CN110597072B - Robot admittance compliance control method and system - Google Patents

Robot admittance compliance control method and system Download PDF

Info

Publication number
CN110597072B
CN110597072B CN201911006791.8A CN201911006791A CN110597072B CN 110597072 B CN110597072 B CN 110597072B CN 201911006791 A CN201911006791 A CN 201911006791A CN 110597072 B CN110597072 B CN 110597072B
Authority
CN
China
Prior art keywords
admittance
sample
robot
neural network
network model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201911006791.8A
Other languages
Chinese (zh)
Other versions
CN110597072A (en
Inventor
周朝政
潘昕荻
凌宇飞
李丹
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shanghai Electric Group Corp
Original Assignee
Shanghai Electric Group Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shanghai Electric Group Corp filed Critical Shanghai Electric Group Corp
Priority to CN201911006791.8A priority Critical patent/CN110597072B/en
Publication of CN110597072A publication Critical patent/CN110597072A/en
Application granted granted Critical
Publication of CN110597072B publication Critical patent/CN110597072B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B25—HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
    • B25J—MANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
    • B25J9/00—Programme-controlled manipulators
    • B25J9/16—Programme controls
    • B25J9/1602—Programme controls characterised by the control system, structure, architecture
    • B25J9/161—Hardware, e.g. neural networks, fuzzy logic, interfaces, processor
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
    • G05B13/048—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators using a predictor

Landscapes

  • Engineering & Computer Science (AREA)
  • Automation & Control Theory (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Medical Informatics (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Robotics (AREA)
  • Mechanical Engineering (AREA)
  • Manipulator (AREA)
  • Feedback Control In General (AREA)

Abstract

The invention discloses a robot admittance compliance control method and a system, wherein the robot admittance compliance control method comprises the following steps: detecting an operating force applied to the robot, and detecting a moving speed and an acceleration of the robot; and inputting the operating force, the movement speed and the acceleration into a trained neural network model, and outputting admittance parameters by the neural network model, wherein the admittance parameters comprise virtual damping and virtual mass. The invention realizes the compliance control algorithm through the neural network model, and can change the admittance parameters of the robot system in real time according to the rigidity change of the human-computer interaction environment, namely different forces externally applied to the robot, thereby improving the compliance of the human-computer interaction and realizing the sensitive and quick response of the robot.

