Limb-assistive device with energy shaping

Abstract

A lower limb exoskeleton includes a controller employing a modular multi-task energy shaping framework for any lower-limb joint configuration, using a convex formulation of a multitask, data-driven optimization problem. In addition to being modular and convex, the framework offers adjustment of the torque goal, distributed processing across devices, and smooth transitions between stance and swing that are incorporated into the stability analysis. The framework can reduce overall muscle activation when considering multiple tasks among multiple individual users.

Description

[0001] LIMB-ASSISTIVE DEVICE WITH ENERGY SHAPING

[0002] This invention was made with government support under 1949869 awarded by the National Science Foundation, and 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 user tasks.

[0005] BACKGROUND

[0006] Backdrivable lower-limb exoskeletons have the potential to assist volitional motions of able-bodied users and people with mild to moderate gait disorders. While such devices have demonstrated the mechanical capabilities necessary to assist these activities in specific contexts, a control framework does not currently exist that can be deployed on any j oint and assist any activity of daily life in a provably stable manner.

[0007] Exoskeletons on the market today in rehabilitation applications provide a user with complete assistance using highly geared actuators that track pre-defined reference trajectories. While such designs may be appropriate for severe impairments like paraplegia, they hinder users from populations with full or remnant volitional control over their limbs. Highly geared actuators introduce high mechanical impedance at the joint(s) and therefore impede the user’s ability to move the device under their own power — i.e., to backdrive the actuators. Although these actuators can still achieve normative joint motions through kinematic control methods for a variety of tasks, the pre-defined trajectories inevitably conflict with the user’s desired motion.

[0008] Low-impedance actuation systems represent a shift from task-specific, kinematic tracking to task-invariant torque-tracking approaches that can deliver partial rather than complete assistance to the user. Backdrivable designs allow augmentation of voluntary human motion, compensation for human-exoskeleton mass / inertia, and direct amplification of human strength. An open-source hardware platform known as the Modular Backdrivable Lower-limb Unloading Exoskeleton (M- BLUE) has recently been developed and exemplifies a lighter, minimalist structure for attaching quasi-direct drive (QDD) actuators to people.

[0009] Although backdrivable exoskeletons may eventually prove useful in everyday life, their control systems are not yet up to the challenge. In particular, there is no general -purpose controller available that operates reliably across all the core activities of daily life, allows for adjustment of the control behavior, avoids instability, and handles the wide diversity of possible joint configurations of actuators.

[0010] Several control approaches for backdrivable exoskeleton systems are based on machine learning, including some that achieve task-invariance. As one example, task-invariant biological joint moments can be directly estimated from angle measurements at the hip, though this has only been demonstrated offline. Stair-specific strategies have been the subject of knee exoskeleton research. Work in this direction can draw inspiration from the gait analysis field, where estimating joint torques and ground reaction forces from inertial sensors is an active area of research. Machine learning has also been explored to estimate gait completion percentage or continuous gait phase using shank- or thigh-mounted sensors. Such phase estimates can be used to look up a predefined assistive torque for a specific task. While some estimators attempt to robustly estimate phase across tasks, they typically require secondary task classifiers to adapt assistance appropriately between tasks. Generally speaking, these black-box strategies based on machine learning offer no guarantees of safety and stability, especially outside the training dataset, requiring empirical validation for every exoskeleton application.

[0011] Model-based phase estimators offer a more analytical alternative to machine learning approaches, but such models are currently limited to walking or a continuum of ramp walking tasks. While there is some potential to track both phase and task variables (e.g., stride length and ground inclination) and to use the phase estimate to apply task-appropriate torque, applications have been limited to a narrow set of periodic walking behaviors.

[0012] Adaptive oscillators are a model-free alternative that purport to offer the ability to track any task with provable convergence properties, but they suffer from multiple disadvantages. For one, adaptive oscillators depend on periodicity in locomotion and do not easily handle non-steadystate tasks outside of the laboratory. Additionally, they cannot automatically adapt the torque profde to be task-appropriate. As such, a task-invariant and trustworthy solution for general- purpose control of backdrivable exoskeletons remains elusive.

[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-invariant passivity -based energyshaping control scheme to modulate torque at the pivot joint, and the control scheme employs a common control law during stance phase and swing phase. The assistive device may include any technically feasible combination of the above-listed features and / or the following features:

[0015] - the control law scales torque using vertical ground reaction force measurements; the control law is based on able-bodied data fit to a model including a kinematic chain having four links interconnected by three revolute j oints, the four links corresponding to a foot, a lower leg, an upper leg, and a torso, and the three revolute joints corresponding to an ankle joint, a knee joint, and a hip joint; the kinematic chain is a first kinematic chain representing an ipsilateral leg

[0016] - the model including the first kinematic chain further includes a second kinematic chain representing a contralateral leg having four links interconnected by three revolute joints, and an interaction wrench linking the first and second kinematic chains at a respective revolute joint of each kinematic chain; the control law employs negative power tapering where net-negative work is performed at the joint in the able-bodied data; a magnitude of the negative power tapering is adjustable; the control law is represented by Weight • LOA%, where τ is the torque at the pivot joint is a basis matrix for the torque, q is a vector including angles corresponding to an orientation of each link of the model, p is a momentum vector corresponding to momentum with respect to each joint of the model, vGRF is the vertical component of a ground reaction force, is the parameter vector that solves an optimization problem applying able-bodied data to the model, Weight is of a user of the device, and LOA% is a level of assistance that scales the torque to a fraction of modeled torque; a sensor system including one or more sensors, an actuator operable to apply the torque at the pivot joint, and a controller receiving an input from each of the one or more sensors and controlling the actuator based on each input and on the control scheme; at least one of the sensors provides a measurement of an unactuated joint;

[0017] - the first and second frame members are adapted for attachment to respective first and second limb portions of a user, and the sensor system includes a first sensor configured to measure a global angle of the first limb portion and a second sensor configured to measure a global angle of the second limb portion, wherein the controller determines an angle between the first and second limb portions as a difference of the measured global angles;

[0018] - the articulated frame includes a third frame member interconnected with the first or second frame member at a second pivot joint and the control scheme modulates torque at both pivot joints;

[0019] - the articulated frame includes a fourth frame member interconnected with the third frame member at a third pivot joint and the control scheme modulates torque at all of the pivot joints;

[0020] The device is a bilateral assistive device, wherein at least one of the first and second frame members is configured for attachment to a first lower limb of a user, the bilateral assistive device further comprising a third frame member and a fourth frame member interconnected at a second pivot joint, at least one of the third and fourth frame members is configured for attachment to a second lower limb of a user, and the taskinvariant passivity-based energy-shaping control scheme modulates torque at the second pivot joint; the control scheme is modular such that the control law can be applied to a hip-assistive joint, a knee-assistive joint, or an ankle-assistive joint; the articulated frame includes two or more frame members, including the first and second frame members, each frame member being connected to another of the frame members at a respective pivot joint, the assistive device further including one or more actuators corresponding in number to the number of pivot joints, each actuator being operable to apply a torque at the corresponding pivot joint, and_a controller employing the task-invariant passivity-based energy-shaping control scheme to modulate torque at each pivot joint, wherein the control scheme is modular and optimizable using able- bodied data to provide the control law for an arbitrary number of pivot joints selected from: a first hip-assist joint, a second hip-assist joint, a first knee-assist joint, a second knee-assist joint, a first ankle-assist joint, and a second ankle-assist joint; each of the first and second frame members is adapted for attachment to respective first and second limb portions of a user, and the assistive device further includes: an actuator operable to apply the torque at the pivot joint, a sensor system including one or more sensors, and a controller receiving an input from each of the one or more sensors and controlling the actuator based on each input and on the task-invariant passivity-based energy-shaping control scheme, _wherein the sensor system includes a first sensor configured to measure a global angle of the first limb portion and a second sensor configured to measure a global angle of the second limb portion, and wherein the controller determines an angle between the first and second limb portions as a difference of the measured global angles.

