Limb-assistive device with modified energy shaping

Abstract

A powered lower limb assistive device such as a knee exoskeleton includes an articulated frame having a pivot joint and employs a task-adaptive energy-shaping control scheme to modulate torque at the pivot joint. The control scheme employs a single control law based on a combination of a task-adaptive stance torque function (Tst) and a task-invariant swing torque function (Tsd) and operates without discrete identification of gait phase. The control law is based in part on torque basis functions for at least one virtual spring, at least one virtual damper, gravity compensation, and / or inertial compensation. A virtual spring can inject energy at the joint at specific times during ambulation making the device particularly suitable for reducing user muscle fatigue during lifting, lowering, and carrying (LLC) tasks and encouraging injury-reducing lifting and lowering techniques.

Description

[0001] LIMB-ASSISTIVE DEVICE WITH MODIFIED ENERGY SHAPING

[0002] This invention was made with government support under EB031166 awarded by the National Institutes of Health. The government has certain rights in the invention.

[0003] TECHNICAL FIELD

[0004] This disclosure is related to limb-assistive devices and, in particular, to controlling operation of such devices among a range of tasks.

[0005] BACKGROUND

[0006] Quadriceps fatigue during lifting-lowering and carrying (LLC) tasks is a critical, unaddressed problem affecting worker performance, posture and, ultimately, lower-back injury risk. According to the National Safety Council, 69% of workers across the construction, manufacturing, transportation, and utility industries experience workplace fatigue, which increases the risk of injuries and incidents on the job. Worker fatigue also causes a decrease in cognitive and physical function in assigned work tasks. This fatigue-induced performance decline becomes crucially deleterious when there is a minimum performance level constraint imposed by factors outside a worker’s control, such as when rate of lifting and placing is determined by the speed of a conveyor belt. More than a million people in the United States suffered a non-fatal work-related musculoskeletal disorder in 2021 — most commonly low back pain (LBP). Such injuries cost employers millions of dollars annually in worker compensation. Repeated lifting-lowering (LL) activities are highly associated with overexertion and LBP incidence, and significant research effort has been put into investigating the effect of LL techniques on LBP.

[0007] There are two common lifting techniques: stoop-lifting with the back and a flexed lumbar spine, and squat-lifting with the legs while maintaining a neutral lumbar spine. Compared to squat-lifting, stoop-lifting consumes less energy and is associated with less perceived lower-limb exertion and fatigue. On the other hand, squat-lifting is associated with less lumbar shear forces and reduced lumbar passive tissue stress compared to stoop-lifting. The squat technique is favored over stooping by people who have previously incurred lower back injuries. This could be an instinctual self-protective choice that prioritizes the safety of the ligaments and discs of the lumbar spine, which are stressed during stooping, over the muscles of the lower limbs, which are stressed during squatting. Moreover, national workplace labor guidelines, military manual-lifting guidelines, and most physical therapists recommend the squat-lifting technique to prevent LBP, especially for lifting weights that can be placed between the feet. Overall, squatting is often considered safer but is more energetically demanding and fatiguing on the quadriceps.

[0008] Quadriceps fatigue has been shown to influence lifting technique, causing a transition from squat-lifting to stoop-lifting as fatigue progresses. Consequently, high quadriceps fatigue is associated with higher effort of the lower back extensor muscles during repeated LL tasks, which can increase the chance of overuse injuries. The quadriceps are also critically involved in carrying tasks while traversing stairs and ramps, which are demanding activities that further contribute to quadriceps fatigue. Fatigue could potentially be mitigated by orthotic devices — i.e., exoskeletons — that assist the wearer during LLC activities over multiple terrains.

[0009] Orthoses have been developed to prevent LBP by directly supporting the lumbar spine rather than the quadriceps. A lower back belt or back support are basic examples of this intervention approach. But the National Institute for Occupational Safety and Health does not recommend back belts due to the lack of supportive evidence for their effectiveness. Passive hip-back orthoses can reduce lumbar moments during forward bending tasks but hinder active hip flexion or lumbar flexion in walking. Newer commercial devices like the HeroWear® Apex exosuit circumvent this issue with a clutch that allows manual disengagement of the device when unrestricted lumbar flexion and / or hip flexion is desired. Powered hip-back orthoses can modulate assistance with software rather than a mechanical clutch as the task changes. However, designing a versatile control strategy for such a device has proven challenging. In particular, while these devices can facilitate LL tasks, they tend to hinder normal walking. Attempts have been made with active devices to explicitly classify tasks using electromyography sensing, which is cumbersome to setup and calibrate. Moreover, these lumbar and hip-back devices, passive and powered alike, do not directly support the quadriceps muscles that are critically involved in both squatting and load carrying, nor do they provide any user assistance at the knee joint.

[0010] Actively providing knee-joint assistance has problems of its own. Knee joint kinetics and kinematics vary significantly among the various activities that fall under LLC. State-of- the-art passive knee orthoses naturally fall short of the versatility required for LLC tasks, because a fixed spring — even with an elaborate mechanical clutch — simply cannot recreate biomimetic assistance over a variety of tasks. Such devices have therefore been limited to supporting only squatting, at best. While active knee orthoses have the potential to provide versatile assistance for multi-terrain LLC tasks, the development of such devices has favored rigid, highly geared actuators that track predefined kinematics to assist severely impaired individuals. Such devices are not suitable for the LLC application, as the combination of rigid actuators and kinematic control hinders the voluntary motion of able-bodied individuals. One commercially available pneumatic knee exoskeleton, for example, must be manually turned off for non-squatting tasks because it impedes knee flexion during walking.

[0011] The overall goal of a partial assistance controller for an application like LLC is to use measurable signals to command a biomimetic assistance torque over a wide range of tasks. One direct approach is to use inverse dynamics to estimate joint torque from measured ground reaction forces (GRFs). But dynamic modeling errors, sensor noise, and missing GRF components present significant issues. While machine-learning methods have been used to classify among distinct activity modes, such discretization of activities does not provide seamless adaptation to continuous variations in the environment or in user behavior. Kalman filter-based approaches have been explored but have not considered stair ascent-descent or LL tasks. Implicit task representation has been proposed to directly predict human hip torques using deep learning with a multi-activity able-bodied dataset, but this black-box method does not provide formal safety guarantees, lacks a biomechanically intuitive means of customization for LLC tasks, and has not been generalized to more complicated joint kinetics like that of the knee.

[0012] Energy shaping is a non-linear control method that alters the dynamics of the human- exoskeleton system in closed loop by applying joint torques as functions of system states, including joint and limb-segment angles and velocities. The solutions to partial differential equations called the “matching conditions” determine the realizable alterations to the dynamics of an underactuated system, which can be used to define a set of admissible torque basis functions that parameterize the controller. WIPO Publication No. WO 2025 / 097003, referred to hereinafter as Gregg, et al., discloses a data-driven approach that can be used to find the optimal coefficients for a large set of torque bases to closely predict normative human torques given normative kinematic inputs over the primary activities of daily life. The drawbacks of this approach for application to specialty tasks such as LLC include a non- intuitive combination of hundreds of torque basis functions, which makes manual controller customization for LLC tasks impossible, and susceptibility to “overfitting,” which results in unhelpful torques for input kinematics that grossly deviate from the training dataset. This approach has also demonstrated inconsistent reductions in muscle effort across tasks. A crucial gap thus exists in state-of-the-art exoskeletons for LLC applications, which involve repetitive, high-stress, and fatiguing tasks. Although existing devices can effectively support the back musculature, especially during stooping, no exoskeleton has thus far demonstrated versatile and effective assistance of the quadriceps during multi-terrain LLC, including the clinically recommended squat technique. Assisting the knee joint over a wide range of activities presents significant challenges in actuation and control.

[0013] SUMMARY

[0014] Embodiments of a powered lower limb assistive device include an articulated frame including a first frame member and a second frame member interconnected at a pivot joint for relative rotation about the pivot joint. The device employs a task-adaptive energy-shaping control scheme to modulate torque at the pivot joint, and the control scheme employs a single control law.