Description

Robot admittance compliance control method and system
Technical Field
The invention relates to the field of control, in particular to a robot admittance compliance control method and system.
Background
Currently, commercial human-computer interaction robots mainly adopt two methods, namely impedance control and admittance control.
Impedance control refers to a control mode of inputting displacement and outputting force, has strong robustness on uncertainty of model parameters, can limit interaction force, and can ensure very good performance and stability in a very rigid environment, however, impedance control cannot provide enough rigid behavior and compensation friction, so that the precision of flexible environment and free motion is poor, and therefore, impedance control is not suitable for the occasion of man-machine interaction.
Admittance control is a problem of researching the relation between input force and output speed, is more suitable for a scene of interacting with a flexible environment or operating in a free environment, is more suitable for a doctor to more intuitively, fully and safely sense the environment of surgical interaction in a surgery, and can cause system instability if the environmental rigidity changes greatly. When a doctor drags the robot to move, errors and shaking cannot be avoided, and in an operation area with high precision requirements, the doctor cannot ensure that the doctor drags the robot to well complete corresponding operations.
Patent CN106618375A has designed the robot of a plurality of modularization joints, and when the operation of the terminal installation multidimension force transducer of robot, the doctor handheld terminal surgical tool of robot, the doctor exerts on the power transmission of instrument to multidimension force transducer, through the matching of power and robot speed, realizes the gentle and agreeable control of robot to realize man-machine interaction in coordination. However, only the robot design of a plurality of modular joints is described in detail, and a specific case for designing the control algorithm is not described.
Patent CN104626168A designs a robot force level compliance control algorithm based on impedance control, which predicts the interaction force between the robot and the environment through a prediction algorithm and compares the interaction force with the actual perception force of the robot to correct the actual force of the control system. And forming servo motor signals of each joint according to the track so as to control the servo motors to realize force and position flexible control. The patent CN104626168A adopts force-position flexible robot control, is suitable for rigid environment interaction and is not suitable for a clinical operation human-computer interaction environment.
Patent CN106695797A designs a compliance control system based on cooperative operation of dual-arm robots, which adopts master-slave and shared strategies to perform load common force decomposition, and further proposes a master-slave force compliance control and shared force compliance control method based on cooperative operation of dual arms. In patent CN106695797A, a dynamic model of an expected motion trajectory and an expected force is established, and an expected pose and force are obtained by satisfying constraint solution, so as to implement a force compliance operation of robot impedance control, and this impedance control compliance control is not suitable for a human-computer interaction mode of clinical surgery, and the intuitiveness of interaction between a doctor and the environment is poor.
Disclosure of Invention
The invention aims to overcome the defect of poor precision and flexibility of robot operation in a human-computer interaction environment in clinical operation in the prior art, and provides a robot admittance compliance control method and system based on a neural network strategy.
The invention solves the technical problems through the following technical scheme:
a robot admittance compliance control method, comprising:
detecting an operation force applied to the robot, and detecting a movement speed and an acceleration of the robot;
Inputting the operating force, the motion speed and the acceleration into a trained neural network model, and outputting admittance parameters by the neural network model, wherein the admittance parameters comprise virtual damping and virtual mass.
Preferably, the step of inputting the operation force, the motion velocity and the acceleration to a trained neural network model is preceded by:
establishing the neural network model;
training the neural network model using a genetic algorithm.
Preferably, the neural network model is a fully-connected multi-layer feedforward network, the fully-connected multi-layer feedforward network comprises a hidden layer and an output layer, the hidden layer comprises a plurality of neurons, and the output of the hidden layer is the input of the output layer;
the step of establishing the neural network model comprises:
setting the output function of the hidden layer as:
Figure BDA0002243014120000031
where s is the sigmoid activation function, wikWeight coefficient vector for the k-th neuron of the hidden layer, bhT is a matrix transposition symbol;
setting the output function of the output layer as:
Figure BDA0002243014120000032
wherein woAs a vector of weight coefficients of the output, boTo output the deviation, y is the admittance parameter and T is the matrix transposition symbol.
Preferably, the step of training the neural network model using a genetic algorithm comprises:
acquiring a plurality of groups of sample operating forces applied to the robot, and corresponding sample movement speeds and sample accelerations;
calculating and obtaining ideal admittance parameters according to an admittance control formula by using the sample operating force, the sample movement speed and the sample acceleration;
respectively setting an energy function, a penalty function and an optimal energy equation;
randomly generating the weight coefficient vector and the deviation amount of the output function of the hidden layer by using a genetic algorithm;
