internal/external rotation. By using a parallel kinematic design, we can place the actuators laterally to the user’s thigh, which reduces the posterior dimension of the exoskeleton from more than 10 to 3 cm while maintaining a lateral profile of 8 cm which is similar to, if not less than, most bilateral hip exoskeletons (see Table V). Depending on the chair and individual, the low-profile design may allow users to sit comfortably without the exoskeleton contacting the arms or back. Moreover, the frontal plane joint of the proposed hip exoskeleton is located distal to the sagittal plane joint. This kinematic configuration reduces the actuator torques that cause hip internal/external rotation. Thus, the proposed parallel actuator enables the proposed hip exoskeleton to provide both propulsive and stabilizing torques without hindering common activities like sitting, arm swinging, and ascending stairs.

In the proposed parallel kinematic design, two linear actuators concurrently generate forces to control the hip abduction/adduction and flexion/extension. In contrast, in a series kinematic configuration, each actuator independently generates abduction/adduction or flexion/extension. As a result, the proposed parallel actuator has torque density of 17.6 N·m/kg in the sagittal plane and 11.8 N·m/kg in the frontal plane. In contrast, the largest actuator torque density for a series exoskeleton is 12.4 N·m/kg (see Table V).

The use of the proposed parallel mechanism can also lead to increases in exoskeleton torque density and a decrease in exoskeleton size. Exoskeletons that utilize a series configuration to assist in both the frontal and sagittal planes are typically constructed from posterior and lateral actuators that are connected by load transferring frame [24], [25], [38], [43], [52]. These actuators add 6.5 to 10 cm lateral and 13 to 17 cm posterior to the user. In contrast, the proposed parallel actuator is only located lateral to the user which decreases the posterior dimension to 3 cm while maintaining a lateral profile of 8 cm. Additionally, linkages in the parallel actuator function as structural frames that transfer load directly to the user. Thus, no additional structural parts other than the pelvis and thigh brace are needed to complete the exoskeleton. This results in an exoskeleton torque density of 5.7 N·m/kg in the sagittal plane and 3.8 N·m/kg in the frontal plane which is 41%–53% larger than the next largest exoskeleton torque density (3.7 and 2.7 N·m/kg, respectively [25], [38], see Table V).

The proposed exoskeleton does not use force sensing and closed-loop control, which is similar to our previous powered exoskeletons and prostheses using linear actuators [30], [32], [33], [34]. Bench top testing shows that the feed forward compensations reduce the reflected damping and inertia by 68% and 35%, respectively, achieving high backdrivability. Analysis of step response performance indicates that the hip actuator has a maximum steady state RMS error of 1.0±0.1N·m for 15 N·m torque steps and an RMS error 0.7±0.0N·m for 5 N·m torque steps. This performance is similar to other hip exoskeletons using open- and closed-loop control [30], [43], [53].

Analysis of the chirp performance on the benchtop indicates a −3 dB crossing of 21.2 Hz in the frontal plane and 19.4 Hz in the sagittal plane [see Fig. 8(b) and (d)], which is similar to other exoskeletons [30], [54], [55], [56], [57]. However, there is substantial phase lag (about 300°) at the −3 dB crossing and signal amplification (about 11 dB). Signal amplification and large phase lag in the torque response of wearable robots are not uncommon. Signal amplification can be as large as 4 to 10 dB and occur at frequencies between 6 and 100 Hz. Moreover, phase lag at the −3 dB frequency can be between −100° and −360°, occurring at frequencies of 10–60 Hz [29], [55], [56], [57], [58]. However, these performance characteristics represent the performance of the torque controller as the exoskeleton interacts against an external frame with substantially higher stiffness than the human body, affecting the frequency spectrum of the torque response. Most importantly, the power spectrum calculated from human biomechanics datasets indicates that 99.7% of kinematics occur at frequencies lower than 4.8 Hz [59], which is substantially lower than the resonance frequency and −3 dB bandwidth. Similarly, the power spectrum of the desired torque control used during our human subject experiments indicates that 90% of the signal energy is below 3 Hz. At this frequency, the hip actuator magnitude is about 1.1 to 1.3 dB while the phase lag is about −25° to −30°. This analysis, combined with the low overshoot and RMS error during torque step response, suggests that the proposed exoskeleton satisfies the necessary requirements for human subject testing.

