System and method for controlling a wearable robotic device

Abstract

Systems and methods related to the operation of wearable robotic systems are disclosed. In one embodiment, a user intention model is developed to accurately identify the intentions of a user in real-time. Utilizing data from multiple wearable sensors, including Inertial Measurement Units (IMUs) and capacitive deformation sensors, a neural network-based approach is employed to capture both kinematic-related and kinetic-related signals. This intention model aims to enhance the accuracy and reliability of movement identification, enabling more effective control of the exoskeleton. In one embodiment, the effects of hysteresis in the exoskeleton system are addressed through the implementation of the Preisach model. By integrating the Preisach model into the control approach, the effects of hysteresis are mitigated, enabling accurate prediction of the required pressure for delivering support to the individual user. By accurately identifying user intentions in real-time and compensating for hysteresis effects, the exoskeleton's performance is optimized.

Description

SYSTEM AND METHOD FOR CONTROLLING A WEARABLE ROBOTIC DEVICERELATED APPLICATIONS

[0001] This Application claims the benefit of priority under 35 U.S.C. § 119(e) of U.S. Provisional Application Serial No. US 63 / 520,926, filed August 21, 2023, and U.S. Provisional Application Serial No. US 63 / 581,575, filed September 8, 2023, the disclosures of each of which are incorporated herein by reference in their entirety.GOVERNMENT SUPPORT

[0002] This invention was made with government support under 2236157 and 2011754 awarded by National Science Foundation (NSF). The government has certain rights in this invention.FIELD

[0003] Disclosed embodiments are related to robotics, specifically to wearable soft robots.BACKGROUND

[0004] Assistive wearable robotics is an emerging trend in the domain of robotics. From medical applications (e.g., stroke rehabilitation, spinal cord injury assistance, etc.) to augmenting the strength or endurance of non-impaired individuals (e.g., to reduce injuries in workplaces, to assist during sport performances, etc.), hundreds of assistive wearable robots have been developed in both universities and industry. Currently, commercially available wearable robots are mostly rigidly framed, with either passively or actively actuated joints, and can provide targeted assistance to one or more human body joints simultaneously.Recently, assistive wearable robots have been made sufficiently portable to be usable in real- world tasks, like for performing activities of daily living in the home or for performing overhead work in a factory. For portable wearable robots being used in these ecological contexts, applying forces to the human user that are appropriate for a given task is an especially important consideration.

[0005] Within medical applications, portable assistive and rehabilitative wearable robots are designed to enable a user to be more independent in their home environment. Purely assistivedevices are designed to help those with conditions such as amyotrophic lateral sclerosis (ALS), spinal cord injury, or muscular dystrophy who have permanent loss of some physical function be able to perform functional tasks independently (eating, self-grooming, walking, etc.). Similarly, rehabilitative wearable robots can also be used for assisting with functional tasks, but it is also critical that the wearable robot can be tuned to provide just enough assistance so that the user is able to perform more functional tasks without over-assisting to the point of negating the goal of rehabilitation. Conventional rehabilitation techniques, such as physical and occupational therapy, have shown effectiveness in promoting recovery. Nevertheless, the limited frequency of therapy sessions restricts the potential for a continuous rehabilitation progress. One population that stands to benefit greatly from wearable rehabilitative robots is stroke, which is a leading cause of long-term disability worldwide which often results in impairments in motor functions of the upper and lower limbs. These functional limitations significantly impact the quality of life of stroke survivors and their ability to perform daily activities independently. One of the key challenges for medical applications of wearable robots is the loss of the users' loss of varying degrees of voluntary control over their affected limb (due to neurological issues, muscle weakness, etc.) leading to difficulty in initiating and fully executing desired movements. This impaired motor control can result in compensatory strategies and abnormal movement patterns.

[0006] There are also many applications for assistive robots for non-impaired individuals. One such example is for factory workers, who frequently develop musculoskeletal disorders from overuse of their bodies during physical labor. In the US alone, there are nearly 70 million physical visits due to musculoskeletal disorders resulting from use in the workplace. Assistive wearable robots have the potential to reduce these injuries by reducing the physical effort required from human workers while maintaining the versatility and precision that makes human workers valuable in many factory processes. Within these applications, the wearable assistive robot must provide appropriate levels of assistance while avoiding getting in the user’s way.SUMMARY

[0007] In one embodiment, a method for controlling a wearable robot comprising one or more actuators on a user, comprises: reading motion data from one or more motion sensors;reading interaction data from one or more interaction sensors; and determining at least one control of the at least one actuator of the wearable robot from the motion data, the force data, and a model of user intent to assist motion of at least one joint of the user.

[0008] In one embodiment, the method further comprises controlling the at least one actuator of the robot based at least in part on the determined at least one control. In one embodiment, the method may further comprise: constructing the model of user intent; determining an intent of the user from multimodal wearable sensor data and dataset of training data; and constructing a model for the at least one actuator; wherein the at least one control of the at least one actuator is determined from the determined intent of the user and the model for the at least one actuator, of the at least one actuator.

[0009] In one embodiment, constructing the model of user intent may comprise: generating at least one training data collection comprising: performing a plurality of trials comprising a plurality of trials in two or more directions and a plurality of static trials where the user keeps the target joint steady while performing another action in parallel. In one embodiment, the method may further comprise: updating the model of user intent while the wearable robot is used. Optionally, the model of user intent is a statistical model. In one embodiment, the method may further comprise: extracting from the at least one data collection a plurality of features relating to the movement of one or more limbs of the user; and constructing the model of user intention using the plurality of features as inputs.

[0010] In one embodiment, the method may further comprise: extracting from the at least one data collection a plurality of features comprising one or more of: joint velocities and change in interaction forces; and constructing a neural network model of user intention using the plurality of features as inputs, a plurality of classes as outputs, and a plurality hidden layers. Optionally, constructing the model for the at least one actuator comprises: constructing a Preisach model comprising a plurality or continuous distribution of relays having a corresponding plurality or continuous distribution of weights; determining values for the plurality or continuous distribution of weights comprising: performing one or more calibration sweeps measuring a plurality of pressures corresponding to a plurality of angles. Optionally, the intent of the user comprises a probability of the user intending to extend a limb, a probability of the user intending to flex a limb, and a probability of the user intending to extend the limb, and a probability of the user intending to hold the limb stationary.Optionally, the intent of the user comprises a probability of the user intending to move the arm upward, a probability of the user intending to move the arm downward, and a probability of the user intending to hold the arm at a desired angle. Optionally, determining at least one control to the at least one actuator comprises: determining a command to the actuator corresponding to an input wearable sensor data value using the Preisach model. In one embodiment, the method may further comprise: controlling one or more valves to modulate a flow of air to the actuator to regulate an error between an actual pressure value and the desired pressure value.

[0011] In another embodiment, a wearable robot system comprises: one or more actuators configured to apply a force to at least one joint of a user wearing the wearable robot system; one or more motion sensors configured to measure a movement of a portion of the user of the robot system; one or more force sensors configured to measure force data applied to the portion of the wearer of the robot system; a processor operatively coupled to the pressure source, the one or more motion sensors, and the one or more force sensors, wherein the processor is configured to execute one or more of the embodiments of the methods described above. Optionally, at least one of the motion sensors is at least one selected from an inertial measurement unit. Optionally, the one or more actuators comprise at least one actuator configured to tension a tether. Optionally, the one or more actuators comprise at least one pneumatic actuator. Optionally, the at least one pneumatic actuator is configured to be positioned in the axilla of the wearer. Optionally, at least a first one of the force sensors is positioned between a torso of the wearer and the pneumatic actuator. Optionally, at least a second one of the force sensors is positioned between an arm of the wearer and the pneumatic actuator. Optionally, the robot is an exosuit and / or an exoskeleton. Optionally, the assisted force is applied by at least one of a pressure, torque, and / or tensile force of the at least one actuator. Optionally, at least one of the interaction sensors is a capacitive deformation sensor configured to measure a proxy of the interaction force between at least one actuator and the human body, and tension sensor configured to measure a tension of the at least one actuator. Optionally, the interaction data includes at least one selected from a tensile force, a pressure, a deformation, and a torque.

[0012] It should be appreciated that the foregoing concepts, and additional concepts discussed below, may be arranged in any suitable combination, as the present disclosure isnot limited in this respect. Further, other advantages and novel features of the present disclosure will become apparent from the following detailed description of various nonlimiting embodiments when considered in conjunction with the accompanying figures.BRIEF DESCRIPTION OF DRAWINGS

[0013] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component that is illustrated in various figures may be represented by a like numeral. For purposes of clarity, not every component may be labeled in every drawing. In the drawings:

[0014] FIG. 1A depicts one embodiment of a wearable robotic system;

[0015] FIG. IB depicts additional details of the embodiment of a wearable robotic system of FIG. 1A;

[0016] FIG. 2A depicts a block diagram showing one embodiment of a wearable robotic system;

[0017] FIG. 2B depicts a block diagram showing another embodiment of a wearable robotic system;

[0018] FIG. 2C depicts an exemplary machine learning training data collection that may be used to produce training data as input to a user intent model of an exemplary embodiment of a wearable robotic system;

[0019] FIG. 2D depicts exemplary limp air pressure sweeps that may be used to produce calibration data as input to a Preisach model of an exemplary embodiment of a wearable robotic system;

[0020] FIG. 3 shows an exemplary Graphical User Interface (GUI) of one embodiment of the present invention.