[0021] BRIEF DESCRIPTION OF THE DRAWINGS

[0022] FIG. 1 is a schematic representation of a powered limb-assistive device.

[0023] FIG. 2 schematically illustrates a human-exoskeleton system modeled as interlinked 4-link sagittal plane monopods.

[0024] FIG. 3 A illustrates a human lower leg at heel contact during locomotion.

[0025] FIG. 3B illustrates the leg of FIG. 3 A at a flat foot position during locomotion.

[0026] FIG. 3C illustrates the leg of FIGS. 3A and 3B at toe contact during locomotion.

[0027] FIG. 4A is a photographic side view of a user wearing a knee-assist embodiment of the powered limb-assistive device.

[0028] FIG. 4B is a schematic side view of the knee-assist device of FIG. 4A.

[0029] FIG. 5 A is a photographic front view of the user and knee-assist device of FIG. 4A.

[0030] FIG. 5B is a schematic front view of the knee-assist device of FIGS. 4A and 5 A.

[0031] FIG. 6A is a photographic side view of a user wearing a hip-assist embodiment of the powered limb-assistive device.

[0032] FIG. 6B is a schematic side view of the hip-assist device of FIG. 6A.

[0033] FIG. 7A is a photographic front view of the user and hip-assist device of FIG. 6A.

[0034] FIG. 7B is a schematic front view of the hip-assist device of FIGS. 6A and 7A. embodiment of the powered limb-assistive device.

[0035] FIG. 8 is a schematic view of an ADL circuit used for experiments with the limb-assistive device.

[0036] FIG. 9 illustrates across-subject comparisons of total muscle effort between bare condition and four different exoskeleton conditions.

[0037] FIG. 10 illustrates across-subject comparisons of the correlation coefficient between EMG muscle activation and the applied joint torques for each ADL task.

[0038] FIG. 11 illustrates EMG activation for four different muscle groups during each ADL task for bare condition and four different exoskeleton conditions. FIG. 12 illustrates hip and knee torques during each ADL task for bare condition and four different exoskeleton conditions.

[0039] DESCRIPTION OF EMBODIMENTS

[0040] 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 framework is a general-purpose, modular, convex-optimization-based framework for multi-task optimized energy shaping (M-TOES) and includes modular energy bases, convex penalties on incorrect torque sign, and unification of stance and swing controllers by use of insole force sensors. Below, the behavior of the target energy resulting from the control framework is rigorously analyzed, including all possible power leaks due to practical relaxations of the matching conditions and passivity. The stability, modularity, and task-invariance of the resulting controllers are also empirically validated with multiple able- bodied participants performing the primary activities of daily living (ADL) with four different configurations of the above-mentioned M-BLUE system. For each modular configuration of the device, the controller can be optimized across multiple tasks, including multi-speed walking, ramps, stairs, start-stop, and sit-to-stand tasks from an able-bodied dataset.

[0041] FIG. 1 is a schematic representation of a powered limb-assistive device 10 including an articulated frame 12 having a plurality of frame members 14-18 interconnected by pivot joints 20, 22. Each frame member 14-18 is rotatable with respect to another frame member about one of the pivot joints 20, 22. The device 10 includes one or more actuators 24, 26 that provide a controllable torque TK, TAat each joint 20, 22. A controller 28 programmed with a control scheme 30 controls each actuator 24, 26 based on the control scheme and inputs from a sensor system 32 that includes one or more sensors. The control scheme 30 may be an energy-shaping control scheme as discussed further below.

[0042] The actuators 24, 26, controller 28, and sensor(s) 32 are illustrated schematically in FIG. 1. Each actuator 24, 26 may be an electric motor (e.g., a servomotor) fixedly mounted along one of the frame members 14-18 and operably coupled with another frame member for relative frame member movement about the shared joint. For example, a static motor housing may be mounted at a fixed position along one frame member on one side of the joint with movement of the rotor transmitted to the frame member 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 away from the corresponding joint if desired, or the actuator may be coaxial with the respective joint.

[0043] The controller 28 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 20, 22 in accordance with the control scheme 30. The controller 28 is depicted as a single unit in FIG. 1 . It should be understood that torque may be modulated at each joint by a dedicated controller and / or that each controller may perform other tasks in addition to torque modulation at a single joint. Examples of sensors included in the sensor system 32 to provide information to the controller 28 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), and electromyography (EMG) sensors. The device 10 may include other nonillustrated components as well, such as an on-board power source to power the actuators 24, 26 and / or controller 28, various housings, cables, bracing portions, and fastening devices to make the device 10 wearable by the user.

[0044] The device 10 disclosed in the experimental methods and results below is an exoskeleton (i.e., a powered orthosis) configured to assist a user with movement of one or more portions of a biological limb about one or more biological joints. In other embodiments, the device may be a powered prosthesis configured to replace one or more biological joints and at least a portion of a biological limb, or a combined exoskeleton-prosthesis configured to assist with movement of a limb portion about a biological joint and to replace another limb portion and / or joint. Each joint 20, 22 of the device thus corresponds to a human joint and is either an assistive joint or a replacement joint.

[0045] In the case of an exoskeleton, each frame member 14-18 is configured for removable attachment to respective and corresponding portions of a user’s limb on opposite sides of each joint. The illustrated example is a lower-limb exoskeleton 10, where a first frame member 14 is an upper leg member configured for attachment to the upper leg of the user, a second frame member 16 is a lower leg member configured for attachment to the lower leg of the user, and a third frame member 18 is a foot member configured for attachment to the foot of the user. Each frame member 14-18 is generally at least as rigid as the portion of the limb to which it is attached and is attached to the respective limb portion in a manner that causes the user’ s limb portion and the frame member to move together — i.e., relative motion between the respective frame member and associated limb portion of the user is minimized by the manner of attachment. For example, the upper and lower leg members 22, 24 may be removably attached to the respective upper and lower leg of the user via a molded brace with multiple adjustable and cinchable straps, and the foot member 18 may be configured to fit within a shoe such that the shoe effectively attaches the foot member to the user’ s foot when the shoe is worn by the user.

[0046] In other examples, the device includes only two frame members and one actuator providing modulated torque at a single joint. For example, a knee-assistive exoskeleton may include a single assistive knee joint 20 interconnecting upper and lower leg members 14, 16 whose relative movement is in coordination with one actuator 24. In another example, an ankle-assistive exoskeleton may include a single assistive ankle joint 22 interconnecting a lower leg member 16 with a foot member 18 whose relative movement is in coordination with one actuator 26. The device 10 may alternatively be in the form of a hip-assistive exoskeleton having a torso member (not shown) configured for attachment to the user’s torso and coupled to the upper leg frame member 14 via an assistive hip joint corresponding to the user’s hip joint. Or the torso member and additional joint can be added to the illustrated device 10 to provide a three-jointed device assisting leg movement about the hip joint, lower leg movement about the knee joint, and foot movement about the ankle joint.