[0015] The assistive device may include any one or more of the following features in any technically feasible combination:

[0016] - the control law is based on a combination of a task-adaptive stance torque function and a task-invariant swing torque function and operates without discrete identification of gait phase;

[0017] - the control law is based in part on torque basis functions for at least one virtual spring, at least one virtual damper, gravity compensation, and / or inertial compensation;

[0018] - the pivot joint is a knee joint and the assistive device is a knee exoskeleton;

[0019] - the control law is based in part on a task-specific torque basis function for a virtual unidirectional spring located at the joint, the virtual spring operating to inject energy after heelstrike, and the injected energy being limited by a maximum angle between the first and second frame members;

[0020] - torque generated by the virtual spring is prevented when a leg angle of the user exceeds a predefined threshold such that joint movement is not hindered during late stance;

[0021] - a task-specific torque basis function is sensitized such that the injected energy is scaled according to a degree of incline or a step height along which the user ambulates and such that energy injection is prevented when the user assumes a symmetric posture or when a leading leg of the user becomes a trailing leg;

[0022] - the control law is based in part on a task-specific torque basis function for a virtual unidirectional spring and damper located at the joint, the virtual spring operating to absorb energy after heelstrike in an amount dependent on a change in angle between the first and second frame members after heelstrike, limited by an elastic limit of the virtual spring, the virtual damper operating to dissipate energy by an amount dependent on an angular velocity of the joint;

[0023] - torque provided by the virtual spring and damper is prevented when a leg angle of the user exceeds a predefined threshold such that joint movement is not hindered during late stance;

[0024] - a task-specific torque basis function is sensitized such that energy absorption is scaled according to a degree of decline or step height along which the user ambulates and such that torque provided by the virtual spring and damper is prevented when the user assumes a symmetric posture or when a leading leg of the user becomes a trailing leg;

[0025] - the control law is based in part on task-invariant torque bases including gravity compensation, inertial compensation, and a virtual unidirectional spring and damper;

[0026] - gravity compensation is modulated based on angular velocity of the joint such that user foot clearance is assisted during early-to-mid swing without hindering forward leg swing during mid-to-late swing;

[0027] - inertial compensation is modulated based on a global angle of the second frame member such that forward leg swing is assisted while the global angle is within a predefined range;

[0028] - a virtual unidirectional spring and damper dissipates energy during late swing to assist in preventing joint hyperextension;

[0029] - the control law is employed during a squatting task performed by the user.

[0030] - the control law modulates the torque such that a peak torque at the pivot joint is provided when the user transitions from a squatting movement to a standing movement during the squatting task; or

[0031] - peak torque is provided via a virtual torsion spring acting on the joint, the stiffness of the virtual spring being a function of joint angle and joint angular velocity such that joint movement is not hindered during the squatting movement.

[0032] BRIEF DESCRIPTION OF THE DRAWINGS

[0033] FIG. 1 is a schematic representation of a powered limb-assistive device fitted to the lower limbs of a user. FIG. 2 schematically illustrates an illustrative method of developing a control scheme based on an energy-shaping framework.

[0034] FIG. 3 illustrates normative versus optimized controller torques for multiple tasks.

[0035] FIG. 4 illustrates experimental conditions used to validate an optimized controller.

[0036] FIG. 5A includes box plots and individual line plots illustrating post-fatigue performance deficits.

[0037] FIG. 5B illustrates ensemble averaged lifting-lowering cycle durations as a function of trial progression.

[0038] FIG. 6 A illustrates a post-fatigue thorax lean deviation distribution.

[0039] FIG. 6B illustrates ensemble averaged deviations as a function of session progression.

[0040] FIG. 7 illustrates deviations in peak knee flexion angles from their maximum values observed during a fatiguing phase.

[0041] FIG. 8 illustrates ensemble averaged thorax lean profiles during lifting-lowering cycles.

[0042] FIG. 9 includes histograms of participant satisfaction ratings.

[0043] FIG. 10 illustrates distributions of quadriceps effort for multiple conditions and tasks.

[0044] FIG. 11 illustrates ensemble-averaged quadriceps electromyography (EMG) profiles alongside exoskeleton torque profiles.

[0045] FIG. 12 illustrates ensemble-averaged vastus medialis oblique (VMO) EMG profiles alongside the exoskeleton torque profiles.

[0046] FIG. 13 illustrates ensemble-averaged vastus lateralis (VL) EMG profiles alongside the exoskeleton torque profiles.

[0047] FIG. 14 illustrates ensemble-averaged rectus femoris (RF) EMG profiles alongside the exoskeleton torque profiles.

[0048] FIG. 15 illustrates ensemble-averaged hamstrings EMG profiles alongside the exoskeleton knee torque profiles.

[0049] DESCRIPTION OF EMBODIMENTS

[0050] Described below is a powered limb-assistive device (e.g., an exoskeleton) and an associated energy-shaping control framework. Energy shaping is a non-linear control approach based on classical dynamics in which a controller reshapes the open-loop plant to have a new Lagrangian (or, equivalently, Hamiltonian) energy function when the loop is closed. Considering the human-exoskeleton system as the plant, energy-shaping control can enable task-invariant assistance in a device with backdrivable actuators. The controller is parameterized by torque basis functions representing virtual springs, dampers, gravity compensation, and inertia compensation to produce knee torques that are harmonious with quadriceps activation across multi-terrain LLC tasks, including squat-lifting and lowering (LL), stair ascent (SA) and descent (SD), ramp ascent (RA) and descent (RD), and level walking (LW). The controller can be implemented on a highly backdrivable bilateral knee exoskeleton and has demonstrated significant improvements in fatigue-induced deficits in user performance of lifting and lowering tasks and improvements in posture, garnering an effectiveness rating among study participants of 4.3 out of 5 for post-fatigue LLC tasks. In a non-fatigued state, the device and controller significantly reduces quadriceps effort for all multi-terrain LLC tasks, other than level walking.

[0051] As used herein, a backdrivable actuator has a static torque (e.g., minimum backdrive torque to begin motion of the motor shaft) less than 20 Nm. Preferably, the backdrive torque is less than 5 Nm, less than 2.5 Nm, or less than 2.0 Nm. A high output torque for an actuator is a peak output torque (e.g., measured over a 1 second time period) of at least about 1.0 Nm. Preferably, the peak output torque is at least 1.5 Nm, at least 2.0 Nm, or at least 4.0 Nm.

[0052] FIG. 1 is a schematic representation of a powered limb-assistive device 10 fitted to the lower limbs of a user. The illustrated device 10 is a bilateral exoskeleton including an articulated frame 12 and actuator 14 for each limb of the user, a controller 16, a power source 18, and a sensor system 20 including one or more sensors. Each articulated frame 12 includes first and second frame members 22, 24 interconnected by a pivot joint 26. The first and second frame members 22, 24 are rotatable with respect to each other about the pivot joint 26. In this case, the first frame member 22 is an upper leg member, including lateral and medial portions 22a, 22b configured for removable attachment to the upper portion of the user’s leg between the knee and hip joints, and the second frame member 24 is a lower leg member, including lateral and medial portions 24a, 24b configured for removable attachment to the lower portion of the user’s leg between the knee and ankle joints. Each joint 26 also has lateral and medial portions.

[0053] Each actuator 14 provides a controllable torque at the respective joint 26. Each actuator 14 may be an electric motor fixedly mounted along one of the frame members 22, 24 and operably coupled with the other frame member for relative frame member movement about the joint 26. For example, a static motor housing may be mounted at a fixed position along one frame member 22 on one side of the joint 26 with movement of the rotor transmitted to the frame member 24 on the other side of the joint. The coupling may include a transmission including one or more gears, pulleys, belts, etc. to permit mounting of the actuator 14 away from the corresponding joint 26 if desired, or the actuator may be mounted at and / or coaxial with the respective joint.