[0021] FIG. 4 depicts a state diagram of one embodiment of a wearable robotic system;

[0022] FIG. 5A depicts graphs of three pressure sensors and one motion sensor over time for different arm elevations over time;

[0023] FIG. 5B depicts graphs of one processed motion sensor signal and one interaction sensor signal over time;

[0024] FIG. 6 depicts an exemplary model of user intent for one embodiment of a wearable robotic system;

[0025] FIG. 7 depicts an exemplary Preisach model of an actuator of one embodiment of a wearable robotic system;

[0026] FIG. 8 depicts a plane of an exemplary Preisach model of an actuator of one embodiment of a wearable robotic system;

[0027] FIGs. 9A-9D depicts boundary behavior of an exemplary Preisach model of an actuator of one embodiment of a wearable robotic system;

[0028] FIG. 10 illustrates one embodiment of a method for controlling an exemplary wearable robot or other device;

[0029] FIG. 11 A depicts an example process for adapting a model to a particular participant, according to one embodiment;

[0030] Fig. 11B illustrates an example calibration loop, according to one embodiment;

[0031] FIG. 12A illustrates a plot of an example linear model and a calibration loop in the pressure-angle plane, according to one embodiment;

[0032] FIG. 12B illustrates a plot of an example nonlinear model and a calibration loop in the pressure-angle plane, according to one embodiment;

[0033] FIG. 12C illustrates a plot of an example binary model and a calibration loop in the pressure-angle plane, according to one embodiment;

[0034] FIG. 12D illustrates a plot of an example Preisach model and a calibration loop in the pressure-angle plane, according to one embodiment;

[0035] FIG. 13 depicts an example process for testing the accuracy of models for patient activities, according to one embodiment;

[0036] FIG. 14A depicts the commanded and measured angles from the testing of the linear model, according to one embodiment;

[0037] FIG. 14B depicts the commanded and measured angles from the testing of the nonlinear model, according to one embodiment;

[0038] FIG. 14C depicts the commanded and measured angles from the testing of the binary model, according to one embodiment;

[0039] FIG. 14D depicts the commanded and measured angles from the testing of the Preisach model, according to one embodiment;

[0040] FIG. 15A depicts the angle tracking accuracy, as the angle RMSE, across four models, according to one embodiment;

[0041] FIG. 15B depicts the angular velocity tracking accuracy as the angular velocity RMSE, across four models, according to one embodiment;

[0042] FIG. 16A depicts the UP and DOWN movements performed for evaluating an intention detection model, according to one embodiment;

[0043] FIG. 16B depicts the HOLD movements performed for evaluating an intention detection model, according to one embodiment;

[0044] FIG. 17 depicts charts displaying the accuracy of the intention detection model, according to one embodiment;

[0045] FIG. 18A depicts the control objectives during the UP intention states on an anglepressure plot, according to one embodiment;

[0046] FIG. 18B depicts the control objectives during the HOLD intention states on an anglepressure plot, according to one embodiment;

[0047] FIG. 18C depicts the control objectives during the DOWN intention states on an angle-pressure plot, according to one embodiment;

[0048] FIG. 19 depicts data associated with a representative participant post-stroke performing isolated shoulder abduction / adduction, according to one embodiment;

[0049] FIG. 20 depicts data associated with a representative participant living with ALS performing a simulated eating task, according to one embodiment;

[0050] FIG. 21A depicts the shoulder elevated internal / external rotation and actuator pressure time series figures for a participant with the controller on, according to one embodiment;

[0051] FIG. 21B illustrates the shoulder elevated internal / external rotation time series for a participant with the controller off, according to one embodiment;

[0052] FIG. 22 depicts representative time series results showing the differences in force, actuator air pressure, and shoulder elevation between a human-robot hysteresis only controller and a controller which combines the human-robot hysteresis and intention detection models during isolated shoulder flexion and abduction tasks, according to one embodiment;DETAILED DESCRIPTION

[0053] Wearable robots are oftentimes equipped with sensors and actuators that detect and respond to a user’s movements, providing customized assistance tailored to their specific needs during a given task. While the sensors may be used to understand the state of the coupled human-robot system in real-time, it is oftentimes difficult to properly determine when a user’s intended movement changes. Thus, understanding the relationship between these sensed parameters and the intention of a user may be important in various applications for appropriately controlling the actuators of these wearable robots.

[0054] As described above, conventional wearable robots are typically rigid. Shortcomings associated with these rigid robots include their size and weight, the need for careful alignment between the user joints and the robot joints, relatively high cost, and limited portability.

[0055] A promising new trend in this field is the creation of assistive wearable robots that use soft actuators to directly engage with the human body and to support its movements. In a soft wearable robot, the wearer’ s skeletal structure may be considered the frame of the robot. The combination of soft actuation methods, like pneumatic textile-based actuators or cable-driven actuators, and the lack of a rigid frame may address many of the aforementioned limits of rigid robots. The inherent compliance of the actuators may allow for a more direct, safer and natural physical interaction between the users and the robot.

[0056] Despite the promising potential of exoskeletons in stroke rehabilitation, several challenges remain. Designing effective control strategies that accurately capture user intentions and seamlessly coordinate with the user’s movements, while also addressing hysteresis effects resulting from non-linear materials in the exoskeleton, is a complex task. Ensuring user comfort, adaptability to different user profiles, and addressing issues such as device weight and bulkiness are also critical factors for successful implementation. To address these challenges, ongoing research focuses on developing advanced control algorithms, integrating sensor technologies, and personalizing exoskeletons based on individual user characteristics.

[0057] Accordingly, there exists a need to overcome these challenges and further refine exoskeleton-based rehabilitation systems and methods to advance stroke rehabilitation and ultimately improve user outcomes.

[0058] Despite these benefits of assistive wearable robots that use soft actuators, the complexity of sensing and controlling the interaction between the robot and the wearer is dramatically increased due to inherent compliance in the combined system of the soft robotic actuators and the user’s body. This interaction is challenging to model due to the many degrees of freedom involved between the soft actuator and the biological tissues of the user. This challenge can be further complicated when dealing with medical applications and impaired populations whose muscle tone and body composition may vary more broadly than a healthy population.

[0059] When considering medical applications such as stroke rehabilitation, the most commonly used conventional wearable robots are rigid exoskeletons and rigid manipulanda or end-effector robots. Exoskeletons and manipulanda have been evaluated intensively with large populations of stroke survivors and with randomized controlled trials. However, the improvements in relevant outcome metrics when comparing robotic assisted therapies with conventional therapies are still limited and therefore the advantage of using these devices is still unproven.

[0060] The relative portability of soft robots may allow for seamless integration with the clinical environment, favoring the adoption of robotic therapy during simulated activities of daily living (ADL) scenarios. This portability can also potentially allow the robots to be brought into the homes of those in need of therapy, enabling greater therapy engagement and a greater volume and duration of therapy. In fact, limited volume of therapy is considered the number one cause of limited outcomes in both conventional and the robot-assisted stroke therapy.

[0061] The literature of soft wearable assistive and rehabilitation robots is still young and evolving. To date, it is believed that very few devices targeting upper-limb assistance (excluding hand) exist and predominantly assist only a single joint. Many of these robots are cable-driven, targeting either the shoulder, the elbow or the wrist. Multi joint robots have typically assisted both the shoulder and elbow. Many of these devices have been evaluated on healthy participants but very few have been evaluated on any impaired users. A pneumatic shoulder robot recently demonstrated reductions in muscle activity and was shown to improve the average range of motion (ROM) across 6 stroke survivors.

[0062] In a study with 5 stroke survivors, it was demonstrated that a pneumatic shoulder robot may both increase their functional ROM while also reducing the fatigue of a therapist when performing rehabilitation exercises. However, the device in that study was manually controlled to modulate the delivered assistance by a member of the research team. Additionally, based on stroke survivor and therapist feedback during that study, it was determined that additional assistance was required at the elbow to fully overcome the flexor synergy, whereby stroke survivors often experience involuntary coupling of shoulder abductor activity with activation of elbow flexors.

[0063] In view of the above, the inventors have recognized and appreciated the benefits associated with sensing and dynamic control strategies to facilitate control of a wearable robot. A rigid wearable robot may include a well-defined structure with definite joint locations. As such, conventional rigid wearable robots may take advantage of certain calibration and / or control strategies that employ inverse kinematics. In contrast, a soft wearable robot may not be associated with well-defined structures and / or joint locations. Accordingly, the inventors have appreciated that an experimental approach to calibration and / or control may be more suitable for a soft wearable robot, as described in greater detail below.

[0064] In some embodiments, a method for controlling a wearable robot comprising one or more actuators on a user, comprises reading motion data from one or more motion sensors, reading interaction data from one or more interaction sensors, and determining at least one control of the at least one actuator of the wearable robot from the motion data, the interaction data, and a model of user intent to assist motion of at least one joint of the user. In one embodiment, the method may further comprise controlling the at least one actuator of the robot based at least in part on the determined at least one control. In one embodiment, the method may further comprise constructing the model of user intent, determining an intent of the user from the motion data, the interaction data, and the model of user intent, and constructing a model for the at least one actuator. Optionally, the at least one control of the at least one actuator may be determined from the determined intent of the user and the model for the at least one actuator. Optionally, the step of constructing the model of user intent may comprise generating at least one data collection. Optionally, the data collection may begenerated by performing a plurality of trials comprising a plurality of trials in two or more directions and a plurality of stationary trials.

[0065] In one embodiment, an intent of the user is determined based partly on a physical attempt of the user to move a joint in a desired direction of joint motion (e.g., extension or flexion), or to hold it stationary. This may lead to movement in the direction of an applied force by the user, a direction against an applied force or holding it stationary when an applied force is greater than, less than, or equal to a force needed to support the joint in the desired pose respectively. Accordingly, the objective is to identify one of these three scenarios and use the identification with other factors to determine the user’s intent.

[0066] Motion data includes motion of a joint which may be characterized in a number of ways (e.g., sense relative motion of two portions of a user’s body on either side of the joint, an extension angle of the joint, any other type of motion data that can be used to characterize movement of the joint). Interaction data includes pressure applied to a portion of the user’s body by the device, tension applied to a flexible tether, deformation applied to a soft sensor, current applied to a motor, or other parameter capable of characterizing a pressure, torque, tensile force, or other type of interaction force data to a user body. The types of joints include any arm, back, leg, shoulder, elbow wrist, hip, knee, or other joint.

[0067] In some embodiments, a wearable robot system comprises one or more actuators configured to apply a force to at least one joint of a user wearing the wearable robot system, one or more motion sensors configured to measure a movement of a portion of the user of the robot system, one or more interaction sensors configured to measure an interaction force (or proxy of interaction force) applied to the portion of the wearer of the robot system, a processor operatively coupled to the pressure source, the one or more motion sensors, and the one or more interaction sensors, the processor configured to execute the methods for controlling the robot system described herein. Optionally, the one or more actuators comprise at least one actuator configured to tension a tether. Optionally, the one or more actuators comprise at least one pneumatic actuator. Optionally, the at least one pneumatic actuator is configured to be positioned in an arm pit area of the wearer.

[0068] Optionally, depending on the type of actuator used to apply a desired force / torque and / or motion to a joint, the motion sensor is at least one selected from an inertialmeasurement unit, and / or any other sensor configured to measure a parameter capable of characterizing the movement of a joint or other portion of a user’s body.

[0069] Optionally, at least one of the interaction sensors is positioned at the surface between a torso of the wearer and the pneumatic actuator. Optionally, at least one of the interaction sensors is positioned at the surface between an arm of the wearer and the pneumatic actuator.

[0070] In one embodiment, the robot is an exosuit and / or an exoskeleton. Optionally, the assisted force is applied by at least one of a pressure, torque, and / or tensile force of the at least one actuator. Optionally, the interaction sensor is a pressure sensor configured to measure a proxy of the interaction force between at least one actuator and the user’s body, and tension sensor configured to measure a tension of the at least one actuator.

[0071] As used herein, the term soft actuator may refer to any actuator that includes flexible, compliant, and / or elastomeric materials and / or structures. For example, a fluidic soft actuator may include an elastomeric body, such as a bladder, that expands or contracts in response to the change in pressure of a fluid (e.g., air, water) within the elastomeric body. A fluidic soft actuator may alternatively be comprised of the textile outer shell with a fluid-impermeable lining which fold and unfold in response to changes in the fluid pressure within the actuator, or application of external forces and torques. Alternatively, a soft actuator may include a flexible tether (e.g., a cable, wire, or other flexible structure cable of transferring a force) that is driven by a more traditional actuator, such as a motor. It should be appreciated that a soft actuator may include components that are flexible enough such that one or more components of the soft actuator engaged with one or more portions of the user’s body deform with, and may conform to, a shape and / or orientation of a portion of the user’s body the components are engaged with. For example, one or more flexible straps, cuffs, bladders, and / or any other appropriate component may be flexible enough to conform to a shape and / or orientation of an associated portion of the user’s body during operation. In some embodiments, a soft actuator may be controlled by modulating a pressure applied to the soft actuator. In some embodiments, a soft actuator may be controlled by modulating a flow rate of fluid applied to the soft actuator. However, embodiments in which a force applied to a tether is used to provide the desired actuation are also contemplated.