[0047] Below, a lower-body sagittal -plane model of a human-exoskeleton system is presented using a port-Hamiltonian formulation, along with interconnection and damping assignment passivity-based control (IDA-PBC) in the context of the ipsilateral leg of the model. The ipsilateral and contralateral legs of the model can be coupled together with energy shaping considered for the whole model and with an arbitrary configuration of assisted joints and available sensors. A solution to the matching conditions (considering contact constraints) is presented, and an optimization is defined that will efficiently produce the corresponding control law for each possible configuration.

[0048] Port-Controlled Hamiltonian Dynamics

[0049] With reference to FIG. 2, each leg of a human-exoskeleton system can be modeled as a 4- link sagittal plane monopod MhMcthat starts from a floating foot and has three revolute joints corresponding to an ankle joint A, a knee joint K, and a hip joint H. Tn FIG. 2, subscripts I and C indicate respective ipsilateral and contralateral monopods and joints. During the ipsilateral heel contact phase, the inertial reference frame (IRF) is coincident with the position of the ipsilateral heel where COP is the center of pressure. The global ipsilateral heel angle <p is defined with respect to a vertical axis (y). The ankle, knee, and hip angles of the ipsilateral leg are denoted , respectively, while the contralateral side follows the same convention with the additional subscript “C” The same convention is used to denote the contralateral heel position and global heel angle . The model’s masses and moments of inertia reflect the combination of the human and exoskeleton masses.

[0050] The dynamics of the ipsilateral and contralateral monopod models are linked by an interaction wrench at the hip center that acts equal and opposite on their respective torso bodies. The six degree-of-freedom (DOF) ipsilateral monopod model has the generalized coordinates in the 6-dimensional configuration space The conjugate momenta are defined by the positive-definite inertia matrix and the velocity vector The port- controlled Hamiltonian dynamics can be characterized by the Hamiltonian (the cotangent bundle of Q), through the equations where the skew-symmetric matrix above is known as the interconnection matrix. The Hamiltonian function is given by the kinetic plus potential energy The gradient is a column vector in as row vectors. The vector of joint torques aggregates the monopod’s exoskeleton input and human input Thum= with the Jacobian matrix J mapping the interaction wrench F into the monopod dynamics. The control inputs and respectively represent the exoskeleton and human torques (at the ankle, knee, and / or hip joints), which are mapped into the dynamics via matrices B , where m denotes the number of the exoskeleton actuators with The system is underactuated with the number of generalized coordinates larger than the number of control inputs — i.e.

[0051] FIGS. 3A-3C illustrate a human lower leg at three different contact configurations, including heel contact (FIG. 3A), flat foot (FIG. 3B), and toe contact (FIG. 3C) during the singlesupport period of human locomotion. To explain the term A in equation (1), it is first noted that the holonomic contact constraints in the human-exoskeleton dynamics can be expressed as as shown in FIGS. 3A-3C. Here, c is the number of constraints when the monopod is in stance, and the subscript { heel, flat, toe } indicates the contact configuration. The constraint matrix given the top row of equation (1). The possible cases are

[0052] Heel Contact

[0053] Flat Foot

[0054] Toe Contact where If is the length of the foot. The Lagrange multiplier represents the ground reaction forces (GRFs), which are mapped into the system through the constraint matrix A. Henceforth, q and p terms are omitted in matrices to simplify notation. The Lagrange multiplier A can be obtained by solving wher denotes the second-order derivative of with respect to p.

[0055] Matching Conditions of the Monopod Controller

[0056] Now assume the ipsilateral feedback loop has been closed for exoskeleton input u, while the human ipsilateral input v remains as an input to the Hamiltonian system. Then consider a desired, closed-loop Hamiltonian represents the new potential energy with shaping term V. The corresponding gravitational vector is Setting simplifies the matching process and passivity proof and avoids complicated calculations of the inertia matrix inverse in the control law. Hence, Instead of modifying the inertia matrix, velocity-dependent shaping is achieved by modifying the interconnection matrix of the closed-loop Hamiltonian system.

[0057] The desired closed-loop dynamics based on are

[0058] (2) where " selects the nudegrees of freedom that are measured but not actuated, and uxrepresents the power leak resulting from their use in the control law (relaxing the matching conditions presented below). Together, these symbols refine the definition of the exogenous input to formalize its restriction to the image space of Bx. The skew- symmetric matrix represents the extra shaping DOF provided in the interconnection structure by the IDA-PBC method, wher is a smooth vector-valued function within the artificial gyroscopic term . Moreover, the closed-loop GRFs in equation (2) are represented by

[0059] Hamiltonian systems (1) and (2) match if

[0060] By plugging in GRFs A and A and following the IDA-PBC method, we have (3) where

[0061] The corresponding matching condition is (4) where is any full-rank left annihilator of (satisfying Note that the as-of-yet unspecified relaxes this matching condition. For example, it allows for to have non-zero values corresponding to the global angles measured by the sensor system — e.g., inertial measurement units (IMUs). Because the global angles lack actuators to physically apply input ux, the target energy shape is not perfectly achieved — i.e., there is a power leak from the target energy as discussed below in Section D.

[0062] The matching condition (4) can be simplified by first decomposing matrix M into four submatrices: where corresponds to the floating base joints corresponds to the joints Then, the following is obtained: where As a result, we have can be expressed as where . Let where and plug in to obtain which has the corresponding left annihilator where is the (full-rank) left annihilator of Plugging into (4), we have

[0063] (5)

[0064] The solution (5) of the matching condition gives the feasible structure of the closed-loop system. Similarly, the six degree-of-freedom (DOF) contralateral monopod has the generalized coordinates

[0065] The port-controlled Hamiltonian dynamics can be characterized by the Hamiltonian through the equations where aggregates the exoskeleton input and the human input with the Jacobian matrix Jc. The constraint matrix is p dynamics based on are with the exogenous input

[0066] Matching Conditions of the Bipedal Controller

[0067] Considering both monopod models together, we have a combined bipedal Hamiltonian , with combined generalized position vector and conjugate momenta

[0068] The open loop dynamics is where with the constraint . The GRFs are given by The closed-loop combined dynamics based on with (6) where and The skew- symmetric interconnection structure is now

[0069] The corresponding matching conditions are given by

[0070] Plugging GRFs into the matching conditions, we have

[0071] 77) where

[0072] The corresponding matching condition to (7) in the bipedal model becomes

[0073] (8) where (full-rank) left annihilator of .

[0074] Following the previous matrix decomposition and simplification, we have

[0075] Plugging into (8), we have

[0076] which shows the feasible structure of the closed-loop system using underactuated energy shaping control.

