JP7432430B2 - Movement support device and movement support method - Google Patents

Description

TECHNICAL FIELD The present invention relates to a technique for a motion support device that supports a user's motions, and more specifically to a power assist robot that uses an actuator device to support motions performed by a user.

With the declining birthrate and aging society becoming a problem in many countries including Japan, demand for assistive devices that apply robotics technology is increasing. On the other hand, robots that can balance and walk are being developed. For example, there is a robot that can optimally distribute the acting force necessary for movement to any plurality of contact points in space and generate torque at each joint in the same way as a human (see Patent Document 1).
With recent advances in robotics, it is anticipated that wearable robots, such as exoskeletons, will be able to physically interact with and assist in human movement during activities.

For example, hand exoskeleton robots and upper and lower body exoskeleton robots have been studied (see Non-Patent Documents 1, 2, 3, and 4).

Techniques using so-called "model predictive control" have also been reported for controlling such exoskeletal robots (see Patent Document 2).

On the other hand, for these applications, the use of body surface electromyography (EMG) can be an approach that allows for intuitive control of the robot by evaluating the user's motor intention (Non-patent Document 5).

FIG. 22 is a conceptual diagram showing a comparison between conventional control methods, ie, a "regression method" and a "discrimination method."

In order to control assistive devices such as exoskeleton robots, regression approaches have often been used to evaluate the user's motor intentions.

In these approaches, motor intention has been evaluated by finding the relationship between EMG signals and joint torques, with linear or non-linear models (see, for example, Non-Patent Document 6).

In this way, the robot can be controlled by continuously monitored muscle activity.

Another frequently used approach to assessing a user's motor intentions uses discriminant methods.

For example, support vector machines (SVM) or linear discriminant analysis (LDA) have been employed to discriminate sensed and measured information from human users (see [7, 8]).

The determination result is used to initiate a pre-designed control output pattern associated with the evaluated class label.

On the other hand, as mentioned above, although there are cases where multiple EMG channels are used to obtain various user's own motor intentions, for example, in everyday industrial application situations, it is difficult to communicate with others as assist control. It is necessary to use it for interactions with the surrounding environment, such as in communication with cars and autonomous driving. In this case, it is necessary to detect the motor intention of the interacting partner.

Note that, for example, a method called inverse reinforcement learning (IRL) is used as a method for modeling the selection of movement routes for objects such as people and cars (for example, Patent Document 3, Patent Document 4). IRL assumes that there is a latent cost in the routes and places people travel, which is an example of an object, and that people are more likely to choose a route with a lower sum of costs over the entire route. It is modeled as follows.

WO2007/139135 publication

JP 2019-25104 Publication

Japanese Patent Application Publication No. 2018-197892

JP2019-82755A

J. Ngeo, T. Tamei, T. Shibata, M. Orlando, L. Behera, A. Saxena, and A. Dutta, “Control of an optimal finger exoskeleton based on continuous joint angle estimation from emg signals,”pp. 338 -341, July 2013.

C. J. Gearhart, B. Varone, M. H. Stella, and B. F. BuSha,“An effective 3-fingered augmenting exoskeleton for the human hand,” in 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC),Aug 2016, pp. 590-593.

J. Furukawa, T. Noda, T. Teramae, and J. Morimoto, “Estimating joint movements from observed emg signals with multiple electrodes under sensor failure situations toward safe assistive robot control,” in IEEE International Conference on Robotics and Automation (ICRA) , May 2015, pp. 4985-4991.

A. J. Young and D. P. Ferris, “State of the art and future directions for lower limb robotic exoskeleton,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 25, no. 2, pp. 171-182, Feb 2017.

K. Nagata and K. Magatani, “Basic study on combined motion estimation using multichannel surface emg signals,” 33rd Annual International Conference of the IEEE EMBS Boston,Massachusetts USA, August 30 - September 3, pp. 7865-7868,2011.

J. Furukawa, T. Noda, T. Teramae, and J. Morimoto, “Human movement modeling to detect biosignal sensor failures for myoelectric assistive robot control,” IEEE Transactions on Robotics, vol. PP, no. 99, pp. 1- 12, 2017.

M. A. Oskoei and H. Hu, “Support vector machine-based classification scheme for myoelectric control applied to upper limb,” IEEE Transactions on Biomedical Engineering, vol. 55, no. 8, pp. 1956-1965, Aug 2008.

D. Zhang, X. Zhao, J. Han, and Y. Zhao, “A comparative study on pca and lda based emg pattern recognition for anthropomorphic robotic hand,” in 2014 IEEE International Conference on Robotics and Automation (ICRA), May 2014 , pp. 4850-4855.

However, since discriminant methods require that all assist controls related to a target task be calculated in advance, such an approach can only be used for a limited number of individual target tasks. be.

On the other hand, in regression approaches, in order to perform a given task based on a simple EMG-based control strategy, a human user needs to generate a highly tuned EMG profile himself in advance. be. Moreover, the calculation load on the system side also increases.

Furthermore, although it is assumed that it will be necessary to assist the user with cooperative exercise/work between the user and a partner, rather than the user's exercise alone, conventionally, there has been insufficient assistance control for such cooperative exercise. The situation is such that it cannot be said that it is being considered.

The present invention has been made to solve the above-mentioned problems, and its purpose is to control an exoskeleton robot that suits the user's purpose of operation while reducing the burden on the user in advance. An object of the present invention is to provide a motion support device and a motion support method that make it possible.

Another object of the present invention is to make it possible to control an exoskeleton robot that suits the purpose of movement of the user and others while reducing the burden on the user in advance during cooperative movements and tasks between the user and others. An object of the present invention is to provide a motion support device and a motion support method.

According to one aspect of the present invention, there is provided a motion support device for assisting movement of a user's joints, the actuator means being attached to a plurality of joints of the user receiving assistance and generating assist torque to the joints. a sensor means for detecting the state of the active joint; and a biosignal detection means for detecting a biosignal generated based on a command from the user's center regarding the movement of the user's joint. , a control unit that generates a control signal for controlling the assist torque generated by the active joint, and the control unit operates to estimate the purpose of the user's movement based on the signals from the sensor means and the biological signal detection means. The motion purpose estimating means includes a purpose estimating means, and the motion purpose estimating means determines a point before the start of the assisting motion, with a point at which the angular velocity of a predetermined active joint detected by the sensor means exceeds a predetermined value as a starting point of the assisting motion. a storage means for pre-storing information for estimating a user's movement objective by learning feature quantities for a predetermined period and specifying a plurality of control strategies for optimal control corresponding to each of the plurality of discrete movement objectives; A mixing means for linearly combining a plurality of control strategies with corresponding weights according to the estimated user's operational purpose to derive a mixed control law corresponding to the estimated operational purpose; The apparatus further includes control signal generation means for generating a signal for controlling the assist torque accordingly.

Preferably, the motion purpose estimating means generates a signal in a low-dimensional feature space using a partial least squares algorithm from the measured signals from the sensor means and the biological signal detection means, and assigns a motion purpose label. The generating and mixing means linearly combines a plurality of control strategies that are previously associated with the operational objective label using corresponding weights.

Preferably, the motion purpose label estimated by the motion purpose estimating means is one that interpolates a plurality of discrete motion purposes.

Preferably, the mixing means derives a control law corresponding to the estimated operating objective by a linear Bellman combination method.

Preferably the actuator means includes a pneumatic air muscle.

According to another aspect of the invention, there is provided a motion support method for assisting joint movements of a user with an exoskeleton robot, the exoskeleton robot being attached to a plurality of joints of the user receiving assistance; The active joint includes an actuator means for generating an assist torque to the joint, and a storage means for storing in advance information specifying a plurality of control strategies for optimal control corresponding to a plurality of discrete movement objectives. A signal from a sensor means for detecting the state of an active joint, and a biosignal detection means for detecting a biosignal generated based on a command from the user's center regarding the movement of the user's joint. the step of estimating the purpose of the user's movement based on the signal of The step includes a step of estimating the purpose of the user's movement by learning the feature values for a predetermined period before the start of the movement to be assisted, using a point exceeding the point as the movement starting point. linearly combining the control strategies with corresponding weights to derive a mixed control law corresponding to the estimated motion objective; and controlling the assist torque generated by the active joint according to the derived mixed control law. and generating a control signal.