[0072] It should be understood that the various methods and systems described herein may be operated to calibrate and operate a system over a plurality of joint angles and a plurality ofpressures or other actuation parameters. Additionally, the various methods of operation may be used either for a single joint and / or may be applied to control separate joints of a user either individually and / or in coordination with one another.

[0073] Turning to the figures, specific non-limiting embodiments are described in further detail. It should be understood that the various systems, components, features, and methods described relative to these embodiments may be used either individually and / or in any desired combination as the disclosure is not limited to only the specific embodiments described herein.

[0074] FIG. 1A depicts one embodiment of a wearable robotic system 100. The wearable robotic system 100 includes a shoulder portion 120 and an elbow portion 140. The system 100 additionally includes a controller 200 and a pressure source (not shown) operatively coupled to the actuators (described below) of the robotic system 100. The portability and compliance of the wearable robotic system may allow for the robot to be used safely outside of clinical settings, where the robot can be donned by untrained caregivers as accurate alignment may not be needed due to the inherent compliance of the robot. The shoulder portion may be made by a single or multiple chambers, assisting a single or multiple degrees of freedom and thus shoulder motions simultaneously. The shoulder portion may provide gravity compensation to the upper limb using one or more actuators located in the axilla, as described in greater detail below. The shoulder portion may also provide horizontal flexion compensation to the upper limb using one or more actuators located around the scapula. The shoulder portion may also provide internal / external rotation compensation to the upper limb using one or more actuators located around the shoulder.

[0075] FIG. IB depicts the elbow portion 140 of the wearable robotic system 100 of FIG. 1A. For applications related to stroke or other neurological rehabilitation, the elbow portion may counteract the effects of the flexor synergy by assisting with elbow extension. The elbow may also be assisted to improve performance during functional tasks. The elbow component 140 shown in FIG. IB includes a sleeve 143. The sleeve may include both flexible and inflexible materials (e.g., textile materials). For example, the sleeve may include an extensible compression material with select inextensible elements for anchoring. A cutout on the posterior side of the sleeve may both locate and provide pressure relief about the olecranon. A zipper may be included on the upper half of the sleeve to assist with donning.The sleeve 143 may additionally include straps, such as proximal strap 142 and distal strap 144, to enable anchoring to the body and / or adjustments.

[0076] The elbow component 140 includes one or more actuators, including a first actuator 146 and a second actuator 148. In some embodiments, the actuators may be textile-based actuators. In some embodiments, the actuators may be fluidic (e.g., pneumatic, hydraulic) actuators. The actuators may include a simple cylindrical geometry, sewn on the anterior side of the elbow, offset on either side of the mid-line, to provide assistance with elbow extension. The actuators may also include a simple cylindrical geometry, sewn on the posterior side of the elbow, offset on either side of the mid-line, to provide assistance with elbow flexion. A pair of actuators may be used, as a single larger actuator located along the mid-line may twist rather than unfold under loading. Furthermore, the inextensible textile of the sleeve between both actuators may act as a sling, evenly distributing the forces of actuation. The anchoring and fit of the actuator may be tuned using straps (not shown) located on either side of the olecranon.

[0077] The wearable robot may use an estimate of the user arm kinematics in order to determine the appropriate assistance to provide through the actuators. This estimation can also be used clinically to quantitatively measure short- and long-term changes in kinematics when the wearer is being assisted by the robot, which may assist the therapists in planning and determining personalized goals of the therapy. To enable estimation, a wearable robot may include one or more sensors (e.g., sensors 150 in FIG. 1A) configured to measure a pose of a portion of a wearer of the wearable robotic system. For example, three inertial measurement units (IMUs) may be daisy chained on a CAN-bus and placed on the wearer’s torso, upper arm and forearm to measure upper limb pose (shoulder elevation, horizontal flexion, internal-external rotation, and elbow flexion / extension) and torso pose. The IMUs may be placed to approximately align their y-axis parallel to the respectively joints rotation axis. The torso IMU may provide a reference orientation for arm angle estimation. Angle estimation may be performed using rotation matrices or quaternions provided by the internal Kalman filter of the IMUs. In some embodiments, calibration may include a simple static calibration of the rest pose (e.g., arms down along the torso) and in a T-pose (e.g., arm abducted at 90 degrees). Such calibration may be performed on startup to zero any angular offsets due to IMU misalignment with the human joints. While the use of IMUs, rotationmatrices and quaternions is described above, it should be appreciated that other appropriate sensors and calculation methods may be used, as the present disclosure is not limited in this regard. As a brief, non-limiting list of examples, appropriate sensors may include strain gauges, stretch sensors, Hall Effect sensors, potentiometers, or encoders either independently or together, including together with IMUs.

[0078] FIG. 10 illustrates one embodiment of a method for controlling an exemplary wearable robot or other device. In step 1002, data such as, for example, motion and / or interaction data is measured. In step 1004, the measured data may optionally be preprocessed. The preprocessing may include taking derivatives of the motion and / or interaction data. In step 1006, the measured data and / or preprocessed data may be input into a model. The model may be any appropriate model. In step 1008, the output of the model may be used to control the exemplary wearable robot or other device. In step 1010, the output from the control device as well as data from the wearable robot or other device may be used to reinforce the model. The process shown in FIG. 10 may be repeated over time.

[0079] FIG. 2A depicts a block diagram showing one embodiment of a wearable robotic system. The system may include deformation or pressure sensors 202 measuring the pressure that a user asserts on the wearable robotic device. The system may also include motion sensors (e.g., IMUs) 204 to measure motion of different portions of the wearable robotic device. The system may also include a user intent model 208 and a Preisach model 206 to model the human-actuator interaction the actuator affects joint angle(s). The system may also include an actuator air pressure controller 210 to control the air pressure to an inflatable actuator 212.

[0080] In some embodiments, the goal of the controller is to augment a user’s ability to rotate a joint and hold at a particular position. For example, in medical application where users have partial volitional control of their joints, the goal of the controller would be to change the state of the actuators to provide force / torque that benefits the user’s intended movement direction. For applications where the robot’s controller is designed to resist a person’s movement (e.g., for strength training), the goal of the controller would be to change the state of the actuators to provide force / torque that counteracts the user’s intended movement.

[0081] In some embodiments, the goal of the controller is to compensate for / counteract the torque about a joint that is caused by gravity as a person moves. The controllers may allowfor adaptation of the robot assistance to the needs of different users and, in the case of medical applications, to the different recovery statuses of the users (e.g., during stroke rehabilitation). This type of controller may assist the user by dynamically supporting the estimated gravitational load of the limb. This control strategy may use the joint angles as inputs to a feedforward term which may determine a desired actuator assistance. The feedforward term may additionally command a pressure profile to a low level controller (LLC). In some embodiments, the feedforward command output can also be scaled to deliver partial gravity compensation. The controller may assist the user in following a predefined reference trajectory. The controller may be modulated in order to increase or decrease the assistance as needed.

[0082] In some embodiments, the goal of the controller is to move a limb along a particular trajectory. This type of controller may be preferable when a wearer has a higher impairment level and would benefit from assistance to initiate and complete any movement. The therapist can easily record the desired trajectory with the wearable robot powered off (due in part to the wearable robot’s mechanical transparency), and the wearable robot may then automatically guide the wearer’s arm through this programmed trajectory. For applications such as in person therapy, the assistance provided by the robot could focus on assisting or demonstrating tasks with other joints than the one being assisted by the robot, thereby enabling a therapist to demonstrate and help with more activities within the constraints of the physical assistance the therapist themself can provide. This trajectory could also be simultaneously recorded and played back to enable teleoperation of the robot. For example, in a telerehabilitation session, a therapist could demonstrate a motion remotely and the robot would actuate to generate this same motion for the user. The desired trajectory could be generated in various ways, including determining the intent (e.g. with imaging, camera, brain computer interface) or having pre-defined trajectories for different tasks.

[0083] 1 , Intention Detection Model

[0084] Estimating human intentions is an important element in the development of wearable systems, enabling robots to effectively predict and respond to human actions. This capability is particularly important in collaborative scenarios where humans and robots work together, requiring mutual understanding and coordination. Various techniques such as sensor-based approaches, cognitive modeling, probabilistic inference, and multimodal fusion methods mayenhance the ability of robots to infer human intentions. Sensor-based approaches utilize data from sensors such as vision, auditory, tactile, and IMU sensors to analyze human behavior and infer intentions. Cognitive modeling focuses on simulating human decision-making processes, while probabilistic inference deals with uncertainty in intention estimation. Multimodal fusion combines information from multiple sensor modalities to enhance intention estimation accuracy.

[0085] Sensor-based approaches may analyze human behavior / movement and infer intentions by leveraging data from sensors such as vision, auditory, tactile, and IMU sensors. Visionbased approaches may use cameras and computer vision techniques to extract features from visual data, while IMU-based approaches may utilize motion and orientation data from IMU sensors. Interaction sensors provide valuable information about physical interactions between the actuator and human user’s body. Additionally, multimodal fusion combines data from multiple sensor modalities to improve the accuracy of intention estimation.

[0086] In one embodiment, a method and system include a sensor-based approach, utilizing IMU and deformation sensors in conjunction with a neural network for human intention estimation. The IMU sensors capture motion-related cues, while the deformation sensors capture interaction force-related cues. By training a neural network on this combined sensor data, human intentions may be accurately estimated. This sensor-based approach provides valuable insights into human behavior and enables robots to interact intuitively with humans.

[0087] One embodiment may include a comprehensive framework for developing a model that detects user intention. The first step involves outlining the data collection process, which may be facilitated through the use of a graphical user interface (GUI) and the application of predefined categories for labeling. Optionally, a feature detection mechanism may be employed to identify the most optimal features from the collected data for use in the subsequent machine learning stage. One embodiment may further include a model building process and a validation.

[0088] In one embodiment, an individualized model for each user accurately predicts the user’s intentions. Due to variations in movement patterns and body sizes among individuals, a universal model cannot be applied to all users. Therefore, initial data collected and labeled, and a unique user intention detection model may be developed for each user.

[0089] 1 , 1 Training Data Collection

[0090] One objective of an exemplary embodiment is to minimize the logistical burden on users such as those with upper limb impairment or industrial workers by minimizing the amount of time necessary to collect individualized training data for the user intention detection model. Different models may be compared based on their performance in simulations and experiments, aiming to identify the minimum amount of data required for training a model that is effective for device control while maintaining high accuracy.