[0054] The controller 16 is programmed with a control scheme 28 and controls each actuator 14 based on the control scheme and on inputs from the sensor system 20. The control scheme 28 may be or may include an energy-shaping control scheme as discussed further below. The control scheme 28 includes one or more control laws 30 governing the torque commands provided to the actuators 14. The control scheme 28 disclosed below employs a common control law 30 during stance and swing phases. The controller 16 may include at least one processor and memory (implemented as one or more non-transitory computer-readable mediums) storing or having instructions that, when executed by the at least one processor, cause the controller to modulate torque at the device joints 26 in accordance with the control scheme 28. The controller 26 is depicted as a single unit in FIG. 1. It should be understood that torque may be modulated at each joint by separate coordinated controllers and / or that each controller may perform other tasks in addition to torque modulation at a single joint.

[0055] Examples of sensors included in the sensor system 20 to provide information to the controller 16 include foot contact sensors (e.g., accelerometer) to detect heel strike at the transition from swing phase to stance phase, vertical ground reaction force (vGRF) sensors, inertial measurement units (IMUs) (e.g., for global thigh angle), and electromyography (EMG) sensors. The illustrated example includes a force sensitive resistor (FSR) system including sensors mounted along the posterior side of the user’s shoes.

[0056] The power source 18 may be a rechargeable battery or other suitable source for powering the actuators 14, controller 16, and / or sensor system 20. The device 10 may include other non-illustrated components as well, such as various housings, power and communications cables, bracing portions, and fastening devices to make the device 10 wearable by the user.

[0057] The device 10 disclosed in the experimental methods and results below is an exoskeleton (i.e., a powered orthosis) configured to assist the user with movement of one or more portions of a biological limb about one or more biological joints. In the disclosed example, the biological joint is a knee joint. Skilled artisans may adapt the control scheme describe below for application to a different joint, such as the ankle joint, or to multiple joints, such as any combination of one or more of hip, knee, and ankle joints.

[0058] FIG. 2 illustrates an illustrative method 100 of developing the control scheme 28 based on an energy-shaping framework. In this case, the control scheme 28 employs a single control law 30, which is a task-adaptive knee torque function. The objective of the task-adaptive knee torque function is to approximate the biological knee moment (torque) across common LLC tasks using kinematic and ground reaction force (GRF) feedback. To make the controller stable and predictable, an energy-shaping framework is implemented, which has previously been used to design energetically passive controllers within the stance or swing phase of user gait. This style of controller alters the dynamics of the user’s legs in an assistive manner without timebased trajectories or explicit classification of locomotion type. A theoretical guarantee of passivity — i.e., a net exoskeleton energy input that is less than or equal to zero — provides certainty that the total energy of the human user’s leg remains under human control. However, such strict energetic passivity allows energy injection only when switching between stance and swing controllers, which precludes continuous energy injection throughout the gait cycle.

[0059] In the presently disclosed control scheme 28, the passivity requirement is relaxed to permit both small continuous injections and large event-based injections of energy as in Gregg et al. As an example of permitting small continuous energy injections, a gravity compensation term can be employed at the knee that depends on the global angle of the lower leg member 24 during phases of underactuation. As an example of the large event-based energy injections, a pre-loaded ascent spring can be activated when the foot touches down in order to help propel the user when ascending stairs or ramps, where the amount of energy injection is a predictable function of the initial knee flexion angle.

[0060] Unfortunately, for purposes of LLC tasks, the Gregg et al. controller relies on purely data-driven optimization to choose dozens of unintuitive controller parameters which, while well-suited to normal everyday ambulation, make it nearly impossible to adjust for different use cases like LLC. To construct an intuitive and customizable controller, a short list of basic physical components that obey the assumptions of Gregg et al. is first compiled. These components include springs, dampers, inertial compensation, and gravity compensation. Since each one of these components maps from measurable quantities to torque, they can serve as basis functions 110 to parameterize the control law. Deviation from the Gregg et al. energy-shaping framework is then required. These core torque bases 110 are heuristically modulated at step 120 of FIG. 2 into specialized task- and phase-specific behaviors. Task-specific stance torque bases 130 (e.g., ascent spring) can be obtained that are suitable for specific task categories (e.g., incline walking), and general swing torque bases 140 (e.g., modified gravity compensation) can be obtained that are universally helpful in swing. Part of this specialization involves “phase sensitization” using a robust phase variable (e.g., the leg angle). This can, for example, suppresses the ascent spring in late stance where it would hinder the transition to swing.

[0061] Next, “task sensitization” 150 can be used to modulate the specialized stance torque basis functions 130 to accommodate differences between activities and variations within activities. For example, the aforementioned ascent spring injects net positive energy only for incline or stair ascent tasks with higher energy injected for steeper inclines, and a non-ascent springdamper absorbs energy only for decline, stair descent, and level walking tasks. Task sensitization 150 essentially involves scaling the specialized basis functions 130 by smooth functions of tasksensitive signals — e.g., by the height differential between the two ankle joint centers at heelstrike which is sensitive to terrain incline. Summing the task-sensitized stance basis functions at step 160 of FIG. 2 results in the task-adaptive stance torque function 170, which is parameterized fully by kinematic and GRF feedback. Similarly, summing the general swing torque basis functions 140 at step 180 of FIG. 2 results in the task-invariant swing torque function 190. Finally, the GRF -weighted convex combination of the task-adaptive stance torque function 170 and task-invariant swing torque function 190 at step 200 of FIG. 2 provides the final task- adaptive knee torque as the control law 30. A detailed example of the development of the control scheme 28 used in the prototype controller 16 and exoskeleton 10 is discussed below.

[0062] Control Scheme Development

[0063] For ramp and stair ascent tasks, the knee produces primarily net positive work and is actively involved in driving the body through the gait in stance phase. The ascent spring can be designed as a virtual unidirectional torsion spring located at the knee joint. This spring has a fixed neutral point at 0 degrees knee flexion, and is unlatched for hyperextension angles. This spring is virtually pre-loaded “free of cost” in the swing phase, and then injects energy at heelstrike depending on the amount of knee flexion at heelstrike. The ascent spring torque has the form where kais the ascent spring constant and 0kis the knee angle, which is positive in flexion. ka= 50 Nm / rad in the disclosed example. The sigmoid function <J is generally defined as mapping 0 from (- inf, inf) 1— (0, 1) given sigmoid slope m and offset d. 0lais the so-called “leg angle” — i.e., the angle of the line between the hip joint center and the ankle joint center with respect to vertical, which is positive when the hip joint center is anterior to ankle joint center. The sigmoid slope m01aand offset d0}aare chosen such that the sigmoid function of the leg angle acts to inhibit the effect of the ascent spring in late stance. In the disclosed example, the slope m01a= -50.0 and the offset d01a= 0.2. Finally, is a binary variable that is 1 when the respective leg is leading — i.e., experienced the latest heelstrike — and 0 when trailing, effectively deactivating the ascent spring where it is not needed.

[0064] For non-ascent tasks, including level walking and stair and ramp descent, the knee produces primarily net negative work and is involved in absorbing impact at heelstrike and subsequently lowering the body smoothly. The non-ascent spring can be designed as a virtual unidirectional torsion spring at the knee with a variable neutral point, which is set as the angle of the knee joint at heelstrike. The spring absorbs energy by loading itself along with the knee flexion after heelstrike. Since the spring is unidirectional, it does not apply any torque for knee flexion angles less than the angle at heelstrike. A special energy absorption limiting feature was added in this case, based on subject feedback, and acts as a taper in torque for large knee flexion angles beyond the angle at heelstrike and allows for a smoother transition to mid-stance. This can be implemented by allowing the virtual spring to undergo “plastic deformation” — i.e., a decrease in stiffness for strains beyond the elastic limit. In this example, to provide damping for impact absorption, a unidirectional virtual damper was added parallel to the spring. The damper provides no resistance for knee extension velocities. The spring damper has the form

[0065] Tna Step (0k)] • Step where Kna(0^ax) is the non-ascent spring stiffness function based on 0^x, the maximum knee flexion (spring deflection) angle achieved for the current gait cycle, where 0kd= 0k— 0khsfor knee angle 0kand its value at heel strike 0khs. In particular, Kna{0^x>) = kna■ ) where knais the stiffness of the non-deform ed non-ascent spring, mBais the slope, and dnais the offset. In the disclosed example, kna= 40 Nm / rad, mna= -15.0, and dna= 0.47. The sigmoid function tapers this spring stiffness to emulate the plastic deformation process beyond the elastic region. Additionally, cnais the damping coefficient for knee velocity 0k, which is 1.0 Nm- s / rad in the disclosed example. Unit step functions step( ) are used to make the spring and damper unidirectional so they only apply knee extension torques and only for knee flexion angles more than the spring neutral point 0khs. Similarly to the ascent spring, the leg angle inhibits the activation of the non-ascent spring in late stance via the last sigmoid function. Similarly to the ascent spring, works to deactivate the non-ascent spring for the trailing leg where it is not needed.

