Powered prosthesis and activity classifier

Abstract

A powered prosthesis includes a high-level activity classifier to detect transitions among a plurality of distinct user activities and switch prosthetic joint control to an associated activity controller. The activity classification is reduced to four states with easily distinguishable features to control transitions among more than four activities by not separately classifying upward and downward incline walking or stand-to-sit transitions. The prosthesis has an inter-leg transition accuracy over 99% under both self-paced and rapid-paced fatiguing conditions with a 100% recovery rate due to backup logic or user-cued resets.

Description

[0001] 7935-3262-WO2 (UM 2025-002-02)

[0002] POWERED PROSTHESIS AND ACTIVITY CLASSIFIER

[0003] GOVERNMENT LICENSE RIGHTS

[0004] This invention was made with government support under HD094772 awarded by the National Institutes of Health. The government has certain rights in the invention.

[0005] TECHNICAL FIELD

[0006] This disclosure is related to limb prostheses and, in particular, to powered prostheses having at least one prosthetic joint.

[0007] BACKGROUND

[0008] Research on powered knee-ankle prostheses has demonstrated significant potential to enhance mobility and restore normative biomechanics for individuals with transfemoral (TF) amputation. This advancement is particularly beneficial during high-demand activities such as stair ascent, incline walking, and sit-to-stand transitions by providing net-positive work via motorized knee and ankle joints. Most research to date has focused on control of steady-state activities, such as walking at a constant speed along an unchanging degree of incline. But transitions among distinct activities are also critical to the continuity, biomimicry, and usability of powered prostheses in real-world environments. This is especially true since different user activities, such as walking, sitting or standing, and stair locomotion can require drastically different joint motions.

[0009] State-of-the-art intent or activity classification methods generally include heuristic rulebased methods and machine learning-based methods. Heuristic methods, while intuitive and predictable, often suffer from classification delays and require domain-specific knowledge for feature selection. While machine learning methods are capable of processing high-dimensional feature spaces to automatically detect subtle differences during activity transitions, they are hindered by their non-intuitive nature and the curse of dimensionality. Machine-learning methods have claimed high offline classification accuracy up to 95% but have seen that accuracy drop to less than 80% when applied to real-time prosthesis control, and that was only with level and ramp walking activities. Classification accuracy can be further compromised by daily conditions that do not reflect the self-paced, non-fatigued conditions of the datasets typically used to train the classifier, as joint biomechanics change with user fatigue. Indeed, declines up to 10% in classification accuracy have been observed after only ten test trials, suggesting that existing classification systems may struggle under more demanding conditions. Both heuristic and machine learning approaches struggle with scalability in the number of activity classes (e.g., level walking, incline and decline walking, stair ascent and descent, sitting, standing, etc.) and depend on extensive able-bodied joint datasets. 7935-3262-WO2 (UM 2025-002-02)

[0010] While some classification methods have demonstrated accurate offline classification, real-time classification remains challenging in practical scenarios involving amputee users fitted with a prosthetic leg attempting to implement the classification strategies. Some attempts at real-time activity classification using on-board mechanical sensors have generally been limited to simple differentiation between walking and standing. Others have used electromyography (EMG) signals to classify user activities, but such methods require custom sockets to fit the EMG electrodes and must be retrained with each use. Still others have classified user activities but were only able to detect a transition led with the prosthetic leg — i.e., such classification is ineffective if the user’s first stair step or first step off of a ramp, for example, is with their intact leg. Some classification methods have tried avoid this limitation by instrumenting both user legs, but such an approach is not suitable for controlling a self- contained prosthetic leg.

[0011] In the field of humanoid and legged robots, a LiDAR-based has been proposed to distinguish among level walking, incline / decline walking, and stairs. However, camera and LiDAR systems are limited by their relatively large computational loads, sensitivity to environmental light conditions, and interference from clothing if mounted above the knee of a human user.

[0012] SUMMARY

[0013] Embodiments of a powered prosthesis are configured to detect a transition from a present user activity to a different user activity and, in response, switch prosthetic joint control from a first controller to a different second controller.

[0014] Embodiments of the powered prosthesis may include one or more of the following features in any technically feasible combination:

[0015] - an activity classifier that receives information from a sensor system and detects the transition based on the received information;

[0016] - the prosthesis is configured to detect the transition based on primary criteria or based on secondary criteria such that, if the prosthesis does not detect the transition based on the primary criteria, the prosthesis has a second chance to detect the transition and switch prosthetic joint control from the first controller to the second controller;

[0017] - primary criteria and / or secondary criteria upon which detection of the transition is based are dependent upon whether a transition stride is completed by the prosthesis or an opposite leg of the user;

[0018] - the prosthesis is configured to switch prosthetic joint control to a default controller in response to a user input; 7935-3262-WO2 (UM 2025-002-02)

[0019] - a sensor that detects or measures one or more of: thigh angle, ground contact, heel strike, toe off, distance to an obstacle, ground incline, wherein the prosthesis is configured to detect the transition based at least in part on information from the sensor;

[0020] - a distance sensor configured to measure a distance from the prosthesis to a nearest object in front of the prosthesis, wherein the prosthesis is configured to use the measured distance to detect the transition when one of the user activities is ascending stairs;

[0021] - a ground incline sensor configured to measure a ground incline angle during prosthesis stance phase, wherein the prosthesis does not consider the transition to be to or from stairs when the ground incline angle is greater than a threshold value;

[0022] - each controller employs impedance control during prosthesis stance phase and kinematic control during prosthesis swing phase;

[0023] - a knee joint and an ankle joint, the prosthesis being configured to switch joint control of both joints to the second controller in response to detecting the transition;

[0024] - each of the first controller and the second controller is a walking controller, a stairascent controller, a stair-descent controller, or a sit-stand controller;

[0025] - each of the user activities is one of the following group of user activities: walking, ascending stairs, descending stairs, or standing, the powered prosthesis further comprising an activity classifier that detects the transition from one of the group of user activities to another of the group of user activities, wherein the first controller corresponds to the present user activity and the second controller corresponds to the second controller;

[0026] - the prosthesis is configured to switch prosthetic joint control from a walking controller to a stair-ascent controller when where dst tis distance between the prosthesis and a stair during prosthesis stance phase, dst t-xis distance between the prosthesis and the first stair during a previous prosthesis stance phase, d is a first threshold value, and d2is a second threshold value that is less than dx;

[0027] - the prosthesis is configured to switch prosthetic joint control from a walking controller to a stair-ascent controller when 0t^HFis greater than a threshold value during prosthesis swing phase, where 0t^HFis thigh angle at maximum hip flexion; 7935-3262-WO2 (UM 2025-002-02)

[0028] - the prosthesis is configured to switch prosthetic joint control from a stair-ascent controller to a walking controller when dswor dstis greater than a threshold value, where dswis distance to a nearest object in front of the prosthesis during prosthesis swing phase and dstis distance to a nearest object in front of the prosthesis during prosthesis stance phase;

[0029] - the prosthesis is configured to switch prosthetic joint control from a stair-ascent controller to a walking controller when 0t^HFis less than a threshold value during prosthesis swing phase, where 0t^HFis thigh angle at maximum hip flexion;

[0030] - the prosthesis is configured to switch prosthetic joint control from a walking controller to a stair-descent controller when, at prosthesis heel strike, 02, where 0sh is a prosthesis shank angle, 0t^HFis thigh angle at maximum hip flexion during prosthesis swing phase prior to heel strike, 0i is a first threshold angle, and 02is a second threshold angle;

[0031] - the prosthesis is configured to switch prosthetic joint control from a walking controller to a stair-descent controller when, at prosthesis heel strike,

[0032] 0sh< 03, 04, where 0thSis thigh angle at heel strike, 3 is a third threshold angle that is greater than 015and 04is a fourth threshold angle;

[0033] - the prosthesis is configured to switch prosthetic joint control from a walking controller to a stair-descent controller when, during prosthesis stance phase,

[0034] 0i < 0th < 02 and 0sh< 03, where 0th is thigh angle, 0sh is prosthesis shank angle, 0i is a first threshold angle, 02is a second threshold angle, and 03is a third threshold angle;

[0035] - the prosthesis is configured to switch prosthetic joint control from a walking controller to a sit-stand controller before switching prosthetic joint control to a stair-descent controller;

[0036] - the prosthesis is configured to switch prosthetic joint control from s walking controller to s stair-descent controller when, at prosthesis heel strike, 05, where 04is a fourth threshold angle, and 05is a fifth threshold angle; 7935-3262-WO2 (UM 2025-002-02)

[0037] - the prosthesis is configured to switch prosthetic joint control from a stair-descent controller to a walking controller when, at prosthesis heel strike, e'T > 0, and (0^ - O < 02, is thigh angle at heel strike, 0t^HFis thigh angle at maximum hip flexion during prosthesis swing phase prior to heel strike, 0i is a first threshold angle, and 02is a second threshold angle;

[0038] - the prosthesis is configured to switch prosthetic joint control from the stair-descent controller to the walking controller when, at prosthesis heel strike, where Pxis horizontal distance between a user hip joint and a foot member of the prosthesis, Px,t-i is horizontal distance between the user hip joint and the foot member at a previous prosthesis heel strike, 0shis prosthesis shank angle, d is a first threshold value, d2is a second threshold value that is less than d , and 03is a third threshold angle;

[0039] - the prosthesis is configured to switch prosthetic joint control from a stair-descent controller to a walking controller when, at prosthesis heel strike, where Pxis horizontal distance between a user hip joint and a foot member of the prosthesis, Px,t-i is horizontal distance between the user hip joint and the foot member at a previous prosthesis heel strike, 0shis prosthesis shank angle, di is a first threshold value, d2is a second threshold value that is less than d , and ! is a threshold angle;

[0040] - the prosthesis is configured to switch prosthetic joint control from a walking controller to a sit-stand controller when, during prosthesis stance phase, for a prescribed amount of time, where 0th is thigh angular velocity, 0th is thigh angle, 0shis prosthesis shank angle, 0Xis a threshold angular velocity, 0! is a first threshold angle, and 02is a second threshold angle;

[0041] - the prosthesis is configured to switch prosthetic joint control from a sit-stand controller to a walking controller when a prosthesis heel strike is detected and

[0042] 0i < 0th < 02 and 0th < 0! 7935-3262-WO2 (UM 2025-002-02) where 0this thigh angle, 0this thigh angular velocity, 0j is a first threshold angle, 02is a second threshold angle, and is a threshold angular velocity;

[0043] - the prosthesis is configured to switch prosthetic joint control from the sit-stand controller to the walking controller when prosthesis ground contact is detected and

[0044] 0th< 03and 0th< 0Xwhere 03is a third threshold angle;

[0045] - the prosthesis is configured to switch prosthetic joint control from a sit-stand controller to a walking controller when prosthesis ground contact is detected and

[0046] 0th < 0i and 0th < 0i where 0tdis thigh angle, 0tdis thigh angular velocity, 0 j is a threshold angle, and 0Xis a threshold angular velocity; and / or

[0047] - the prosthesis is configured to switch prosthetic joint control between a stair-ascent controller and a stair-descent controller when A0headis greater than a threshold value, where A0headis a heading direction of the prosthesis.