[0091] The full training data set consists of trials where the user is instructed to rotate a joint to a particular angle or perform a certain movement. The movements the user is instructed to perform are classified into dynamic and static training data trials, where dynamic trials require a user to move the joint that the actuator can actively support, and static trials require the user to keep the joint that the actuator can actively support at a constant angle while performing another motion simultaneously. A GUI may provide real-time feedback on the current angle of the user’s joint and provide a ground truth for the user’s intended action at any given time during the data collection. After a trial is completed, the corresponding time series data may be segmented and labeled according to the cues given to the user and / or when the user achieved the goal as indicated on the GUI.

[0092] Trials are described below in Tables 1 and 2 for the training data collection protocol for study visits with end-users such as people who have had a stroke. Each of these tables represents a collection of trials.

[0093] Exemplary training data sets were obtained for upward and downward movement trials as well as hold movement trials. The dataset used for training a user intention detection model for a device that provides antigravity (gravity compensation) assistance at the shoulder joint. Dynamic trials were recorded using a gravity compensation controller which was defined by a Preisach model that mapped from arm elevation angle to required pressure in the inflatable actuator that is required to support the weight of the arm at this angle. The pressures that were commanded were limited by the physical constraints of the hardware, and manually limited by setting constraints on the pressure in the controller’s software. The maximum pressure commanded to the actuator which is comfortable for the user can be determined by slowly increasing the actuator air pressure to the maximum and stopping if the user reports any discomfort. The pressure which is commanded during the training can then be saturated to just below the maximum comfortable pressure. These example trials consideran inflatable actuator with a maximum pressure that can provide support up to 90 degrees of shoulder elevation, where the shoulder elevation is defined as the angle between the torso and the upper arm. In this exemplary training dataset, the number of trials considered sufficient for balancing the trade-off between the user intention detection model performance’s and the amount of time necessary for collecting this training data was -15-20 minutes (including the time spent between trials and giving the user instructions).Table 1

[0094] The dynamic trials of Table 1 consist of 13 trials characterized by a starting kinematic pose, an ending kinematic pose, and the presence or absence of an object held in the hand. These trials are selected to be representative of distribution of possible wearable sensor datathat could be observed when the user performs or attempts to perform a certain movement. For each trial, the user is cued to 1) move from Pose A to Pose B, 2) hold steady at Pose B, and 3) return back to Pose A. Each of these three phases of the movement within the trial correspond to a separate label of the intention trial.Table 2

[0095] The static trials of Table 2 consist of 11 sub-trials that are each performed with the appropriate pressure necessary to support the arm at the shoulder elevation state. The user is instructed to maintain the specified shoulder elevation state while performing the simultaneous action. In this example, each trial was performed for 3-5 seconds and only performed once. Optionally, data augmentation strategies can be applied to the collected training dataset. For example, all of the collected data can be duplicated and time warped (condensed or stretched in time) so the model is not overfit to the particular speeds at which the user moved during the training data collection.

[0096] 1.2 Graphical User Interface

[0097] FIG. 3 shows an exemplary Graphical User Interface (GUI) of one embodiment of the present invention. In one embodiment, the GUI incorporates various settings for the controller and pump, where another person (researcher, therapist, etc.) can give a user cues and press a button at each time when a cue is given to the user. The user or another person may select settings and control the controller and pump via the control menu 302. The time points when this button is pressed can then be used for segmenting and labeling the time series sensor data that is used for training the user intention detection model. In another embodiment, the GUI incorporates various settings for the controller and pump, the real-time angle feedback is displayed GUI directly to the user to enable a user to know whether or not they have reached the target angle during the dynamic trials. As users perform the intended movements, they can visually observe their joint angles displayed in real-time on the graph. This immediate feedback allows them to make necessary adjustments and have confirmation when they have successfully completed the instructed task. Moreover, the real-time feedback graph not only helps in accurate data collection but also contributes to the subsequent labeling, learning, and training phases of the user intention detection model. The GUI may additionally include graphs for real time sensor measurements, including the current PSI in the system in graph 304, and the capacitance of pressure pads in graph 306 A user or other person may select which graphs are displayed on the GUI during use. The GUI can be used as a way of automating the way instructions and cues are given to the user and indicating when the user has completed the task. The times at which these instructions and cues are given can be synced with the recorded sensor data so that the data can be automatically segmented. This labeled data set facilitates training the user intention detection model effectively. Overall, the integration of the real-time feedback graph within the GUI enhances the accuracy and reliability of data collection while facilitating the subsequent processes of data labeling and model training. It allows a separate person or the user themself to collect the training data as cues and instructions are given either manually or automatically contributing to the development of an intuitive and adaptive model of user intention.

[0098] 1,3 Data LabelingFollowing data collection, the data may be automatically processed and labeled based on the predetermined classes. During the trials, various movements are performed such as those outlined in Tables 1 and 2 above. In some embodiments, these movements are classified into three distinct states: “up,” “hold,” and “down.” The states are labeled by automatically refining the time period between synced data of when movement cues were given to the user (either from another person or through instructions provided on the GUI). Within the two time points that indicate an example movement segment as shown by the dots in FIG. 5A, the initial noticeable elevation velocity in either the upward or downward direction can be detected, marking the start of a shaded rectangle in the figure. When the elevation velocity falls below a certain threshold, it marks the end of a shaded rectangle. This labeling process is used to identify the “up” and “down” states. In contrast, if no arm elevation velocity above a certain threshold is detected, the segment is classified as a “hold” state. In this case, the entire area within the triggers is labeled as “hold.” In addition to these states, the transition periods between them may also be categorized. In some embodiments, the transition periods are named “up-to-hold,” “hold-to-up,” “hold-to-down,” and “down-to-hold.” The transitions serve as an illustrative tool to depict the relationships between the “up,” “hold,” and “down” states. In some embodiments, these transitions are used for the desired elevation angle that is provided to the Preisach model to determine the appropriate pressure setpoint utilized for the controller, as explained in detail below.

[0099] The states and transitions are shown in the directed graph of FIG. 4. States “up,” “hold,” and “down” with their transitions “up-to-hold,” “hold-to-up,” “hold-to-down,” and “down-to-hold.” The arrows looping back into the states indicate the possibility of staying in the current state. Transitions between the “up” and “down” states may involve a “hold” state in between. The model may predict user intention in each time step that it is run as a component of the controller. As a result, the transition states depicted in FIG. 4 have arrows that loop back to the initial state, indicating the possibility of staying in the current state. An example of how the upward, downward and hold trials are segmented can be seen in FIG. 5A. Every shaded rectangle in FIG. 5A corresponds to a sub-trial from Table 1. To maximize the efficiency of the collected data, the period between an upward and downward movement is designated as additional hold data.

[0100] The upward, hold and downward movements are represented by the rectangles labeled 510A-C, 512A-C and 514A-C, respectively. They refer to specific segments of data collection that involve moving the arm upwards from a position by the side of the body using a gravity compensation controller and then downwards again. The arm’s elevation was tracked using an IMU (line 502), while Pressure Pad 1, Pressure Pad 2, and Pressure Pad 3 (shown as line 504, line 506, and line 508, respectively) measured the deformation.

[0101] During Sub-trial 0, as shown in FIG. 5A, the arm underwent an upward movement, transitioning from the side of the body to 80° abduction, with elbow extended. This movement was achieved using a gravity compensation controller. In Sub-trial 1, the arm was similarly moved from the side of the body, but this time to a 40° flexion position. Subtrial 2, shows the arm movement from the side of the body to 80° in abduction with elbow flexed. In all sub-trials, a consistent pattern emerges during the “up” movement in terms of the arm elevation angle and the pressure pad deformation. As depicted in FIG. 5A, the initiation the “up” movement is marked by the left boundary of the rectangles labeled 510A- C, with rectangle 510A corresponding with sub-trial 0, 510B corresponding with sub-trial 1 and 510C corresponding with sub-trial 2. Following this initiation, the arm elevation angle remains constant for a brief period and subsequently begins to increase gradually. Prior to the movement initiation, the capacitive deformation sensors record a deformation value in pF. As the subject moves the arm upwards, the pressure pad deformation value in pF decreases because the pressure exerted by the arm on the deformation sensors, which are located between the subject’s arm and the underlying actuator, reduces. During the ongoing upward movement (within the middle portion of the rectangles 510A-C), the IMU data demonstrates a continuous increase in the arm elevation angle, while the deformation values of the three deformation sensors remain relatively constant. Towards the end of the movement (approaching the right boundary of the rectangles 510A-C), the arm elevation angle transitions from an increasing trend to a stable value, indicating the completion of the upward motion. Concurrently, a slight increase in the deformation values of the deformation sensors is observed.

[0102] The rectangles labeled 514A-C in FIG. 5A displays the collected data and labeling for the downward sub-trials mentioned in Table 1. During Sub-trial 0, the arm was moved downward from 80° abduction to the side of the body with the elbow extended.Similarly, during Sub-trial 1, the arm was moved downward from 40° abduction to a position by the side of the body using the gravity compensation controller. While in Sub-trial 2, the arm was moved downwards from 80° abduction to the side of the body with the elbow flexed. In all sub-trials, several characteristics associated with the downward movement can be observed. As illustrated in FIG. 5A, the start of the downward movement is indicated by the left boundary of the rectangles labeled 514A-C. Following this initiation, the arm’s elevation angle exhibits a decreasing trend along with an increase in the deformation values of the deformation sensors. This behavior arises from the user’s exertion of downward force, resulting in the application of pressure onto the deformation sensors. During the progression of the downward movement (within the middle segment of the rectangles labeled 514A-C), pressure continues to be applied to all three deformation sensors, although gradually diminishing over time, while the arm’s elevation angle consistently decreases. Approaching the end of the movement (towards the right boundary of the rectangles labeled 514A-C), a transition from a decreasing elevation angle to a constant value becomes evident, accompanied by a further reduction in the deformation values of the deformation sensors. By partitioning the data into the aforementioned states and solely utilizing the information shown within the shaded rectangles, the model may be used to discern the user’s intentions based on the distinctive characteristics of each state. If the model can accurately identify the relevant features of each state, and work in conjunction with the corresponding controller designed for each state, the user experience while wearing the device will undergo a substantial improvement.

[0103] FIG. 5B depicts graphs of one processed motion sensor signal and one interaction sensor signal over time. More particularly this figure depicts graphs of a shoulder elevation angle and capacitive deformation sensor signal over time as a user moves an arm up in regions 520A and 520B, down in regions 522A and 522B, and holds the arm in regions 524A and 524B .

[0104] 1 ,4 Feature ExtractionIn one embodiment, the most important features processed from the wearable sensor data are extracted from the raw data set with the aim of increasing system accuracy and decreasing computational cost due to the overuse of data. The selection process involves identifying relevant features that correlate with the target output in the system. To enable efficientcomputation and minimize the latency of the user intention detection model, the number of features from the collected data set are reduced to the most relevant ones. Statistical methods may be used for feature selection. In one embodiment, the Chi-Square method was chosen due to its good performance, particularly in multi-class data. This method has been successfully applied in various applications, such as tumor classification, network intrusion detection, text classification, and disease diagnosis. The statistical value that indicates the correlation strength of each feature individually is calculated by Chi-Square, as expressed in the following equation:X2=L(Oi - Ei)2 / Ei with Oi representing the observed values and Ei the expected values.