[0077] Note that the energy shaping framework can be applied to the bipedal model with an arbitrary configuration of assisted joints. For simplicity, considering an example of bilateral knee exoskeletons a symmetric single-joint configuration), we have

[0078] By zeroing the unactuated rows of i e., those associated with — the matching condition (8) is satisfied with the flexibility to design the skew-symmetric matrix

[0079] Control Law with Relaxed Passivity

[0080] Energetic passivity is defined as follows. Consider a general mechanical system (10) where is the input and is the output. Let be a continuously differentiable, positive semi-definite function, then the system (10) is passive from input u to output y if

[0081] By strict use of energy shaping control, the target Hamiltonian will satisfy or, in other words, serve as a passivity certificate for the closed loop system with respect to the remaining human input (ensuring the human controls energy injection). However, for various practical reasons, relaxations can be introduced to this property, which are described together through a “power leak,” (1 1) where represents the ‘leak’ torque vector. The three contributions to this leak are 1) use of global angle information without associated actuators (i.e., relaxing matching conditions using the exogenous input ) use of vertical ground reaction force to scale torques, and 3) nonlinear negative power tapering which lets the human, rather than the exoskeleton, store mechanical energy in behaviors like initiating a squat.

[0082] Normally, the use of the global information from IMUs to define a change in the Hamiltonian results in ( / -partial derivatives, or torques, that cannot be produced by the underactuated input. This means the target behavior cannot be achieved by the underactuated input. However, by introducing an exogenous input that directly cancels these unachievable torques, we can instead say that the target system has been achieved with a power leak due to the fictitious exogenous input. Since the target system is not perfectly achieved, the underactuated control law can generate non-zero net work where the power leak accounts for the opposite of the non-zero net work.

[0083] Vertical ground reaction force (vGRF) scaling prevents excessive torque as weight transfers from the assisted leg to the contralateral leg during double support. This scaling (J (vGRF, f ) is defined via the sigmoid functions, where with as scalar constants and the linear combination of

[0084] The incorporation of vGRF scaling may be considered an essential aspect of the control scheme. Depending on the defined basis in Section E, below, the vGRF scaling can take the form of either or and is utilized to ensure a smooth transition between the stance and swing phases. This approach also enables employment of a single controller for both the stance and swing phases.

[0085] The negative power tapering strategy is included for the comfort of the user. The controller can perform negative work (which can then be released as positive work), and such behavior is key to energetic passivity. However, based on feedback from some users during pilot testing, negative work assistance was undesirable during certain phases of stair descent, ramp descent, and stand-to-sit transitions. Therefore, a negative power tapering strategy may be applied, where the ultimate torques provided by the controller are scaled down by negative power (pointwise operator) as (12) where 6^ is the zth joint velocity and is the tapering coefficient.

[0086] Altogether, the control law for the feasible shaping structure satisfying (7)-(9) becomes (13) with being the left pseudoinverse of Note that velocity dependence is introduced via the conjugate momenta Here, the exogenous input is extended to which now includes the composite “power leak” associated with the combination of unactuated global variables, vGRF, scaling, and negative power tapering. As a result, matching condition (8) can still be satisfied and the target dynamic equation can be used to describe the system while incorporating the (unactuated) global variable into the actuated part of and Given the similar structure of (6), relaxed input-output passivity can be proved accordingly. Moreover, if the human is assumed to modulate joint impedance and provide the exogenous input stability during small movements can be shown in the sense of Lyapunov.

[0087] Constructing a Modular Basis for the Controller

[0088] To guarantee satisfaction of the matching conditions, the controller (the target energy and target interconnection matrix) is simply parameterized using a functional basis where each element satisfies the conditions. This section demonstrates how such a basis can be built using simple primitives. For example, the sine-cosine primitive is used to express behavior resembling the gravitational potential energy of a pendulum. But to more fully express the potential for a single joint to modify potential energy, this is expanded. The following single degree of freedom Hamiltonian modification primitive describes potential energy alterations that only affect one active joint, qx, with one unactuated angle reference a global angle measured by an IMU — available:

[0089] In our notation, to use this basis for the ipsilateral knee we would substitute . Choosing the target Hamiltonian as would then result in an active joint torque and a power leak equal to

[0090] Adding a second active joint introduces the potential for not only two copies of the first basis, but also additional coupling potential energy terms. For example, with one unactuated angle reference we can define the basis and with two unactuated angle references,

[0091] This second unactuated angle reference may be from an IMU on the other leg.

[0092] Multiple joints also introduce the possibility of a non-trivial interconnection matrix, . There is great freedom in parameterizing the upper triangular elements of this (skew-symmetric) matrix. Unlike the modifications to potential energy, the basis for does not need a power leak through to make use of unactuated measurements (like global link angles). Any skew- symmetric matrix will similarly conserve energy. However, for simplicity, may be parameterized to depend on the same measurements as the two-joint Hamiltonian basis above, Here, the single degree of freedom helper basis is and the helper basis for combinations of two joints (and one inertial reference) is

[0093] These primitives are then combined to create bases for the control law. For example, the law we apply for a bilateral single-joint exoskeleton configuration with measurements and vGRFxdenoting actuated angle, actuated joint angular rate, unactuated angle (global angle reference), and vertical component of the ground reaction force for each side x a, b, the basis for the control torque is defined as (14) where the decoupled terms are and the coupled terms are

[0094] For the single-joint unilateral case, only would be needed to construct the basis. While the inclusion of global angle references into the design complicates the expression of a simple pattern for scaling this basis to larger configurations of joints, the essential technique is to include terms like for each joint, and to leave the interconnection terms free of the influence of the vGRF. The restriction of the basis to those degrees of freedom which are either actuated or measured but unactuated terms) is enough to guarantee the satisfaction of the relaxed matching conditions. Design Optimization

[0095] The optimization problem is designed to calculate the parameters more efficiently and enforces disciplined convex programming rules. The expression is designed as a linear combination of the basis functions with the constant coefficients and vGRF scaling where w basis functions follow the structure of (9). The control law (13) is thus given as where and

[0096] An assistive torque profile proportional to the average biological torque may not be the optimal assistance torque for human subjects. Moreover, based on feedback from users during pilot testing, biomimetic knee extension during late stance resists the users lifting their legs. Instead of fitting the target joint torques to normalized able-bodied joint torques, the constant coefficients a may be optimized so the outputs of control law best fit a weighted combination of the normalized able-bodied joint torques gravity-shaping joint torques Yg, and zero (passive) joint torques where are diagonal weighting matrices for different phases. The optimization problem may be defined as where the subscript j represents the number of different locomotor tasks, including level-ground walking, ramp walking, stair climbing, and stand-to-sit. The state vectors comprise samples over time (n total) for the given task The two-norm of a vector with weighting matrix is denoted as

[0097] The objective function comprises three parts, where scalar corresponds to the least squares error of the exoskeleton control inputs and the target joint torques with the weighting diagonal matrix of different tasks. Scalar represents “LI regularization” to enforce sparsity in the model by zeroing the least important parameters in vector a, which avoids over-fitting and improves the prediction of untrained tasks. The third par corresponds to the cost of opposite signs between the exoskeleton control torques and the target torques to emphasize the importance of assisting rather than resisting human torques. Vector represents the slack for the sign difference, and denotes the pointwise product with The terms A, weight the different costs.

[0098] “CVX” in MATLAB can be used to find the optimal solution a* , where the kinematic and kinetic data from nine subjects over level-ground, ramps, stairs walking, and stand-to-sit from open-source able-bodied datasets are applied. The vGRFs during locomotion tasks are normalized by body weight. The training gaits include level treadmill walking at 0.5, 1.5m / s, ascending / descending ramps with inclines of 5.2°, 11°, and ascending / descending stairs with step heights of 4 and 7 inches. The corresponding controller provides assistance torques where LOA% (level-of-assistance) scales down the controller to a desired fraction of normative torque. The optimal parameters a* are ultimately used in the real-time implementation presented next.