[0048] BRIED DESCRIPTION OF DRAWINGS

[0049] Preferred exemplary embodiments of the invention will hereinafter be described in conjunction with the appended drawings, wherein like designations denote like elements.

[0050] FIG. 1 schematically illustrates an example of a powered lower limb assistive device.

[0051] FIG. 2 schematically illustrates examples of eight transitions between stair locomotion and walking.

[0052] FIG. 3 schematically illustrates an example of an activity classifier for use with the assistive device of FIG. 1.

[0053] FIG. 4 summarizes illustrative conditions for detecting each of the transitions of FIG. 3.

[0054] FIG. 5A depicts a stair circuit used in self-paced experiments using the activity classifier.

[0055] FIG. 5B depicts a full circuit used in rapid-paced endurance experiments using the activity classifier.

[0056] FIG. 6A is a plan photographic view of an outdoor multi-terrain setting used to test the activity classifier.

[0057] FIG. 6B is a landscape photographic view the outdoor multi -terrain setting of FIG. 6A. 7935-3262-WO2 (UM 2025-002-02)

[0058] FIG. 7 illustrates confusion matrices for activity classification among multiple activity controllers.

[0059] FIG. 8 A illustrates a confusion matrix for one user’s transitions detected both with the prosthesis as the leading leg and the intact leg as the leading leg.

[0060] FIG. 8b illustrates a confusion matrix for another user’s transitions detected both with the prosthesis as the leading leg and the intact leg as the leading leg.

[0061] FIG. 9 illustrates average prosthetic joint kinematics and kinetics for two users during sit-to-stand transitions, followed by transitions to level walking, followed by stand-to-sit transitions.

[0062] FIG. 10 illustrates average prosthetic joint kinematics and kinetics for two users transitioning from level walking to ramp ascent and back to level walking.

[0063] FIG. 11 illustrates average prosthetic joint kinematics and kinetics for two users transitioning from level walking to ramp descent and back to level walking.

[0064] FIG. 12 illustrates average prosthetic joint kinematics and kinetics for two users transitioning from level walking to stair ascent and back to level walking.

[0065] FIG. 13 illustrates average prosthetic joint kinematics and kinetics for two users transitioning from level walking to stair descent and back to level walking.

[0066] FIG. 14 illustrates average prosthetic joint kinematics and kinetics two users during self-paced inter-leg stair transition experiments.

[0067] DESCRIPTION OF EMBODIMENTS

[0068] Described below are the components and features of an illustrative powered knee-ankle prosthesis equipped with a high-level, intuitive user activity classifier that automatically detects transitions from one user activity to another and, in response, switches prosthetic joint control from one mid-level controller to another, some with task adaption built-in the dedicated control scheme for the identified activity. The automatic classifier is user-correctable and facilitates timely transitions among multiple distinct user activities and corresponding prosthesis modes, in some cases detecting an upcoming change in the user activity even before the user has initiated the transition stride. In the examples described below, the activity classifier facilitates transitions among four distinct user activities, prosthesis modes, and corresponding controllers: walking, stair ascent, stair descent, and sit-stand. In the walking mode, a walking controller is employed and is capable of controlling the prosthetic joints on level ground and adapting to upward or downward inclines without separate classification of the incline walking. In the stairascent and stair-descent modes, respective stair-ascent and stair-descent controllers are employed and are capable of adjusting to stair height without separate classifications. In the 7935-3262-WO2 (UM 2025-002-02) sit-stand mode, a sit-stand controller is employed and is compatible with multiple seat heights without separate classification and controlling the prosthetic joints during sitting and standing configurations, and during stand-to-sit and sit-to-stand transitions without separate classification of those configurations and transitions. The user activity and prosthesis mode associated with the sit-stand controller may be referred to more simply as “standing” since the transitions to and from the other three activities or modes are from the standing position.

[0069] The classification space is consolidated by integrating prosthetic joint control during level walking, upward incline walking, and downward incline walking into a single mid-level walking controller and by integrating prosthetic joint control during sitting and standing motions into a single mid-level sit-stand controller, effectively eliminating three user activity classifications and their corresponding controllers — i.e., upward-incline walking, downwardincline walking, and sitting — while allowing more adaptability, such as a continuous range of incline angles and user-paced sit-stand transitions. The classifier leverages user-controllable features, including thigh and shank orientations, as well as environmental data, including distance to a nearest object from an ankle-mounted sensor and ground incline from a footmounted incline sensor. As used herein, “incline” refers to an angle with respect to horizontal and includes upward or positive inclines along with downward or negative inclines (also referred to as “declines”). Simple heuristic rules enable both prosthetic-led and intact-led transitions that switch prosthetic joint control among the distinct mid-level controllers at appropriate times, with backup logic to correct early misclassifications. Real-time implementation has been validated with user participants navigating stools, ramps, stairs, and level ground, at self-paced conditions over short distances and at rapid-paced conditions over long distances. The classifier and classifier-equipped prosthesis has achieved an inter-leg classification accuracy of over 99%, with nearly 95% accurate classification of the leading leg to achieve ideal switch timing, enabling over 100 non-stop cycles on a multi-activity circuit. The classifier and prosthesis have also demonstrated accurate activity classification in an outdoor multi-terrain setting.

[0070] FIG. 1 schematically illustrates an example of a powered lower limb assistive device 10 including an upper leg member 12, a lower leg member 14, and a foot member 16. The assistive device 10 may be a powered knee-ankle prosthesis, as illustrated in FIG. 1, or it may be a powered exoskeleton having equivalent artificial knee and / or ankle joints interconnecting equivalent structural limb members configured for removable attachment to an intact lower limb (e.g., a powered knee brace). While described and illustrated below in the context of an exemplary knee-ankle assistive device, this disclosure is applicable to any prosthesis or 7935-3262-WO2 (UM 2025-002-02) exoskeleton having at least one joint, including, for example, any combination of a hip joint, a knee joint, and / or an ankle joint. As used herein, a “powered” limb-assistive device includes any limb-assistive device in which a power source is used to actively assist or resist movement at a device joint. This includes but is not limited to devices equipped with electric motors that apply torque at a device joint and devices that use electric motors to indirectly change joint behavior, such as devices with variable-stiffness springs that partly define joint behavior. A powered limb-assistive device 10 may also include one or more passive joints or structures that passively emulate a joint, such as a knee-ankle prosthesis with a dedicated knee-joint actuator and a passive ankle joint.

[0071] In the illustrated example, one end of the upper leg member 12 is adapted for attachment to the end of the residual portion of the upper leg of an above-knee amputee user. An opposite end of the upper leg member 12 is coupled with the lower leg member 14 at a prosthetic knee joint 18. The knee joint 18 is a rotational joint that provides for rotational movement of the lower leg member 14 relative to the upper leg member 12 about a knee axis 20. The lower leg member 14 may be referred to as the “shank” and extends from the knee joint 18 to a prosthetic ankle joint 22 at which the foot member 16 is coupled with the lower leg member 14. The ankle joint 22 is also a rotational joint that provides for rotational movement of the foot member 16 relative to the lower leg member 14 about an ankle axis 24. The foot member 16 may be adapted to accommodate a shoe to provide device protection and cushioning.

[0072] The device 10 includes a first actuator 26 (e.g., a motor) configured to provide a knee torque in a rotational direction at the knee joint 18 and a second actuator 28 configured to provide an ankle torque Tain a rotational direction at the ankle joint 22. As used herein, a positive knee torque corresponds to knee flexion, and a negative knee torque corresponds to knee extension. Similarly, a positive ankle torque Tacorresponds to dorsiflexion, and a negative ankle torque Tacorresponds to plantarflexion. The first actuator 26 may be rigidly mounted along the upper leg member 12, and its rotational output may be converted to the torque applied at the knee joint 18 via a transmission member rigidly attached to the lower leg member 14. Or the first actuator 26 may have an integral transmission and be concentric with the knee joint 18 with a housing and rotor rigidly affixed to the respective upper leg member 12 and lower leg member 14, or vice versa. Similarly, the second actuator 28 may be rigidly mounted along the lower leg member 14 or at the ankle joint 22 with its rotational output converted to the torque Tavia one or more transmission members mounted on the foot member 16 or integrated with the actuator 28. Other arrangements are possible to provide torque at one or more of the prosthetic joints. 7935-3262-WO2 (UM 2025-002-02)

[0073] The device 10 includes a joint control system 30, including a high-level activity classifier 32 and a set 34 of mid-level controllers. The classifier 32 is programmed or otherwise configured to detect a transition from a present user activity to a different user activity and to switch control of the prosthetic joints 18, 22 among the mid-level controllers in response to detecting the transition. Each controller of the set 34 of controllers is configured to store and employ at least one control scheme according to which the control system 30 operates each actuator 26, 28 (e.g., via a motion controller). Each control scheme may be a hybrid control scheme employing impedance control and kinematic joint control at different times during each distinct user activity. When the control scheme employs impedance control, such as during prosthesis stance phase, the active controller may operate to provide the desired torque Tk, raat each joint 18, 22 according to the impedance control relationship T = —K(d — 0eq) — B0, where K, 0, 0eq, B, and 6 are joint stiffness, angle, equilibrium angle, damping, and velocity, respectively. When the control scheme employs kinematic control, such as during prosthesis swing phase, the active controller may operate to provide a desired joint angle 0kne, 0ankand joint velocity 0kne, 0ankat each joint 18, 22. Each controller of the set 34 of controllers may be programable and / or programmed with respective control scheme and in communication with each actuator 26, 28 to control its respective output torque or joint angle and velocity. A lower- level classifier may switch the active controller between impedance and kinematic control based on a ground contact sensor or phase variable (e.g., global thigh angle). In the illustrated example, the set 34 of controllers is in two-way communication with each actuator 26, 28 to receive one or more inputs from the actuators, such as a real-time encoder position which can be used to determine real-time joint angles and / or angular velocities at each joint 18, 22, among other parameters.