[0105] Because the data set contains both numerical and categorical features, the Chi- Square test may be used for feature selection. To determine the relevance of each feature with respect to the target variable, the Chi-Square statistic is computed using the equation above. In one embodiment, the target variable is time because the process conducts time series forecasting. The process may involve predicting future values or patterns based on historical time series data. The observed values Oi may be obtained from the collected log file, while the expected values Ei may be derived through linear regression analysis on the collected values. Moreover, the correlation of the significant value with the X2value may be calculated from the Chi-Square distribution table. If the signed value is below a critical point of 0.05, the feature is considered relevant, meaning that it is an important feature.

[0106] In one embodiment, the most relevant features computed from the data set are the velocity of the arm elevation; the derivatives of the deformation measured by the three deformation sensors; the shoulder horizontal flexion angle; and the velocity of the elbow movement.

[0107] Experimental evaluations were conducted to determine the optimal number of features for the data set. After selecting the most relevant features, a machine learning algorithm, specifically a neural network, may be used to further improve the model’s performance, as discussed in detail below.

[0108] 1,5 Model BuildingTo construct a model that can accurately predict user intentions based on predetermined classes, tools for embedded machine learning such as Tensor Flow Lite, PyTorch Edge, or Edge Impulse may be used. These tools can generate optimized models for running on microcontrollers or microcomputers. In a first step in the model building process, the labeled data is input into the application. For each classification of the processed data, a window size of 180 ms is selected in some embodiments. In a second step of the model building process, features that will be used as input for the neural network are determined. Optionally, these features may be calculated using the Chi-Squares method, as described above. Features could represent a summary statistic (e.g., average, maximum, minimum, skewness, etc.) over the window size, or the features could be the raw data points during this window (e.g., if the window contains 6 data points of IMU elevation angle, these are 6 unique features passed into the neural network). The number of features passed as inputs into the neural network is determined by the window size, which processed or raw sensor data are selected, and the number of summary statistics applied to these windows.

[0109] To reduce the dimensions of the high-dimensional data while maintaining its under-lying structure, methods such as Uniform Manifold Approximation and Projection (UMAP) may be used in some embodiments. This technique is based on a theoretical framework grounded in Riemannian geometry and algebraic topology, resulting in a scalable and practical algorithm that can be applied to real- world data sets. FIG. 6 illustrates an exemplary neural network architecture diagram with an input layer consisting of 36 features, hidden layer 1 with 24 neurons, hidden layer 2 with 12 neurons and an output layer with 3 features. Once the user intention detection model is developed, it may be evaluated with unseen test data to determine its classification accuracy. This initial evaluation may serve as a preliminary assessment before proceeding to test the model on the actual device. For the following validation of the user intention detection model, a test set may be recorded on the device.

[0110] 1.6 ValidationTo validate the model’s performance, a test set on the device using the gravity compensation controller was collected. This test set enables evaluation of various aspects, including the effectiveness of the chosen training data set, the suitability of the selected features used asinput to the neural network, and the impact of different parameters of the neural network. The evaluation was based on the prediction accuracy of the models across various movements.

[0111] In addition to evaluating the parameters required for constructing a reliable user intention detection model, some embodiments further include a dependable model for the interaction between the soft wearable robot (particularly the actuator) and the human user. In particular, the soft nature of the wearable robot and human user’s body can result in a hysteretic relationship between the input control command (e.g. the pressure in an inflatable actuator, the tension in a cable, etc.) and the output motion of the human user (e.g. the angle of a user’s joint. A model of this hysteresis, in conjunction with the user intention model, is utilized for controlling a wearable robotic device, such as a soft exosuit.

[0112] 2 Human-Robot Hysteresis Model

[0113] Some embodiments include a human-robot hysteresis model to ensure reliable real-time control in wearable robots. This hysteresis model addresses the challenges resulting from dissipation in both the actuator and the human body. In some embodiments, this dissipation may result from sliding friction between the yams of a textile used to construct the soft actuators. In some embodiments, hysteresis, in the context of soft actuators, arises from the non-linear behavior exhibited by the materials, including the fabric and the TPU bladder, as well as the internal dynamics of the actuators and human body. Through a combination of modeling, calibration, and control techniques, the hysteresis can be predicted in real-time so that the effect of a control action on the human-robot system can be known a priori.

[0114] 2,1 Preisach Model for Modeling Human-Robot Hysteresis

[0115] In some embodiments, the human-robot hysteresis model is a Preisach model. This hysteresis model is integrated into a control framework of the soft robotic exoskeleton, thereby improving overall performance and enabling more accurate and precise control. This model includes significant advantages such as:• Accuracy: The Preisach model accurately approximates a wide range of hysteresis loops observed in various physical systems.• Flexibility: The Preisach model is highly flexible and can represent various hysteresis behaviors.• Nonlinear Representation: The Preisach model takes the nonlinearity in hysteresis phenomena into consideration. It captures the relationship between the system’s output and both the present input and past history.• Computational Efficiency: The Preisach model can be run in real-time in the wearable robot’s controller since it is sufficiently efficient in terms of both computation and memory.

[0116] 2,1.1 Background: The Preisach model represents an output quantity (e.g., a command to an actuator), Y, as a summation of non-ideal relay operators, yaP(X) 706, where X is an input variable (e.g., measured data from a sensor or collection of sensors), as illustrated in FIG. 7. Each relay is defined by a pair of switching values (a,P) with a > p. The output transitions from an “off’ state to an “on” state at X=a and from an “on” state to an “off’ state at X=p. In some embodiments, these “on” and “off’ states are denoted with values yaP = +1 and yap = -1, respectively. All possible relays can be represented as points within the half-plane T={(a,P) | a > PJ, as illustrated in FIG. 8. The outputs of these relays are scaled by a system-specific weighting function p(a, P) 702. In some embodiments, a continuous distribution of relays and weights over T are used. In this case, the overall output is given as:Y(t) = fT p(a, P) yaP(X(t)) da dp

[0117] The relays can be divided into two time-varying regions, T- and T+, defined as follows:T-(t) = {(a, P) G T | yaP(X(t)) = - l], T+(t) = {(a, P) G T | yaP(X(t)) = + l], so that T- U T+ = T at all times. As the input increases, all relays with a < X(t) switch to the “on” state, resulting in a horizontal boundary between T- and T+. As the input decreases, all relays with P > X(t) switch to the “off’ state, resulting in a vertical boundary. As such, the overall boundary is a combination of alternating horizontal and vertical lines, as depicted in FIGs. 9A-9D. This boundary stores the memory of the system, with the vertices representing input reversals. In some embodiments, the domain, T, is limited to a minimum and maximum X value, Xmin and Xmax. In these cases, the model is said to be at negative saturation whenX=Xmin (which means that T- = T) and positive saturation when X=Xmax (which means that T+ = T).

[0118] In some embodiments, pneumatic textile soft actuators are employed, in which case the appropriate choice for X is the joint angle (which has been shown to correlate to the observed hysteresis) and the appropriate choice for Y is the pneumatic air pressure in the actuator required to support the joint at that angle.

[0119] 2,1,2 Calibration

[0120] The weighting function, p(a,P) is system specific and must be fit from experimental data. In some embodiments, when a continuous distribution of relays is employed, a related function, known as the Everett function, E(a,P), can be defined as:E(a,P) = fpa fp’a p(a’, P’) da’ dp’The output of the Preisach model can equivalently be constructed using the Everett function as follows:Y(t) = - E(Xmin,pi) + 2 k=ln E(ak,pk) - E(ak,pk+1) where (ak,pk) are the vertices of the boundary between T- and T+ . This approach prevents the need for online integration of p(a,P) and also has the advantage that E(a,P) can be easily constructed from experimental observations as:E(a,P) = (Y+(a) - Y-(a,P)) / 2Where Y+(a) is the value of Y after a trajectory beginning at X=Xmin and increasing to X=a and Y-(a,P) is the value of Y after a trajectory beginning from X=Xmin, increasing to X=a, and then decreasing to X=p. A plurality of calibration sweeps can be conducted to observe these two functions. In some embodiments, the first calibration sweep covers the entire range of motion, beginning at negative saturation, moving to positive saturation, and then returning. This sweep provides a direct measurement of Y+(a), and Y-(Xmax,P). Subsequence calibration sweeps can be utilized to obtain more observations of Y- at different a values. Since an infinite number of sweeps cannot be feasibly performed, interpolation can be used topredict intermediate a values. In some embodiments, as few as three sweeps are enough to define a good approximation of the Everett function.

[0121] In some embodiments, the history property of the Preisach representation is leveraged to accurately predict the pressure necessary to fully support the user’s joint at a specific angle. By tracking the historical information (as encoded in the boundary between T- and T+) , informed predictions may be made based on the recorded history. Upon receiving new angle data, the stored history may be reviewed and utilized to estimate the corresponding pressure. Subsequently, in the next time step, the history may be updated to incorporate the most recent data, ensuring a dynamic and adaptive prediction process. This utilization of the history property and continuous updating of the model enables accurate modeling of the behavior of the actuators.

[0122] In some embodiments, the Preisach model may be calibrated to a user of a wearable robotic system using a calibration process. In some examples, the calibration process involves the user relaxing their arm while keeping their elbow extended, during which the actuator is inflated and deflated cyclically. The shoulder elevation angle may be tracked during this inflation and deflation, yielding calibration loops in the pressure-angle plane. These calibration loops may be used to encode the Everett function such that the Preisach model is calibrated to the user. The Everett function can be written as the difference between the loading support pressure from zero to a first angle and the unloading support pressure from the first angle to a second angle. These pressures can be taken directly and interpolated from the calibration loops.

[0123] 3 Control SystemThe models representing the user’ s intended direction of movement and the human-robot hysteresis behavior may be integrated into a control algorithm as shown in FIG. 2A. This integration enables the predictions generated by these models to be utilized for guiding the robot’s executed movements and controlling the force exerted by the actuators. In some embodiments, the way this force is applied is by changing the air pressure within a pneumatic actuator. The system of FIG. 2A includes a human component 212 and a wearable exosuit or exoskeleton component 210 representing the bi-directional physical interaction between the user and the exosuit or exoskeleton, thereby defining the input and output of the overall system. The input to the system is the user’s muscle torque Tmuscie, which defines the desiredmovement that the system responds to and supports. This muscle torque Tmuscie may be combined with the torque generated by the actuator rato cause a change in the shoulder’s state. The system output includes the arm elevation angle a;u-m, which indicates the resulting position of the arm.