The maximum signal amplification for the mechanism presented here is slightly higher than previously reported values (10.5–11.6 dB compared to 10 dB) [58], while the phase lag is higher than some previously reported values [57], [60]. The observed signal amplification and phase lag may be due to several factors, including test bench setup, backlash in the transmission mechanism, implementation of low-level control, or parallel actuation. However, inferences derived from a comparison of results presented by different authors are limited due to different control strategies (open- versus closed-loop), test conditions (measurement of rotary or linear actuator performance versus assembled exoskeleton), and measurement methods (external load cells versus internal load sensing). Future studies should investigate how different factors affect the frequency response of a powered exoskeleton.

Level ground walking with five human subjects verified that the exoskeleton can produce up to 30 N·m of sagittal plane torque and 20 N·m of frontal plane torque concurrently [see Fig. 11(a) and (b), top plot]. These torque peaks correspond to providing 34% and 20% of the biological sagittal and frontal plane torque of a 95th percentile male [17], [37]. Studies with clinical populations and healthy individuals indicate that these torques are sufficiently large to reduce metabolic consumption in overground walking for a 95th percentile male through applying either frontal or sagittal plane torques [7], [46], [50], [51]. Similarly, studies with healthy individuals indicate that this is sufficient torque to improve stance width and measures of balance in both steady state and transient (as-needed) operations [43], [46]. In agreement with these studies, our human testing shows that abduction frontal plane assistance during swing increased step width from the baseline condition by an average of 24%, whereas adduction frontal plane assistance during swing decreased step width from the baseline condition by an average of 14% [see Fig. 12(b)]. These results show that the proposed exoskeleton can modify stride width and, indirectly, the base of support, suggesting that the proposed exoskeleton may positively impact balance [49].

Frontal plane assistance during swing impacted step width (see Fig. 12). Specifically, frontal plane abduction torques during swing increased step width by 24% [see Fig. 12(b)] while frontal plane adduction torques during swing decreased step width by 14%. A recent study using an autonomous exoskeleton with only powered frontal plane hip assistance demonstrated the ability to increase step width by an estimated 57% and decrease step width by an estimated 31% [26]. These changes in step width are much higher than the ones we observed in this study. However, their exoskeleton applied admittance control to both legs for the entire duration of the gait cycle with a virtual equilibrium position of either ±15° and a stiffness of 60 N·m/rad. In contrast, we applied frontal plane torques to only one leg during either the swing or stance phase of the gait cycle [see Fig. 11(b), top plot]. Specifically, in our experiment, the abduction torque applied in swing lasted for about 40% of the gait cycle and increased step width by 24% [see Fig. 12(b)], which is roughly half of the change in step width observed with constant bilateral admittance control used in prior studies [26]. Similarly, in our study, the adduction torque applied in swing lasted for about 40% of the gait cycle and decreased step width by 14% [see Fig. 12(b)], which is, roughly half the change in step width observed in prior studies [26]. This comparison suggests that assistance magnitude, duration, and control strategy may impact the relative change in step width.

Frontal plane assistance during stance impacted step width [see Fig. 12(a)]. Specifically, abduction and adduction assistance applied to the stance limb increases step width by 14% and 5%, respectively. Similar to prior work [46], all five subjects increased their step width when receiving abduction assistance during stance, suggesting that abduction assistance during stance has the potential to improve balance. In contrast, adduction assistance decreased step width for only one of the five subjects. Moreover, the variation in subject response was large. One subject increased step width by 23%, three subjects had little change in step width, and one decreased step width by 17%. This variation in result may be related to balance. Exoskeleton torques applied during the stance phase may impact both foot placement and the center of mass trajectory. As a result, the center of mass may deviate from its natural course and the subject may take wider steps to increase their base of support. Thus, frontal plane torques applied during stance may positively or negatively influence measures of balance or gait efficiency. Additional biomechanical studies are needed to better understand the relationship between frontal plane torques applied during the stance phase of walking and gait adaptations.