According to still another aspect of the present invention, there is provided a motion support device for assisting joint motion of a user in a coordinated exercise with a partner, the motion support device being attached to the joint of the user receiving the assistance, and providing support to the joint of the user. An active joint having an actuator means for generating an assist torque, a sensor means for detecting the state of the active joint, and detecting a biologically derived signal generated based on a command from the user's center regarding the movement of the user's joint. an image sensor that images the partner, and a control unit that generates a control signal for controlling the assist torque generated by the active joint, and the control unit is configured to detect a biological signal based on the signal from the image sensor. a movement purpose estimating means for estimating the partner's movement purpose, a movement intention estimation means for detecting the user's intention to start exercising based on the signal from the biological signal detection means, and each corresponding to a plurality of discrete movement purposes. storage means for pre-storing information specifying a plurality of control strategies for optimal control based on the estimated motion objective; and a control signal generating means that generates a signal for controlling the assist torque according to the derived mixed control law in response to detection of an intention to start movement.

Preferably, the motion purpose estimating means generates a motion purpose label using an inverse reinforcement learning algorithm from the captured image sensor signal, and the mixing means generates a plurality of control measures that are associated in advance with the motion purpose label. , linearly combined with corresponding weights.

Preferably, the mixing means derives a control law corresponding to the estimated operating objective by a linear Bellman combination method.

According to still another aspect of the present invention, there is provided a motion support method for assisting a joint movement of a user by an exoskeleton robot in a coordinated exercise with a partner, wherein the exoskeleton robot Active joints that are attached to joints and have actuator means that generate assist torque to the joints, and information that specifies multiple control strategies for optimal control corresponding to multiple discrete motion objectives, respectively. a step of estimating the purpose of the partner's movement based on a signal from an image sensor that images the partner; a step of detecting a user's intention to start an exercise based on a signal from a biological signal detection means for detecting a biological signal derived from a biological signal; and a step of determining a weight corresponding to a plurality of control measures according to an estimated purpose of movement of the partner. and an assist torque generated by the active joint in response to the detection of the movement initiation intention in response to the derived mixture control law. and generating a control signal for controlling.

According to the motion support device and motion support method of the present invention, it is possible to control an exoskeletal robot that suits the user's motion purpose while reducing the user's prior burden.

According to the motion support device and motion support method of the present invention, an exoskeleton that is compatible with the purpose of movement of the user and others while reducing the user's prior burden during cooperative exercise or work between the user and others. It becomes possible to control type robots.

It is an example of a block diagram of an exoskeletal robot for an upper limb. FIG. 3 shows the kinematic configuration of the exoskeleton for the right upper arm. FIG. 2 is a diagram illustrating the configuration of a pneumatic actuator system with one degree of freedom according to the first embodiment. FIG. 2 is a conceptual diagram for explaining an overview of control based on a control law (control strategy) according to the present embodiment. FIG. 2 is a conceptual diagram showing the flow of a procedure for determining a mixed control policy. 6 is a conceptual diagram for explaining the contents of each procedure in FIG. 5. FIG. FIG. 2 is a conceptual diagram showing a specific motor task for evaluating the proposed approach. FIG. 3 is a diagram showing the position of an EMG electrode on the user's body surface whose muscle activity was measured. FIG. 6 is a diagram showing simulation results of state transitions generated by an optimal control strategy. FIG. 4 is a diagram illustrating state transitions in a strategy obtained from a mixed method of optimal control strategies for a 2m goal shot. FIG. 3 is a diagram illustrating a typical tendency of one subject's feature amount obtained by the PLS method and the corresponding user's motion purpose. Figure 2 shows the mean and standard deviation of all EMG signals generated during the 10 trials in which the ball was thrown, %maximal voluntary contraction. FIG. 4 is a diagram showing the goal shooting rate when each subject throws the ball 10 times in each condition. FIG. 7 is a conceptual diagram for explaining a method of controlling an exoskeleton robot according to a second embodiment. FIG. 7 is a diagram illustrating an example of a block diagram of an exoskeletal robot for an upper limb according to a second embodiment. FIG. 3 is a functional block diagram for explaining the configuration of a controller 134 according to the second embodiment. FIG. 2 is a conceptual diagram showing a basketball passing task in which an experiment was conducted. FIG. 7 is a diagram showing the time-series average and standard deviation of the mixing ratio when the partner moves from the starting point toward the target position. FIG. 4 is a diagram comparing the sum of amplitudes of EMG signals between the present control method and control without assistance. FIG. 6 is a diagram showing the catch position from the center of the partner's face when the partner catches the ball. FIG. 7 is another diagram showing the catch position from the center of the partner's face when the partner catches the ball. FIG. 2 is a conceptual diagram illustrating a comparison between a "regression method" and a "discrimination method" which are conventional control methods.

The configuration of an assist device using an exoskeletal robot according to an embodiment of the present invention will be described below with reference to the drawings. In the following embodiments, components and processing steps with the same reference numerals are the same or equivalent, and the description thereof will not be repeated unless necessary.

Further, as an example of an actuator for driving the joints of an exoskeletal robot, a "pneumatic actuator (pneumatic artificial muscle)" described below will be described as an example.

Therefore, in the present embodiment, an exoskeletal robot for assisting the movement of the upper limbs will be described below.

However, the exoskeleton robot of the present invention can be used not only as an exoskeleton robot for assisting the movement of the upper limbs, but also as an exoskeleton robot for assisting the movement of the lower limbs.

In addition, in the following explanation, an exoskeleton robot that assists the movement of one of the upper limbs will be explained, but an exoskeleton robot that assists the movement of a pair of upper limbs, or one or a pair of the lower limbs. It is also possible to use

Furthermore, if the exoskeletal robot of the present invention is to assist the musculoskeletal system of a target human, it may be necessary to perform at least one of the upper limbs or at least one of the lower limbs. For example, it may assist only the movement of the lower back of the target human, or it may assist the movement of the lower back in conjunction with the movement of the lower limbs when walking or running. It may also be something that assists. In this specification, such assistance for the exercise of a target human will be collectively referred to as "assistance for the musculoskeletal system exercise of a target human."

The exoskeleton robot of this embodiment has an exoskeleton. An “exoskeleton” is a skeletal structure that a robot has that corresponds to the human skeletal structure. More specifically, an "exoskeleton" refers to a frame structure that externally supports a part of the human body on which an exoskeleton robot is attached.

This frame structure is further provided with joints for moving each part of the frame structure in response to movements based on the human skeletal structure.

In particular, an exoskeletal robot that assists in the movement of an upper limb is a robot that has a shoulder, an elbow, and a wrist, and has joints at least at the positions of the shoulder, elbow, and wrist. Further, the joint can be a pneumatic actuator-driven joint. Hereinafter, in the exoskeleton robot, the joints that are driven by actuators to provide support force to the user's joints will be referred to as "active joints." Furthermore, joints that passively move due to the user's movements are called "passive joints."

[Embodiment 1]
FIG. 1 is an example of a block diagram of an exoskeleton robot for an upper limb.

A command for controlling the exoskeleton robot 40 is given to the exoskeleton robot from the external control device 20 via the communication path. Although not particularly limited, a general-purpose personal computer can be used as the external control device 20, and an Ethernet (registered trademark) cable can be used as the communication path. Of course, as a communication path, in addition to a wired communication path of other standards, a wireless communication path, such as a wireless LAN (Local Area Network) or a wireless communication standard of another communication standard, may be used.

The external control device 20 includes an input unit 208 that receives instructions input from the user, and a nonvolatile storage device in which data necessary for control such as programs for generating commands and various control parameters are recorded. 206, a memory 204 including a ROM (Read Only Memory) in which firmware for starting the external control device 20 is stored, a RAM (Random Access Memory) that operates as a working memory, and a memory 204 that executes commands according to the program. An arithmetic unit 210 that executes processing to generate an exoskeleton robot, an interface (I/F) unit 202 that sends commands to the exoskeleton robot via a communication path, and an The robot 1 includes a display device 212 for displaying information regarding the control state of the robot 1 and the like.

As described above, when the external control device 20 is a general-purpose personal computer, the arithmetic device 210 is configured with a CPU (Central Processing Unit), and the nonvolatile storage device 206 is a hard disk drive, solid state drive, etc. can be used. However, some or all of the functional blocks of the external control device 20 may be configured by dedicated hardware.

Further, the external control device 20 performs a process of configuring, during calibration, a model for estimating joint torque from detected joint angles and contractile forces, for example, for a user to whom an exoskeletal robot is attached.