[0124] In some embodiments, the system may include two types of sensors: IMUs and capacitive compression sensors, as previously discussed. These sensors measure the shoulder state, denoted by ameas and Cmeas. The information obtained from these sensors may be utilized for feature processing. Additionally, the measured angles ameas may serve as an input to the Preisach model. In some embodiments, both ameas and Cmeas are vectors which represent multiple different measurements from either sensor placed in different parts of the body (e.g., IMUs placed on different segments, deformation sensors measuring different interaction forces between the human body and actuator) or multiple derived measurements resulting from the combination of one or more sensor signals. In some embodiments, these individual elements may correspond to different components of the arm and shoulder angle, including the torso angle atorso, the elevation angle aeiev, the flexion angle atiex, and the extension angle aextn. By considering these distinct elements, the system gains a comprehensive understanding of the user’s shoulder configuration.

[0125] In some embodiments, the system includes a high-level controller which plays a role in modeling both the user intention 208 and the hysteresis characteristics 206 of the human robot system. This information is utilized to determine the desired pressure Pdes for the actuator 210. The user intention model 208 may use the processed data features from the sensors and generates a movement decision vector probintent, represented as: probintent = (probhoid, probdir_i, probdir_2, ... probdir nJ, where there are n possible directions of movement that the intention model may classify. In some embodiments, probintent= (probhoid, probup, probdown) as is the case for the example of a robot designed to support the elevation angle of a user’s shoulder. The user intention model 208 informs the controller about the probabilities of the user intending to move the arm upwards, downwards, or maintaining it at a specific angle.

[0126] In some embodiments, the desired input Xdes that is passed into the Preisach Model is not equal to Xmeas, and is instead determined in part by the output of the intention model. In some embodiments, the Preisach Model 206 receives the measured angles ades as input (i.e., Xdes = ades). As an example the concept of using the intention model todetermine the Xdes within the case of Xdes = ades, if probdown were high enough that “down” is used as the current intention state, then ades could be set as ameas minus an offset adown, with the result being the Preisach Model would output a commanded actuator air pressure that is lower than what would be predicted if the current ameas were used instead. In some embodiments, such an offset (e.g., adown) is set to a constant value. In some embodiments, such an offset is a function of some other variable (e.g., the value of adown is determined by the velocity multiplied by a tunable gain). Optionally, the Preisach Model 206 may utilize the elevation angle aeiev to calculate a pressure value Pcaic. In some embodiments, the Preisach Model can be used to calculate the output Y for hypothetical scenarios, and then these hypothetical calculations could be used for later control actions. In these embodiments, the hypothetical Preisach model calculations are not used to update the internal values stored in the non-hypothetical Preisach model that is keeping track of the system’s true state. In some embodiments, the classifications output from the user intention model determine how quickly the Ydes should be changed in a given moment. For example, it may be undesirable for Ydes to be changed when probhold is high enough that it is selected as the user’s current intended state. In some embodiments, a hypothetical Preisach model may generate predictions of Ydes that are different from the currently commanded Ydes and the future Ydes are selected to slowly approach the value found by the hypothetical model. This is useful in cases where the human-robot hysteresis of the actuator can be taken advantage of to change Ydes so the robot is in a state that is more likely to be desirable for future actions that the user may perform. For example, when the user is in the hold state, a hypothetical Preisach model may identify a lower pressure than continue to support the arm at a similar elevation angle, and approaching this pressure will make it so the device will require less effort for the user to move downward in the future.

[0127] During an upward movement, for example, the desired pressure aligns with the inflation curve computed by the Preisach Model 206. For a hold movement, the hypothetical Preisach Model determines the minimum pressure required to support the arm at a given angle within an angle tolerance of ±1.5 degrees. In contrast, during a downward movement, the hypothetical Preisach Model identifies a lower pressure based on the deflation curves mapping. Smooth transitions between states are ensured by gradually adjusting the pressure.

[0128] The value of Ydes may be modulated before it is commanded to the low-level controller. In some embodiments, kinematic information that describes the human-robot system pose may affect the level commanded actuator force necessary to counteract an external force (e.g. counteract gravity), so Ydes can be scaled based on this information. In one embodiment, the angle of a user’s elbow has a significant effect on the torque about the shoulder necessary to support the arm to counteract gravity, and the kinematic relationship between the reaction force to do gravity at the shoulder joint can be considered as a way to scale Ydes. In some embodiments, Ydes can be scaled by a constant factor to control the “level of support” or “level of assistance” that the device will be controlled at. For example, if Ydes were scaled by 0.8 in the case of a device designed to counteract gravity and support shoulder elevation, then the user would receive 80% of the torque necessary to support their arm, and their muscles would need to generate the remaining 20% of the torque necessary to lift the arm. This is particularly helpful for rehabilitative applications, where it may be beneficial for the robot to slowly reduce the support afforded by the device as a user gains function.

[0129] In some embodiments, a low-level control serves as an internal pressure loop responsible for controlling the valves and maintaining the desired pressure Pdes within the actuator. To achieve this, the low-level control may employ a Proportional-Integral- Derivative (PID) controller, which continuously regulates the error Perr between the desired pressure Pdes and the actual pressure P measured by pressure sensors, which are distinct from the deformation sensors. The deformation sensors are used to measure the deformation during arm movements and represent the force exerted by the arm on the actuator, as discussed previously. The control output u produced by the PID controller effectively adjusts the valves, modulating the airflow into the actuator. As a result of this regulated airflow, the actuator undergoes inflation, generating the actuator torque ra. This low-level control is crucial for ensuring precise and responsive pressure control.

[0130] FIG. 2B depicts a block diagram showing another embodiment of a wearable robotic system. This embodiment may include machine learning data collection 216 that may be used to produce training data as input to a user intent model. FIG. 2C depicts this exemplary machine learning training data collection. The exemplary embodiment of FIG. 2B may also include exemplary limp air pressure sweeps that may be used to produce calibrationdata as input to a Preisach model of an exemplary embodiment of a wearable robotic system. FIG. 2D depicts this exemplary limp air pressure sweeps.

[0131] Example: Human-Robot Hysteresis Model Comparison

[0132] The following figures and discussion are related to an experiment testing the accuracy of human-robot hysteresis models, according to some embodiments. An experiment was conducted to compare the accuracy of different models, including the Preisach model. The participants of the experiment were outfitted with a soft wearable robot, which included a shirt with a woven textile pneumatic soft actuator integrated into the axilla that pushes upward on the arm when inflated, such as that described above with reference to FIGs. 1A-B.

[0133] FIG. 11 A depicts an example process for adapting a model to a particular participant, according to one embodiment. The models were personalized to each participant at the start of each test. During personalization, the participant was instructed to relax their arm to the best of their ability while keeping their elbow extended. The actuator was then inflated and deflated cyclically with decreasing amplitude, as shown in FIG. 11 A, while the shoulder elevation angle was tracked. This yielded a set of calibration loops in the pressureangle plane, which were used to calibrate the models. An example calibration loop is shown in FIG. 1 IB. The model was then adapted to a participant as described herein, for example by using a GUI such as that of FIG. 3.

[0134] The models tested on the participants included a linear model, a non-linear model, a binary model, and a Preisach model. Examples of these models are shown in FIGs. 12A-D.

[0135] FIG. 12A illustrates a plot of an example linear model and a calibration loop in the pressure-angle plane. The linear model was fit through the endpoints of the calibration loops, to guarantee that the full pressure range was realized, ensuring a consistent range of arm motion for comparison with the other models.

[0136] FIG. 12B illustrates a plot of an example nonlinear model and a calibration loop in the pressure-angle plane. The nonlinear model was personalized such that it approximates the response with the average of the outer loading / unloading curves.

[0137] FIG. 12C illustrates a plot of an example binary model and a calibration loop in the pressure-angle plane. The binary model was personalized such that the system response was assumed to jump between the outer loading and unloading curves instantaneouslydepending on the current direction of arm motion. As seen, the binary model aligns with and covers the shape of the calibration loop.

[0138] FIG. 12D illustrates a plot of an example Preisach model and a calibration loop in the pressure-angle plane. The Preisach model was configured by encoding the Everett function using the pressure loops. The Everett function was written as the difference of Pf(0a) the primary loading support pressure when moving from 0min(at zero pressure) to 0aand pu(0a, Op) was the primary unloading support pressure when subsequently moving down to a new Op, giving:Pi>(0a) was directly observed from the inflation data during the first calibration loop, and pu(0a, Op) was interpolated between the experimentally observed unloading curves.

[0139] FIG. 13 depicts an example process for testing the accuracy of hysteresis models for patient activities, according to one embodiment. The process began with model personalization, which was performed as described above with reference to FIGs. 11A-B. The process additionally included breaks, in which a simple IMU de-drifting procedure was performed, where the estimated shoulder angle in the horizontal plane was tared when the arm was resting at the side of the body. This reduced the effect of yaw drift of the IMUs which can introduce error for estimating kinematic angles.

[0140] Following the model personalization for the participants, open-loop trajectory tracking evaluations were performed for each of the four system models in a randomized order. Since the hysteretic state of the device was unknown after each break, a single maximal hysteretic loop trajectory (shown as a star) was performed before each trial to clear any existing system memory, and consequently the data from these portions was not analyzed. Subsequent maximum hysteric loop trajectories were included in the analysis and used to assess the consistency of the participant’s arm relaxation, since the resulting motion should be identical if the biological torque of the arm is consistent.

[0141] Following the experimental procedure, statistical analyses were performed to assess differences in angle root mean square error (RMSE) and angular velocity RMSE between the four system models. A repeated measures design was utilized to account fordifferences in participants' ability to consistently relax their muscles, resulting in lower or higher RMSE values across all four model conditions.

[0142] First, the normality of the data was assessed using the Shapiro-Wilk test (or the Shapiro-Francia test for platykurtic samples). For normally distributed data, the main effect was then evaluated with a repeated measures ANOVA test (with the Greenhouse- Geiger correction for non- sphericity, if desired, as assessed with Mauchly's sphericity test). For non-normally distributed data, the Friedman test was used to assess the main effect. In each case, if the main effect was found to be significant, post-hoc pairwise comparison tests were performed (with the Bonferroni correction) to assess individual differences between each of the four model conditions.

[0143] FIGs. 14A-D depict the performance of the system models by comparing the measured and commanded angles. FIG. 14A depicts the commanded and measured angles from the testing of the linear model, according to one embodiment. FIG. 14B depicts the commanded and measured angles from the testing of the nonlinear model, according to one embodiment. FIG. 14C depicts the commanded and measured angles from the testing of the binary model, according to one embodiment. FIG. 14D depicts the commanded and measured angles from the testing of the Preisach model, according to one embodiment.

[0144] The data of FIGs. 14A-D was collected using an open loop trajectory tracking procedure, such as that described above with respect to FIGs. 11A-13, performed on an articulated mannequin. The articulated mannequin allows for the isolation of the effect of the robot on the overall system because it does not include biological factors, which are present in human subjects, such as non-volitional muscle activation.