[0074] The control system 30 may receive information from a sensor system 36 to classify the present user activity, to detect a transition from one user activity to another, and to implement the control scheme of each of the set 34 of activity controllers. The sensor system 36 includes one or more sensors, each of which collects information pertinent to the device 10, a user of the device, user gait, and / or the surrounding environment. The illustrated sensor system 36 includes a distance sensor 38, a ground incline sensor 40, and a thigh-angle sensor 42. The sensor system 36 may include other sensors, such as a heel-strike sensor (e.g., a load cell), toe- off sensor, joint angle encoders, and / or sensors that are not attached to the prosthesis 10. The sensor system 36 may include a sensor that detects or measures one or more of the following characteristics: thigh angle, ground contact, heel strike (HS), toe off (TO), distance to an obstacle, or ground incline. A single sensor may detect or measure more than one characteristic, 7935-3262-WO2 (UM 2025-002-02) each characteristic may have a dedicated sensor, or multiple sensors can be combined to detect or measure a single characteristic. In a non-limiting example, the sensor system 36 includes a single sensor (e.g., a load cell) that detects HS, TO, and ground contact. Alternatively, ground contact can be detected with separate HS and TO sensors. For example, a gyroscope or IMU affixed to the lower leg member 14 can be used to detect HS, and an accelerometer or IMU affixed to the foot member 16 can be used to detect TO. With detection of HS and TO, foot contact can be inferred.

[0075] The distance sensor 38 is configured to measure a distance dstbetween the prosthesis 10 and a nearest object O located forward of the prosthesis. In FIG. 1, the distance sensor 38 is an ultrasonic sensor mounted along an ankle end of the lower leg member 14 above and proximate to the ankle joint 22 of the device 10. The ultrasonic sensor 38 emits sounds waves and can determine a distance during prosthesis stance dstbased on a time factor of reflections of the sound waves. The illustrated object O is an upward extending step. While an ultrasonic distance sensor 38 may be preferred due to its low cost, low energy consumption, and low computational burden, other types of distance sensors such as LiDAR, radar, and vision systems could be used with the disclosed classifier 32 and control system 30. In addition to providing the control system 30 with information pertinent to the existence of an impending transition, the distance sensor 38 can be used during a given user activity (e.g., stair ascent) to modify a baseline kinematic control scheme to add additional knee flexion to help avoid a toestub by the prosthesis.

[0076] The incline sensor 40 may be mounted along the foot member 16 of the device 10 and operates to determine an incline angle a of the surface along which the user is ambulating. One suitable incline sensor 40 is an inertial measurement unit (IMU), which can be used to measure ground incline angle and provide additional information. The thigh-angle sensor 42 may be mounted along upper leg member 12 operates to determine a global thigh angle (e.g., with respect to vertical) in real time for use by the classifier 32 and the group of activity controllers 34. The thigh-angle sensor 42 may for example be configured to provide real-time measurements of global thigh angle 0thto the control system 30 to estimate the gait phase of the device 10 and / or user during locomotion or to help identify a present or impending activity in which the user is engaged. One suitable thigh-angle sensor 42 is an IMU. The sensor system 36 may include other sensors, such as additional IMUs, to provide kinetic and kinematic information about the user, prosthetic members 12-16, prosthesis joints 18, 22, etc. to the control system 30 to implement activity classification and the respective activity controller control schemes. 7935-3262-WO2 (UM 2025-002-02)

[0077] Other control schemes are contemplated, with the control system 30 switching between and / or implementing the control scheme of the active controller in real-time as necessary based on information received from the sensor system 36. A finite state machine (FSM) may be employed to switch between kinematic and impedance controllers based on detection of swing or stance portions of the gait cycle, or to switch between a baseline control scheme and a stubavoidance control scheme depending on a real-time value measured by the distance sensor 38. The device 10 may include dedicated controllers for each joint as well. The device 10 may further include additional unillustrated components, such as power sources for the actuators and controller, cables, brackets, etc.

[0078] It should be understood that each of the control system 30, the activity classifier 32, and the various activity controllers 34 may all be considered a “controller” in its own right but are given different nomenclatures here for purposes of clarity in description. Also, each of these controllers is not necessarily embodied as a separate physical component, although this is of course possible. The control system 30 may be considered a master controller and embodied as a computer with a processor and memory capable of storing and executing instructions according to programming code. The classifier 32 may be embodied as high-level programming by which the control system 30 detects transitions among a finite number of user activities and implements prosthetic joint control via one of a corresponding finite number of control schemes — i.e., each “controller” of the group 34 of controllers may exist as a subroutine selected for execution by the activity classifier 32 based on a detected user transition to a different user activity.

[0079] Throughout this disclosure, the designations “W,” “SA,” “SD,” and “SS” respectively refer to walking, stair ascent, stair descent, and sit-stand modes, controllers, and user activities. The designation “2” is shorthand for “to” and indicates the direction of a transition from one user activity to another. For example, W2SA indicates a transition from walking to stair ascent, while SD2W indicates a transition from stair descent to walking. The letters “P” and “I” respectively indicate whether the transition is led by the prosthesis 10 or by the intact leg.

[0080] Examples of eight transitions between stair locomotion and walking are illustrated in FIG. 2. Starting at the top-left of FIG. 2, the illustrated transitions include an intact-led walking- to-stair ascent transition (IW2SA), a prosthesis-led stair ascent-to-walking transition (PSA2W), a prosthesis-led walking-to-stair ascent transition (PW2SA), an intact-led stair ascent-to-walking transition (ISA2W), an intact-led walking-to-stair descent transition (IW2SD), a prosthesis-led stair descent-to-walking transition (PSD2W), a prosthesis-led walking-to-stair descent transition (PW2SD), and an intact-led stair descent-to-walking 7935-3262-WO2 (UM 2025-002-02) transition (ISD2W). The curved arrows generally represent movement of the foot portion 16 of the prosthesis 10 during swing, with the solid arrows indicating steady-state strides and the dashed arrows indicating the first stride of the new user activity by the prosthesis 10. The lead leg (P or I) for a W2SA transition is the leg taking the first step up, and the lead leg for a SA2W transition is the leg taking the last step up. The lead leg for a W2SD transition is the leg taking the first step down, and the lead leg for a SD2W transition is the leg taking the last step down. Elevation changes are indicated in units of stair height (h), and the illustrated gait is a step- through gait (i.e., only one foot per stair). Factors to be considered for the detection and timing of each transition of FIG. 2 are described further below.

[0081] Embodiments of the activity classifier 32 may be in the form of a FSM, an example of which is schematically illustrated in FIG. 3. This high-level classifier 32 receives input information from the sensor system 36, detects the prosthesis user’s transition from one user activity to another user activity among a plurality of distinct activities based at least in part on the sensor information, and switches control of the prosthetic joints 18, 22 to the appropriate mid-level controller among a corresponding plurality of activity controllers 44-50. In the illustrated example, the plurality of prosthesis activity modes includes four discrete activity modes, including a walking mode (W), a stair-ascent mode (SA), a stair-descent mode (SD), and a sit-stand mode (SS), each with a corresponding mid-level controller, including a walking controller 44, a stair-ascent controller 46, a stair-descent controller 48, and a sit-stand controller 50.

[0082] Each of the controllers 44-50 also receives or uses information from the sensor system 36 and controls torque, angle, and / or angular velocity at each of the prosthetic joints 18, 22 based at least in part on the sensor information and on the control scheme associated with each controller. Each control scheme may, for example, employ impedance control during stance phase — while the prosthesis is in contact with the ground or other support surface — and kinematic control during swing phase — while the prosthesis is out of contact with the ground or other support surface. The sit-stand controller may employ only impedance control.

[0083] The walking controller 44 may be configured to vary torque at each prosthetic joint as a function of impedance parameters (e.g., 0eq, K, B) that vary as continuous functions of stance phase progression and one or more walking task condition, including ground inclination a. While not used in the experiments described here, the walking task conditions may also include walking speed. Phase progression may be defined as a fraction between 0 and 1, which represent stance phase endpoints (e.g., 0 at HS and 1 at TO), swing phase endpoints (e.g., 0 at TO and 1 at HS), or the same point within sequential complete gait cycles (e.g., 0 at HS and 1 7935-3262-WO2 (UM 2025-002-02) at next HS). The continuous impedance parameter functions can be derived from able-bodied data collected at discrete walking speeds and inclines and fit to continuous functions such that the walking controller 44 need not identify and switch between any particular discrete regions of stance phase progression with which constant impedance parameters are used. Instead, the instant impedance parameter values are based on smooth functions of stance phase progression and walking task conditions such that the impedance parameters automatically adapt to walking speed and / or ground incline. Kinematic control may be used during the swing portion of each gait cycle. A suitable walking controller 44 is described in international patent application publication number WO 2024 / 178129, which is hereby incorporated by reference. Information from the ground incline sensor 40 can also enhance the robustness of activity transitions. For example, while ramps (upward incline and downward incline walking) may not be distinctly classified by the classifier 32, estimated ground incline permits an inference of ramp walking. Consequently, certain transitions can be disabled when ground incline is outside a threshold range. For example, the W2SA can be disabled if ground incline exceeds a predetermined threshold based on the reasonable assumption that transitions between a steep ramp and stairs are unlikely in real-world locomotion.

[0084] The stair-ascent and stair-descent controllers 46, 48 may also be configured to vary torque at each prosthetic joint as a function of the same impedance parameters that vary as a continuous function of stance phase progression. Phase progression for these controllers may be defined as a fraction between 0 and 1, representing either stance phase endpoints (e.g., 0 at foot strike (FS) and 1 at TO), swing phase endpoints (e.g., 0 at TO and 1 at FS), or the same point within sequential complete gait cycles (e.g., 0 at FS and 1 at next FS). The impedance parameters may also be continuous functions of step height. These controllers may also add additional knee flexion to a baseline kinematic control function during stair ascent based at least in part on information from the distance sensor 28. Suitable stair-ascent and stair-descent controllers 46, 48 are described in international patent application number PCT / US2024 / 040356, filed July 31, 2024, which is hereby incorporated by reference.

[0085] The sit-stand controller 50 may also be configured to vary torque at each prosthetic joint as a function of the same impedance parameters that vary as a continuous function of stance phase progression. Phase progression for the sit-stand controller may be defined as a fraction between 0 and 1 when the user is respectively seated and standing, or vice versa. The impedance parameters may also be continuous functions of seat height. A suitable sit-stand controller is described in PCT / US2024 / 035847 filed June 27, 2024, which is hereby incorporated by reference. 7935-3262-WO2 (UM 2025-002-02)

[0086] Each change by the classifier 32 from the present activity mode to a different activity mode is represented in FIG. 3 as a transition. In this example, a total of nine identifiable transitions among four user activities and prosthesis activity modes is indicated. Here, “identifiable” means that the classifier includes instructions for detecting the respective transition. Six of the transitions are transitions to or from walking (W), including W2SS, W2SA, W2SD, SS2W, SA2W, and SD2W. Seven of the transitions are transitions to or from one of the stair modes or activities (SA, SD), including W2SA, W2SD, SA2W, SD2W, SD2SA, SA2SD, and SS2SD.