The exoskeleton robot further includes an exoskeleton section 121 and an internal control device 10. Note that in the figure, only the exoskeleton portion 121 of the right arm is illustrated.

The exoskeleton section 121 includes frames corresponding to the upper arm, forearm, and palm of the upper limb, respectively, an active joint 122, a passive joint 123, and a detection mechanism 124. Furthermore, the active joints 122 include a shoulder joint SFE and an elbow joint EFE. The active joint includes a pulley (not shown) that is driven by a contractile force transmitted from an air muscle (not shown) by a driving force transmission cable, as will be described later. Furthermore, a passive joint WR is provided on the forearm.

In other words, the active joint 123 is a hybrid actuator that receives driving force from the air muscle. Note that the actuator has a function of receiving a torque value serving as a control target value as a drive signal and performing control based on the received torque value. For example, highly accurate torque control can be achieved by disposing a torque sensor in the active joint 123 and feeding back the value detected by the torque sensor to the drive circuit.

The internal control device 10 also includes an I/F section 11, a recording device 131, a storage device 132, a measuring device 133, a control section 134, and an output device 135.

The I/F section 11 can receive torque or position commands, etc., commanded from the external control device 20.

The detection mechanism 124 detects the joint angle of each joint and the torque related to each joint. The detection mechanism 124 is, for example, an angle sensor arranged at each joint that detects the joint angle, a load cell that detects the driving force of each air muscle, or the like. Furthermore, the detection mechanism 124 detects myoelectric signals based on signals from electrodes attached to the user's body surface, as will be described later.

In the following, a myoelectric signal will be explained as an example of a biological signal generated based on a command from the user's center regarding the movement of the user's joints. However, for example, as such a biologically derived signal, other signals such as an electroencephalogram or a near infrared detection signal (NIRS signal: Near Infrared Spectroscopy signal) for the brain may be used.

Internal control device 10 operates active joint 123. The internal control device 10 operates the active joint 123 in response to the target torque, position command, etc. received by the I/F unit 11.

The measuring device 133 receives various signals (data) indicating detection results from the detection mechanism 124 such as a sensor. The control unit 134 performs various calculations such as calculating a control target value.

The output device 135 outputs a control signal to the drive section 30. For example, the output device 135 outputs a target pressure value of the air muscle to the air valve 34 to drive the air muscle 36. Compressed air compressed by the compressor 40 and stored in the air tank 32 is supplied to the air muscle 36 via the air valve 34, and the contractile force of the air muscle 36 is transmitted as driving force to the active joint by the driving force transmission cable.

The power supply 50 supplies power to the internal control device 10 and the drive unit 30.

The drive unit 30 and the power source 50 may be fixed on the ground, for example, or when the subject is in a wheelchair, they may be mounted on the rear of the wheelchair.

In addition, a "driving force transmission cable" is a metal wire that is passed through an outer case, similar to those used in bicycle brake mechanisms, and is flexible while transmitting force. A standard Bowden cable can be used. In the following description, it is assumed that the driving force transmission cable is a Bowden cable.

Further, the internal control device 10, the drive section 30, and the power source 50 may be mounted as an integrated controller unit 1, for example, in the rear part of the wheelchair as described above.

In addition, the pneumatic air muscle described above is lightweight and can generate large amounts of force by converting the energy of compressed air (or compressed gas; hereinafter collectively referred to as "compressed fluid") into contraction force using a rubber tube. can be produced.

The principle behind how Air Muscle generates force is that a spiral fiber in which a pneumatic air bladder is embedded contracts in the longitudinal direction (vertical direction) when compressed air is pumped into the air bladder and the air bladder expands. be.

To explain in more detail, it has a structure in which a rubber tube with plugs at both ends is covered with fibers spread in a spiral shape so as to restrict the radial direction. When air is pumped into this rubber tube, the rubber tube expands due to the pressure of the air. However, since the radial direction is restricted by the fibers, it cannot expand, and is pulled by the radial expansion and contracts in the vertical direction. Artificial muscles resemble animal muscles in the way they expand and contract.

The actuator itself is light and soft. Furthermore, since the entire inner surface of the rubber tube contributes to the contraction of the actuator, it is easier to achieve a larger power-to-weight ratio than a general air cylinder, etc., which is structured to receive pressure only through its cross-sectional area.

Note that the term "air bag" is more generally referred to as a "fluid bag," since what flows into the bag is not limited to air, as long as it is inflated or deflated by fluid.

Therefore, pneumatic air muscles are also referred to as "pneumatic artificial muscles," and more generally, when fluids other than air are included, they are referred to as "fluid pressure artificial muscles."

However, below, a pneumatic air muscle will be explained as a specific example.

FIG. 2 shows the kinematic configuration of the exoskeleton for the right upper arm.

As shown in Figure 2(a), the shoulder and elbow active joints of the exoskeleton are powered by pneumatic artificial muscles (PAMs) with Bowden cables.

As described later, the control unit 134 controls the two joints of the upper limbs of the exoskeleton robot according to the calculated optimal policy (mixed control policy).

The shoulder active joint assists the user's shoulder flexion/extension movement (Shoulder E/F), and the elbow active joint assists the user's elbow flexion/extension movement (Elbow E/F).

As depicted in Figure 2(a), the robot has two additional degrees of freedom: shoulder abduction/adduction (Shoulder A/A) and wrist flexion/adduction (Wrist E/F). ). However, in the configuration of FIG. 2(a), these are free joints (passive joints).

FIG. 2(b) shows the appearance of such an exoskeleton robot for upper limbs.
[1 degree of freedom pneumatic actuator]
Below, in order to explain the configuration and operation of the pneumatic actuator of this embodiment, a pneumatic actuator with one degree of freedom will be described as an example.

FIG. 3 is a diagram illustrating the configuration of a pneumatic actuator system with one degree of freedom according to the first embodiment.

Two opposing pneumatic air muscles 302a and 302b generate opposing contractile forces that are transmitted to pulley 320 by inner wires 308a and 308b in Bowden cables 310a and 310b, respectively.

Bowden cables consist of an inner wire and a flexible outer case. By using Bowden cables for force transmission, the weight of the exoskeleton robot 40 worn on a human being can be reduced and the structural space of the robot skeleton can be saved.

The torque τ at pulley 320 is the pneumatic air muscle torque (τ _PAM ).

Note that it is also possible to configure an electric motor to be connected to the pulley 320, if necessary. Due to the torque of the electric motor and the time response of the pneumatic air muscle, the pneumatic air muscle covers a large torque for gravity compensation or low frequency torque generation, and the motor covers high frequency torque, but its torque is For back-drivability, it is also possible to use a small reduction gear with a low reduction gear.

Load cells 304a and 304b for detecting the contraction force of the pneumatic air muscles are provided between the bottom of the external frame in which the pneumatic air muscles are housed and the lower ends of the pneumatic air muscles 302a and 302b, respectively.
[Overview of control using control law (control strategy)]
FIG. 4 is a conceptual diagram for explaining an overview of control based on the control law (control policy) of this embodiment.

As will be described later, when controlling the pneumatic air muscle by the control unit 134, assist control laws optimized for a plurality of discrete motion objectives (motion objectives #1 to #n) are prepared in advance. do. It is assumed that information for specifying such an assist control law is stored in the storage device 132 in advance.

As described later, the control unit 134 estimates the user's movement purpose (not limited to movement purposes #1 to #n, but may be an intermediate purpose among them) based on myoelectric signals, etc., and makes preparations in advance. The optimal control laws created are combined using a linear Bellman combination method to generate continuous optimal control outputs for the exoskeleton robot.

Such a "linear Bellman combination method" is disclosed, for example, in the following documents.

Reference 1: M. da Silva, F. Durand, and J. Popovic, “Linear bellman combination for control of character animation,” 2009, pp.82:1-82:10.
In this type of control method, the optimal policy for a given task evaluated from measurement data caused by user actions is determined by a linear Bellman combination method (mixed control law) as described below. ”, it is possible for a human user to perform control corresponding to various operational purposes without having to carefully generate his or her own EMG signal in advance.

Furthermore, the control approach of this embodiment has a low computational load because it simply combines multiple pre-optimized control laws in order to derive the corresponding optimal policy for a given task. It is lightweight and can easily derive real-time policies.
(Control method)
In the following, an approach for determining a blended control strategy that supports the user's operational objectives is described.

FIG. 5 is a conceptual diagram showing the flow of procedures for determining such a mixture control strategy.