[0066] The LL spring is a virtual torsion spring at the knee with a fixed neutral point of 0 degrees knee flexion. The spring stiffness is however tapered in a nonlinear fashion dependent on knee angular velocity and knee angle itself. At a high level, the spring in this case is designed to become more compliant with higher magnitudes of knee flexion angular velocity such that it minimally impedes the intentional lowering (knee flexion) portion of LL but also provides a strong boost for the intentional lifting (knee extension) portion.

[0067] Further, the angular velocity dependence may be modified by the depth of the squat — i.e., the amount of knee flexion angle. For deeper squats, the spring stiffness is less sensitive to knee flexion angular velocity. This feature can help provide a bracing effect at the bottom of the squat, which is a biomechanically compromised position, as the knee flexion velocity decreases. The LL spring has the form where kLLis the spring constant of the LL spring, x}and x2scale and shift the knee anglemodulating sigmoid, respectively, and F^1is the GRF normal to the ipsilateral foot normalized by body weight. In the disclosed example, k = 76.0 Nm / rad, mLLi= -1.0, x = 2.0, mLLz= 5.0, dLL= 1.0, and x2= -1.5. Note that a personalized quasi-stiffness LL controller can be seamlessly integrated into this framework in place of the given LL spring.

[0068] The basis function for partial gravity compensation during stance has the form step (0th), (5) where gstscales the magnitude of gravity compensation for stance, and 0this the global thigh angle, which is positive for thigh anterior to vertical. In the disclosed example, gst= 34 Nm. The unit step function can restrict the torque to only flexion torques which are found to be helpful in propelling the body forward in late stance and allow a smooth transition to the flexion torque required for foot clearance in early swing.

[0069] In the step of task sensitization 150, the task-specific stance torque bases 130 can be scaled by task functions — i.e., sigmoid functions parameterized by important “task variables” to make their amplitudes sensitive to the current task. To modulate the torque basis functions based on terrain incline, and implicitly distinguish between ascent and non-ascent tasks, the height (y- coordinate) of the ankle joint center (AJC) of the leading leg relative to the trailing leg’s AJC at the latest heel-strike can be used. This task variable is denoted here as <5Ajcy- In this case, a subject height of 180 cm was assumed, with thigh and shank lengths of 46.8 cm and 44.6 cm, respectively.

[0070] To modulate the gait basis functions Ta, Tna, TgraVst, versus non-gait basis function TLL, four task variables can be used: 1) the distance between the AJCs of the two legs (<5Ajcdist)> the GRF at the ipsilateral and contralateral heels normalized by body weight (F^ andhce°etra), and the GRF at the contralateral foot (FQRFtra). Prior to operating on the torque basis functions, all task variables are transformed to a number in the range [0, 1] by intuitively constructed sigmoid functions, such that they only work to scale down the torque of the torque basis functions they act on and thereby preserve boundedness of torque. The modulated (task sensitive) stance basis functions are presented next.

[0071] The modulated LL torque basis, after task sensitization in this example, is where suitable values include mF= 75.0, dF= 0.15, mAjcdlst= 5.0, and dAjcdlst= 14.0. It is noted that the bipedal model is co-planar with a common hip joint center (HJC), such that zero distance between the two AJCs represents perfect symmetry of the bilateral joint configuration, assuming either knee cannot be hyper-extended. It is assumed this case implies a squatting posture. It may also be assumed that both heels will be in contact with the ground when squatting assistance is required, although this assumption can be relaxed via the modulating sigmoid for users not able to keep their heels on the ground during squatting.

[0072] The modulated ascent spring torque basis, after task sensitization in this example, is where suitable values include mAjCy=5.0 and dAjCy= 8.0. The first sigmoid modulates the torque based on height of the leading leg’s AJC relative to the trailing leg’s AJC, both measured at the latest heelstrike to obtain a constant value throughout the gait cycle. This mimics a higher support torque and corresponding higher energy injection for climbing steeper inclines. The last sigmoid prevents activation of the ascent spring for symmetric postures such as squatting. Note that the modulated ascent spring also supports activity transitions. For transitions from level walking to ascents, the stance phase of the trailing leg experiences the level-walking spring stiffness, determined by bAjcyof level walking, which is appropriate as the body does not need to be lifted in this phase. The subsequent stance phase of the leading leg experiences the ascent stiffness to lift the body upwards. For the opposite transition direction, the stance phases of the trailing leg and the leading leg both experience the ascent stiffness, determined by bAjcyof ascent, as needed to lift the body upward.

[0073] The modulated non-ascent spring torque basis, after task sensitization in this example, is min([cr(<5AjCdist, mAJCdist, dAJCdist) + a(FGc^tra, -mF, dF)], 1.0) ’ where suitable values include 7ftAjcy=_10.0 and dAjcy= 7.0. The first sigmoid serves a similar purpose as it does for the ascent spring with the difference being that energy is absorbed rather than injected in the first half of the stance cycle. The second sigmoid down-modulates the spring torque when significant weight transfer to the leading leg is unlikely — i.e., when the trailing heel is on the ground (contralateral heel GRF is high). The third term ensures the non-ascent spring remains inactive during LL (low AJC distance at latest heelstrike), but transitions to an active state as the contralateral foot is lifted off (low contralateral GRF). This enables LL-to-carrying transitions. Note that the modulated non-ascent spring similarly supports activity transitions. For transitions from level walking to descent, the stance phase of the trailing leg experiences a lower stiffness than steady-state descent, determined by bAjCyof level walking. But this spring still supports the lowering of the body till the heelstrike of the leading leg on the staircase or ramp. The subsequent stance phase of the leading leg experiences the steady-state descent stiffness as desired. For the opposite transition direction, the stance phases of the trailing leg and the leading leg both experience the steady-state descent stiffness, determined by <5AJCyfor descent to support lowering the body. The modulated stance gravity compensation torque basis is

[0074] Here, both sigmoids ensure the gravity compensation sigmoid is not active during non-gait tasks, such as LL tasks. Note that the modulated stance gravity compensation is invariant to changes in gait tasks and their transitions.

[0075] Finally, the stance torque is the sum of the modulated basis functions, given as

[0076] — > —mod i —mod i —mod i —mod lst ‘• LL '1alna ‘•grav -

[0077] For the swing phase of the gait cycle, angular velocity-modulated gravity compensation, inertial compensation, and a unidirectional virtual spring / damper can be provided for all tasks. The purpose of the gravity compensation is to provide assistance to lift the shank in early swing for leg clearance. To prevent gravity compensation from hindering the free pendular downswing in mid-swing, it can be modulated by angular velocity. The angular velocity-modulated gravity compensation basis function has the form where gswscales the magnitude of gravity compensation for swing, and 0shis the global shank segment angle, which is positive when the shank is anterior to vertical. In the disclosed example, w = 8.0Nm.

[0078] The purpose of inertial compensation is to provide additional assistance to accelerate the shank forward during mid-swing. The inertial compensation basis function has the form Step ( 0sh) ’ Step 0k)< (8) where aswis the maximum inertial compensation torque, x3modifies the sensitivity of the torque to knee angular acceleration 0k, and the step functions ensure the inertial compensation torque is zero for angular acceleration in flexion and for shank orientation anterior to vertical. Suitable values include asw= 0.2 Nm and x3= 0.2.