[0087] As noted above, ground incline information received from the sensor system 36 may used by the classifier 32 to selectively enable or disable the seven transitions to or from one of the stair modes and is used by the walking controller 44, when active, as an input that the stance impedance parameters are partly based on. For example, when the ground incline sensor 40 indicates that the absolute value of ground incline a is greater than a threshold value <zT(— <zT> a > aT), the classifier 32 may disable all of transitions to or from one of the stair modes (SA, SD) or otherwise ignore the programmed conditions for those transitions. In the experiments described herein, the threshold value aTwas set to 7°. The W2SS and SS2W transitions are permitted on inclines greater than the threshold value to allow the user to stop walking while on ramps at their discretion.

[0088] It is noted that FIG. 3 does not explicitly illustrate inputs from other sensors (e.g., distance sensor 38, thigh-angle sensor 42, etc.) of the sensor system 36 to the classifier 32 or individual controllers 44-50, but it should be understood activity classification by the classifier and prosthesis control by each of the individual controllers may be based at least in part on information from these and other sensors of the sensor system. For example, the classifier 32 may use information (e.g., dst) from the distance sensor 38 to monitor the approach toward stairs and the exit from stairs to provide early W2SA and SA2W transitions with the user’s maximum thigh angle (e.g., from thigh angle sensor 42) confirming the transitions. In another example, a vertical or near-vertical shank angle at HS triggers the W2SD transition and / or an increase in stride length L signals the SD2W transition. Shank angle 0sh may be calculated as the difference between 0th, based on information received from the thigh-angle sensor 42, and knee angle based on information from a knee joint encoder (e.g., as part of the knee joint actuator 26). An increased stride length may be detected by comparing a present horizontal position Pxof the foot member 16 relative to the hip joint, which can be estimated using forward kinematics, to one or more preceding horizontal positions Px, for example. Additionally, other types of sensors or sensors positioned elsewhere on the device 10 or user or in a different 7935-3262-WO2 (UM 2025-002-02) orientation or direction are contemplated. A downward- or rearward-facing distance sensor may be used to provide the control system with information useful for detecting a W2SD transition, for example.

[0089] As shown in FIG. 3, nearly all transitions among the four activity modes, or “states” of the FSM classifier 32, are bidirectional. In addition, the rules for each transition may vary based on whether the prosthesis or the intact leg is the leading leg. The classifier 32 can also detect SA2SD and SD2SA transitions when the user turns around to reverse course while on a staircase. These transitions may be indicated by changes in the heading direction (0head)asdiscussed below. Additionally, a user reset logic based on a predefined user movement of the prostheses (e.g., a hip abduction maneuver) may be implemented, resetting all feature history and the current activity mode to walking mode as a final (manual) recovery measure if misclassifications cannot be resolved by the automatic backup logic described below.

[0090] The features selected for activity classification by the classifier 32 may include kinematic features derived from thigh and shank angles 0th, 0sh, as well as environmental features measured by the distance sensor 28 and ground incline sensor 40, for example. The kinematic features allow prosthesis users to volitionally control changes in the activity mode, while the environmental features facilitate timely transitions and enhance the robustness of the activity classifier 32.

[0091] In some cases, the features may include thigh orientation-based instantaneous characteristic features (ICFs). These ICFs monitor the thigh angle 0thor changes in thigh angle at specific gait events such as maximum hip flexion (MHF) and heel strike (HS) to identify prosthesis-led transitions. An MHF event occurs when the prosthesis thigh reaches its highest point after being raised — i.e. at maximum thigh angle 0^. To identify a PW2SA transition, for example, a first ICF (ICF-1) can be used to assess the thigh angle at MHF (0t^HF) to determine if it exceeds a threshold value. To identify a PW2SD transition, a second ICF (ICF-2) can be used measure the difference between the thigh angle at MHF and at HS (0t^HF —^thS)t0determine if it exceeds a threshold value. Identification of transitions in the opposite directions (PSA2W, PSD2W) may rely on the same respective ICFs, where transitions back to walking are identified when the respective ICF falls below a different, lesser threshold value. The ICF thresholds used in the experiments described below are given in TABLE I and were tuned from those of Cheng et al. to make the transitions easier for participants in the experiments, with the same thresholds used for all participants. 7935-3262-WO2 (UM 2025-002-02)

[0092] TABLE I

[0093] Detection of transitions between walking and sit-stand modes may be based on thigh angle 0tk, thigh angular velocity 0tk, knee angle 0ne, and / or whether the foot member 16 is in contact with the ground or other support surface — obtainable via the same IMU used for incline estimations or via a load cell at the distal end of the ankle joint 22. A PSS2W transition may be detected when a prosthesis-led heel strike occurs within specified ranges for 0thand — e.g., when 0thand indicate sufficient forward movement and speed of the user’s residual thigh at prosthesis HS. A ISS2W transition may be detected when prosthesis-side ground contact is broken during the equivalent of late stance, based on different specified ranges for 0th and 0th — e.g., when 0thand indicate sufficient rearward movement of the user’s residual thigh just before TO. A W2SS transition may be identified when 0th, 0th, and 0kneare all within prescribed limits indicative of upright stance while the foot member 16 is in contact with the ground. A time threshold may be added to enhance robustness when integrated with the stair transition rules.

[0094] Global shank angle 0sk, calculated from the thigh and knee angles 0tk, 0kne, may be useful as a kinematic feature to identify some transitions because users can volitionally control leg orientation during late swing phase when the phase variable controlling the prosthetic joints ceases to progress — i.e., as the phase progression approaches or reaches 1. This volitional signal is especially useful for transitions from stair descent to walking, including both prosthetic-led (PSD2W) and intact-led (ISD2W) transitions. Due to physical constraints, the user is typically limited to small steps during stair descent but can take much longer strides off the stairs, allowing the transition to be characterized by increases in both shank angle 0skand the horizontal position Pxof the foot member relative to the hip joint, which can be estimated using forward kinematics. Additionally, shifting upper body weight during stance phase can alter the knee angle 0kne,ar|d thus shank orientation, when backdrivable motors are employed. Considering that both stair descent and sitting motions involve considerable hip flexion and thigh motion, shank angle may be more useful at distinguishing between these activities than thigh angle. 7935-3262-WO2 (UM 2025-002-02)

[0095] Shank angle 0shmay be particularly useful for classifying the IW2SD transition — a transition during which the activity mode requires switching in early- to mid-stance phase. This particular transition is relatively difficult to detect because the prosthetic knee must flex around mid-stance to lower the user down stairs, while the walking controller has the knee extended throughout stance phase. The IW2SD transition can be implemented by first switching from walking to sit-stand, and then switching to stair descent from sit-stand. This is effective because it is much simpler to differentiate between sit-stand and stair descent than between walking and stair descent. In sit-stand mode, the shank angle remains mostly vertical, whereas in stairdescent mode the shank exhibits significant flexion. Moreover, similar thigh flexion is observed during mid- to late-stance phase in both sitting and stair descent, allowing more time to differentiate the two modes. Shank angle may also serve as a distinctive feature for PW2SD since the angle at HS is nearly vertical but remains positive (more horizontal) during walking strides. When combined with ICF-2 as backup logic, this can offer a more robust feature space for classifying the transition from walking to stair descent.

[0096] Finally, hip abduction is frequently observed among prosthesis users during step-by- step stair ascent, but this behavior is uncommon or minimal during step-over-step stair ascent and other activities. Frontal-plane thigh angle, detectable from a thigh-mounted IMU (e.g., thigh-angle sensor 42) and which indicates hip abduction, can be used as a feature to implement an intuitive reset mechanism into the default system logic, serving as a final recovery strategy. Specifically, when the user intentionally abducts their hip to bring the frontal plane thigh angle beyond a threshold angle (e.g., 30°), all classification decisions and feature histories can be reset, and the activity mode can be set to walking mode.

[0097] Distance measurement by the distance sensor 38 is useful for facilitating timely transitions between walking and stair ascent. For both intact-led and prosthetic-led W2SA transitions, the aforementioned ICF-1 feature (at MHF in mid-swing) may not detect the transition early enough to provide sufficient knee flexion to clear the first stairstep without excessive user hip flexion, hiking, and / or abduction. Additional knee flexion must be provided just after toe-off (TO) to reliably clear the stairstep. Accordingly, distance measurements can be used to switch controllers before toe-off As the prosthesis user approaches a staircase, stance-phase measurements from the distance sensor decrease with each subsequent step. Decreasing distance from a stair or other obstacle in front of the user can be used to detect the staircase prior to toe-off of the W2SA transition stride. Additionally, ICF-1 can serve as backup logic or secondary criteria if the early transition fails to trigger. 7935-3262-WO2 (UM 2025-002-02)

[0098] This approach is also applicable to transitions from stair ascent to walking. For ISA2W, when the measured distance-to-obstacle is greater than that of the previous step — due to the absence of another stairstep — the transition to walking can be made in early- to mid-stance, which is much earlier than the mid-swing transition provided by the ICF-1 backup logic. For PSA2W, the usefulness of distance-based switching is limited during stance phase as the prosthesis remains on the stairs and distance measurements remain low until mid-swing. In such cases, distance-based logic may serve as a secondary strategy if the primary ICF-1 -based transition fails. To ensure accurate distance measurement to the upcoming terrain, distance measurements may be taken and averaged only while the global shank angle is within a threshold range (e.g., ± 5°) with respect to vertical during either stance or swing phase.

[0099] The validity of the distance sensor reading can be checked using heading direction ^head, which can be estimated from a yaw measurement from a thigh-mounted IMU, for example. If a turning motion is detected throughout the gait cycle, the distance reading may be deemed invalid, and preceding distance measurements can be discarded and not used for classification. A turning motion may be detected when 0headis greater than a threshold value (e.g., 0head> 20°) during stance and greater than a different threshold value (e.g., 0head> 30°) during swing. In addition, heading direction can be used to identify a direct transition between SA and SD when a 180° or near- 180° turn is detected.

[0100] The timing of each switch from one activity mode to another can have a considerable impact on how comfortable the transition is for the user. The switch timing may be defined as the latest point of the prosthesis transition stride at which the classifier 32 must switch to the mode associated with the detected activity to prevent problems such as stubbing, gait pauses, or instability.

[0101] For the W2SA transition, attention should be given to clearance with the terrain to avoid stubbing (i.e., collision of the foot member 16 with an obstacle). At this transition, the terrain includes the first step of a staircase. Accordingly, the timing of the activity mode change from walking to stair ascent may be set to occur at a time no later than toe-off (TO) of the prosthesis transition stride for both the prosthesis-led transition (PW2SA) and the intact-led transition (IW2SA).

[0102] For the SA2W transition, it is useful for the lower leg member 14 of the prosthesis 10 to be in an extended position before the post-transition heel strike (HS) to prepare the prosthesis for body weight acceptance. Accordingly, the timing of the activity mode change from stair ascent to walking may be set to occur at a time no later than HS of the prosthesis transition stride for both the prosthesis-led transition (PSA2W) and the intact-led transition (ISA2W). 7935-3262-WO2 (UM 2025-002-02)

[0103] For transitions between walking and stair descent led by the prosthesis (PW2SD, PSD2W), the timing of the activity mode change may be set to occur at a time no later than HS of the prosthesis transition stride due to similarities in joint kinematics before HS for both walking and descending stairs.