FIG. 6 is a conceptual diagram for explaining the contents of each procedure in FIG. 5.

(A. Evaluation of user movement objectives)
Referring to FIGS. 5 and 6, the computing device 210 or the control unit 134 is configured to determine how to combine pre-optimized control outputs to accommodate continuous changes in the user's operational objectives. , a low-dimensional feature space is first found.

This embodiment uses a partial least squares (PLS) algorithm to find a low-dimensional feature space relevant to the task.

Such a partial least squares method is disclosed in the following documents.

Reference 2: H. Wold, “Soft modeling by latent variables: the nonlinear iterative partial least squares approach,” in Perspective in Probability and Statistics, Paper in Honor of MS Bartlett, pp. 520-540, 1975.
Reference 3: S. Wold, M. Sjostrom, and L. Eriksson, “Pls-regression: a basic tool of chemometrics,” Chemometrics and Intelligent Laboratory Systems, vol. 2, pp. 109-130, 2001.
First, is the user's state ψ=[ξ ₁ , ξ ₂ , ..., ξ _m ] associated with the action purpose label y ^? is observed.

Here, a training data set of the user's state ψ and the action objective label y according to the passage of time for n samples is defined as follows.

In the PLS algorithm, for a given sample set, a projection matrix W∈ into the j-dimensional feature space is used such that the sample covariance between ψ _i and y _i is maximized in the feature space. R ^m×j is determined as follows.

Projection from the original m-dimensional feature space to a j-dimensional space (j<m) can be performed as follows.

As described above, since the subspace is determined in consideration of the covariance of class labels, it is possible to obtain low-dimensional features that reflect label information.

In this embodiment, the user's action objective will be directly evaluated by a regression model between the low-dimensional feature μ and the label y.
Note that, although not particularly limited, it is desirable that the estimated motion purpose is one that interpolates a preset motion purpose.

The evaluated motion objective label y (hat: hereinafter, the letter X with ^ attached to the beginning indicates an estimated value and will be written as "X (hat)") is calculated by the formula ( 3) is used to determine the coefficients w _i for the linear Bellman combinations.

To map y (hat) as a weight w _i whose mixing weight is between 0 and 1, we use a sigmoid function as follows.

Here, a (>0) and b are parameters adjusted in advance by experiment.
(B. Optimal control based on motor intention)
In this embodiment, a control strategy is derived as follows according to the user's operational purpose.

Here, π ^*' indicates the optimal control strategy (mixed control strategy) to achieve the final operating objective, x indicates the state variable, and ψ reflects the user's operating objective.

Therefore, the controller 134 executes the process of finding the optimal policy to achieve the user's operational purpose as follows.

Note that at first glance, a simple method is to set an objective function corresponding to the user's final operational objective and optimize it. However, generating a controller that achieves a goal typically requires either empirical, time-consuming manual tuning or cost-intensive computational effort for optimization. Furthermore, if the goals change continuously, it is difficult to generate a controller corresponding to each goal.

Therefore, the controller 134 controls the pneumatic actuator using the optimal control strategy obtained for the new purpose by combining them as follows.

Here, α(x, t) is the mixing coefficient. In addition, π ^* _n is the sub-optimal control policy, which is the element corresponding to the n-th operational objective task, obtained by minimizing the cost function (or objective function) v(・) as follows. show.

Here, g(·) and l(·) represent the terminal cost and instantaneous cost, respectively.

The terminal cost represents the position error etc. in the final state, and the instantaneous cost represents the state of torque change during the operation.

Assuming that it is possible to collect N finite horizon controls that share all settings except the terminal cost g(x(T)) in Equation (5), using the linear Bellman combination method , and α(x, t) can be written as the following equation (7).

Such a "linear Bellman combination method" is disclosed in the following documents in addition to the above-mentioned document 1.

Reference 4: E. Todorov, “Compositionality of optimal control laws,” 2009, pp. 1856-1864.

Here, z _i (x, t) = exp(-v ^* (x, t)), v ^* (x, t) = vπ* (x, t) expresses the feasibility of the policy for the current state. where w _i is a fixed weighting factor for determining the target of the controller.

To derive each of the sub-optimal control strategies π ^* _n , as the optimal control, we use the iterative linear quadratic regulator method (iLQR) to the dynamics of the nonlinear system described as follows:

Here, x indicates the state and u indicates the input.

Such "linear quadratic regulator method (iLQR)" is disclosed in the following documents.

Reference 5: E. Todorov, C. Hu, A. Simpkins, and J. Movellan, “A generalized iterative lqg method for locally-optimal feedback control of constrained nonlinear stochastic systems,” 2005, pp. 300-306.
The optimization problem is to find the optimal control strategy π ^* that minimizes the objective function (5). The control strategy is calculated as follows.

Here, k is the sampling number and Δt is the sampling time.

Each iteration of the iLQR method consists of a nominal control sequence u (bar) (t) (hereinafter the letter Starting from the trajectory x(t), it is obtained by applying u(t) to the equation of motion (8) for x(0).

The iLQR method linearizes the system dynamics and the objective function around the nominal control sequence u(k) and the corresponding nominal trajectory x(k) up to second order. It solves optimization problems for nonlinear systems by approximation.

Therefore, in the iLQR method, the problem is solved using the Riccati equation, and the control law δu(k) is as follows.

Here, δx(k)=x(k)−x(k) (bar), l(k)=−H ⁻¹ (k)g(k) is the open-loop term, and L(k )=−H(k) ⁻¹ G(k) is the feedback term.
(Configuration of experimental equipment)
To test the approach described above, we conducted an experiment on three healthy right-handed subjects after obtaining their informed consent.
(A. Upper limb exoskeleton robot)
In this robot experiment, the optimal policy was derived to control the two joints of the upper limbs of the exoskeleton robot.

Joint torque is generated by a pneumatic actuator (PAM) as follows.

where r is the radius of the pulley and f _pam is the force of the pneumatic artificial muscle generated by the pathwise shortening of the helical fiber that was embedded in the air bladder.

Details of the mechanical design and pressure force model are introduced in the inventors' previous papers:

Reference 6: J. Furukawa, T. Noda, T. Teramae, and J. Morimoto, “Human movement modeling to detect biosignal sensor failures for myoelectric assistive robot control,” IEEE Transactions on Robotics, vol. PP, no. 99, pp .1-12, 2017.
In this embodiment, we consider the dynamics of a two-link robot model.

When deriving control strategies to support user movement, we consider dynamics as a human-robot system as follows.

1) The weight of the user's link is determined by adding the center of gravity of the link mass calculated from the subject's physical parameters to the robot link weight.

2) In the dynamic equation (8), the state variable x is the joint angle θ, the angular velocity θ (dot: Hereinafter, the time differential of the variable It includes the pressure P of the pneumatic artificial muscle PAM, which is the measured internal pressure of the air muscle, and is expressed as follows.

x = [θ ₁ , θ ₂ , θ ₁ (dot), θ ₂ (dot), P ₁ , P ₂ ] ^?
Here, subscripts 1 and 2 on the lower right represent the shoulder joint and elbow joint, respectively.

The control output u is the target pressure P _ref for the pneumatic actuator, and is expressed as follows.

u= [ _Pref1 , _Pref2 ] ^?
Although the dynamics between the target pressure P _ref and the measured internal pressure P is non-linear due to aerodynamics, a system of first-order lags can approximate this dynamics as follows.

Here, t _c is a time constant and is set as follows:

Furthermore, the maximum pressure is set at 0.8 MPa for safety.

The terminal cost g(·) for minimization is set as follows.

Or, if expressed correctly as a matrix, it is as follows.

Here, for the task of throwing a basketball as described below, T is the final time at which the ball is released, θtarget is the target angle, θtarget (dot) is the target angular velocity, and T is the final time at which the ball is released. Calculated according to distance.

The instantaneous cost l(·) was set as follows:

Here, too, the exact representation as a matrix is as follows.

Here, Ca, Cv, Cp, and Cpd can be manually selected through experiments, as in the case of normal optimal control.
(B. Motor task)
FIG. 7 is a conceptual diagram showing a specific motor task for evaluating the proposed approach.

As shown in FIG. 7, a basketball goal shot is set as an exercise task.

In this task, subjects threw a basketball in an underhand motion while seated, and the distance of the goal was varied from 1 m, 2 m, and 3 m.