[0079] The virtual spring-damper prevents knee hyperextension and provides bracing at small knee flexion angles in late swing to mimic the flexion torque pulse seen in normative gait. The spring / damper has the form

[0080] Tsdsw= [-fcsw • ex^~e^ + csw• 0k• step • (-0fc)] • step (0kq- 0k), (9) where kswis the spring constant, x4modifies the sharpness of the spring torque, is the constant spring neutral angle, and cswis the damping coefficient. The step functions make the spring and damper unidirectional, such that they only apply knee flexion torques and only for knee flexion angles less than the spring neutral point. Suitable values for the constants include fcsw= 0.6 Nm / rad, x4= 3.0, 0.17 rad, and csw= 0.2 Nm s / rad.

[0081] The task-invariant swing torque is the sum of the swing basis functions and is given as Tsd— Tgravsw+ Lnertialsw+ %dsw(10)

[0082] Note that none of the individual components of this swing controller prescribe taskspecific behavior, nor are they modulated by any task-sensitive signal. Thus, the swing controller is truly task-invariant and can assist the swing phase of any task or task transition.

[0083] The final control torque is a convex combination of the stance and swing torques parameterized by the ipsilateral GRF as follows: where a = dGRFu); mGRFu= 30.0 and dGRFu= 0.2. For comfort purposes, especially at heel-strike during ascent tasks, the rate of torque increase in the extension direction (slew rate) was limited to 200 Nm • s ' in the disclosed example. The torques were limited to ±25 Nm for the safety of the motors.

[0084] The coefficients for the springs and dampers, gravity compensation gains, and slopes and offsets of important sigmoids were optimized within physically sensible bounds, using FMINCON in MATLAB. The optimization minimized the L2 loss between the corresponding control torques and normative human torques, scaled to provide 25% assistance for an 80 kg male subject of 1.8 m height, for multiple tasks: level walking at 0.5 m / s and 1.5 m / s, ramp ascent at 5.2° and 11°, ramp descent at 5.2° and 11°, stair ascent at 4 inch and 7 inch step heights, and lifting-lowering at a fast speed. The optimization was formalized as: mini Amize 2) subject to where A contains the parameters to be optimized, Tc tand TN tG IR1O1X1comprise the control torques and normative torques respectively over a 101 point normalized task cycle for the tthtask, VFcycleG D01x101is a diagonal weighting matrix for each point in the task cycle, and wtis a weighting for the tt / ltask. Finally, the matrix-weighted norm || x ||wis defined as / xTWx. The optimization variable vector A is constrained to be within manually chosen, physically sensible bounds Alband Aub. FIG. 3 shows the fitting results, and TABLE I shows the variation accounted for percentage (VAF%) for each task, assessing the correlation between optimized control torques and normative torques.

[0085] TABLE I

[0086] The optimized controller was then implemented on a bilateral knee exoskeleton as described below. This provided a starting point for a manual coefficient tuning process based on subjective feedback in pilot testing. After a satisfactory controller was obtained, the same set of coefficients were used for all participants in the study.

[0087] Experimental Conditions

[0088] The optimized controller was implemented on an improved, bilateral version of the M- BLUE knee exoskeleton module (C. Nesler et al., Enhancing Voluntary Motion With Modular, Backdrivable, Powered Hip and Knee Orthoses, IEEE Robotics and Automation Letters 7, 6155— 6162 (2022)) as illustrated schematically in FIG. 1. Each knee module has a highly backdrivable (< 2 Nm) commercial actuator 14 (T-Motor AK80-9), including a high-torque motor and an internal 9:1 planetary gearset. The motor is driven by the FASTER motor controller (Dephy, Inc., Boxborough, MA, USA) with custom firmware to bypass the default thermal limits. A thermal model-based torque limiter was implemented that smoothly tapers the actuator torque based on estimated coil temperature, allowing short, 1 to 2 second bursts of much higher peak torques than the default setting — up to 25 Nm. A ground reaction force sensor (IEE Sense) based on a matrix of force sensitive resistors (FSRs) was implemented as part of the sensor system 20. Also, comfort and practicality were improved by attaching the Raspberry Pi computation unit on-board as the controller 16, waist-mounted plug-and-play power tool batteries were implemented as the power source 18, and a waist suspension strap was included to prevent exoskeleton shifting on the user.

[0089] The effects of the controller 16 and exoskeleton 10 on performance, posture, muscle activity, and user perception were studied during exoskeleton-assisted multi-terrain LLC tasks. Ten able-bodied participants took part in the study, including five men and five women aged 24- 26 who had prior knowledge and experience of the squat-lifting technique. The study included two sessions and was performed by the participants at two conditions. The two conditions were with the exoskeleton and without the exoskeleton, respectively referred to as the “exo” condition and the “bare” condition. Session 1 was a “fatigued” session, and Session 2 was a “non-fatigued” session, performed on a separate day at least one week after Session 1.

[0090] With reference to FIG. 4, Session 1 was divided into two sub-sessions. Sub-session 1.1 tested the hypothesis that the exo condition mitigates fatigue-induced LL performance deficit compared to the bare condition when squatting posture is enforced. Sub-session 1.1 included a fatiguing phase (also referred to as the pre-fatigue phase) and a post-fatigue phase. In the fatiguing phase, participants performed repeated squat LL cycles until fatigue-induced failure. In the post-fatigue phase, initiated immediately after the fatiguing phase for each participant, participants performed ten squat LL cycles. The post-fatigue squat cycles were timed cycles.

[0091] Sub-session 1.2 immediately followed sub-session 1.1 and involved each participant traversing a multi -terrain LLC circuit based on FIG. 4. In the exo condition, the exoskeleton was unpowered during the pre-fatigue LL cycles and powered during the post-fatigue LL cycles and during the post-fatigue multi-terrain LLC circuit of sub-session 1.2. Session 2 included each participant performing the individual tasks of the multi-terrain LLC in the bare and exo conditions without performing any pre-fatigue cycles.

[0092] Before beginning Session 1, participants underwent exoskeleton acclimation for approximately 15 minutes, during which they traversed the multi -terrain circuit until they felt comfortably attuned to the behavior of the exoskeleton. The participants were informed regarding expected exoskeleton behaviors during different tasks. For the ascent tasks, the participants were informed to expect effects comparable to a pre-loaded spring, likened to a spring-loaded toy car. For descent and level walking tasks, the participants were informed to expect effects comparable to a bracing knee spring that prevents knee buckling. For the liftinglowering task, the participants were informed to expect effects comparable to a velocity dependent torsion spring. For the lowering portion of the LL task, participants were informed that the knee spring will resist downward motion less if it detects a stronger intention to lower — i.e., faster knee bending. For the lifting portion of the LL task, participants were informed that the controller would feel similar to a conventional torsion spring.

[0093] After acclimation with the exoskeleton, important gait parameters were acquired from each participant in the bare condition while traversing each portion of the multi-terrain circuit except level walking. For each participant, maximum knee extension velocity in stance was acquired during ascent tasks, maximum knee flexion velocity in stance was acquired during descent tasks, and maximum knee extension velocity was acquired during the lifting portion of the LL task. These parameters later served as experimental controls in Session 2.

[0094] The purpose of Session 1 was to assess the mitigating effect of the exoskeleton on postfatigue LL performance deficit when enforcing the squat form and to acquire a subjective assessment of its effectiveness during post-fatigue multi -terrain LLC. During the fatiguing phase of session 1.1, participants performed continuous squat LL cycles with a 9 kg kettlebell until participant declaration of fatigue-induced “failure,” with no pauses between cycles. Participants were instructed to verbally declare failure when they felt they could not complete the next cycle with proper squat form without a pause. Declaration of failure marked the end of the fatiguing phase and the beginning of the post-fatigue phase of sub-session 1.1. The post-fatigue phase was a time trial of ten LL cycles during which participants were instructed to pause between successive cycles long enough to be able to complete the next cycle with perceived good squat form. For the exo condition, the fatiguing cycles were performed with the exoskeleton in passive mode, for which the exoskeleton has an imperceptible approximately 1 Nm backdrive torque for the joint accelerations encountered during LL. The exoskeleton was remotely changed to active mode immediately upon participant declaration of fatigue-induced failure. The participants were not aware of the timed nature of the post-fatigue phase and were instructed to focus on their posture.