[0099] Experimental Example

[0100] The control scheme was implemented on a controller of the modular M-BLUE exoskeleton as illustrated in FIGS. 4A-7B. Details of the M-BLUE device can be found in C. Nesler et al., “Enhancing voluntary motion with modular, backdrivable, powered hip and knee orthoses,” IEEE Robot. Autom. Lett., vol. 7, no. 3, pp. 6155-6162, 2022. FIGS. 4A and 5A are respective photographic side and front views of a user wearing a knee-assist version of the device 10, with FIGS. 4B and 5B being schematic representations of those views. FIGS. 6A and 7A are respective photographic side and front views of a user wearing a hip-assist version of the device 10, with FIGS. 6B and 7B being schematic representations of those views. The components of each device 10 are annotated with reference numerals corresponding to those of FIG. 1.

[0101] The knee-assist device 10 of FIGS. 4 and 5 includes an upper leg member 14 affixed to the user’s upper leg via an upper leg brace 34 and a lower leg member 16 affixed to the user’s lower leg via a lower leg brace 36. The frame members 14, 16 are connected at a two-sided (medial and lateral) pivot joint 20 coaxial with the actuator 24. The lower leg member 16 in this example extends down to the user’s foot and is affixed at the shoe to provide support for the end of the user’s lower limb. Describe differently, the device 10 includes a foot member to support the user’s foot, but it is rigid and moves with the lower frame member 16 — i.e., there is no pivot joint at the ankle. The device also include a pair of contact sensors 32a, with one mounted at each of the user’s heels. Each shoe, including the shoe of the contralateral limb, includes an insole force sensor 32b. The controller 28 and a power source 38 are mounted to the upper leg member 14. The unilateral knee-assist device 10 of FIGS. 4 and 5 weighs 2.36 kg, including the power source 38.

[0102] The hip-assist device 10 of FIGS. 6 and 7 includes an upper leg member 14 affixed to the user’s upper leg via an upper leg brace 34 and a torso member 40 affixed to the user’s torso via a torso brace 42. The frame members 14, 40 are connected at a lateral pivot joint 44 coaxial with the actuator 46. The device also include a pair of contact sensors 32a, with one mounted at each of the user’s heels. Each shoe includes an insole force sensor 32b. The controller 28 is mounted to the upper leg member 14, and the power source 38 is mounted to the torso member 40. The unilateral hip-assist device 10 of FIGS. 6 and 7 weighs 2.36 kg, including the power source 38.

[0103] The devices 10 of FIGS. 4-7 combine commercially available orthoses with a quasi-direct drive actuator. In this example, each actuator 24, 46 is a T-motor® AK80-9 which includes a high- torque electric motor with an internal 9: 1 plenary gearbox. The actuator is easily backdrivable with less than 0.5 N-m static backdrive torque and can provide 9 N-m continuous torque and 18 N-m peak torque, according to the manufacturer. In practice, up to 30 N-m peak torque was measured in a bench-top calibration. This modular exoskeleton facilitates bilateral / unilateral knee and / or hip configurations to match different use cases.

[0104] The high-level control loop ran at ~200 Hz on an 8 GB RAM Raspberry Pi® 4B as the controller 28 for each leg. Bilateral configurations communicated with each other through ZeroMQ, a TCP-based package. Each device 10 was powered by a 24V, 2 A-h Kobalt power tool battery (~470 g) attached to a 3D-printed adapter mounted on the lateral side of each orthosis. Sagittal-plane joint angles and global segment (i.e., frame member) angles were measured by two 6-axis Microstrain® IMUs attached to the brace straps around each limb portion. Joint angle measurements were obtained by taking the difference between the global angles of adjacent limb portions (e.g., upper leg and lower leg), which bypassed compliance between the actuator and limb segment to give more accurate joint measurements than provided by the encoders of the actuators. Soft-tissue and strap compliance caused vibrations that introduce error into the angle measurements using the joint encoders.

[0105] Accordingly, embodiments of the limb assistive device include at least two sensors (e.g., IMUs), each configured to measure a global angle of different limb portions of the user. The controller determines the angle between the different limb portions as a difference of these measurements. The two limb portions may be portions of the same limb (e.g., right or left leg) on opposite sides of a joint of the user. For purposes of this disclosure, the user’s torso is also considered a limb portion joined with the user’s upper leg at the hip joint. In one implementation of a knee-assist device, one sensor affixed along the user’s upper leg (i.e., not directly to the corresponding frame member of the device) measures the global angle of the user’s upper leg, another sensor affixed along the user’s lower leg (i.e., not directly to the corresponding frame member of the device) measures the global angle of the user’s lower leg, and the controller determines the angle between the upper and lower leg as the difference between the measured global angles.

[0106] The vGRF was measured using a commercially available sensor (IEE Smart Footwear) placed beneath the shoe insole (sensors 32b in FIGS. 4-7). Similar to zero-order hold, a parallel thread was created to read vGRF continuously at ~55 Hz, which gave the latest vGRF every 0.02 seconds and avoided slowing down the main control loop at 200 Hz. The vGRF sensor was calibrated using a predefined calibration procedure before each use to achieve a final readout normalized to body weight in the same manner as the vGRFs from the normative dataset used for the controller simulation. An infinite impulse response (IIR) second-order low-pass filter (50 Hz cutoff frequency) was applied to the vGRF for noise-reduction. The negative power tapering coefficient / ? in equation (12) was adjusted for user comfort during several practice trials and fixed for all subjects during data collection.

[0107] Safety features such as mechanical hard stops, thermal protectors, software program interventions, and current limiters were present at all joints. A motor current limiting policy was also implemented to prevent overheating the motor windings.

[0108] Eight able-bodied (AB) human subjects (TABLE I) were enrolled to demonstrate the ability of the orthosis and control scheme to assist multiple tasks. Muscle activation was assessed via wireless electromyography (EMG) (Delsys Inc.) of vastus medialis oblique (VMO), rectus femoris (RF), biceps femoris (BF), and gluteus maximus (GLUT), which function as a knee extensor, knee extensor / hip flexor, knee flexor, and hip extensor, respectively. Neonatal sensors were used for VMO, RF, and BF.