[0104] For transitions between walking and stair descent led by the intact leg (IW2SD, ISD2W), the timing of the activity mode change may be set to occur much earlier in the gait cycle, as kinematics and kinetics begin to differ early in the stance phase. Specifically, the timing of both the IW2SD and ISD2W transitions may be set to occur at a time no later than before mid-stance of the prostheses. This facilitates knee flexion necessary for descending stairs for IW2SD and ensures that the lower leg member 14 remains nearly extended throughout the stance phase for ISD2W to prevent the lowering action of the stair descent controller.

[0105] For transitions between walking and sit-stand modes, transition timing is informed by the necessity for the control system 30 to be in the correct mode to facilitate the sitting motion for W2SS and walking for SS2W. While transitioning to walking from standing, the sit-stand controller 50 keeps the prosthetic knee extended with damping behavior, whether the prosthesis 10 is in stance or swing. This allows the prosthesis 10 to step forward, initiating walking regardless of the leading leg. Accordingly, the timing of the transition from sit-stand to walking may be set to occur at a time no later than HS of the first step moving forward. Conversely, since both the walking and sit-stand controllers 44, 50 manage stationary standing, the timing of the transition from walking to sit-stand may be set to occur at a time no later than the time at which the user initiates a sitting motion to sit down.

[0106] Prosthesis and Classifier Experiments

[0107] A powered knee-ankle prosthesis 10 equipped with a control system 30 implementing a high-level classifier 32 and four mid-level controllers 44-50 consistent with the abovedescription has been fitted to and tested by two trans-femoral amputee users, both of whom are congenital amputees and whose demographics are given in TABLE II.

[0108] TABLE II

[0109] The control system 30 was implemented on a powered knee-ankle prosthesis 10 featuring low-impedance actuators 26, 28 and equipped with G-SOLO Twitter R80A / 80VDC 7935-3262-WO2 (UM 2025-002-02) drivers (Elmo Motion Control, Petah Tikva, Israel). These high-torque motors were paired with 22: 1 single-stage stepped-planet compound planetary gear transmissions to deliver up to 180 N-m torque at each of the knee and ankle joints 18, 22. Thigh orientation angle 0th and foot orientation angle were measured using thigh- and foot-mounted IMUs (3DM-CX5-25LORD, Microstrain, Williston, VT, USA). Joint angles 0kne> #ank were monitored using E5 optical quadrature encoders (LORD Microstrain, Williston, VT, USA). A 6-axis load cell (M3564F, Sunrise Instruments, Nanning, China) located above the prosthetic foot member 16 measured ground reaction forces and moments, which are used to calculate the center of pressure. Additionally, a forward-facing ultrasonic distance sensor 38 (LV-MaxSonar-EZTM, MaxBotix, Brainerd, MN) was attached above the ankle joint 22 on the lower leg member 14. The ultrasonic sensor 38 is inexpensive ($34 USD), is 19.9x22.1x15.5 mm, weighs 4.3 g, consumes 11 mW, and samples at a maximum frequency of 40 Hz. Sensors were sampled and control actions were computed at a frequency of 500 Hz using a myRIO microcontroller (National Instruments, Texas, USA). The high-level classifier 32 and mid-level controllers 44- 50 described herein were developed within the National Instruments Lab VIEW software environment and subsequently deployed to the microcontroller.

[0110] The control system 30 and set of controllers 34 among which the classifier 32 detects and implements transitions was based in part on a control system detailed by Best, et al. in “Improving amputee endurance over activities of daily living with a robotic knee-ankle prosthesis: A case study” (2023 IEEE / RSJ Int. Conf. Intelligent Robots and Systems (IROS), 2023, pp. 2101-2107), which enabled a range of activities including level walking, upward and downward incline walking, sitting and standing, and stair ascent and descent. That control system could switch among four user activity modes, but the switching was performed manually. With the automatic classifier described herein, transitions from one activity mode to another were smoothed using time-based linear interpolation over a 100 to 300 millisecond window, depending on the activity transition.

[0111] Prosthesis control in each activity mode has a similar architecture. A phase variable s is calculated based on the user’s residual thigh angle 0th, enabling volitional control over the gait cycle progression of the prosthesis. The controller calculates a commanded torque T depending on whether the prosthesis is in stance or swing phase, which is determined based on a ground reaction force. Impedance control is used in stance phase based on: 7935-3262-WO2 (UM 2025-002-02) where K, B, and 0eqdenote stiffness, damping, and equilibrium joint angle, respectively. Each of these terms is a function of phase s, task / (e.g., walking speed / incline or stair height), and activity mode During swing phase, kinematic position-based control is used based on: and using proportional-derivative control (with gains kpand kd) to track the desired kinematic trajectory 0dwith a constant viscous damping coefficient ft to enhance the stability of the low- impedance drivetrain of the prosthesis. The reference kinematics are adjusted to avoid potential toe stubs ultrasonically detected at distance z by applying an additional joint offset 0s(z) for extra knee flexion and ankle plantarflexion during stair ascent to enhance safety and comfort.

[0112] Because the sit-stand mode involves only stance phase, the associated controller 50 uses only impedance-based control. Both the sit-stand and walking controllers can manage stationary standing when the thigh is relatively vertical. However, due to kinematic differences between late-stance walking and sitting, the walking controller cannot facilitate sitting. Accordingly, the sit-stand controller is favored for stationary standing to provide seamless volitional control over sit-stand transitions based on the thigh-based phase variable.

[0113] A distinguishing feature of the walking activity mode is its capability to manage a continuous range of inclines. Unlike other known control schemes that treat upward and downward incline walking as distinct modes requiring separate classification, the prosthesis disclosed herein measures and automatically adapts to the ground angle at mid-stance of every step. This approach unifies ramp and level walking, thereby reducing the size of the activity space and complexity of the classifier.

[0114] In WO 2024 / 178129, ground angle was estimated using thigh angle 0th, forward kinematics, and a foot-bending model. While advancing the state-of-the-art at the time, this method could exhibit relatively large variations and relied heavily on averaging, which could delay adaptation to discrete changes in ground angle (e.g., at ramp transitions). The footbending model could also exhibit user-dependent bias. To help overcome these challenges, an IMU was incorporated directly within the foot member 16 of the prosthesis. The foot-mounted IMU directly measures ground angle through foot member orientation within the sagittal plane during mid-stance, where mid-stance is defined as an interval during which the center of pressure is within a manually tuned range (e.g., from 0.05 m to 0.10 m) anterior to the ankle. The average of the ground angles measured through mid-stance is then used to update the controller’s incline task variable at the end of mid-stance.

[0115] Foot-member compliance and other hardware imperfections can result in a groundangle bias, which can be corrected via a calibration offset. To reduce variability, a moving 7935-3262-WO2 (UM 2025-002-02) average of three strides is taken when the ground angle used by the controller is updated with small changes (e.g., differences less than 5°). If the change in ground angle estimation exceeds this threshold, the moving average is bypassed and reset, allowing for quick adaption when transitioning to or from ramps. Additionally, the incline angle estimate may be reset to zero when the following three conditions are met: 1) incline angle exceeds 5°; 2) the distance sensor 38 does not detect an obstacle within a distance threshold (e.g., 1 meter) during swing phase; and 3) a smaller maximum thigh angle 0th relative to the previous stride. This optimization enhances controller performance when transitioning from incline walking to level walking. Finally, the incline estimate is saturated within ±10°, corresponding to the design range of the walking controller.

[0116] Although WO 2024 / 178129 also describes a walking speed estimator to update the impedance and kinematic references, this feature was not implemented in the experiments described herein, given the relatively constant walking speed for each participant. The speed task variable was instead fixed at 1 m / s in the walking controller. The phase variable, however, inherently synchronizes the progression through these references with the user’s hip movements, allowing seamless starting and stopping. Similarly, a fixed step height was assumed in the stair ascent-descent controller described in PCT / US2024 / 040356.

[0117] FIG. 4 summarizes illustrative conditions used to detect each of the transitions of FIG. 3 for both prosthesis-led (P) transitions and intact-led (I) transitions, where applicable. These conditions may be referred to as the primary criteria for transition. Automatic back-up logic designed to compensate for misclassification is also given for each transition in the table of FIG. 4. The backup logic may be referred to as the secondary criteria for transition. Additionally, in the “Classification Timing” columns of FIG. 4, the timing of the transition detection for the primary criteria (Normal), the timing of the transition detection for the secondary or backup criteria (Backup), and the latest permissible switch timing for each transition are given as references to compare the classification timings associated with the primary and backup logics. FIG. 4 also includes instructions that can be used to familiarize new users with the classification rules, such as “raise the thigh high” or “walk normally.” The asterisk preceding some of the transitions listed in the table of FIG. 4 indicates that a cornercase situation associated with that transition may arise. Methods of addressing these corner cases are discussed below, where appropriate.

[0118] The primary criteria for a PW2SA transition were, during walking mode, the distance during stance dst tbetween the prosthesis 10 and an approaching stairstep being less than a threshold value and a difference between that distance dst tand the distance dst t-! measured 7935-3262-WO2 (UM 2025-002-02) during the previous stance being greater than a threshold value. In this example, the threshold value for dstis 1 meter, and the threshold value for its difference from the previous measurement is 0.2 meters. ICF-1 was employed as the secondary criteria for this transition.

[0119] The primary criteria for a IW2SA transition were the same criteria as the PW2SA transition except that the distance dstthreshold is lower — in this case 0.5 meters. ICF-1 was employed as the back-up criteria for this transition.

[0120] The primary criteria for a PSA2W or a ISA2W transition were, during stair-ascent mode, the distance dstexceeded a threshold value, indicating no more stairs. In this example, the threshold value is 1 meter for both prosthesis-led and intact-led transitions. ICF-1 was employed as the secondary criteria for these transitions. A comer-case situation may also be addressed for the SA2W transitions. For the PSA2W transition, it is possible that the distance sensor 38 could see over the last stair step and trigger the transition to walking mode during late-swing of the last stair ascent stride (very close to HS). To handle this case, thigh angle at heel strike can be checked after detecting a PSA2W transition. If 0t^sis greater than a threshold value (e.g., 40°), then a false transition has been made, and the thigh angle will remain high.