Subjects were instructed to throw the ball with only those arms and were asked to throw the ball under two conditions;
・No-assist (NA) condition with an exoskeleton robot attached ・Assistant (A) condition with an exoskeleton robot attached The subject threw the ball 10 times in each condition.

One tactile sensor was attached to the tip of the subject's middle finger to measure the timing of ball release.
(C. Feature selection and parameter identification based on PLS method)
In order to evaluate the movement purpose from the user's very early movements, the rectified and filtered EMG signal e and the angles of the shoulder and elbow joints as well as the angular velocities generated by the shoulder and elbow joints are used as features. used. Therefore, the following 12-dimensional features are used.

ψ = [e ₁ , e ₂ , e ₃ , e ₄ , e ₅ , e ₆ , e ₇ , e ₈ , θ ₁ , θ ₂ , θ ₁ (dot), θ ₂ (dot)] ^T
Electromyography (EMG) signals were measured from the subject's right arm.

FIG. 8 is a diagram showing the position of the EMG electrode on the user's body surface whose muscle activity was measured.

The myoelectric signals are transmitted through eight sensor channels (Ch1: e ₁ , Ch2: e ₂ , Ch3: e ₃ , Ch4: e ₄ , Ch5: e 5 , Ch6: e ₆ , _Ch7 : e ₇ , Ch8: e ₈ ).

The relationship between each channel and the muscles to be measured is as follows.

Ch1: Front deltoid Ch2: Biceps Brachii Ch3: Back deltoid Ch4: Triceps Ch5: Flexor carpi ulnaris Ch6: Flexor carpi radialis Ch7: Extensor carpi radialis Ch8: Extensor carpi radialis logic
Ag/AgCl bipolar surface EMG electrodes can be used.

The angle and angular velocity of the shoulder and elbow joints are simultaneously acquired using the encoder of the upper limb exoskeletal robotic system.

During the movement of shooting a basketball to the goal, the point at which the angular velocity of the shoulder joint exceeds a threshold is defined as the movement starting point.

This is because preliminary tests have shown that the shoulder joint tends to move faster than the elbow in this task.

Additionally, the threshold was set at 0.2 from preliminary testing.

Although not particularly limited, for example, the average value of each data for several tens of milliseconds before the start of the movement is used as the feature value ψ.

Specifically, the EMG signal is activated 60-100 milliseconds in advance of the actual limb movement, and this delay is set by cross-validation of the training data.

In this embodiment, assuming that the goal is known, the sensor values are obtained as training data when the exoskeleton robot assists the movement to shoot at a distance of 1 m and 3 m from the goal. It was done.

With the corresponding labels y = 1 and y = 3 (1: 1m throw, 3: 3m throw), the injection matrix W for computing the one-dimensional feature μ∈R was derived by PLS.
(result)
In the following, we will first show the simulation results, and then show the results of online assist control based on the user's motion objective evaluation.
(A. Simulation results)
FIG. 9 is a diagram showing simulation results of state transitions generated by the optimal control strategy.

The errors between the target and the generated angle at the final release point, and the errors in the angular velocity were:
In the 1 m ball throw state, the shoulder joint angle was 0.003 rad, the angular velocity was 0.034 rad/s, the elbow joint angle was 0.003 rad, and the angular velocity was 0.013 rad/s.
On the other hand, in a ball throw state of 3 m, the shoulder joint angle is 0.0076 rad, the angular velocity is 0.0066 rad/s, the elbow joint angle is 0.0023 rad, and the angular velocity is 0.0036 rad/s. .

Similar results were obtained for other intra-user movements, with small errors between users.

From these results, it can be said that we were able to obtain an optimal control strategy to achieve the target state required for the basketball release point.

FIG. 10 is a diagram showing state transitions in a strategy obtained from a mixed method of optimal control strategies for a 2m goal shot.

To calculate the mixed control strategy for the 2m goal shooting condition, the weight parameters were set as w ₁ =0.5 and w ₂ =0.5. Here, w ₁ and w ₂ represent the weights for the mixture of optimal control strategies for 1 m and 3 m goal shooting conditions, respectively.

For comparison, we also show the state transition results with a mixed optimal control strategy for a 2m goal shot.

The errors between the joint angles and joint angular velocities at the target and generated release points were as follows:
The optimal control strategy had shoulder joint angle, 0.019 rad; angular velocity, 0.0031 rad/s; elbow joint angle, 0.0081 rad and angular velocity, 0.0042 rad/s.

On the other hand, for the mixed control strategy, the shoulder joint angle was 0.0038 rad; the angular velocity was 0.0067 rad/s; the elbow joint angle was 0.015 rad and the angular velocity was 0.09 rad/s.

The final state was close to the target.

These results demonstrate the utility of the mixed strategy for new goal shooting conditions.
(B. Feature extraction and weight determination based on PLS)
Below, we examine the trends in PLS results and explain the results.

FIG. 11 is a diagram illustrating a typical tendency of one subject's feature amount obtained by the PLS method and the corresponding user's motion purpose.

The parameters for the PLS method were trained using 1 m and 3 m goal shot data and tested on 2 m goal shot data.

FIG. 11 shows that the feature amounts and goal labels are ordered according to distance conditions.

This trend was similarly observed in other subjects.

Therefore, it was shown that the evaluated goal label y (hat) can be used as the user's action objective calculated by equation (3).

In this embodiment, based on these results, the purpose of the user's movement for a 1m goal shot is set as w ₁ =1-w ₂ , and w 2 =1/1+exp(-ay(hat)-b). Set up a 3m goal shot.

If the user intends to throw 1 meter away, the weight w ₁ is large and w ₂ is small.

Conversely, if the user intends to throw 3 meters away, the weight w ₁ decreases and w ₂ increases.
(C. Basketball Goal Shoot Online Control Results)
In the following, we will describe the experimental results for a basketball goal shooting test using the proposed assist-control approach.

To verify the effect of the assist, the amplitude of the rectified and filtered EMG signal was observed.

FIG. 12 shows the mean and standard deviation of all EMG signals generated during the 10 trials in which the ball was thrown, %MVC (%maximal voluntary contraction).

%MVC is the normalized muscle activity for each individual muscle by using the local maxima of the rectified and low-pass filtered EMG signal.

Welch's t-test between MVC at condition NA and condition A was applied.

We found that for most subjects, the average MVC was significantly reduced under condition A compared to condition NA.

It was confirmed that the average MVC tended to decrease compared to the NA condition, although the 3m distance goals shot in Condition A by Subject 1 did not show significant differences.

These results show that the control strategy derived by the proposed approach can assist the user in the ball throwing movement.

FIG. 13 is a diagram showing the goal shooting rate when each subject throws the ball 10 times under each condition.

When the ball was thrown 1 meter away, the shooting percentage under the NA and A conditions did not differ much for all subjects. Therefore, it is thought that the task was easy because the goal distance was close.

However, it can be said that it was observed that the assist control did not interfere with the movement of throwing the ball. This was the result of an accurate evaluation of the user's movement objectives.

At 2m goal distance, there was a higher shooting percentage for subjects 2 and 3 compared to the NA condition.

This indicates that goal shots can be successfully achieved with higher accuracy and less effort in condition A compared to the NA condition.

In contrast, in Subject 1, although the user's movement purpose was accurately evaluated, the shooting accuracy was lower in condition A than in the NA condition.

It is possible that a goal shooting distance of 2m was an easy task for this user.

Therefore, for this subject, the assistance provided by the exoskeleton robot may have reduced the burden of movement, but it may not have affected goal shooting accuracy.

At a goal distance of 3 m, subjects 1 and 3 demonstrated higher shooting percentages under condition A than under condition NA, with lower effort.

For subject 2, the shooting percentage was 0% even under condition A, and this goal distance task seemed quite difficult even with the robot assisting in movement.

These results demonstrate that the proposed approach is effective in supporting control for movements that require speed and precision.

That is, by using the exercise support device of this embodiment, it is possible to realize control of an exoskeleton robot that suits the user's purpose of movement while reducing the user's prior burden.

[Embodiment 2]
In the first embodiment, a mode has been described in which a mixed control strategy is used for an exoskeleton robot that assists a user in exercising alone.

In Embodiment 2, an example will be described in which a mixed control strategy is used for an exoskeleton robot in order to assist not the user's independent exercise but the cooperative exercise/work between the user and a partner.

As for the movement intention of the user of the exoskeletal robot, use of an electromyograph (EMG) on the body surface is assumed as in the first embodiment.