[0109] TABLE I

[0110] Participants performed the same activities of daily living (ADLs) with five exoskeleton conditions: bare (no exoskeleton), active bilateral hip exoskeleton (HipB), active unilateral hip exoskeleton (HipU), active bilateral knee exoskeleton (KneeB), and active unilateral knee exoskeleton (KneeU). The LOA% for the active modes was set based on each subject’s comfort level during practice trials and fixed for the entire experiment. Each trial was performed in two parts Pl, P2 of an ADL circuit 100 at a self-selected speed, as illustrated in FIG. 8. The ADL circuit included a stand-sit station (A), a 12° ramp (B), a level platform (C), stairs (D) with 6-inch risers, and a level walkway (E). Part Pl of each trial included five tasks, starting with a stand-sit cycle (SS), followed sequentially by incline walking (II) along the ramp (B), level walking on the platform (C), stair descent (SD), and level walking (LL). Part P2 of each trial included the same five tasks in reverse order, starting with a stand-sit cycle (SS), followed sequentially by level walking (LL), stair ascent (SA), level walking on the platform (C), and decline walking (DD) along the ramp (B). More concisely, and excluding the transitional platform (C) walking, Pl =

[0111] All incline, decline, ascent, and descent tasks started with the right foot contacting the ramp and stairs first to obtain the maximum number of strides for the right leg. Starting with the right leg was not required for the level walking task. Five trials were performed for each exoskeleton condition, which provided a minimum of 20 gait cycles of level walking, 10 gait cycles per stair task, 10 gait cycles per ramp task, and 10 stand-sit cycles. At least five minutes of acclimation time was provided for each exoskeleton condition, and a five minute break was provided between the trials of exoskeleton conditions. Subjects were instructed not to use the handrails (omitted in FIG.8) except to prevent a fall. The walking trials were separated into different tasks using a stopwatch and VICON video, and cropped into gait cycles by detecting heelstrike with a heel-mounted accelerometer (sensors 32a in FIGS. 4-7). Stand-sit cycles were cropped into individual repetitions using a thigh-mounted accelerometer built into the EMG sensor. Each muscle’s EMG was demeaned, bandpass filtered (20-200 Hz), smoothed with a moving 100 ms window RMS, and then normalized with respect to the maximum peak of the ensemble averages (across repetitions) of all the active modes. This was done for each task and muscle separately, resulting in the signals being converted to a percentage of the maximum voluntary contraction level (%MVC) for consistent comparison across subjects.

[0112] The subject-wise muscular efforts analysis involved a linear mixed model (LMM) in MATLAB with restricted maximum likelihood estimation of parameters. Data from the eight subjects were tabulated with information including muscle effort change, exoskeleton condition (Bare, HipB, HipU, KneeB, KneeU), weight, LOA, and gender. Muscular effort (%MVC.s) were quantified by integrating normalized EMG over time from the beginning to the end of five repeat trials for each exoskeleton condition. The difference between the %MVC.s of active conditions and the %MVC.s of the bare condition was used to determine the effort change. Exoskeleton conditions were defined as categorical variables and a LMM was fit, where the condition, weight, LOA, and gender are fixed effects:

[0113] Effort Change ~ Controller + Weight + Gender + (l|Muscle)+ (1 |Task)+ ( 11 Subj ect), where ( 11 ) represents random effects. Statistical significance of each fixed effect parameter was determined by a two-tailed t-test. As a secondary analysis, a LMM without the random effects of task and muscle,

[0114] Effort Change ~ Controller + Weight + Gender + (1 (Subject), was applied to each muscle and task separately.

[0115] Though it was not the goal to strictly reproduce biological human torque profiles, the similarity between applied torque and this biomechanical reference was analyzed. The cosine similarity metric (SIM) was defined as

[0116] This similarity was then measured for each combination of exoskeleton condition and task, comparing the average human torque at the actuated joint (from the dataset) to the average applied torque (from the experiment). This metric was included to compare the behavior of the applied task-invariant controller to a well-studied task-varying reference. The average net work done per joint was also calculated for each task and condition to investigate the generation of non-zero net work due to the power leak and passivity relaxation of the control framework.

[0117] To quantify controller synchronization to the user, the Pearson correlation coefficient between the EMG results and the applied exoskeleton joint torques was additionally calculated. The formula of the coefficient for two vectors X, Y is where are the mean values an yare the standard deviations. The use of correlation in place of similarity helps account for the unknown offset between EMG and torque due to cocontraction.

[0118] Finally, the kinematics of actuated j oints between unilateral and bilateral conditions were compared to identify any systematic behavioral changes between these configurations. Overlaid phase plots of the actuated angles we specifically examined against the global thigh angle for each task.

[0119] The experimental outcomes of the study are presented below. The bilateral knee exoskeleton conditions of participants AB02 and AB03 were excluded due to a failure in the synchronization between the left and right controllers. Additionally, EMG data for the BF muscle of participant AB04 was excluded because of a sensor failure, which was detected after completion of the experiment.

[0120] FIG. 9 includes across-subject comparisons of total muscle effort during five repetitions. Muscle effort is compared between the bare (no exoskeleton) condition and the four different exoskeleton conditions for combined muscles. A positive value represents the total muscle effort increment with respect to the bare mode. represents statistical differenc represents represents Both of the unilateral configurations (KneeU and HipU) significantly reduced the muscular effort required to complete the ADL circuit, with statistically significant fixed effects in the primary LMM. The HipU configuration reduced effort by an average of 2.71% MVC.s, 95% CI [1.16, 4.27], and the KneeU configuration reduced effort by 3.40% MVC.s, 95% CI [1.85, 4.95], There was also a significant gender effect equivalent to a penalty of 3.00% MVC.s, 95% CI [0.46, 5.55] for women in all exoskeleton-bare comparisons (p = 0.021). This penalty is comparable in magnitude to the benefits from the unilateral controllers. Thus, there was only a net benefit for male subjects. A correlation was also observed between weight and muscular effort, leading to an average reduction of 0.02-mass %MVC.s, 95% CI [-0.02, 0.06], However, this correlation was not statistically significant, indicating that gender and weight were distinct effects.

[0121] The two bilateral configurations (KneeB and HipB) had a statistically null effect on muscular effort (p > 0.05). Considering the gender effect, this amounts to a net penalty for women. No significant effect was found for subject mass (p > 0.05). The kinematics indicated a slight reduction in global thigh angle range of motion for bilateral configurations as compared to unilateral configurations.

[0122] Correlation between extensor muscle activation and controller torque was high for the set of non-descent tasks — that is, excluding stair descent and decline walking. FIG. 10 includes crosssubject comparisons of the correlation coefficient between the EMG muscle activation and the applied joint torques for each task {SA, SD, DD, LL, II, SS}, with normative able-bodied (AB) human joint torques calculated and included for illustration. In the top chart of FIG. 10, each condition is presented in the same left-to-right order for each task: able-bodied (A), unilateral knee (U), and bilateral knee (B). In the bottom chart of FIG. 10, each condition is presented in the same left-to-right order for each task: able-bodied (A), unilateral hip (U), and bilateral hip (B). As a baseline, normative biological profiles for hip and knee extension torque showed a modest to high correlation with the “bare” exoskeleton condition GLUT and VMO signals (FIG. 10, AB). The correlation between controller torque and EMG measurements in the four exoskeleton conditions (HipB, HipU, KneeB, KneeU) was task-dependent, but in several cases comparable. For the knee torque and VMO, the SA, LL, II, and SS tasks were comparable to the baseline AB correlation. The correlation was lower for the SD and DD tasks, which require negative work. Similarly, for the hip torque and GLUT, the SA, LL, II, and SS tasks were of similar correlation to the baseline AB correlation, while the correlations in the SD and DD tasks were reduced, with an inverse correlation for the SD case.

[0123] TABLE II includes the across-subject muscle activation results for each task, configuration, and muscle and offer a very detailed analysis of the effect of the controller. Muscle effort comparisons were made between the bare (no exoskeleton) condition and the various exoskeleton conditions. A positive value represents each muscle pair’s effort increment (%MVC) with respect to the bare condition. * represents statistical difference (p < 0.05), ** represents p < 0.01, and *** represents p < 0.001. Standard deviations are omitted for brevity.