[0121] The primary criteria for a PW2SD transition were, during walking mode, HS was detected while the shank angle 0sh was less than a threshold value and 0t^JHFwas greater than a threshold value. In this case, the respective threshold values are 5° and 18°. ICF-2 with 0shless than a different threshold value (e.g., 11°) was employed as the secondary criteria for this transition. A corner-case situation may also be addressed for the PW2SD transition. For this transition, if the prosthesis stubs the ground during early-swing phase of walking, an early HS may be detected. The shank angle could be very flexed or close to 5°, which would inaccurately trigger the PW2SD transition. To handle this case, shank angle pattern can be checked to ensure the maximum shank angle throughout the gait cycle is greater than 0° and is decreasing (0smhax> 0° and 0smhax> 0sh).

[0122] The primary criteria for a SS2SD transition were, during sit-stand mode, foot member contact with the ground being detected while thigh angle 0th is within a prescribed range and 0shis less than a threshold value. In this example, the prescribed range is between -8° and 30° and the threshold value is -30°. The secondary criteria for this transition are the same as the primary criteria for the PW2SD transition. As noted above, the IW2SD transition may use these transition conditions as the second part of a 2-stage transition from walking to stair descent, with the first part of the 2-stage transition being a W2SS transition. 7935-3262-WO2 (UM 2025-002-02)

[0123] The primary criteria for a PSD2W transition were, during stair-descent mode, the ICF- 2 conditions being met and 0tIkHFbeing greater than a threshold value, which in this case is 7°. The secondary criteria for this transition were the same as the primary criteria for the ISD2W transition.

[0124] The primary criteria for a ISD2W transition were, during stair-descent mode, the prosthesis shank angle 0shexceeding a threshold value, the distance Pxbetween the user’s hip joint and the foot member 16 of the prosthesis 10 exceeding a threshold value at HS, and the difference between Pxand the preceding Pxat HS being greater than a threshold value. The respective threshold values in this case are 8°, 0.55 meters, and 0.07 meters. The secondary criteria for this transition are the same as the primary criteria for the PSD2W transition. A corner-case situation may also be addressed for the SD2W transition. For the SD2W transition, the leading leg can sometimes be difficult to distinguish because the transition conditions for PSD2W and ISD2W can both be satisfied. To address this, global shank angle 0shcan be checked at the post-transition HS to determine if it is greater than a threshold value (e.g., 11°). If so, the SD2W transition is classified as a PSD2W transition. Otherwise the transition is classified as a ISD2W transition.

[0125] The primary criteria for a W2SS transition were, during walking mode, the following conditions being met for a threshold amount of time: thigh angular velocity 0this less than a threshold value, thigh angle 0shis within a prescribed range, knee joint angle 0kneis less than a threshold value, and contact between the foot member and the ground is detected. In this case, the threshold time is 50 ms, the threshold is 11.5 deg / s, the threshold 0sh range is between -15° and 15°, and the threshold 0kneis 20°. There are no secondary criteria for this transition, but there is a corner-case situation that may be addressed. Specifically, it is possible for the prosthesis to enter sit-stand mode while the user is performing a W2SA transition if the user is walking very slowly. To prevent this, the time threshold can be increased (e.g., to 200 ms) if the measured di stance-to-ob stacl e during stance phase dstis less than a threshold value (e.g., 1 meter).

[0126] The primary criteria for a PSS2W transition were, during sit-stand mode, a heel strike being detected while thigh angle was within a prescribed range and thigh angular velocity 0th being less than a threshold value. In this case, the prescribed range is between 0° and 40°, and the threshold value is -17.2 deg / s. The secondary criteria for this transition were the same as the primary criteria for the ISS2W transition.

[0127] The primary criteria for a ISS2W transition were, during sit-stand mode, foot member contact with the ground being detected while thigh angle was less than a threshold value 7935-3262-WO2 (UM 2025-002-02) and 0thwas less than a threshold value. In this case, the respective threshold values are -15° and -17.2 deg / s. The secondary criteria for this transition are the same as the primary criteria for the PSS2W transition.

[0128] The primary criteria for a SA2SD transition were, during stair-ascent mode, a change in heading direction 0headis greater than a threshold value, which in this case is 150°. The primary criteria for a SD2SA transition were the same criteria occurring during stair-descent mode.

[0129] A certified prosthetist fitted the powered prosthesis to each participant. Each participant was trained on the mid-level controllers employed in each of the different activity modes and on the proposed classification system in several acclimation sessions totaling more than 5 hours prior to any data collection. This allowed the participants to become familiar with the prosthesis, to try to reduce compensatory movements they may have become accustomed to with their passive prostheses, and to identify comfortable foot placement for initiating transitions with different legs. Auditory and visual feedback was provided to participants regarding real-time mode switching during a self-paced inter-leg transition experiment and a rapid-paced endurance experiment. For safety, participants were offered to wear an overhead harness in the self-paced experiment and were required to wear it in the rapid-paced experiment. Parallel bars were available on all segments of the activity circuit depicted in FIGS. 5A and 5B, which are, respectively, an elevation view of the stair and ramp portion of the activity circuit and a plan view of the entire activity circuit.

[0130] The self-paced experiment evaluated inter-leg transition classifications among the four activity modes under self-paced conditions. Incline (ramp) walking was omitted from the self- paced experiment to focus on evaluation of different transition-leading legs and multiple stair heights. The stair circuit used in the self-paced experiment is depicted in FIG. 5A and included a five-step staircase interconnecting a ground-level level portion and a raised level portion, allowing the participants to take a few walking strides before and after ascending and descending the staircase. A stool was positioned at the far end of the ground-level portion, away from the stairs. Each trial began with the participant seated on the stool and then proceeding through the following sequence of activities: sit-to-stand — level walking toward stairs — stair ascent — level walking away from stairs — turn around — level walking toward stairs — stair descent — level walking away from stairs — turn around — stand-to-sit. Each cycle thus included the following sequential transitions: SS2W, W2SA, SA2W, W2SD, SD2W, W2SS. This sequence was repeated for a total of 40 cycles, split equally between prosthetic- led and intact-led stair transitions. These 40 cycles were performed at two stair inclinations, 7935-3262-WO2 (UM 2025-002-02) including 22.6° (5-inch steps) and 31.7° (7-inch steps). To minimize fatigue, participants were allotted three minutes of rest every five cycles.

[0131] The rapid-paced endurance experiment evaluated the robustness of the classifier and controller, including incline adaptation, during long-duration ambulation at each of the participant’s fastest sustainable pace until fatigue-induced failure. This experiment was conducted on a different day from the self-paced experiment with the same participants using the full circuit depicted in FIG. 5B. Relative to the stair circuit, the full circuit adds a ramp at the end of the elevated level portion opposite the stairs, a half-circle turn on level ground, a relatively long straight portion, and a second stool. The stair and ramp inclines were fixed at 22.6° (5-inch step height) and 11.1°, respectively. Each cycle began with the participant seated on the second stool, and then proceeding through the following sequence: sit-to-stand — level walking away from 2ndstool — half-circle turn — upward incline walking toward elevated portion — level walking toward stairs — stair descent — level walking away from stairs — turn around — stand-to-sit on 1ststool — sit-to-stand — level walking toward stairs — stair ascent — level walking away from stairs — downward incline walking away from elevated portion — half-circle turn — level walking toward second stool — turn around — stand-to-sit on second stool. Participants were instructed not to pause at the first stool.

[0132] At least three days prior to the endurance experiment, rapid-paced baseline times were recorded for each participant completing the circuit. To encourage maintaining a high speed during the experiment, participants were asked to traverse the circuit at a rapid but sustainable pace within a fixed time interval equal to 10% longer than the baseline time. Participants were instructed to remain seated at the second stool after each cycle during any excess time until an audio cue was given to begin the next cycle. Completing a cycle too quickly with at least 5 seconds of rest was discouraged via instruction during the experiment. A cycle exceeding the time limit was considered a failed cycle causing the interval timer to be reset. The experiment concluded after three consecutive failed cycles, which indicated a fatigued state. If the experiment exceeded 90 minutes, 1.8 kg (4 lbs) weights were incrementally added to a vest worn by the participant every three cycles to progressively increase the challenge. Participants chose their preferred leading leg for each transition, and researchers were permitted to manually correct misclassifications only with participant consent (though this was never needed). To ensure continuity of the experiment, a high-capacity battery pack (2800 mAh) was used to power the prosthesis and was located inside the weight vest rather than on the prosthesis.

[0133] One of the participants demonstrated prosthesis and classifier functionality in an outdoor multi-terrain setting shown in the plan view of FIG. 6A and the landscape view of FIG. 7935-3262-WO2 (UM 2025-002-02)

[0134] 6B, which included multiple portions of level walking, stair ascent and descent at a step height of about 5.5 inches, and upward and downward incline walking at a ramp angle varying between 1.6 to 8.0 degrees.

[0135] Activity classification was deemed correct when the classifier accurately identified the next activity within the designated transition strides depicted in FIG. 2. Classifications that occurred before or after the designated transition stride were deemed incorrect (i.e., a misclassification), with the exception of the ISD2W transition, which is intentionally classified earlier than the start of the transition stride, as described in conjunction with FIG. 4. The timing of the switching from one activity mode to another was deemed a late if the switch occurred within the designated stride (or sitting initiation) but later than the switch timing discussed above and illustrated in FIG. 4 — i.e., when the backup logic was relied on rather than the primary conditions listed in FIG. 4.

[0136] For both the self-paced and rapid-paced sets of experiments, classification accuracy was defined as the number of correct classifications divided by the total number of classifications. The results are reported here in terms of overall accuracy, steady-state accuracy, and transition accuracy. Additionally, misclassification recovery rate was assessed by calculating the ratio of successful recoveries (without the researcher manually overriding the misclassification) to total misclassifications. Successful recoveries from a misclassification included automatic implementation of the backup logic or user-initiated reset initiated by thigh abduction, as discussed above.

[0137] In the self-paced experiment, classification accuracy for the leading leg of the transition was also evaluated, as this impacts the switch timing between distinct activity controllers. In the rapid-paced experiment, the accuracy of incline angle estimation during level and ramp walking was evaluated by calculating the error between the controller-estimated and actual values of the incline angle. Controller performance was evaluated in terms of biomimicry via qualitative comparison of the kinematics and kinetics of the prosthetic joints to mean able- bodied data, which is data that was not used to train the respective mid-level activity controllers. For each participant, strides were parsed and averaged to obtain representative strides for each steady-state activity and the associated transitions. Data from the rapid-paced experiment was used to assess the entire activity sequence, whereas data from the self-paced experiment was used to assess inter-leg stair transitions.

[0138] FIG. 7 includes confusion matrices for overall activity classification among the midlevel controllers 44-50, and TABLE III, below, includes a summary of the overall, steady-state, and transitional accuracies, as well as the recovery rate following misclassifications. 7935-3262-WO2 (UM 2025-002-02)

[0139] Classification results for the 5-inch and 7-inch stair heights in the self-paced experiment are combined, as no significant differences in performance between them were observed.