[0095] Immediately after completing the ten post-fatigue LL cycles, the participants traversed the multi-terrain circuit while carrying the 9 kg mass. The circuit included a 3.7 meter ramp at a 15° incline, a 2 meter level platform, and a five-step staircase with 18 cm step height. The circuit was traversed in a continuous “freestyle” fashion in both directions and included a squat LL cycle at both ends of the circuit to emulate a workplace multi-terrain LLC scenario. Two roundtrip laps of the circuit were completed for each condition. After completing the circuit with the second condition, participants rated the effectiveness of the exoskeleton assistance over the multi-terrain circuit by filling out a modified QUEST questionnaire. The questionnaire gathered discretized ratings from 1 through 5 (not satisfied at all, not very satisfied, somewhat satisfied, quite satisfied, and very satisfied, respectively) for all steady-state portions of the circuit (ramp and stairs ascent / descent, level walking, LL). The participants only considered the effectiveness of the exoskeleton during the post-fatigue circuit traversal.

[0096] The order of the bare and exo conditions was alternated between participants during Session 1, and a minimum break of 15 minutes was enforced between the two conditions.

[0097] The purpose of Session 2 was to assess the effect of the exoskeleton on quadriceps effort on the six individual tasks of the multi-terrain circuit in a non-fatigued state. Participants traversed each portion of the circuit along with a 10 meter level walkway multiple times until at least ten gait / task cycles we obtained for each task in which the corresponding maximum velocities in stance were within ±10% of their baseline values collected during acclimation. Since level walking is relatively more common and natural than the other tasks, a velocity constraint was not enforced. Instead participants were instructed to walk at their natural, selfselected speed. The tasks were repeated with both bare and exo conditions, with the order of the conditions alternated between participants. For the stair and ramp descent tasks, participants were instructed to carefully descend and not skip or hop down in order to emulate safe carrying etiquette. All participants naturally used a knee-dominant stair ascent style during acclimation and data collection trials.

[0098] Post-fatigue LL performance deficit was evaluated by the percent increase in time to complete ten post-fatigue LL cycles with respect to a baseline in Session 1. The baseline was defined as the time required to complete the first ten LL cycles during the pre-fatigue phase at the bare condition for each participant. A single LL cycle is defined as follows. Each cycle started with the participant standing in the upright posture while not holding the weight, followed by the participant squatting and then lifting the weight from the ground to again attain the upright posture, followed by lowering the weight back to the ground and standing upright once again without the weight. Because some participants declared fatigue midway through a cycle, the start time for the post-fatigue phase was defined when the last fatiguing cycle was fully completed. Since pausing between cycles was permitted during the post-fatigue cycles, most participants took a 2 to 3 second pause immediately after declaring fatigue. This initial post-fatigue pause was included in the time-to-complete metric. The time-to-complete analysis was performed offline using sagittal-plane video recordings.

[0099] Although squatting posture was an experimental control in Session 1, an exploratory analysis of lifting posture (peak thorax lean) was performed to study whether the exoskeleton helped participants maintain better squat form. A Vicon motion capture system (Oxford Metrics, Oxford, UK) was used to collect three-dimensional marker trajectories for the Session 1 LL cycles. Retro-reflective markers on the torso were placed at C7, T10, CLAV, and STRN to define the thorax segment. Three additional backup markers were placed on the shoulders and lower back to aid in post-process gap filling. After appropriate data cleanup, gap filling, and filtering, the sagittal plane global thorax angle was calculated with respect to upright standing (zero degrees). First, the peak global thorax angle was determined for each lifting and lowering cycle. Next, the deviation in thorax lean was calculated by subtracting the value of the lowest peak obtained in the pre-fatigue bare condition. Finally, the mean of the deviations from 20 postfatigue squats (10 lifts and 10 lowers) was determined for each subject and condition.

[0100] In Session 2 five electromyography (EMG) electrodes (Trigno Avanti and Snap, Delsys, Massachusetts, USA) were secured to the participant’s right vastus medialis oblique (VMO), vastus lateralis (VL), rectus femoris (RF), biceps femoris (BF), and semitendinosus (ST) to assess muscle activation. Participants performed an MVC procedure comprising explosive jump squats, eliciting maximal dynamic contraction, and maximal isometric contraction against manual resistance, which enabled EMG data to be normalized to %MVC.

[0101] To assess muscular effort the mean of the MVC-normalized RMS signal was calculated for each gait / task cycle. The RMS signal provided by the acquisition software was used for this purpose (RMS window length: 125 ms, 122 ms overlap). Since gross quadriceps effort was of interest, and to reduce the number of degrees of freedom in the statistical analysis, the weighted average of the three quadriceps muscles that were recorded were used to obtain gross quadriceps effort metric. The weighting was based on the respective PCSA of each muscle.

[0102] The power analysis (for 80% power, alpha=0.05) on pilot data for LL performance and EMG returned a sample size of n=10. After collecting data from all ten participants, normality of the data was first confirmed using QQ-plots. Muscular effort was analyzed using a linear mixed model (LMM) in MATLAB with restricted maximum likelihood estimation of parameters. Data from the ten participants were tabulated with information comprising log- transformed quadriceps muscle effort, condition (bare, exo), and sex (male, female). Condition and sex were defined as categorical variables, and a separate LMM was fit for each task, where the condition and sex were fixed effects and subject was a random effect:

[0103] Effort ~ Condition + Sex + (1 (Subject).

[0104] The LMM provided the effect sizes and uncorrected p-values for the fixed factors. The p-values were corrected for multiple comparisons (six tasks) using the Holm Bonferroni correction.

[0105] For the post-fatigue performance and posture metrics, a similar LMM included additional fixed effects of order (bare vs. exo) and pre-fatigue workload (number of LL repetitions before fatigue-induced failure) without having to correct for multiple comparisons. Workload was log- transformed prior to analysis and mean-centered for each sex and condition combination.

[0106] Experimental Results

[0107] For all metrics, linear mixed models (LMMs) were used, with condition (bare, exo) and sex (male, female) being fixed effects and the participant being treated as a random effect. For post-fatigue performance and posture metrics, order (for the bare vs. exo condition) and prefatigue workload (number of fatiguing cycles required to reach fatigue) were treated as additional factors. Sex, order, and workload were found to be non-significant factors in all statistical tests reported below.

[0108] The results confirm the hypothesis that the knee exoskeleton decreases post-fatigue LL performance deficit. Compared to the bare condition, the bilateral knee exoskeleton induced a significant (p < 0.01) 43% reduction in the post-fatigue time deficit, corresponding to 44% (bare condition) and 1% (exo condition) increases in post-fatigue completion times with respect to pre-fatigue bare completion time.

[0109] FIG. 5A includes box plots and individual line plots illustrating the post-fatigue performance deficit for each of participants S1-S10 in the bare and exo conditions. The performance deficit is given as the percent increase in time required for each participant to complete 10 post-fatigue LL squat cycles, where the baseline for the percent increase is the time required for each participant to complete 10 pre-fatigue LL squat cycles. FIG. 5B shows the ensemble averaged LL cycle durations as a function of percent trial progression, where the fatiguing cycles are represented from -100% to 0% progression and the post-fatigue cycles are represented from 0% to 100% progression.