However, in the second embodiment as well, a myoelectric signal will be explained as an example of a biologically derived signal generated based on a command from the user's center regarding the movement of the user's joints. However, for example, as such a biologically derived signal, other signals such as an electroencephalogram or a near infrared detection signal (NIRS signal: Near Infrared Spectroscopy signal) for the brain may be used.

By the way, using EMG for partners who perform cooperative work/exercise will limit its practicality.

Therefore, in the second embodiment, as described below, the intention of the partner's movement is detected without using EMG. As such a method, a vision-based approach is adopted that uses an image of the partner captured by an image sensor.

Consider the situation in which the partner in collaborative work/cooperative action is moving, and consider the situation in which the user interacts with the other party through rapid movements, such as passing an object to the moving partner. In such situations, predicting the movement intentions of the "other" (partner), which has been generally ignored by traditional approaches, is a way to modify the robot's movements and base them on currently observed environmental information. This is important for robot control because it is difficult to synchronize interaction timing.

In the first embodiment, an example has been described in which the linear Bellman combination method is used to derive a control strategy for the activity of the person to be assisted.

On the other hand, in Embodiment 2, the coefficients for mixing the control measures calculated in advance are derived based on the predicted state of the partner, taking cooperative work/cooperative action into consideration. This approach derives a control strategy for a robot that interacts with its partner by transmitting the user's movement intentions through a simple EMG interface that uses a small number of sensor channels to detect motor start timing. , can start collaborative work/cooperative movements.

FIG. 14 is a conceptual diagram for explaining a method of controlling the exoskeleton robot according to the second embodiment.

As shown in FIG. 14, the partner's movements are visually observed by an image sensor 220. Given the goal of movement of partner 202, the probability of the movement path is derived based on the principles of the maximum entropy approach. From the derived path probabilities with multiple candidate targets, estimate the posterior probability of the target intended by partner 202.

The posterior probabilities are then used as a mix of weights to combine optimal control strategies for multiple goals. The controller 134 of the exoskeleton robot 121 derives a control strategy by linear Bellman combination using the derived mixed weights.

In the second embodiment as well, in such a control method, the optimal policy for a given task evaluated from the measurement data caused by the user 201's motion and the prediction of the partner's 202 motion goal is calculated using a linear Bellman method. Since it is derived as a "mixed control law (mixed control strategy)" using a combinatorial method, the user 201 can implement control corresponding to various operational purposes without having to carefully generate his or her own EMG signal in advance. It is possible.

Furthermore, in order to derive the corresponding optimal policy for a given task, multiple control laws (hereinafter referred to as "sub-optimal control strategies") that have been optimized in advance for multiple discrete operation objectives are used. Since the control approach of this embodiment has a light calculation load, it is possible to easily derive real-time measures.

FIG. 15 is a diagram showing an example of a block diagram of an exoskeleton robot for an upper limb according to the second embodiment.

The basic configuration is the same as the block diagram of the first embodiment shown in FIG. In the embodiment, information from the image sensor 220 is input to the control unit 134.

Then, as will be described later, the control unit 134 estimates the movement goal of the partner 202, derives an optimal policy according to the estimated movement goal, and controls the exoskeleton robot 121.

That is, to explain a more specific example, an example will be described below in which the control unit 134 accurately assists the partner 202 in passing a basketball while estimating the goal of the partner 202's movement. Of course, the control by the control unit 134 is not limited to this case, but may be performed by combining pre-calculated sub-optimal control measures when a plurality of motion targets exist for the partner 202. It is also possible to control other collaborative tasks and actions.

FIG. 16 is a functional block diagram for explaining the configuration of the controller 134 according to the second embodiment.

As explained below, the controller 134 determines a control strategy to support coordinated movement between the exoskeleton robot user 201 and partner 202.

(Prediction of partner's movement intention)
As shown in FIG. 16, the controller 134 includes a tracking processing unit 1342 that tracks the movement of the partner 202 based on a signal from the image sensor 220, and a target position prediction unit 1344 that predicts the target position of the movement of the partner 202. including.

As described below, the target position prediction unit 1344 uses the maximum entropy inverse reinforcement learning method (MaxEnt IRL) to predict the partner's exercise target (specifically, the exercise target position). The probability distribution of the target position G is estimated.

That is, first, the tracking processing unit 1342 monitors the movement trajectory of the partner 202. Specifically, the target position prediction unit 1344 assumes that the movement of the partner 202 observed for a certain period of time attempts to minimize the posterior distribution for the movement target (position), that is, the goal-dependent cost function Cg. , estimate the target position G based on the partially monitored movement trajectory.

Such a "goal-dependent cost function" is disclosed in the following documents.

Literature: AD Dragan and SS Srinivasa, “A policy-blending formalism for shared control,” Int. J. Rob. Res., vol. 32, no. 7, pp.790-805, June 2013.
Literature: K. Muelling, A. Venkatraman, J.-S. Valois, JE Downey, J. Weiss, S. Javdani, M. Hebert, AB Schwartz, JL Collinger, and JA Bagnell, “Autonomy infused teleoperation with application to brain computer interface controlled manipulation,” Auton. Robots, vol. 41, no. 6, pp. 1401-1422, Aug. 2017.
As disclosed in the above literature, in the maximum entropy inverse reinforcement learning method, the conditional probability distribution of the movement path of the partner 202 given a specific goal G is
P(ξ|G)∝ exp(-Cg(ξ))
It is expressed as Here ξ is the observed trajectory. Using this model, given a particular goal, the path probabilities of observed partial movements of partner 202 can be derived.

Here, Gi is the i-th target position among n candidate target positions, ξ ^* _{sp → cp} is the optimal trajectory that achieves the minimum cost, and s _p and c _p are the starting positions of partner 202. and the current position.

Next, the conditional probability of each target can be derived as the following equation (22).

Here, P(G _i ) is the prior probability of the i-th target position.

As an example, although not particularly limited, it is assumed that a predetermined number of target position candidates are first defined. Next, the posterior probability for each candidate target position is derived by equation (22) based on the observed path.

(Derivation of control policy)
Dynamic collaboration requires that control outputs be derived in real time, taking into account the partner's 202 situation. However, it is practically difficult to optimize the exoskeleton controller at each control time step considering the dynamically changing partner 202 situation.

Therefore, the controller 134 includes a mixed control policy calculation unit 1346 that calculates a control policy for the exoskeleton robot 121 as described below.

The mixed control policy calculation unit 1346 mixes (blends) sub-optimal control policies that have been optimized in advance using the linear Bellman combination method as described in the first embodiment.

In a specific example such as the above-mentioned basketball toss, the mixed control policy calculation unit 1346 first calculates in advance the optimal control strategy for the exoskeleton robot for the candidate target position, and Information on the control policy is stored in the storage device 132. Next, the mixed control policy calculation unit 1346 combines the pre-calculated sub-optimal control policies as shown in equation (23) below, according to the output of the value function for each sub-optimal control policy.

Here, π ^* ' indicates the control strategy for generating the exoskeleton robot's motion with respect to the target position to which the partner 202 intends to move, α _i (x, t) is the mixing coefficient, and x is the user's Denote the state variable and π ^* _n denote the sub-optimal control strategy for the nth component of the motor task obtained by minimizing the total cost (or objective function) v(·) of the following equation:

Here, g(·) and l(·) represent the terminal cost and instantaneous cost, respectively.

If the sub-optimal control policy for n finite horizon control problems that share all settings except the terminal cost g(x(T)) in Equation (24) has been calculated in advance, the mixed control policy can also be used here. Calculator 1346 can calculate α(x) using a linear Bellman combination approach as follows.

Here, z _i (x) expressed by the following equation indicates the feasibility of the policy for the current state.

w _i is a constant weight that takes into consideration the conditions of the partner in deriving the policy. Here, w _i is determined as the posterior probability of the partner's target candidate as follows.

Note that for each sub-optimal control policy π ^* _n (u|x) for each target position candidate of the partner 202, the mixed control policy calculation unit 1346 calculates the relationship between the user 201 and the exoskeleton robot in the same manner as in the first embodiment. 121 using an iterative linear quadratic controller (iLQR) method. Here, x is the state of the exoskeleton robot derived from the following equation (28), and u is the control output to the robot.

(Control based on user's exercise intention)
The movement intention of the partner 202 is derived based on visual information, whereas the user intention detection unit 1348 detects the movement start timing based on information monitored from the user 201 of the exoskeleton robot 121.