inputting the sample operation force, the sample movement velocity, and the sample acceleration as input parameters to the neural network model;
processing the input parameters by using the energy function, the penalty function and the optimal energy equation to obtain sample admittance parameters;
judging whether the difference value of the ideal admittance parameter and the sample admittance parameter is in a preset range, if so, finishing the training; if not, adjusting the weight coefficient vector and the deviation amount, and returning to the step of inputting the sample operation force, the sample motion speed and the sample acceleration as input parameters to the neural network model.
Preferably, the first and second electrodes are formed of a metal,
setting the energy function as:
Figure BDA0002243014120000041
where p is the sample coordinate position, prefSetting the coordinate position as a preset coordinate position, and setting T as a response characteristic, wherein the response characteristic comprises at least one of response time, a steady-state value and rise time;
and/or setting the penalty function:
Figure BDA0002243014120000042
Figure BDA0002243014120000043
wherein R isforwardAs a forward penalty function, RbackwardIs a backward penalty function, where N is the sample operating force, JfiAnd JbiRespectively a forward energy equation and a backward energy equation of the step i;
and/or setting the optimal energy equation as follows:
F＝max(Rforward,Rbackward)。
the robot admittance compliance control system comprises a detection module and an admittance module;
the detection module is used for detecting the operation force exerted on the robot and detecting the movement speed and the acceleration of the robot;
the admittance module is used for inputting the operating force, the movement velocity and the acceleration into a trained neural network model, and the neural network model outputs admittance parameters which comprise virtual damping and virtual mass.
Preferably, the robot admittance compliance control system comprises a model building module and a model training module;
the model establishing module is used for establishing the neural network model;
The model training module is used for training the neural network model by using a genetic algorithm.
Preferably, the neural network model is a fully-connected multi-layer feedforward network, the fully-connected multi-layer feedforward network comprises a hidden layer and an output layer, the hidden layer comprises a plurality of neurons, and the output of the hidden layer is the input of the output layer;
the model building module is further configured to set an output function of the hidden layer as:
Figure BDA0002243014120000051
where s is the sigmoid activation function, wikWeight coefficient vector for the k-th neuron of the hidden layer, bhT is a matrix transposition symbol;
and further configured to set the output function of the output layer to:
Figure BDA0002243014120000052
wherein woAs a vector of weight coefficients of the output, boTo output the deviation, y is the admittance parameter and T is the matrix transposition symbol.
Preferably, the model training module is further configured to acquire a plurality of sets of sample operating forces exerted on the robot, and corresponding sample movement speeds and sample accelerations;
the model training module is also used for calculating and obtaining ideal admittance parameters according to an admittance control formula by utilizing the sample operating force, the sample motion speed and the sample acceleration;
the model training module is also used for respectively setting an energy function, a penalty function and an optimal energy equation;
The model training module is further used for randomly generating the weight coefficient vector and the deviation amount of the output function of the hidden layer by utilizing a genetic algorithm;
the model training module is further used for inputting the sample operation force, the sample motion speed and the sample acceleration as input parameters to the neural network model;
the model training module is further used for processing the input parameters by using the energy function, the penalty function and the optimal energy equation to obtain sample admittance parameters;
the model training module is further used for judging whether the difference value between the ideal admittance parameter and the sample admittance parameter is within a preset range, and if so, the training is finished; if not, adjusting the weight coefficient vector and the deviation amount, and returning to the step of inputting the sample operation force, the sample motion speed and the sample acceleration as input parameters to the neural network model.
Preferably, the first and second electrodes are formed of a metal,
the model training module is further configured to set the energy function to:
Figure BDA0002243014120000053
where p is the sample coordinate position, prefSetting the coordinate position as a preset coordinate position, and setting T as a response characteristic, wherein the response characteristic comprises at least one of response time, a steady-state value and rise time;
And/or the model training module is further used for setting the penalty function:
Figure BDA0002243014120000061
wherein R isforwardAs a forward penalty function, RbackwardIs a backward penalty function, where N is the sample operating force, JfiAnd JbiRespectively a forward energy equation and a backward energy equation of the step i;
and/or the model training module is further configured to set the optimal energy equation to:
F＝max(Rforward,Rbackward)。
the positive progress effects of the invention are as follows:
the invention realizes the compliance control algorithm through the neural network model, and can change the admittance parameters of the robot system in real time according to the rigidity change of the human-computer interaction environment, namely different forces externally applied to the robot, thereby improving the compliance of the human-computer interaction and realizing the sensitive and quick response of the robot.
Drawings
Fig. 1 is a flowchart of a robot admittance compliance control method according to embodiment 1 of the present invention.