[0145] The data in the top plots of FIGs. 14A-D depicts the commanded and measured angles across a 30 second time period of the test performed on the mannequin. The dark lines represent the commanded angles and the light lines represent the measured angles. The data in the bottom plots of FIGs. 14A-D depicts the commanded and measured angles across the full range of angles tested. The dark lines represent the commanded angles and the light lines represent the measured angles. The measured and commanded values were observed to align better as the complexity of the model increased. Though the angle RMSE values were similar between the binary and Preisach models, the binary response was observed to be more jumpy as seen in the associated plot in the top plots, with the armdropping or rising as much as ±17° over 0.5 second intervals. This jumpiness of the binary model was reflected in the angular velocity RMSE values in Table 3, where the binary model performed significantly worse than all other models.

[0146] Table 3 below includes the angle RMSE values and angular velocity RMSE values measured during the testing of the mannequin. As seen in the table, the Preisach model performed best overall, with the lowest error for both the angle RMSE and the angular velocity RMSE.Table 3

[0147] FIGs. 15A-B depicts the performance of the system models during testing performed with human participants. The testing was performed on 11 participants according to the process described above, with respect to FIGs. 11 A- 12. In FIGs. 15A-B, individual participant results are shown at the left, with group-level distributions at the right. FIG. 15A depicts the angle tracking accuracy, as the angle RMSE, for the four models. FIG. 15B depicts the angular velocity tracking accuracy as the angular velocity RMSE.

[0148] As with the mannequin data described with reference to FIGs. 14A-D, a steady improvement in angle RMSE was observed as model complexity increased. Likewise, the binary model exhibited much higher angular velocity errors than the other models. It was observed that there was significant interparticipant variability, where some participants had consistently higher or lower RMSE values than other participants.

[0149] From the analysis of the human participants, the main effects were found to be statistically significant for both angle RMSE (p=0.013, Friedman test) and angular velocity RMSE (p=4.3xlOA(-5), repeated measures ANOVA test with Greenhouse-Geiger correction). This indicates that the Preisach model outperformed the three other system models in its ability to predict the support pressure needed to achieve the commanded angular trajectory.

[0150] Though the Preisach model was modeled during limp arm, open-loop trajectory tracking tasks, this modeling approach can also be readily applied toward improving physical human-robot interactions when the human user volitionally engages their muscles to initiate motion. When both the human and robot move simultaneously, distinguishing their respective contributions to the overall movement is a challenge with soft wearable robots, since it is difficult to directly measure the interaction forces between them on the body. As such, it is desirable to have a reliable system model that accurately estimates the amount of assistance required to support the user, accounting for the full complexity of the human-robot system. The Preisach model enabled the robot's controller to accurately quantify the effect of the robot's actuation on the output movement, thereby allowing for a decoupling of the contributions to movement from the human and robot.

[0151] Example: Personalized real-time control of wearable assistive robotics

[0152] The following figures and discussion are directed to an evaluation of the effectiveness of a control strategy for a soft wearable shoulder robot in detecting the intended direction of shoulder movement and providing the necessary level of support for individuals either post-stroke or living with ALS during isolated joint movement and functional tasks, according to one embodiment. The control strategy combined 1) a machine learning model to decode a user’s motion intention from IMU and custom compression sensors and 2) the Preisach model to capture hysteretic relationship between the arm angle and actuator assistance, such as described herein.

[0153] The intention detection model was structured as a fully connected neural network with two dense layers (20, 10 neurons) and ReLU as the activation function. Regularization was applied to both layers to prevent overfitting. Specifically, LI regularization was used with a regularization strength of 0.00001 to penalize large weights in the neural network. A softmax layer was used to output confidence thresholds for the predictions in the three classes (UP, HOLD, and DOWN). The architecture for the model was selected to ensure that sufficient patterns in data were captured while keeping the time to perform inference sufficiently fast at every control interaction to not cause delays in controller response. While a particular model architecture was used in this example, in some embodiments, other model architectures may be used for the intention detection model as well.

[0154] Joint velocities and changes in deformation of the soft deformation sensors attached to the human-robot interaction surfaces were selected as input features for the intention detection model since these signals captured residual movement while being minimally sensitive to the sensors’ exact placement after redonning the robot. In total, six input signals were selected to calculate the features to train the intention detection models: 1) shoulder elevation / depression velocity, 2) shoulder horizontal abduction / adduction velocity, 3) elbow flexion / extension velocity, 4) derivative of torso soft deformation sensor signal, 5) derivative of in medial upper arm the soft deformation sensor signal, and 6) derivative of in lateral upper arm the soft deformation sensor signal. In some embodiments, a greater or lesser number of signals may be used.

[0155] The signals underwent processing before being input into the intention detection model. The signals were first smoothened by applying a short moving average filter with a 20 ms window. To minimize the effect of noise in the sensor data and learn patterns from temporal trends, data was included from the current timestep and historical data from the preceding 50ms (5 control iterations at 100Hz). In total, the intention detection models included 36 input features that were used to perform each time inference: the 6 current features and 5x6 historical features. This window size and the number of features were selected to balance trade-offs related to the sensitivity of the model to noise, the latency in performing the inference, and the amount of training data required.

[0156] The intention detection model was adapted to each of the participants using a guided training data collection session. During the training process, participants were instructed to perform a variety of upper limb movements to collect IMU and deformation sensor data labeled based on the instructed shoulder elevation: lifting the arm (UP), maintaining the arm elevation (HOLD), and lowering the arm (DOWN). This data was recorded and labeled, and used to train the intention detection model to each participant. The training data collected from the movements were 1) joint angular velocities estimated in realtime from the IMUs attached to the torso, upper arm, and forearm and 2) changes in compression at the human-robot interaction surfaces estimated in real-time from the soft compression sensors placed on the side of the torso and underside of the upper arm. Training data for each participant’ s personalized model intention detection model was collected by performing the aforementioned movements, with the dynamic shoulder elevation movementsperformed while the robot was controlled purely based on gravity compensation defined by the hysteresis model-with the output pressure scaled to ensure the participant received sufficient support and could still push down-and quasistatic shoulder elevation movements performed while the actuator was set to a constant pressure necessary to maintain the shoulder elevation. The training data was windowed with a stride of 10ms so that each window contained the 36 features that were input to the neural network. Tables 4 and 5 below show the movements used to train the intention detection model for the participants:Table 4Table 5

[0157] The intention detection model accuracy was evaluated during an evaluation session where the intention detection model’s real-time predictions (100Hz) were used for the robot’s control as participants performed joint individuation (n=7) and simulated functional tasks (n=6, excluding one participant with ALS who could not complete the tasks due to fatigue). FIG. 16 depicts an example process for the analysis of the intention detection model.

[0158] FIGs. 16A-B depicts the movements performed during the evaluation. The movements were classified as UP, DOWN or HOLD, depending on the direction of the movement. FIG. 16A depicts the UP and DOWN movements performed during an evaluation. The UP movements included shoulder flexion, lifting a spoon, lifting a bottle, shoulder abduction, lifting to brush teeth and lifting to comb. The DOWN movements included shoulder extension, lowering a spoon, lowering a bottle, shoulder adduction, lowering to brush teeth, and lowering to comb.

[0159] FIG. 16B depicts the HOLD movements performed during an evaluation. During the HOLD movements, the participants were instructed to maintain their shoulders at a 90° angle. The HOLD movements included horizontal shoulder adduction and abduction, shoulder external and internal rotation, elbow flexion and extension, forearm supination and pronation, and wrist extension and flexion. In the HOLD class, intention detection model performance was analyzed for only shoulder elevated joint individuation movements, since these are more challenging to classify and more akin to functional use than static holding at a shoulder elevation.

[0160] The intention detection model predictions of the intended shoulder elevation direction were compared to the instructed upper limb movement cues given to participants that were time- segmented and labeled from videos. These labels were compared to the plurality intention detection model class predicted within each movement period to calculate the accuracy. This allowed for separation between movements that were predominately classified correctly at each controller iteration versus those with longer periods of misclassified timesteps which have a greater impact on the control. This definition of accuracy does not, however, imply that the intention was not identified at all during the movement. Across all participants and tasks, the participants’ intention was identified for a sufficient period of time to trigger actuation.

[0161] On average, the accuracy of the intention detection models was 93.5+1.6% across the seven participants, corresponding to 92.4+9.9%, 91.9+6.9%, and 96.3+4.9% for UP, HOLD, and DOWN, respectively this is shown in FIG. 17. In the context of the intention detection models’ use for robot control, misclassified movements manifested as either delayed or discontinuous actuation for UP and DOWN or perturbations in the robot-provided support during HOLD.

[0162] The personalized intention detection models’ responsiveness was assessed in terms of time and angle to the first identification UP and DOWN intention. The time until the first correct prediction of UP and DOWN was 253 ± 222 ms and 203 ± 128 ms, respectively, from the video labeled initiation time. These times correspond to small changes in shoulder elevation (estimated from optical motion capture) before intention was first predicted by the intention detection models: 3.4 ± 2.2° shoulder elevation and 0.3 ± 0.9° shoulder depression for UP and DOWN, respectively. Given that the average change in shoulder elevation was 10.3 ± 3.9° within each cued HOLD movement, the small displacements required to trigger UP / DOWN predictions imply the intention detection models could reliably distinguish between wearable sensor signal changes caused by volitional changes in shoulder elevation as compared to changes caused by involuntary movements.

[0163] FIGs. 18A-C depict the control objectives during the UP, HOLD, and DOWN intention states, selected to highlight how the controller addresses hysteresis. FIG. 18A depicts the control objectives during the UP intention states on an angle-pressure plot. FIG. 18B depicts the control objectives during the HOLD intention states on an angle-pressure plot. FIG. 18C depicts the control objectives during the DOWN intention states on an anglepressure plot. Arrows show the pressure trajectory the controller would command for a given initial state (white circle) to the target state (dark circle) given a particular intention state prediction and Preisach calibration curves. In the UP and DOWN states, the measured elevation angle was (optionally) offset by a constant angle in the direction of movement, allowing the robot to anticipate and promote a user’s future movement in that direction. In the HOLD state, the controller’s objective was to maintain antigravity support while maximizing the transparency of the system (i.e., at the minimum actuator pressure). Given the typical shape of the hysteretic curves for the system, the actuator pressure could be reduced significantly in the HOLD state while still maintaining support of the arm, as shown FIG. 18B.

[0164] The hysteresis model’s ability to predict the minimum pressure for the commanded control of the system was analyzed by evaluating the change in angle during the steady elevation at 90° portion of the isolated shoulder flexion and abduction tasks (n=7). A representative figure of the controller response during the isolated shoulder abduction task is shown in Fig. 19. FIG. 19 depicts a representative participant post-stroke performing isolatedshoulder abduction / adduction. On the left, a time series plot of shoulder elevation 1901 is shown, with the mean change from baseline compression across the three compression sensors shown as 1902. The above video frames depict the actions of the users from time synced video. The intention detection model real-time predicted intention states 1903 is shown compared to the ground truth 1904. Finally, FIG. 19 depicts the measured 1905 and commanded 1906 actuator pressure during the representative action. The right side of FIG. 19 depicts the elevation angle vs commanded pressure response with labels corresponding to the video frames.