Specifically, the exercise start timing is extracted according to the actual movement of the user 201. Normally, an exoskeleton robot with limited degrees of freedom cannot assist the movement of the entire joint of the user 201, so synchronization of movement between the user and the exoskeleton robot requires accurate coordinated movement. is very important for the production of Therefore, the user intention detection unit 1348 outputs the control signal u so that the exoskeleton robot starts supporting the movement of the collaborative work at the timing of the start of the movement of the user 201. As a specific example, EMG signals are used to detect the onset of a user's movement.
[Experimental setup]
Below, in order to verify the assist control of the second embodiment, a basketball passing experiment was conducted with four healthy subjects. The results of this experiment will be explained below.

Before the actual session, all subjects used the method of Embodiment 2 (hereinafter referred to as "this control method") and the baseline method until they got used to generating movements with the assistance of the exoskeleton robot 121. Tossed the basketball a few times.
(Upper limb exoskeleton robot)
Similar to Embodiment 1, the derived strategy is used to control the three joints of the upper limb exoskeleton robot: shoulder flexion/extension (SFE), shoulder abduction/adduction (SAA), and elbow controlled flexion/extension (EFE) joint. Each joint torque is generated by a pneumatic artificial muscle (PAM) according to the following formula:

Here, as in the first embodiment, r is the pulley radius, and f _pam is the PAM force generated by path contraction of the spiral fiber embedded in the air bladder.

On the other hand, the wrist joint can be moved freely.

In other words, user 201 needs to generate the wrist joint torque himself in each trial. Thus, although the movement of the ball path cannot be fully automated for the user 201, the movement is generated by joint control between the user 201 and the exoskeleton robot 121.

The state variables include the measured internal pressure P of the PAM for joint angle θ, angular velocity θ (dots), and dynamics (28).

Here, subscripts 1, 2, and 3 represent joints SFE, SAA, and EFE, respectively.

The encoder of the upper limb exoskeleton robot system 121 is used to obtain both angle and angular velocity data. The control output u is the pressure input Pref required for the pneumatic actuator, and is expressed by the following formula.

Although the dynamics between the reference pressure Pref and the measured internal pressure P have nonlinearity due to air dynamics, this dynamics is approximated by the following first-order delay system.

Here, t _c is a time constant and is set as follows.

The terminal cost g(·) for minimization is set as follows.

Here, T is the final time, θ _target is the target angle, and θ (dot) _target is the target angular velocity at the basketball release point calculated based on the target distance. The instantaneous cost l(·) is set as follows.

Here, Ca, Cv, Cp, and Cpd are constant values of the objective function. Here, too, it can be manually selected through experimentation, as in the case of normal optimal control.

(motor task)
FIG. 17 is a conceptual diagram showing the basketball passing task in which the experiment was conducted.

As shown in FIG. 17, three right-handed subjects participated as an exoskeleton robot user 201 who toss a ball, and another person participated as a partner 202 who caught the ball. In this task, as shown in FIG. 17(a), the user 201 tosses a basketball while sitting, and as shown in FIG. 17(b), the partner 202 is randomly selected in a fan-shaped area. move from the designated starting point to any one of several target positions on the goal line. Note that, as will be described later, this target position of the partner's movement is not a candidate target position (a position for which a sub-optimal control policy has been calculated in advance) used to derive the optimal control law.

As shown in FIG. 17(a), in this experiment, the center of the face, that is, the nose of the partner 202, was assumed as the target point for throwing the basketball. The partner moves to one of the target locations unknown to user 201 before the session.

The distances from the user 201 to the unknown target positions (on goal line 1, g1 to g3, on goal line 2, g4 to g6) are 1.5 m, 2 m, and 2.5 m at each goal line, The subject tosses the ball to partner 202 five times as partner 202 reaches each target location.

Using Ag/AgCl bipolar EMG electrodes, five The EMG signal is measured to assess the starting point of the user's ball toss movement.

The optimal policy for the target distances at a 1 m distance position (G1) on goal line 1, and a 1 m distance position (G3) and a 3 m distance position (G4) on goal line 2 is determined from the known target position from the user 201. Obtained as a control measure against.

Then, the final optimal control policy π ^* ' for the unknown target position (g1, g2, ..., g6) is obtained by combining these four sub-optimal control strategies corresponding to the four known targets G1, ..., G4. derived.
(Image recognition of partner's movements)
A depth camera is used as the image camera 220 to monitor the partner's movements. The camera is placed near the user, detects the face position of the partner 202 with a facial recognition tool, and tracks the trajectory ξ of the partner 202.
(Detection of start of user movement)
The start timing of the user's movement is detected from the muscle activity of the user 201 of the exoskeleton robot 121. Specifically, EMG signals are measured from the anterior deltoid muscle. Motor intention is detected by a simple detector with a predetermined threshold. When a muscle is activated above a threshold, a control output is generated from the exoskeleton robot's controller 134. Thresholds are adjusted for each subject before the experiment.
[Experimental result]
In the following, we first show the variation of the mixing rate of the sub-optimal control strategy based on the partner's movement, and then show the analysis results of the EMG signal to verify the load on the user's movement and the performance of tossing the ball.
(mixing ratio)
FIG. 18 is a diagram showing the time-series average and standard deviation of the mixing ratio when the partner moves from the starting point toward the target position.

As shown in Figures 18(a), (b), and (c), the mixing ratios w ₁ and w ₂ increased as the partners moved toward the target on goal line 1 compared to w ₃ and w ₄ . .

From FIG. 18(a), when moving towards the target position g1, the mixture ratios w ₁ and w ₂ increase in this order, and the usage rate of the sub-optimal control policy for the candidate target position G1 is the highest, while for the candidate target position G2 It was shown that the proportion of using sub-optimal control strategies was the second highest.

On the contrary, from FIG. 18(c), the mixing ratios w ₂ and w ₁ increase in this order, indicating that the ratio of using the sub-optimal control strategy is opposite to condition (a).

Regarding Fig. 18(b), the mixing ratio w ₃ decreases and w ₁ increases compared to condition (c), which means that the ratio of using the sub-optimal control strategy is lower than that of conditions (a) and (c). It was shown that there is a difference between

On the other hand, from Fig. 18(d), (e), and (f), the mixing ratios w ₃ and w ₄ change as the other partner moves toward the target position on the goal line 2 opposite to the goal line 1. , increased more than w ₁ and w ₂ .

Similarly, the mixing ratios w ₃ and w ₄ in FIGS. 18(d), (e), and (f) are equal to w ₁ under the above conditions shown in FIGS. 18(a), (b), and (c). and w ₂ showed the same behavior.

These results showed that the partner's motor intention was successfully predicted and the mixing ratio was calculated appropriately.

(Analysis of user exercise load)
FIG. 19 is a diagram comparing the total amplitude of EMG signals between the present control method and control without assist.

To verify the user's exercise load with this control method, the amplitude of the rectified and filtered EMG signal was observed and compared with the unassisted condition.

Figure 19 shows the mean and standard deviation of the total EMG signals of all subjects on the two goal lines of 1.5 m (g1, g4), 2 m (g2, g5), and 2.5 m (g3, g6). There is. To compare inter-subject EMG signals, each muscle's rectified and low-pass filtered EMG signal was normalized separately for each subject.

According to FIG. 19, Welch's t-test was applied between the EMG signal without assistance and the present control method for the distance to each target position.

The mean summed EMG signal was significantly reduced under the present control method conditions than under the no-assist condition at all goal distances. These results indicate that the control strategy derived by the proposed approach successfully reduced the load of the user's ball tossing motion.

(ball toss performance)
Below, we examine the control method on ball tossing performance by measuring the distance from the center of the face in the coronal plane to the ball when the partner catches the ball, and by analyzing the recorded video. .

When a user threw a ball that his partner could not catch, the following measurements were taken. If the ball does not reach the opponent and falls in front of the partner, set the distance to 1.6m, which is the distance from the center of the player's face to the ground. If the ball is thrown out of reach of the opponent without falling, the distance between the center of the face and the ball when the ball passes is added.

This control method was compared with two schemes: 1) baseline method and 2) autonomous (automatic) method. In the baseline method, the partner's movements were tracked only by a camera without predictive processing, and the control strategy was derived from the most appropriate optimal strategy as determined by a distance-based criterion without blending.

This baseline method used a strategy that corresponds to the closest known target based on the time when a threshold-based detector identified the user's movement intent.