[0110] Compared to the bare condition, the bilateral knee exoskeleton also induced a significant (p < 0.01) 4.4° reduction in peak thorax lean deviation (from the minimum lean angle at the bare condition during the fatiguing phase) during the post-fatigue phase. This corresponds to average deviations of 10.4° (bare) and 6.1° (exo), and average absolute peak thorax lean angles of 53.3° (bare) and 48.9° (exo). FIG. 6A shows the post-fatigue thorax lean deviation distribution for the bare and exo conditions. FIG. 6B shows the ensemble averaged deviations as a function of subsession 1.1 progression. As shown in FIG. 6B, the fatiguing phase of the trial reveals a progressive increase in peak thorax lean with similar slopes and relative magnitudes between the exo and bare conditions. In the post-fatigue phase, the thorax lean trajectories for the two conditions diverge, with the exo condition tending to return to pre-fatigue levels.

[0111] FIG. 7 shows deviations in peak knee flexion angles from their maximum values observed during the fatiguing phase for the respective bare and exo conditions. The fatiguing phase of the trial reveals a progressive decrease in peak knee flexion angles with similar slopes and relative magnitudes between the exo and bare conditions. In the post-fatigue phase, the kneeangle deviation trajectories for the two conditions diverge, on average, with the exo condition more clearly returning to pre-fatigue levels. FIG. 8 shows the ensemble averaged thorax lean profiles during LL cycle, verifying that the peaks in thorax lean occurred at the bottom of the squat LL cycles, as expected.

[0112] FIG. 9 includes histograms of subjective participant satisfaction ratings based on their experiences with the effectiveness of the exoskeleton in aiding post-fatigue multi-terrain LLC tasks. Histograms for lifting-lowering (LL), ramp ascent (RA), stair ascent (SA), level walking (LW), ramp descent (RD), and stair descent (SD) are shown individually. The average participant feedback rating for effectiveness of exoskeleton assistance in traversing the various tasks in post-fatigue LLC was 4.3 out of 5. The individual task averages were 4.8 for LL, 4.8 for stair ascent, 4.3 for stairs descent, 4.2 for ramp ascent, 4.2 for ramp descent, and 3.6 for level walking (LW).

[0113] Quadriceps effort was estimated during the six different LLC tasks for the bare and exo conditions in a non-fatigued state during session 2. Muscle effort was calculated as the mean RMS EMG, normalized to percent of maximum voluntary contraction (MVC) over the gait / task cycle. Quadriceps effort was estimated as the weighted mean of the efforts of the vastus medialis oblique (VMO), vastus lateralis (VL), and rectus femoris (RF) based on their respective physiological cross-sectional areas. The results confirmed the hypothesis that the knee exoskeleton reduces quadriceps effort during non-fatigued, multi-terrain LLC, with the caveat of LW having a non-significant reduction. Compared to the bare condition, the exoskeleton significantly reduced mean quadriceps effort by 22% for LL (p<0.05), 17% for RA (p<0.01), 24% for SA (p<0.01), 10% for RD (p<0.05), 11% for SD (p<0.05), and 3% for LW (p=0.75). FIG. 10 shows the distributions of quadriceps effort for the conditions and tasks tested for each participant.

[0114] A good qualitative match can be observed between the ensemble-averaged quadriceps EMG profiles and the exoskeleton torque profiles in FIG. 11, demonstrating appropriate timing and magnitude of the assistance torque. FIGS. 12-14 show similar trends with the individual quadriceps muscle profiles for the VMO, VL, and RF, respectively.

[0115] FIG. 15 illustrates hamstrings EMG and exoskeleton knee torque during LLC. The figure shows the ensemble averaged hamstrings EMG profiles, including biceps femoris (BF) and semitendinosus (ST) for bare and exo conditions, along with the exoskeleton torque profiles, for all LLC tasks. The exoskeleton controller did not cause any noticeable increase in hamstring activation compared to the bare condition.

[0116] Prior exoskeletons fall short on providing versatile assistance to the critically-involved quadriceps during multi -terrain LLC, which comprise varied tasks including high-torque closed- chain squat lifting as well as open-chain gait tasks such as level walking, ramps, and stairs. In particular, prior LLC studies have not demonstrated holistic (multi-terrain) and multifaceted (muscular, performance, postural, and perceptual) exoskeleton benefits, especially in high- fatigue physical states which are correlated with LBP incidence. As shown in FIG. 3 and TABLE I, the resulting control torques are well-matched to biological torques in-silico and to quadriceps activations in-vivo, as shown in FIG. 11, for all tasks in multi -terrain LLC. The results demonstrate that a knee exoskeleton equipped with a properly designed controller can provide: 1) significant performance and postural benefits to squat LL in a highly -fatigued condition, as shown in FIGS. 5A-6B, and 2) holistic muscle effort reductions during multi-terrain LLC in nonfatigued conditions, as shown in FIG. 10. These results suggest the disclosed controller and exoskeleton can be applied to mitigate fatigue in real-world workplace conditions in which lower back injuries tend to occur.

[0117] The peak knee torque for a fast squat is about 1.4 Nm kg-1, or about 100 Nm for an average person. During the fast LL squats described above, the peak knee torque is required at the bottom of the squat to redirect the body’s downward momentum upward. As shown in FIG. 8, the bottom of the squat is also where the peak in post-fatigue thorax lean was observed — a likely consequence of participants trying to reduce the moment arm on the knee and stress on the fatigued quadriceps in this biomechanically compromised position. The peak of approximately 25 Nm in the experimental exoskeleton assistance torque — about 25% of the normative squat torque — was indeed well aligned with this high-torque squat phase, resulting in good alignment with the quadriceps EMG peaks in the non-fatiguing LL task, as shown in FIG. 11. Accordingly, a 22% reduction in mean quadriceps effort and a 23% decrease in peak activation for non-fatigued LL with knee exoskeleton assistance was realized, when compared to the bare condition. The 23% savings in peak quadriceps force afforded by the exoskeleton was likely sufficient to fully overcome the estimated 25% deficit in force from the induced fatigue. This could explain how the exoskeleton assistance enabled the participants to significantly reduce their fatigue-induced deficit in time to complete 10 LL cycles by 43% and with significantly better posture — i.e., a 4.4 degree decrease in global thorax angle compared to the bare condition, as illustrated in FIGS. 5A and 6A.

[0118] In fact, the exoskeleton assistance limited the fatigue-induced time deficit to a mere 1% increase relative to pre-fatigue levels, as shown in FIG. 5B. On the other hand, the LL durations in the bare condition increased substantially post-fatigue, indicating a persistent fatigue-induced performance deficit. The mean LL cycle durations prior to the participant declaring fatigue have similar magnitudes for both conditions, suggesting the similarity in bare and passive-exo conditions. After participants declared fatigue, both conditions had a peak LL duration at the 1st post-fatigue cycle because they were instructed to pause just enough to recover and continue. The exo condition likely had a smaller pause because participants had more confidence continuing post-fatigue when assistance was expected. The post-fatigue LL feedback rating is also the highest amongst the tasks — almost a perfect 4.8 / 5 — indicating the effectiveness of the exoskeleton assistance at a perceptual level.

[0119] The exoskeleton’s replenishment of the diminished quadriceps force in post-fatigue LL explains the significant reduction in peak thorax lean — i.e., the amount of stooping — for the exo condition compared to the bare condition. A similar rate and magnitude of increase in peak thorax lean can be observed between the bare and passive exo conditions in the fatiguing phase of the session 1.1, as shown in FIG. 6B, indicating similar intensities of fatigue. This observation agrees with a classic study showing quadriceps fatigue leads to a compensatory change in posture. Upon onset of exo assistance in the post-fatigue phase, the peak thorax lean returns toward a pre-fatigue level quickly — within 2 post-fatigue repetitions — indicating mitigation of fatigue. In contrast, peak thorax lean continues to gradually increase in the bare condition postfatigue phase, reaching its highest point at the third post-fatigue repetition before leveling off at a noticeably higher value than the pre-fatigue level, suggesting retention of high-fatigue postural effects. Without exoskeleton assistance, the participants likely offloaded their highly fatigued quadriceps by leaning more forward and thus reducing the moment arm on their knees. In contrast, exoskeleton assistance enabled participants to use more of their legs thus precluding stooping, which in indicated in the ensemble average deviation in peak knee flexion angle in FIG. 7. Moreover, while fatigue decreases peak knee flexion and knee range of motion at similar rates and magnitude in both conditions, exoskeleton assistance is able to partially recover the loss in range of motion, thus allowing for a straighter and safer back posture.