Fig. 2 is a flowchart of a robot admittance compliance control method according to embodiment 2 of the present invention.
Fig. 3 is a schematic structural diagram of a neural network model in the robot admittance compliance control method according to embodiment 2 of the present invention.
Fig. 4 is a flowchart of step 100 in the robot admittance compliance control method according to embodiment 2 of the present invention.
Fig. 5 is a flowchart of step 101 in the robot admittance compliance control method according to embodiment 2 of the present invention.
Fig. 6 is a schematic block diagram of a robot admittance compliance control system according to embodiment 3 of the present invention.
Fig. 7 is a schematic block diagram of a robot admittance compliance control system according to embodiment 4 of the present invention.
Detailed Description
The invention is further illustrated by the following examples, which are not intended to limit the scope of the invention.
Example 1
This embodiment provides a method for controlling compliance of robot admittance, as shown in fig. 1, the method for controlling compliance of robot admittance includes:
and 11, detecting the operation force applied to the robot, and detecting the movement speed and the acceleration of the robot.
And step 12, inputting the operation force, the movement speed and the acceleration into a trained neural network model, and outputting admittance parameters by the neural network model, wherein the admittance parameters comprise virtual damping and virtual mass.
The robot in the embodiment can be used for assisting in human-machine cooperation such as medical treatment and art carving, is particularly suitable for assisting in medical treatment, and the embodiment of the invention does not limit the specific application field of the robot. In addition, the robot should be understood in a broad sense, for example: the robot may be a robotic arm capable of human-machine interaction and intelligent control.
The operation force may be an interaction force given to the robot by an operator during human-computer interaction. For example: in the process of clinical operation, the tail end of the mechanical arm is connected with one end of the sensor, the other end of the sensor is connected with the clamp, an operation tool is fixed in the clamp, an operator controls the operation tool to perform displacement in multiple directions by hands, and then the operation process is executed, at the moment, the interaction force applied to the operation tool by the operator can be collected by the sensor, namely, the operation force is the operation force.
The effect of admittance parameters (virtual mass and virtual damping) on robot kinematics is explained below by a specific control principle of admittance control.
The admittance control law proposed by the present embodiment is based on force monitored using position, desired velocity and sensors.
Because the influence of system rigidity on admittance control can be ignored, the mass and the damping of admittance parameters control the overall movement speed and the interactive performance of the robot, and an admittance model is as follows:
Figure BDA0002243014120000071
wherein F is an operating force vector measured by the sensor, m is a virtual mass parameter, c is a virtual damping parameter,
Figure BDA0002243014120000081
respectively velocity and acceleration vectors.
For the formula (1), the operation force is used as the system input quantity, the speed is used as the system output quantity to carry out Laplace change, and the obtained equation is as follows:
Figure BDA0002243014120000082
H(s) determining the value of H(s) for the admittance transfer function, the virtual damping c and the ratio m/c of the virtual mass to the virtual damping; wherein the virtual damping c affects the steady state value of the system response and the ratio m/c of the virtual mass to the virtual damping affects the dynamic performance of the system. It can also be said that the virtual damping c affects the steady state value of the response, while the virtual mass to virtual damping ratio m/c affects the dynamic performance of the entire robot system.
When the admittance parameters (virtual mass and virtual damping) are fixed values, the dynamic performance of the robot remains unchanged, but coordination between fine small-range operation movement and coarse large-range operation movement cannot be performed, i.e. the force required to move the robot and the fine movement cannot be coordinated; when the admittance parameter is set to a high value, a larger operation force is generally required to move the robot at a given speed or acceleration, at which point the robot performs a fine motion more easily, because the robot is less reactive and the human-machine interaction is smooth. When the admittance parameter is set to a low value, it is easier to move the robot at a high speed or acceleration, but it is more difficult to perform a fine action.
Based on the principles of the above formulas (1) and (2), the embodiment can change admittance parameters in real time according to the change of the human-computer interaction environment stiffness, that is, different forces externally applied to the robot, by using the neural network model to realize the compliance control algorithm, thereby improving the output compliance of the robot and realizing the sensitive and quick response of the robot.
Example 2
Compared with embodiment 1, the present embodiment is different from embodiment 1 in that, as shown in fig. 2, the step of inputting the operation force, the movement velocity, and the acceleration to a trained neural network model further includes:
and step 100, establishing a neural network model.
Step 101, training a neural network model by using a genetic algorithm.