[0140] TABLE III

[0141] In the self-paced experiments, the classifier achieved an overall accuracy of 98.58% for participant TF01 and 99.67% for participant TF02. In the rapid-paced experiments, the classifier achieved an overall accuracy of 99.41% for participant TF01 and 97.83% for participant TF02. The steady-state classification accuracy surpassed the transition identification accuracy for all cases except in the self-paced experiment for participant TF01, where the two were comparable. Remarkably, the classifier achieved steady-state accuracies of 100% and 99.76% in participant TF02’s respective self-paced and rapid-paced experiments, indicating an exceptionally low misclassification rate during steady-state walking and stair locomotion.

[0142] Participant TF01 completed 117 cycles within two hours in the rapid-paced endurance experiment, with additional weight added to the vest for the final nine cycles. In contrast, Participant TF02 completed only 18 rapid-paced cycles, hindered by self-described cardiovascular fatigue and, ultimately, a socket issue where the attachment came loose. The socket issue started during cycle 17, so only the first 16 cycles were analyzed. To examine the temporal effect of fatigue accumulation, TABLE IV presents the overall classification accuracy (mean ± standard deviation) during early, middle, and late cycles of the rapid-paced 7935-3262-WO2 (UM 2025-002-02) experiments. The classifier maintained high classification accuracy over the entire rapid-paced experiment for participant TF01, with the lowest mean accuracy occurring during the early cycles. For participant TF02 the classifier had its highest accuracy during the late cycles, with a slight decrease from the early to the middle cycles.

[0143] TABLE IV

[0144] As illustrated in TABLE III, the classifier maintained consistent steady-state and transition accuracies across both the self-paced and rapid-paced experiments for participant TF01, while steady-state and, particularly, transition accuracy for the classifier declined in the rapid-paced experiments for TF02. This may be related to participant TFOl’s higher endurance fitness level. Notably, the classifier exhibited a 100% recovery rate for both participants across all experiments such that no misclassification interventions by researchers were required. During the outdoor terrain demonstration, participant TF01 comfortably navigated all terrains with seamless activity transitions and adaptation to ramp inclines.

[0145] The compliance of the hybrid impedance-based mid-level controllers allowed the participants to continue transition strides despite some late classifications by the classifier, most of which were resolved by the automatic backup logic to successfully complete the transition stride. Late classifications in the experiments occurred in four situations, including at W2SA, ISD2W, and W2SS transitions and at one non-transition.

[0146] Late classification occurred in 5% to 10% of self-paced W2SA transitions due to the distance sensor not detecting the approaching stairs before TO of the transition stride. This rate respectively decreased to 0% to 1% during the rapid-paced experiments. In all of those cases, the backup logic successfully identified the transition and switch from the walking controller to the SA controller. Failure of the distance sensor to detect the upcoming stairs typically occurred when the participant attempted to make a large transition stride — i.e., positioning the prosthesis far enough from the stairs to ensure sufficient foot clearance.

[0147] Late classification occurred in 10.5% to 11% of ISD2W self-paced transitions where the participants failed to make the longer stride required to exit SD. The amount of late ISD2W classifications increased to 12.5% to 20% during the rapid-pace experiments. In these cases, the stair descent controller caused the prosthetic leg to initiate toe-off earlier than usual. 7935-3262-WO2 (UM 2025-002-02)

[0148] However, both participants were able to adjust to the delay and preferred using the backup logic for transitioning rather than resetting the control mode through the hip abduction maneuver.

[0149] Late classification occurred in 2% to 8% of rapid-paced W2SS transitions. Due to the experiment’s rapid pace and time constraints, participants sometimes took an extra step to turn around while simultaneously trying to sit down, prompting the classifier to briefly revert to walking mode. In these cases, after a brief pause, the prosthesis transitioned back to SS mode, allowing the participant to then sit down. This type of late classification did not occur in any cycles of the self-paced experiment for either participant.

[0150] The fourth observed late classification occurred on the rare occasion that the distance sensor detected a nearby object that was not a stair step (e.g., one of the stools), which triggered transition to stair ascent after TO during level walking. However, ICF-1 backup logic, which checks the maximum thigh angle as described above, quickly overrode the unintended transition in each case and switched the prosthesis back to walking before the subsequent heel strike. Because the participants sometimes paused to visually confirm the prosthesis was in the right mode before resuming walking, these incidents were categorized as late transitions during level walking. The occurrence rates were 0-1% across all experiments.

[0151] The classifier enables transitions between walking and stair ascent or descent whether the transition stride is made by the prosthesis or by the intact leg. The classification results are illustrated for each participant in the confusion matrices of FIGS. 8A and 8B, where leading leg errors represent non-ideal timing when switching to the correct activity in all cases. The classifier successfully distinguished between leading legs for each transition with overall accuracies of 93.81% for TF01 and 94.04% for TF02. The SD2W transitions exhibited lower accuracy than the other transitions due to the similar transition conditions used for both I- and P-led transitions. However, this did not cause any practical issues during the experiments as the switch timings during the transition stride are the same for both conditions. Participants were able to walk through other leading leg misclassifications despite the resulting non-ideal switching timing.

[0152] TABLE V illustrates ground incline estimation errors (mean ± standard deviation) for participants TF01 and TF02 across different walking conditions: ramp upward incline walking, ramp downward incline walking, and level walking. Taken together, the overall error was 0.03 ± 0.86°. For each incline condition, the mean absolute error was below 1.7°, which is not much greater than the variation in absolute ground angle, which measured upwards of 0.8° in some places on the full circuit. In addition, this magnitude of incline deviation is considered 7935-3262-WO2 (UM 2025-002-02) manageable for maintaining effective gait dynamics. The experimental results from participant TF01 generally provided closer estimates of the actual ground inclination compared to those from participant TF02, except in the case of downward incline walking. The variability in estimation accuracy across different walking scenarios may be attributed to factors such as uneven ground or differences in foot bending during walking compounded by the substantial weight differences between the two participants.

[0153] TABLE V

[0154] FIGS. 9-13 illustrate average prosthetic joint kinematics and kinetics from the rapid- paced experiments for both participants, highlighting various sub-sequences of the full circuit and including all inter-leg stair transitions. The sign convention is positive for (dorsi)flexion displacements and torques. The kinematics and kinetics resemble the mean AB biomechanics (in dashed lines in FIGS. 9-13) during both steady-state and transition strides as well as during sit-stand transitions. In particular, the unified high / mid-level controller enabled natural, continuous transitions to and from ramps without classification and to and from stairs with classification. The continuously varying joint kinematics and kinetics during ramp transitions (FIGS. 10-11) demonstrate rapid adaptation of the walking controller via real-time incline estimation, as detailed above. Similarly, the sit-stand controller enabled biomimetic sit-to-stand and stand-to-sit transitions before and after walking.

[0155] FIG. 14 illustrates average prosthetic joint kinematics and kinetics from the self-paced, inter-leg stair transition experiments respectively. The self-paced transitions exhibit similar biomimicry as the rapid-paced transitions. Notably, the knee and ankle joint angles and torques are dependent on which leg leads the transition, highlighting the ability of the prosthesis to identify leading leg of a transition and implement different transition timings when necessary. For the intact-led transition from walking to stair descent, the classifier initially switches the prosthetic leg to the SS controller before descending, as discussed above. Despite this two- stage mode transition, the resulting knee and ankle trajectories are still continuous and approximate the mean joint trajectories of able-bodied individuals.

[0156] The classifier achieved greater than 99% classification accuracy on average. As illustrated in FIG. 7, the most common type of misclassifications were the transitions between 7935-3262-WO2 (UM 2025-002-02) walking and stair ascent / descent. For PSA2W, the distance sensor occasionally scanned over the last stair step and prematurely triggered the transition to the walking controller during late swing of the last stair ascent stride. This scenario occurred a few times in the self-paced experiment at the 5-inch stair height for participant TF01 but otherwise did not occur in any other experiment. The shallow incline when this issue occurred permitted participant TF01 to finish the few early PSA2W transitions with the walking controller. A few times, the prosthesis exhibited the opposite misclassification at the PSA2W transition when: 1) the distance sensor continued to detect a small distance (< 1 meter) during the swing phase of the transition (as if there was another step), and 2) the user’s maximum thigh angle exceeded the threshold for walking. In this situation, both the normal and backup logic may fail to recognize and trigger the transition and which can force the user to heel strike and walk with slightly flexed knees, due to the continued use of the SA controller. At the 5-inch stair height, the prosthesis recovered in the cases during the next stride. However, when this situation occurred on the steeper 7-inch stair height, the participants sometimes had to employ the reset logic, especially since increased knee flexion at HS could pose challenges with additional flexion on the 7-inch stair height.

[0157] For the transitions between walking and stair descent, misclassifications sometimes occurred when participants made a “fake” step with the prosthetic leg by raising the residual thigh forward and then directly back down. This type of maneuver was erroneously recognized by the classifier as a PW2SD transition, and the resulting change to the SD controller had to be corrected with the reset logic. The classifier also occasionally failed to recognize a PW2SD transition for both participants when they made relatively long transition strides, which resulted in a large shank angle at HS with a minor difference between the maximum and HS thigh angle. However, the back-up logic resolved all of these instances without incident.

[0158] When misclassifications did occur, the backup logic was effective in capturing and correcting those misclassifications, especially after the participants became familiar with the logic rules. This efficacy is underscored by the 100% recovery rate illustrated in TABLE III, highlighting the classifiers robustness in handling classification errors without third-party intervention. For participant TF01, 44.44% of misclassifications were automatically corrected by the backup logic during the self-paced experiment, and 62.50% of misclassifications were automatically corrected during the rapid-paced experiment. This indicates that approximately half of the misclassifications were corrected by participant TF01 using the reset logic, which involves a maneuver including slight hip abduction, either because the backup logic failed to correct the misclassification or because the participant immediately noticed the misclassification and corrected it before the backup logic could. Participant TF02 relied less 7935-3262-WO2 (UM 2025-002-02) on the reset logic maneuver, with 100% of misclassifications automatically corrected during the self-paced experiment and 83.33% during the rapid-paced experiment. This suggests that the backup logic has a range of usefulness depending on the user, and, when combined with manual reset cues, enables a 100% recovery rate.

[0159] In the self-paced experiments, all misclassifications were corrected within the first transition stride for both participants. In the rapid-paced endurance experiments, a respective 83.33% (20 out of 24) and 91.67% (11 out of 12) of misclassifications were corrected within the first stride for participants TF01 and TF02. For the misclassifications not automatically corrected within the first stride, both participants took an additional stride before recognizing the error — a consequence of the fast walking speed imposed by the experiment. However, none of these misclassifications compromised participant balance during the additional step before recovery, thanks to the compliance of the impedance-based mid-level controllers. The tolerance for misclassification in the mid-level controllers, coupled with the ability of the prosthesis to recover using straightforward and intuitive backup and reset logic, marks a significant advancement in preparing the powered prosthesis for outdoor and clinical use.