[0165] FIG. 20 depicts a representative participant living with ALS performing a simulated eating task. On the left, a time series plot of shoulder elevation 2001, and mean change from baseline compression across the three compression sensors 2002 are shown with callouts to frames from time synced video corresponding to the phases of the user tasks. The intention detection model real-time predicted intention states 2003 are shown compared to the ground truth 2004 for the states. Additionally, the measured 2005 and commanded 2006 actuator pressure are shown. On the right, the elevation angle versus commanded pressure response with labels corresponding to the video frames is depicted.

[0166] FIGs. 21A-B depict shoulder elevated intemal / external rotation representative time series figures for one participant. FIG. 21A depicts the shoulder elevated intemal / external rotation and actuator pressure time series figures for a participant with the controller on. As shown, with the controller on, misclassifications in the intention detection model during external rotation (confused as DOWN) led to pressure drops. However, in all the cases, the controller identified HOLD as the correct intention and slowly inflated up to the pressure needed to support the arm. FIG. 21B illustrates the shoulder elevated intemal / external rotation time series for a participant with the controller off. With the controller off, the participant’s movement was much less smooth and below the target shoulder elevation angle.

[0167] The controller’s ability to enable transparent arm lowering was evaluated by comparing the multimodal ML controller with the intention detection model to a baseline controller that performs pure kinematic-based gravity compensation with the user-specific hysteresis model (i.e., there is no information incorporated from an intention detection model in the baseline controller). A representative comparison of these conditions during onerepetition of isolated shoulder flexion / extension is shown in FIG. 22, which depicts representative time series (n = 1) results showing the differences in force (estimated from compression sensors), actuator air pressure, and shoulder elevation between a baseline gravity compensation controller and the multimodal controller incorporating the intention detection model during isolated shoulder flexion and abduction tasks.

[0168] There was a notable reduction in the amount of force required for the participants to lower their arms when using the multimodal ML controller with the intention detection model. From the compression sensor on the torso, the multimodal ML controller is estimated to reduce the peak lowering force during DOWN, despite there being no significant difference in initial angle at the start of DOWN.

[0169] While the present teachings have been described in conjunction with various embodiments and examples, it is not intended that the present teachings be limited to such embodiments or examples. On the contrary, the present teachings encompass various alternatives, modifications, and equivalents, as will be appreciated by those of skill in the art. Accordingly, the foregoing description and drawings are by way of example only.

[0170] The above-described embodiments of the technology described herein can be implemented in any of numerous ways. For example, the embodiments may be implemented using hardware, software or a combination thereof. When implemented in software, the software code can be executed on any suitable processor or collection of processors, whether provided in a single computing device or distributed among multiple computing devices. Such processors may be implemented as integrated circuits, with one or more processors in an integrated circuit component, including commercially available integrated circuit components known in the art by names such as CPU chips, GPU chips, microprocessor, microcontroller, or co-processor. Alternatively, a processor may be implemented in custom circuitry, such as an ASIC, or semicustom circuitry resulting from configuring a programmable logic device. As yet a further alternative, a processor may be a portion of a larger circuit or semiconductor device, whether commercially available, semi-custom or custom. As a specific example, some commercially available microprocessors have multiple cores such that one or a subset of those cores may constitute a processor. Though, a processor may be implemented using circuitry in any suitable format.

[0171] Further, it should be appreciated that a computing device may be embodied in any of a number of forms, such as a rack-mounted computer, a desktop computer, a laptop computer, or a tablet computer. Additionally, a computing device may be embedded in a device not generally regarded as a computing device but with suitable processing capabilities, including a Personal Digital Assistant (PDA), a smart phone, tablet, or any other suitable portable or fixed electronic device.

[0172] Also, a computing device may have one or more input and output devices. These devices can be used, among other things, to present a user interface. Examples of output devices that can be used to provide a user interface include display screens for visual presentation of output and speakers or other sound generating devices for audible presentation of output. Examples of input devices that can be used for a user interface include keyboards, individual buttons, and pointing devices, such as mice, touch pads, and digitizing tablets. As another example, a computing device may receive input information through speech recognition or in other audible format.

[0173] Such computing devices may be interconnected by one or more networks in any suitable form, including as a local area network or a wide area network, such as an enterprise network or the Internet. Such networks may be based on any suitable technology and may operate according to any suitable protocol and may include wireless networks, wired networks or fiber optic networks.

[0174] Also, the various methods or processes outlined herein may be coded as software that is executable on one or more processors that employ any one of a variety of operating systems or platforms. Additionally, such software may be written using any of a number of suitable programming languages and / or programming or scripting tools, and also may be compiled as executable machine language code or intermediate code that is executed on a framework or virtual machine.

[0175] In this respect, the embodiments described herein may be embodied as a computer readable storage medium (or multiple computer readable media) (e.g., a computer memory, one or more floppy discs, compact discs (CD), optical discs, digital video disks (DVD), magnetic tapes, flash memories, RAM, ROM, EEPROM, circuit configurations in Field Programmable Gate Arrays or other semiconductor devices, or other tangible computer storage medium) encoded with one or more programs that, when executed on one or morecomputers or other processors, perform methods that implement the various embodiments discussed above. As is apparent from the foregoing examples, a computer readable storage medium may retain information for a sufficient time to provide computer-executable instructions in a non-transitory form. Such a computer readable storage medium or media can be transportable, such that the program or programs stored thereon can be loaded onto one or more different computing devices or other processors to implement various aspects of the present disclosure as discussed above. As used herein, the term "computer-readable storage medium" encompasses only a non-transitory computer-readable medium that can be considered to be a manufacture (i.e., article of manufacture) or a machine. Alternatively or additionally, the disclosure may be embodied as a computer readable medium other than a computer-readable storage medium, such as a propagating signal.

[0176] The terms “program” or “software” are used herein in a generic sense to refer to any type of computer code or set of computer-executable instructions that can be employed to program a computing device or other processor to implement various aspects of the present disclosure as discussed above. Additionally, it should be appreciated that according to one aspect of this embodiment, one or more computer programs that when executed perform methods of the present disclosure need not reside on a single computing device or processor, but may be distributed in a modular fashion amongst a number of different computers or processors to implement various aspects of the present disclosure.

[0177] Computer-executable instructions may be in many forms, such as program modules, executed by one or more computers or other devices. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Typically the functionality of the program modules may be combined or distributed as desired in various embodiments.

[0178] The embodiments described herein may be embodied as a method, of which an example has been provided. The acts performed as part of the method may be ordered in any suitable way. Accordingly, embodiments may be constructed in which acts are performed in an order different than illustrated, which may include performing some acts simultaneously, even though shown as sequential acts in illustrative embodiments.

[0179] Further, some actions are described as taken by a “user.” It should be appreciated that a “user” need not be a single individual, and that in some embodiments,actions attributable to a “user” may be performed by a team of individuals and / or an individual in combination with computer-assisted tools or other mechanisms.

Claims

CLAIMSWhat is claimed is:

1. A method for controlling a wearable robot comprising one or more actuators on a user, the method comprising: reading motion data from one or more motion sensors; reading interaction data from one or more interaction sensors; and determining at least one control of the at least one actuator of the wearable robot from the motion data, the interaction data, and a model of user intent to assist motion of at least one joint of the user.

2. The method of claim 1, further comprising controlling the at least one actuator of the robot based at least in part on the determined at least one control.

3. The method of claim 1 further comprising: constructing the model of user intent; determining an intent of the user from multimodal wearable sensor data and dataset of training data; and constructing a model for the at least one actuator; wherein the at least one control of the at least one actuator is determined from the determined intent of the user and the model for the at least one actuator, of the at least one actuator.

4. The method of claim 3 wherein constructing the model of user intent comprises: generating at least one training data collection comprising: performing a plurality of trials comprising a plurality of trials in two or more directions and a plurality of static trials where the user keeps the joint steady while performing another action in parallel.

5. The method of claim 3 further comprising: updating the model of user intent while the wearable robot is used.

6. The method of claim 3 wherein the model of user intent is a statistical model.

7. The method of claim 4 further comprising: extracting from the at least one data collection a plurality of features relating to movement of one or more limbs of the user; and constructing the model of user intention using the plurality of features as inputs.

8. The method of claim 4 further comprising: extracting from the at least one data collection a plurality of features comprising one or more of: joint velocities and change in interaction forces; and constructing a neural network model of user intention using the plurality of features as inputs, a plurality of classes as outputs, and a plurality hidden layers.

9. The method of claim 3 wherein constructing the model for the at least one actuator comprises: constructing a Preisach model comprising a plurality or continuous distribution of relays having a corresponding plurality or continuous distribution of weights; determining values for the plurality or continuous distribution of weights comprising: performing one or more calibration sweeps measuring a plurality of pressures corresponding to a plurality of angles.

10. The method of claim 3 wherein the intent of the user comprises a probability of the user intending to extend a limb, a probability of the user intending to flex a limb, and a probability of the user intending to extend the limb, and a probability of the user intending to hold the limb stationary.

11. The method of claim 3 wherein the intent of the user comprises a probability of the user intending to move the arm upward, a probability of the user intending to move an arm downward, and a probability of the user intending to hold the arm at a desired angle.

12. The method of claim 9 wherein the determining at least one control to the at least one actuator comprises: determining a command to the actuator corresponding to an input wearable sensor data value using the Preisach model.

13. The method of claim 1 further comprising: controlling one or more valves to modulate a flow of air to the actuator to regulate an error between an actual pressure value and a desired pressure value.

14. A wearable robot system, the system comprising: one or more actuators configured to apply a force to at least one joint of a user wearing the wearable robot system; one or more motion sensors configured to measure a movement of a portion of the user of the robot system; one or more force sensors configured to measure force data applied to the portion of a wearer of the robot system; a processor operatively coupled to the one or more actuator, the one or more motion sensors, and the one or more force sensors, the processor configured to execute the method of any one of claims 1-4.

15. The wearable robot system of claim 14 wherein at least one of the motion sensors is at least one selected from an inertial measurement unit16. The wearable robot system of claim 14 wherein the one or more actuators comprise at least one actuator configured to tension a tether.

17. The wearable robot system of claim 14 wherein the one or more actuators comprise at least one pneumatic actuator.

18. The wearable robot system of claim 17 wherein the at least one pneumatic actuator is configured to be positioned in an axilla of the wearer.

19. The wearable robot system of claim 17 wherein at least a first one of the force sensors is positioned between a torso of the wearer and the pneumatic actuator.

20. The wearable robot system of claim 17 wherein at least a second one of the force sensors is positioned between an arm of the wearer and the pneumatic actuator.

21. The wearable robot system of claim 14 wherein the robot is an exosuit and / or an exo skeleton.

22. The wearable robot system of claim 14 wherein the assisted force is applied by at least one of a pressure, torque, and / or tensile force of the at least one actuator.

23. The wearable robot system of claim 19 wherein at least one of the interaction sensors is a capacitive deformation sensor configured to measure a proxy of the interaction force between at least one actuator and the human body, and tension sensor configured to measure a tension of the at least one actuator.

24. The method or wearable robot system of any one of the preceding claims, wherein the interaction data includes at least one selected from a tensile force, a pressure, a deformation, and a torque.

Images