The proposed kinematic model and simulation framework indicate that the exoskeleton hip actuator consumes 24.5 J of electrical energy per stride. In contrast, the electrical energy consumption observed in the subject testing was between 29.3 and 71.8 J per stride. The difference between the simulated and real energy consumption is primarily attributed to the exoskeleton torque timing and increase in exoskeleton hip joint range of motion. Specifically, in the sagittal plane, the experimental torque profile is timed such that extension torques are aligned with extension velocity and flexion torques are aligned with flexion velocity [see Fig. 11(a)]. Similarly, the experimental frontal plane torque caused large deformation of the exoskeleton interfaces in the direction of the applied torque [see Fig. 11(b), top and middle plot]. Thus, during the human subject experiments, the exoskeleton primarily performed positive power (net mechanical energy of 24.9 to 55.6 J per stride) at the cost of 29.3 to 71.8 J of electrical energy. This equates to an average efficiency of about 84%. In contrast, the simulation was based on torque profiles from biomechanical datasets of level walking, which include both positive and negative power phases [17]. The simulation predicts that the exoskeleton will inject about 11 J of mechanical energy while absorbing about 7 J of mechanical energy. In total, accounting for thermal losses, this costs 24.5 J of electrical energy, which equates to an expected efficiency of 15%. Thus, the increase in electrical energy consumption can be largely explained by the increase in net mechanical power produced by the exoskeleton. Moreover, this analysis highlights the large difference between simulation and experimental performance, and indicates that device efficiency can be highly impacted by negative power phases.

Human adaptation to the exoskeleton assistance and motor driver inefficiency may also contribute to an increase in electrical energy consumption of the hip actuator. Results from the human subject experiments [see Fig. 11(a), bottom plot] indicate that, on average, the human hip joint exhibited a greater range of motion in the sagittal plane when assisted by the exoskeleton. Fig. 11 also shows subtle changes in the frontal plane human hip joint angles. In contrast, the simulation assumed the actuator followed human biomechanics data sets which did not account for human adaptation to the exoskeleton assistance. Furthermore, the simulation framework assumes 100% efficiency of the 4-Q servo drives during both positive and negative power phases. However, small inefficiencies may have contributed to increased energy consumption. Further refinement of the simulation framework can be done to better predict exoskeleton actuator energy consumption.

We use an adaptive oscillator and finite state machine to provide a continuous estimate of gait-phase evolution. This control approach has been previously applied to generate sagittal plane assistance in walking, running, and stair ascent [30]. Here we show, for the first time, that this control approach can be used to reliably control frontal plane assistance across different walking speeds (1.5 to 2.4 mph) and assistance conditions (abduction or adduction assistance timed during stance or swing). The timing of assistive torques across the experimental conditions and subject-specific walking speeds was consistent (average timing of peak flexion torque of 61.1±0.3% gait phase across powered conditions), indicating that the adaptive frequency oscillator and finite state machine are robust to differences in assistance profile and walking speed. This result suggests that adaptive oscillators with finite state machines can be used to develop controllers aiming to improve balance.

Powered hip exoskeletons that generate torque in both the sagittal and frontal planes are fundamental to improving gait economy and balance in individuals with poor mobility. However, existing technologies are too bulky and heavy to use in the real world. This article contributes the kinematic model, mechatronic design, and benchtop and human subject testing of a hip exoskeleton with a parallel actuator. This parallel actuator enables a powered hip exoskeleton that is substantially lighter, more compact, and more ergonomic than previous devices, while still generating similar levels of torque. Human studies with five healthy young adults show that the proposed powered hip exoskeleton, controlled with a state machine and adaptive frequency oscillators, can consistently generate torques in the frontal and sagittal plane. Moreover, the application of torques in the frontal plane alters step width, a key component of balance. Future work will explore control strategies that intentionally alter step width to improve balance in clinical populations.