[0160] The above-described heuristic, rule-based classifier is very computationally efficient, as its real-time implementation primarily involves threshold checking on directly measured quantities. This results in a very low computational demand and virtually negligible classification latency after the relevant gait event. During both the self-paced and rapid-paced experiments, the classifier operated at a frequency of 500 Hz on the on-board microprocessor (myRIO, National Instruments, Texas, USA) of the powered prosthesis. This setup resulted in a maximum classification latency of only 2 ms to identify a transition and the leading leg of each transition.

[0161] This efficiency is particularly advantageous when compared to state-of-the-art learning-based methods, which often require additional hardware such as GPUs (e.g., Jetson Nano) or mobile phones to run complex neural networks. The computational latency for these methods ranges from 44 to 140 milliseconds — i.e., two to three orders of magnitude longer than the classifier described herein. The classifier also does not burden the prosthesis with the additional weight and power consumption associated with GPUs, making it more suitable for daily use. These benefits are important for the practical deployment of advanced prosthetic legs.

[0162] In contrast to non-adaptive EMG-based classification systems, where accuracy significantly declines after just 10 trials, the classifier described herein maintained high accuracy throughout the experiments for both participants. As outlined in TABLE III, the 7935-3262-WO2 (UM 2025-002-02) classifier achieved approximately 99% accuracy during the rapid-paced endurance experiment of participant TF01, even after 80 cycles. And, although participant TF02 exhibited signs of fatigue as early as trial 10 of the rapid-paced experiment, the classifier maintained high classification accuracy throughout the entirety of the experiment.

[0163] The protocol for the rapid-paced endurance experiment added extra weight to the participant if the fast-paced experiment lasted longer than 90 minutes. Accordingly, weights were incrementally added to participant TF01 after 108 cycles, totaling an additional 5.4 kg (12 lbs) over three increments. Despite the added burden, overall classification accuracy remained high at 98.38% during the final 9 cycles. Notably, most misclassifications occurred toward the very end of the fast-paced experiment when participant TF01 was significantly challenged by the combined effects of the extra weight and fatigue.

[0164] These results, as well as the outdoor demonstration, suggest that the above-described classifier and powered prosthesis may be robust enough for sustained and variable locomotive tasks including extra weight (e.g., a backpack). Hence, this control approach may enable powered prosthetic legs to support users in demanding real-world scenarios.

[0165] 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.

[0166] 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

1. 7935-3262-WO2 (UM 2025-002-02)CLAIMS1. A powered prosthesis configured to detect a transition from a present user activity to a different user activity and, in response, switch prosthetic joint control from a first controller to a different second controller.

2. The powered prosthesis of claim 1, further comprising an activity classifier that receives information from a sensor system and detects the transition based on the received information.

3. The powered prosthesis of claim 1 or claim 2, wherein the prosthesis is configured to detect the transition based on primary criteria or based on secondary criteria such that, if the prosthesis does not detect the transition based on the primary criteria, the prosthesis has a second chance to detect the transition and switch prosthetic joint control from the first controller to the second controller.

4. The powered prosthesis of claim 3, wherein the primary criteria and / or the secondary criteria are dependent upon whether a transition stride is completed by the prosthesis or an opposite leg of the user.

5. The powered prosthesis of any one of claims 1 to 4, wherein the prosthesis is configured to switch prosthetic joint control to a default controller in response to a user input.

6. The powered prosthesis of any one of claims 1 to 5, further comprising a sensor that detects or measures one or more of thigh angle, ground contact, heel strike, toe off, distance to an obstacle, ground incline, wherein the prosthesis is configured to detect the transition based at least in part on information from the sensor.

7. The powered prosthesis of any one of claims 1 to 6, further comprising a distance sensor configured to measure a distance from the prosthesis to a nearest object in front of the prosthesis, wherein the prosthesis is configured to use the measured distance to detect the transition when one of the user activities is ascending stairs.

8. The powered prosthesis of any one of claims 1 to 7, further comprising a ground incline sensor configured to measure a ground incline angle during prosthesis stance phase, wherein the prosthesis does not consider the transition to be to or from stairs when the ground incline angle is greater than a threshold value.

9. The powered prosthesis of any one of claims 1 to 8, wherein each controller employs7935-3262-WO2 (UM 2025-002-02) impedance control during prosthesis stance phase and kinematic control during prosthesis swing phase.

10. The powered prosthesis of any one of claims 1 to 9, further comprising a knee joint and an ankle joint, the prosthesis being configured to switch joint control of both joints to the second controller in response to detecting the transition.

11. The powered prosthesis of any one of claims 1 to 10, wherein each of the first controller and the second controller is a walking controller, a stair-ascent controller, a stair-descent controller, or a sit-stand controller.

12. The powered prosthesis of any one of claims 1 to 11, wherein each of the user activities is one of the following group of user activities: walking, ascending stairs, descending stairs, or standing, the powered prosthesis further comprising an activity classifier that detects the transition from one of the group of user activities to another of the group of user activities, wherein the first controller corresponds to the present user activity and the second controller corresponds to the second controller.

13. The powered prosthesis of any one of claims 1 to 12, wherein the prosthesis is configured to switch prosthetic joint control from a walking controller to a stair-ascent controller whenwhere dst tis distance between the prosthesis and a stair during prosthesis stance phase, dst t-xis distance between the prosthesis and the first stair during a previous prosthesis stance phase, d is a first threshold value, and d2is a second threshold value that is less than d .

14. The powered prosthesis of claim 13, wherein the prosthesis is configured to switch prosthetic joint control from the walking controller to the stair-ascent controller when 0t^HFis greater than a threshold value during prosthesis swing phase, where 0t^HFis thigh angle at maximum hip flexion.

15. The powered prosthesis of any one of claims 1 to 14, wherein the prosthesis is configured to switch prosthetic joint control from a stair-ascent controller to a walking controller when dswor dstis greater than a threshold value, where dswis distance to a nearest object in front of the prosthesis during prosthesis swing phase and dstis distance to a nearest object in front of the prosthesis during prosthesis stance phase.7935-3262-WO2 (UM 2025-002-02)16. The powered prosthesis of claim 15, wherein the prosthesis is configured to switch prosthetic joint control from the stair-ascent controller to the walking controller when 0t^HFis less than a threshold value during prosthesis swing phase, where 0t^HFis thigh angle at maximum hip flexion.

17. The powered prosthesis of any one of claims 1 to 16, wherein the prosthesis is configured to switch prosthetic joint control from a walking controller to a stair-descent controller when, at prosthesis heel strike,02, where 0sh is a prosthesis shank angle, 0t^HFis thigh angle at maximum hip flexion during prosthesis swing phase prior to heel strike, 0i is a first threshold angle, and 02is a second threshold angle.

18. The powered prosthesis of claim 17, wherein the prosthesis is configured to switch prosthetic joint control from the walking controller to the stair-descent controller when, at prosthesis heel strike,0sh< 03,04,is thigh angle at heel strike, 03is a third threshold angle that is greater than 0i, and 04 is a fourth threshold angle.

19. The powered prosthesis of any one of claims 1 to 18, wherein the prosthesis is configured to switch prosthetic joint control from a walking controller to a stair-descent controller when, during prosthesis stance phase, i < 0th < 02 and 0sh< 03, where 0th is thigh angle, 0sh is prosthesis shank angle, 0i is a first threshold angle, 02is a second threshold angle, and 03is a third threshold angle.

20. The powered prosthesis of claim 19, wherein the prosthesis is configured to switch prosthetic joint control from the walking controller to a sit-stand controller before switching prosthetic joint control to the stair-descent controller.

21. The powered prosthesis of claim 19 or claim 20, wherein the prosthesis is configured to switch prosthetic joint control from the walking controller to the stair-descent controller when, at prosthesis heel strike,05,7935-3262-WO2 (UM 2025-002-02) where 04is a fourth threshold angle, and 6Sis a fifth threshold angle.

22. The powered prosthesis of any one of claims 1 to 21, wherein the prosthesis is configured to switch prosthetic joint control from a stair-descent controller to a walking controller when, at prosthesis heel strike,where 0t^sis thigh angle at heel strike, 0t^HFis thigh angle at maximum hip flexion during prosthesis swing phase prior to heel strike, 6 is a first threshold angle, and 02is a second threshold angle.

23. The powered prosthesis of claim 22, wherein the prosthesis is configured to switch prosthetic joint control from the stair-descent controller to the walking controller when, at prosthesis heel strike,where Pxis horizontal distance between a user hip joint and a foot member of the prosthesis, horizontal distance between the user hip joint and the foot member at a previous prosthesis heel strike, 0shis prosthesis shank angle, di is a first threshold value, d2is a second threshold value that is less than d , and 03is a third threshold angle.

24. The powered prosthesis of any one of claims 1 to 23, wherein the prosthesis is configured to switch prosthetic joint control from a stair-descent controller to a walking controller when, at prosthesis heel strike,where Pxis horizontal distance between a user hip joint and a foot member of the prosthesis, horizontal distance between the user hip joint and the foot member at a previous prosthesis heel strike, 0shis prosthesis shank angle, di is a first threshold value, d2is a second threshold value that is less than di, and 6, is a threshold angle.

25. The powered prosthesis of any one of claims 1 to 24, wherein the prosthesis is configured to switch prosthetic joint control from a walking controller to a sit-stand controller when, during prosthesis stance phase,for a prescribed amount of time, where 0th is thigh angular velocity, 0th is thigh angle, 0sh is prosthesis shank angle,Xis a threshold angular velocity, 0! is a first threshold angle, and 027935-3262-WO2 (UM 2025-002-02) is a second threshold angle.

26. The powered prosthesis of any one of claims 1 to 25, wherein the prosthesis is configured to switch prosthetic joint control from a sit-stand controller to a walking controller when a prosthesis heel strike is detected and01 < 0th< 02and 0th< 0Xwhere 0this thigh angle, 0this thigh angular velocity, ! is a first threshold angle, 02is a second threshold angle, and 0Xis a threshold angular velocity.

27. The powered prosthesis of claim 26, wherein the prosthesis is configured to switch prosthetic joint control from the sit-stand controller to the walking controller when prosthesis ground contact is detected and0th< 03and 0th< 0Xwhere 03is a third threshold angle.

28. The powered prosthesis of any one of claims 1 to 27, wherein the prosthesis is configured to switch prosthetic joint control from a sit-stand controller to a walking controller when prosthesis ground contact is detected and0th < 0i and 0th< 0Xwhere 0tdis thigh angle, 0tdis thigh angular velocity, 6, is a threshold angle, and 0Xis a threshold angular velocity.

29. The powered prosthesis of any one of claims 1 to 28, wherein the prosthesis is configured to switch prosthetic joint control between a stair-ascent controller and a stairdescent controller when A0headis greater than a threshold value, where A0headis a heading direction of the prosthesis.

Images