In the autonomous method (hereinafter referred to as the "autonomous method"), the partner's movement intention was estimated by the present control method, but the timing of throwing the ball was determined automatically without detecting the start of the user's movement. Ta. The tossing movement is initiated when the distance between the partner's position and the goal line falls below a certain threshold, and the time of the user's movement and the time for the partner's movement to reach the goal line match. It was calculated as follows.

FIG. 20 is a diagram showing the catch position from the center of the partner's face when the partner catches the ball.

In Figure 20, the mean values and standard deviations of all subjects under the baseline condition and the proposed method were measured for targets of 1.5 m (g1, g4), 2 m (g2, g5), and 2.5 m (g3, g6). Show value.

For each goal distance, Welch's t-test was applied between the distance under the baseline method and the distance under the present control method. center) was found to have significantly decreased.

FIG. 21 is another diagram showing the catch position from the center of the partner's face when the partner catches the ball.

In Figure 21, under the conditions of the autonomous method and the present control method, the average value and standard deviation of all subjects are set to 1.5 m (g1, g4), 2 m (g2, g5), and 2.5 m (g3, g6). This is the value measured five times for all subjects. Again, for each goal distance, Welch's t-test was applied between the distance in the autonomous method and the distance in the proposed conditions, and the average distance (center of the ball) in the present control method compared to the autonomous method. was found to be significantly reduced.

In this control method, considering that the ball toss by the user is stably controlled, the baseline method uses only the optimal strategy for the known target without mixing, and the autonomous method uses The control method is not consistent with the user's intention because it forced the user's arm to move without estimating the partner's movement intention.

These results suggest the importance of predicting the partner's movement intention and detecting the user's movement initiation timing. This control method can be said to effectively support motor control that requires speed and precision for collaborative work.

As explained above, in the second embodiment, the control strategy for the exoskeleton robot is derived by mixing the optimal control methods calculated in advance, and the mixing ratio is predicted by the maximum entropy inverse reinforcement learning method. Determined by the intention of the partner's movements.

To verify this control method, we conducted a ball-throwing experiment in which the user threw a ball to a moving partner in a specific interaction movement task. This control method significantly reduced user effort and improved ball toss performance compared to the baseline method. These results showed that the present control method effectively supports the exoskeleton robot user in interacting with a partner, enabling motion control that requires speed and precision.

The embodiments disclosed herein are examples of configurations for concretely implementing the present invention, and do not limit the technical scope of the present invention. The technical scope of the present invention is indicated by the claims rather than the description of the embodiments, and includes changes within the literal scope and equivalent meaning of the claims. is intended.

1 Controller unit, 10 Internal control device, 20 External control device, 40 Exoskeleton robot, 121 Exoskeleton part, 122 Active joint, 123 Passive joint 124 Detection mechanism, 220 Image sensor, 302a, 302b Pneumatic air muscle, 308a, 308b inner wire, 310a, 310b Bowden cable, 320 pulley, 321 transmission mechanism.

Claims (10)

A movement support device for assisting a user's joint movements,
an active joint having actuator means attached to a plurality of joints of the user receiving the assist and generating an assist torque to the joints;
sensor means for detecting the state of the active joint;
a biological signal detection means for detecting a biological signal generated based on a command from the user's central nervous system regarding movement of the user's joints;
a control unit that generates a control signal for controlling the assist torque generated by the active joint, the control unit comprising:
comprising a movement purpose estimation means for estimating the movement purpose of the user based on signals from the sensor means and the biological signal detection means,
The motion purpose estimating means sets a point at which the angular velocity of a predetermined active joint detected by the sensor means exceeds a predetermined value at an initial stage of the assisting motion as a motion starting point, and calculates a predetermined period of time before the start of the assisting motion. Estimate the purpose of the user's movement by learning the features,
a storage means for storing in advance information specifying a plurality of control strategies for optimal control corresponding to a plurality of discrete operation objectives;
Mixing means for linearly combining the plurality of control strategies with corresponding weights according to the estimated user's motion purpose to derive a mixed control law corresponding to the estimated motion purpose;
An operation support device further comprising: control signal generation means for generating a signal for controlling the assist torque according to the derived mixture control law. The motion purpose estimating means generates a signal in a low-dimensional feature space using a partial least squares algorithm from the measured signals from the sensor means and the biological signal detection means, and generates a motion purpose label. death,
2. The motion support device according to claim 1, wherein the mixing means linearly combines a plurality of control strategies that are associated with the motion purpose label in advance with corresponding weights. 3. The motion support device according to claim 2, wherein the motion purpose label estimated by the motion purpose estimating means is an interpolation of the plurality of discrete motion purposes. 3. The motion support device according to claim 1, wherein said mixing means derives a control law corresponding to said estimated motion purpose by a linear Bellman combination method. The motion support device according to any one of claims 1 to 4, wherein the actuator means includes a pneumatic air muscle. A motion support method for assisting a user's joint movements with an exoskeletal robot, the method comprising:
The exoskeleton robot has active joints that are attached to a plurality of joints of the user receiving the assist and have actuator means for generating assist torque to the joints, and each corresponds to a plurality of discrete motion purposes. and storage means for storing in advance information specifying a plurality of control strategies for optimal control,
A signal from a sensor means for detecting the state of the active joint, and a biosignal detection means for detecting a biogenic signal generated based on a command from the user's center regarding the movement of the user's joint. estimating the purpose of the user's movement based on the signal of the
The step of estimating the purpose of the user's movement includes starting the assisting movement by setting a point at which the angular velocity of a predetermined active joint detected by the sensor means exceeds a predetermined value as a starting point in the initial stage of the assisting movement. The method includes a step of estimating the purpose of the user's movement by learning feature quantities of a previous predetermined period;
linearly combining the plurality of control strategies with corresponding weights according to the estimated operation purpose of the user to derive a mixed control law corresponding to the estimated operation purpose;
A motion support method further comprising the step of generating a control signal for controlling the assist torque generated by the active joint according to the derived mixed control law. A movement support device for assisting a user's joint movements in a coordinated exercise with a partner,
an active joint having an actuator means attached to a joint of the user receiving the assist and generating an assist torque to the joint;
sensor means for detecting the state of the active joint;
a biological signal detection means for detecting a biological signal generated based on a command from the user's central nervous system regarding movement of the user's joints;
an image sensor that images the partner;
a control unit that generates a control signal for controlling the assist torque generated by the active joint, the control unit comprising:
Movement purpose estimating means for estimating the movement purpose of the partner based on a signal from the image sensor;
Movement intention estimation means for detecting the user's intention to start exercise based on the signal from the biological signal detection means;
a storage means for storing in advance information specifying a plurality of control strategies for optimal control corresponding to a plurality of discrete operation objectives;
Mixing means for linearly combining the plurality of control strategies with corresponding weights according to the estimated operation purpose to derive a mixed control law corresponding to the estimated operation purpose;
and control signal generation means for generating a signal for controlling the assist torque according to the derived mixed control law in response to the detection of the exercise start intention. The motion purpose estimating means generates a motion purpose label from the imaged signal of the image sensor using an inverse reinforcement learning algorithm,
8. The motion support device according to claim 7, wherein said mixing means linearly combines a plurality of control strategies associated with said motion purpose label in advance with corresponding weights. 9. The motion support device according to claim 7, wherein said mixing means derives a control law corresponding to said estimated motion purpose by a linear Bellman combination method. A motion support method for assisting the movement of a user's joints by an exoskeleton robot in a coordinated exercise with a partner, the method comprising:
The exoskeletal robot has an active joint that is attached to a joint of the user that receives the assist and has an actuator means that generates an assist torque to the joint, and an active joint that is attached to a joint of the user that receives the assist, and has an active joint that has an actuator means that is attached to the joint of the user that receives the assist, and has an active joint that has an actuator means that is attached to the joint of the user that receives the assist. and storage means for pre-storing information specifying a plurality of control strategies for control,
estimating a movement objective of the partner based on a signal from an image sensor imaging the partner;
Detecting the user's intention to start exercising based on a signal from a biosignal detection means for detecting a biosignal generated based on a command from the user's center regarding the movement of the user's joints; ,
linearly combining the plurality of control strategies with corresponding weights according to the estimated operational objective of the partner to derive a mixed control law corresponding to the estimated operational objective;
A motion support method comprising the step of generating a control signal for controlling the assist torque generated by the active joint according to the derived mixed control law in response to the detection of the movement start intention.