[0124] TABLE II ACROSS-SUB JECT COMPARISONS OF MUSCLE EFFORT CHANGE (MEAN %MVC. S)

[0125] These results can be interpreted through the measured control torques and ensemble- averaged VMO, RF, BF, and GLUT EMGs for subject AB 01 in FIGS. 11 and 12. Moreover, the Cosine Similarity (SIM) analysis in TABLE III shows agreement between the exoskeleton torques and average human torques from the datasets (standard deviations omitted for brevity). Agreement was strongest for incline tasks (SA and II) and weakest for decline tasks (SD and DD), likely due to negative power tapering.

[0126] TABLE III

[0127] SUBJECT 1 COSINE SIMILARITY ANALYSIS (MEAN %)

[0128] Finally, the average net work analysis in TABLE IV shows the knee exoskeleton controllers provided net positive work for ascent tasks and net negative work for descent tasks, whereas the hip configurations only provided net positive work (standard deviations omitted for brevity). Torques are normalized by LOA% and body weight. Note that AB hip work is also positive during descent tasks.

[0129] TABLE IV

[0130] ACROSS- SUBJECT AVERAGE NET WORK PER JOINT (J / KG).

[0131] Incline walking and stair ascent are primarily associated with positive power via concentric muscle contractions in the quadriceps. During II and SA tasks, the exoskeleton was able to facilitate clear reductions in the quadriceps (RF and VMO) EMG activation during stance phase (i.e., the first 50% of each cycle in FIG. 11). However, these quadriceps effort reductions in stair and ramp ascent tasks were highly variable across all subjects and device conditions, with bilateral conditions failing to achieve statistically significant reductions in particular (see TABLE II). Observations suggest that the subjects who were less familiar with the system struggled to anticipate the assistance of the devices and that the bilateral conditions interfered with natural motion more than the unilateral conditions. All conditions provided either knee or hip extension torques in the stance phase of these tasks, as expected. Unilateral hip conditions (providing hip extension torques) were also capable of reducing GLUT EMG compared to the bare condition (see FIG. 11). However, this effect was statistically insignificant across the participants (see TABLE II).

[0132] Stair descent and decline walking tasks are primarily associated with negative power and involve eccentric quadriceps contractions. Commonly, a double peak quadriceps activation profile occurs in stance — firstly to absorb the impact of heel strike, and secondly to lower the COM. However, due to the negative power tapering strategy, all active conditions showed a minor effect on EMG reductions compared to the bare condition. Both knee and hip active conditions provided minor knee and hip extension torques during early stance to absorb the impact. Sit-to-stand and stand-to-sit tasks primarily require knee extension torques, specifically concentric contractions during sit-to-stand and eccentric contractions during stand-to-sit. All active knee conditions (KneeB and KneeU) provided substantial knee extension torques, resulting in a noticeable reduction in quadriceps activation (VMO and RF) (see SS in FIG. 11). This effect appeared in the aggregate results across all subjects by the KneeU condition, with VMO and RF reduction (p < 0.01), while the effect was not replicated across the participants for the KneeB condition (see TABLE II). The hip-assist devices also were capable of reducing quadriceps activation during sit-to-stand (see FIG. 11). However, this effect was not detected by the secondary LMM analysis (see SS in TABLE II).

[0133] Knee flexors like the BF are responsible for lifting the foot in swing. The knee-assist devices could be expected to assist with this flexion torque in level walking. However their reproduction of human-like flexion torques in this task was weak (see negative knee torque behavior for in task LL, FIG. 12). Across subjects, the controller was only able to reduce swingphase BF activation using knee-assist devices during the ramp incline task II (see TABLE II). For this task, the controller appears to create a pulse of knee flexion in late-stance, agreeing with able- bodied torques in FIG. 12. The controller was also able to reduce BF activation in the SS task with the KneeU configuration, which is interesting since the predominant torque in SS is extension with notable co-contraction (see the competing BF flexor and RF, and VMO extensor activation in late SS in FIG. 11).

[0134] The assistance torques provided by the hip modules are capable of reducing the hip extensor EMG (GLUT) for most of the tasks (see FIG. 11). This result was also not detectable in the population, with high variance in the EMG results (TABLE II). However, a GLUT EMG penalty with bilateral knee-assist devices was significant in SA, SD, II (p < 0.001), and DD, LL (p < 0.01). Interestingly, the effect was much reduced for the unilateral condition of the knee exoskeleton (KneeU), with only a penalty in LL being significant (p < 0.01). For AB01, increased GLUT activation occurs in stance phase for the KneeB and KneeU conditions (FIG. 11).

[0135] The primary analysis revealed that the task-invariant exoskeleton was demonstrably beneficial only in unilateral configurations with male users. However, there are two important caveats to this direct interpretation of the LMM’s statistical test. The first is the unintended correlation between acclimation experience and gender. Two male subjects (including AB01) were quite experienced, and also received large benefits from the device. A more experienced user can anticipate the exoskeleton’s assistance behavior and thereby exploit it, whereas an inexperienced user may frequently co-contract and incompletely relax the relevant muscles. Due to the large number of tasks and exoskeleton conditions in the study, the subjects were provided only approximately 5 minutes of acclimation time for each task. Previous research has shown that 30 minutes of acclimation time is required before observing EMG reductions with exoskeleton assistance. Moreover, becoming an expert exoskeleton user may take considerable practice, with around 109 minutes of training required for full adaptation, according to some studies. Unfortunately, the study was not designed to accommodate the same level of acclimation for all subjects and all exoskeleton conditions as earlier pilot-test subjects received. It seems highly plausible that additional acclimation would improve the significance of the muscle activation benefits and reduce / eliminate the gender penalty.

[0136] The second caveat is the unmeasured physical effect of simply wearing the bilateral configurations, which impede out-of-plane motion and add twice the mass. The large GLUT penalty in the SA task imposed by KneeB but not by KneeU could be due to the increased mass of the contralateral leg when it is swinging, or anterior torso lean in stance. There could also have been indirect penalties on muscle effort due to changes in kinematics, as the study revealed that unilateral conditions exhibited subtle kinematic differences (e.g., higher range of motion) compared to the bilateral conditions. Given the similarity of command torques between bilateral and unilateral conditions for the same joints, it is suspected these direct or indirect muscle activation penalties likely result from the form factor of the device itself. The mechanism of a penalty from the device itself is well-established in the exoskeleton literature. In fact, participants informally noted their motion was physically constrained by the bilateral devices. Similarly, the smallest size of the off-the-shelf hip braces was notably uncomfortable for the shortest participants, and this effect could have penalized women at a higher level than it did the men. This discomfort could also explain their preference for lower values of the LOA% during the tuning process. In light of these caveats, direct interpretation of the primary analysis may understate the potential benefit of the multi-task optimal energy shaping controller.

[0137] The detailed secondary analyses show the devices produced helpful torque outputs across tasks and exoskeleton conditions. Among all the active conditions, the unilateral hip module produced the greatest EMG reductions among all tasks for AB01 as shown in FIG 11, but muscle effort reduction was different for each subject over the various exoskeleton conditions. The benefits of the unilateral configurations on RF and VMO were clear in several tasks. Only AB02, AB03, and AB04 had greater reductions in muscle efforts in the bilateral cases over the unilateral cases. This discrepancy may depend on how subjects acclimate to the various configurations.