[0120] The observed changes in posture were likely subconscious phenomena, because participants were reminded to: 1) maintain proper squat posture during the fatiguing phase, and 2) take a longer pause if they perceived an imminent compromise in posture during the postfatigue phase. This may resemble real workplace scenarios, where fatigue-induced injurious changes in posture occur subconsciously to override worker training and guidelines. Importantly, this study is the first to show a partial reversal of compensatory changes in posture via exoskeleton assistance. This intervention addresses the root cause of postural changes — fatigued quadriceps — rather than simply alerting the user of detected posture changes.

[0121] A holistic decrease in quadriceps effort was observed without negatively affecting the antagonistic hamstrings for all carrying tasks in non-fatigued Session 2, as illustrated in FIGS. 10 and 11. It is believed that one of the main factors leading to these consistent findings is the hybrid in-silico and in-vivo approach to controller development, whereas prior task-invariant controllers are limited by and restricted to a normative dataset that informs their data-driven optimization and learning processes.

[0122] The peaks in assistive torques are well-aligned with the peaks in quadriceps activations across tasks, as shown in FIG. 11, explaining the significant reductions in quadriceps effort in the exo condition. It is noted that human muscles are much more efficient in performing negative work, as required in descent tasks, compared to positive work, as required in ascent tasks. It is therefore understandable that participants found the assistive torques the most helpful in postfatigue session 1.2 in the tasks requiring high positive work, as shown in FIG. 9. Indeed, stair ascent, liftin-lowering, and ramp ascent, in that order, resulted in the highest reductions in quadriceps effort. The level walking task of session 1.2 resulted in the lowest knee moments of all tasks, as shown in FIG. 3, and accordingly received the lowest assistance torque magnitude compared to other tasks, as shown in FIG. 11. Although the assistive torque was well-aligned to the quadriceps activation profile, the level walking received the lowest subjective effectiveness rating of all tasks in post-fatigue at 3.6 out of 5. The most common feedback was that the exoskeleton did not help in level walking as much as the other tasks. Consistent with this finding, level walking was also the only task that had a non-significant decrease in quadriceps effort with exoskeleton assistance.

[0123] It is suspected that soft-tissue compliance prevents effective energy transfer from the relatively smaller exoskeleton torques associated with level walking — i.e., the knee simply does not need much assistance in this task. Nonetheless, the disclosed controller and exoskeleton satisfactorily eliminated any potential problem with high peaks in hamstrings activation profiles in late stance in level walking, as shown in FIG. 15. This is likely the result of: 1) adding knee flexion torques via gravity compensation to aid the hamstrings in late stance, and 2) the use of a relatively light-weight exoskeleton configured to actuate only one joint.

[0124] It is to be understood that the foregoing description is of one or more embodiments of the invention. The invention is not limited to the particular embodiment(s) disclosed herein, but rather is defined solely by the claims below. Furthermore, the statements contained in the foregoing description relate to the disclosed embodiment(s) and are not to be construed as limitations on the scope of the invention or on the definition of terms used in the claims, except where a term or phrase is expressly defined above. Various other embodiments and various changes and modifications to the disclosed embodiment(s) will become apparent to those skilled in the art.

[0125] As used in this specification and claims, the terms “e.g.,” “for example,” “for instance,” “such as,” and “like,” and the verbs “comprising,” “having,” “including,” and their other verb forms, when used in conjunction with a listing of one or more components or other items, are each to be construed as open-ended, meaning that the listing is not to be considered as excluding other, additional components or items. Other terms are to be construed using their broadest reasonable meaning unless they are used in a context that requires a different interpretation.

Claims

CLAIMS1. A powered lower limb assistive device (10), comprising: an articulated frame (12) including a first frame member (22) and a second frame member (24) interconnected at a pivot joint (26) for relative rotation about the pivot joint, wherein the device employs a task-adaptive energy-shaping control scheme (28) to modulate torque at the pivot joint, and wherein the control scheme employs a single control law (30).

2. The assistive device (10) according to claim 1, wherein the control law (30) is based on a combination of a task-adaptive stance torque function (rst) and a task-invariant swing torque function (Tsd) and operates without discrete identification of gait phase.

3. The assistive device (10) according to claim 1 or claim 2, wherein the control law (30) is based in part on torque basis functions for at least one virtual spring, at least one virtual damper, gravity compensation, and / or inertial compensation.

4. The assistive device (10) according to any one of the preceding claims, wherein the pivot joint (26) is a knee joint and the assistive device is a knee exoskeleton.

5. The assistive device (10) according to claim 4, wherein the control law (30) is based in part on a task-specific torque basis function (ra) for a virtual unidirectional spring located at the joint (26), the virtual spring operating to inject energy after heelstrike, the injected energy being limited by a maximum angle between the first and second frame members (22, 24).

6. The assistive device (10) according to claim 5, wherein the torque generated by the virtual spring is prevented when a leg angle (0[a) of the user exceeds a predefined threshold such that joint movement is not hindered during late stance.

7. The assistive device (10) according to claim 5 or claim 6, wherein the task-specific torque basis function (ra) is sensitized such that the injected energy is scaled according to a degree of incline or a step height along which the user ambulates and such that energy injection is prevented when the user assumes a symmetric posture or when a leading leg of the user becomes a trailing leg.

8. The assistive device (10) according to any one of claims 4 to 7, wherein the control law is based in part on a task-specific torque basis function for a virtual unidirectional spring (ra) and damper (rna) located at the joint, the virtual spring operating to absorb energy after heelstrike in an amount dependent on a change in angle between the first and second frame members (22, 24) after heelstrike, limited by an elastic limit of the virtual spring, the virtual damper operating to dissipate energy by an amount dependent on an angular velocity (0k) of the joint (26).

9. The assistive device (10) according to claim 8, wherein torque (Ta, Tna) provided by the virtual spring and damper is prevented when a leg angle (0[a) of the user exceeds a predefined threshold such that joint movement is not hindered during late stance.

10. The assistive device (10) according to claim 8 or claim 9, wherein the task-specific torque basis function (ra, Tna) is sensitized such that energy absorption is scaled according to a degree of decline or step height along which the user ambulates and such that torque provided by the virtual spring and damper is prevented when the user assumes a symmetric posture or when a leading leg of the user becomes a trailing leg.

11. The assistive device (10) according to any one of claims 4 to 10, wherein the control law (30) is based in part on task-invariant torque bases including gravity compensation (Tgravswinertial compensation (Tinertiaisw), and a virtual unidirectional spring and damper12. The assistive device (10) according to claim 11, wherein the gravity compensation (Tgravsw) is modulated based on angular velocity (0k) of the joint (26) such that user foot clearance is assisted during early-to-mid swing without hindering forward leg swing during mid-to-late swing.

13. The assistive device (10) according to claim 11 or claim 12, wherein the inertial compensation (Tinertiaisw) is modulated based on a global angle (0sh) of the second frame member (24) such that forward leg swing is assisted while the global angle is within a predefined range.

14. The assistive device (10) according to any one of claims 11 to 13, wherein the virtualunidirectional spring and damper (Tsdsw) dissipates energy during late swing to assist in preventing joint hyperextension.

15. The assistive device (10) according to any one of the preceding claims, wherein the control law (30) is employed during a squatting task performed by the user.

16. The assistive device (10) according to any one of the preceding claims, wherein the control law (30) modulates the torque such that a peak torque at the pivot j oint (26) is provided when the user transitions from a squatting movement to a standing movement during a squatting task.

17. The assistive device (10) according to any one of the preceding claims, wherein a peak torque is provided via a virtual torsion spring (TLL) acting on the joint (26), the stiffness of the virtual spring being a function of joint angle (0fc) and joint angular velocity (0k) such that joint movement is not hindered during the squatting movement.

Images