[0138] During sit-to-stand, large reductions in VMO (knee extensor) and BF (hip extensor) activations were observed with the unilateral knee exoskeleton condition when compared to the bare condition (see TABLE II for mean effort changes). The reductions in VMO and GLUT activations were also aligned with the exoskeleton assistance torques, as shown in FIG. 10. Although KneeU and KneeB provided no direct hip assistance torque, a reduction in BF activation was observed, which can be due to a change of strategy to a more knee-dominant one.

[0139] Stair and ramp climbing during stance have similar biomechanics to sit-to-stand, whereby knee extension and hip extension torques are required during early- to mid-stance to elevate the body’s center of mass. Accordingly, reductions in VMO and RF activations were observed during early- to mid-stance (see FIG. 11 for ensembled EMG averages). The correlation between dominant muscle EMG and torque profiles in FIG. 10 shows a harmony of the assistive torques with the nervous system in stair ascent and incline walking. Overall, reductions in muscle activation aligned with the respective assistance torque, showing the contribution of the device toward the net joint torque. This was made possible by relaxing passivity with power leak terms (sub-section D, above) to allow net positive or negative work over the gait cycle (TABLE IV).

[0140] The above-noted results could have been positively or negatively impacted by insufficient experimental control over kinematic variability between bare and exoskeleton conditions. Although the study design aimed to strike a balance between practicality and experimental control, such as averaging several repeat trials to reduce noise, there were unanticipated aspects of the user experience that may have affected the results. Specifically, introducing the exoskeleton and its torque may have influenced the kinematics / kinetics of subjects, potentially altering EMG signals. For example, subjects may have unconsciously changed their gait style, shifting their load from the knees to hips or vice versa to leverage (or fight) the device assistance. In subsequent pilot testing, it was found that stricter enforcement of gait style leads to a clearer exoskeleton effect on muscle effort, suggesting this effect may have been hidden by kinematic / kinetic variation between conditions in this study. The relatively low peak torque values observed during ascent tasks also may have led to weak EMG trends. In order to protect the actuator, the torque was saturated at 25 N-m, although this level was never reached during ascent tasks. For instance, FIG. 12 shows that a torque of approximately 10 N-m during ascent tasks resulted in EMG reduction trends for subject AB01. This peak torque is a consequence of the assistance fraction of the controller, which was tuned to optimize comfort for each participant during acclimation trials.

[0141] Tuning for comfort tended to sacrifice the peak assistance torque, especially for the female participants. These tests have only covered a small subset of the exoskeleton configurations that are possible using the energy shaping control framework. While the above study considered knee and hip joints separately, the combination of knee and hip joints would introduce even more predictive power into the basis function set. The framework could extend to arbitrary uni- and bilateral ankle, knee, and hip configurations of the device for complete assistance over lower limbs. It could even extend beyond this, to joints out of the sagittal plane like the frontal plane hip abduction / adduction torque.

[0142] In various embodiments, a lower limb assistive device includes 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 one or more device actuators to modulate torque at one or more device joints in accordance with the above-described control scheme.

[0143] 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.

[0144] 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, comprising: 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, wherein the device employs a task-invariant passivity-based energy-shaping control scheme to modulate torque at the pivot joint, and wherein the control scheme employs a common control law during stance phase and swing phase.

2. The assistive device of claim 1 , wherein the control law scales torque using vertical ground reaction force measurements.

3. The assistive device of claim 1, wherein the control law is based on able-bodied data fit to a model including a kinematic chain having four links interconnected by three revolute joints, the four links corresponding to a foot, a lower leg, an upper leg, and a torso, and the three revolute joints corresponding to an ankle joint, a knee joint, and a hip joint.

4. The assistive device of claim 3, wherein the kinematic chain is a first kinematic chain representing an ipsilateral leg, the model further including: a second kinematic chain representing a contralateral leg having four links interconnected by three revolute joints; and an interaction wrench linking the first and second kinematic chains at a respective revolute joint of each kinematic chain.

5. The assistive device of claim 3, wherein the control law employs negative power tapering where net-negative work is performed at the joint in the able-bodied data.

6. The assistive device of claim 5, wherein a magnitude of the negative power tapering is adjustable.

7. The assistive device of claim 3, wherein the control law is represented bywhere T is the torque at the pivot joint, is a basis matrix for the torque, qis a vector including angles corresponding to an orientation of each link of the model, p is a momentum vector corresponding to momentum with respect to each joint of the model, vGRF is the vertical component of a ground reaction force, a* is the parameter vector that solves an optimization problem applying able-bodied data to the model, Weight is of a user of the device, and LOA% is a level of assistance that scales the torque to a fraction of modeled torque.

8. The assistive device of claim 1, further comprising: a sensor system including one or more sensors; an actuator operable to apply the torque at the pivot joint; and a controller receiving an input from each of the one or more sensors and controlling the actuator based on each input and on the control scheme.

9. The assistive device of claim 8, wherein at least one of the sensors provides a measurement of an unactuated joint.

10. The assistive device of claim 8, wherein the first and second frame members are adapted for attachment to respective first and second limb portions of a user, the sensor system comprising a first sensor configured to measure a global angle of the first limb portion and a second sensor configured to measure a global angle of the second limb portion, wherein the controller determines an angle between the first and second limb portions as a difference of the measured global angles.

11. The assistive device of claim 1, wherein the articulated frame includes a third frame member interconnected with the first or second frame member at a second pivot joint and the control scheme modulates torque at both pivot joints.

12. The assistive device of claim 11, wherein the articulated frame includes a fourth frame member interconnected with the third frame member at a third pivot joint and the control scheme modulates torque at all of the pivot joints.

13. A bilateral assistive device according to claim 1, wherein at least one of the first and second frame members is configured for attachment to a first lower limb of a user, the bilateral assistive device further comprising a third frame member and a fourth frame member interconnected at a second pivot joint, wherein at least one of the third and fourth frame members is configured for attachment toa second lower limb of a user, and wherein the task-invariant passivity-based energy-shaping control scheme modulates torque at the second pivot joint.

14. The assistive device of claim 1, wherein the control scheme is modular such that the control law can be applied to a hip-assistive joint, a knee-assistive joint, or an ankle-assistive joint.

15. The assistive device of claim 1, wherein the articulated frame includes two or more frame members, including the first and second frame members, each frame member being connected to another of the frame members at a respective pivot joint, the assistive device further comprising one or more actuators corresponding in number to the number of pivot joints, each actuator being operable to apply a torque at the corresponding pivot joint; and a controller employing the task-invariant passivity-based energy-shaping control scheme to modulate torque at each pivot joint, wherein the control scheme is modular and optimizable using able-bodied data to provide the control law for an arbitrary number of pivot joints selected from: a first hip-assist joint, a second hip-assist joint, a first knee-assist joint, a second knee-assist joint, a first ankle-assist joint, and a second ankle-assist joint.

16. The assistive device of claim 1, wherein each of the first and second frame members is adapted for attachment to respective first and second limb portions of a user, the assistive device further comprising: an actuator operable to apply the torque at the pivot joint; a sensor system including one or more sensors; and a controller receiving an input from each of the one or more sensors and controlling the actuator based on each input and on the task-invariant passivity-based energy-shaping control scheme, wherein the sensor system includes a first sensor configured to measure a global angle of the first limb portion and a second sensor configured to measure a global angle of the second limb portion, and wherein the controller determines an angle between the first and second limb portions as a difference of the measured global angles.

Images