In this embodiment, the neural network model structure adopts an adaptive system, and the adopted structure is a fully-connected multi-layer feedforward network with a single hidden layer. The fully-connected multilayer feedforward network comprises an input layer, a hidden layer and an output layer, wherein the hidden layer comprises a plurality of neurons, and the output of the hidden layer is the input of the output layer. The number of layers per layer and the number of neurons are a result of tradeoffs between simplicity, performance, and training time. Since the input layer of the neural network model receives as input the robot's speed of motion and acceleration and the operating force F, it consists of several nodes, which depend on the dimensions of the problem under consideration. In this embodiment, the neural network model structure is specifically shown in fig. 3, the hidden layer and the output layer are respectively composed of 5 neurons and 1 neuron, and the reference standard for the number of the hidden layers is selected (the performance of neurons with hidden layers is not significantly improved by adding hidden layers).
The neural network model may also be other commonly used models, and is not limited herein. However, in a large number of experiments, compared with other models, the control accuracy of the fully-connected multilayer feedforward network is higher, the response is sensitive, the response time is short, and the effect is better.
More specifically, as shown in fig. 4, step 100 includes:
step 1001, an output function of the hidden layer is set.
The method specifically comprises the following steps:
Figure BDA0002243014120000091
where s is the sigmoid activation function, wikWeight coefficient vector for k-th neuron of hidden layer, bhT is a matrix transposition symbol;
step 1002, setting an output function of an output layer.
The method specifically comprises the following steps:
Figure BDA0002243014120000092
wherein woAs a vector of weight coefficients of the output, boTo output the deviation, y is the admittance parameter and T is the matrix transposition symbol.
As shown in fig. 5, step 101 includes:
step 1011, several sets of sample operating forces applied to the robot are collected, along with corresponding sample motion velocities and sample accelerations.
Under the general condition, different application scenes have different operating forces and moving spaces, an operating force range and a moving space position range can be preset in the specific application scenes, a series of data and process data processing are convenient to collect, discretization processing can be carried out on the collected continuous sample operating force in the training process, and the data range after the discretization processing extends over the whole preset operating force range.
And 1012, calculating and obtaining ideal admittance parameters according to an admittance control formula by using the sample operating force, the sample movement speed and the sample acceleration.
And step 1013, respectively setting an energy function, a penalty function and an optimal energy equation.
The energy function is set as:
Figure BDA0002243014120000101
where p is the sample coordinate position, prefSetting the coordinate position as a preset coordinate position, and setting T as a response characteristic, wherein the response characteristic comprises at least one of response time, a steady-state value and rise time;
setting a penalty function:
Figure BDA0002243014120000102
Figure BDA0002243014120000103
wherein R isforwardAs a forward penalty function, RbackwardIs a backward penalty function, where N is the sample operating force, JfiAnd JbiRespectively a forward energy equation and a backward energy equation of the step i;
setting the optimal energy equation as follows: ,
F＝max(Rforward,Rbackward)。
1014, randomly generating a weight coefficient vector and a deviation amount of an output function of the hidden layer by using a genetic algorithm;
and randomly generating a weight coefficient vector and a deviation amount of the output function for each neuron of the hidden layer.
Step 1015, the sample operation force, the sample motion speed and the sample acceleration are input into the neural network model as input parameters.
And step 1016, processing the input parameters by using the energy function, the penalty function and the optimal energy equation to obtain sample admittance parameters.
And calculating admittance parameters of the system and forward and backward penalty functions according to the detected operating force, speed and acceleration for each neuron of the hidden layer. An energy function is calculated for each sub-neuron.
Step 1017, judging whether the difference value between the ideal admittance parameter and the sample admittance parameter is within a preset range, if so, finishing the training; if not, go to step 1018.
And step 1018, adjusting the weight coefficient vector and the deviation amount, and returning to the step 1015.
The main body of the experiment of this example is the UR5 robot manufactured by the priority robot of denmark, weighing 18.4kg (kg), maximum load 5kg, maximum operating space 850mm (mm) of the robot. The controller was a CX2030 controller manufactured by German Kyoto company, and the sampling frequency was 1000Hz (Hz).
Firstly, the environmental rigidity interacting with the UR5 robot is obtained according to experience, the stress value (discrete value 10-200N (Newton)) of the sensor at the moment is measured, the default virtual mass value is set to be 5-10kg, the maximum value of the virtual damping is 1000Ns/m, and the minimum value is set to be 250 Ns/m. The response of the system in the ideal state is calculated from the environmental stiffness (operating force) according to equations (1) and (2) to ensure that the system is in a steady state when the admittance parameters change.
Secondly, a neural network model is constructed, and the output quantity is the trained admittance parameter because the input layer receives the motion speed, the acceleration and the operation force of the robot as input.
Training a neural network model by using a set energy function, a penalty function and an optimal energy equation in a preset operating force range and a preset moving space position range in a specific application scene, randomly generating a population consisting of individuals representing different neural network weight sets by using a genetic algorithm, respectively simulating the operation of the robot by aiming at the operating force in each discrete range, and then minimizing the energy function by using the genetic algorithm; different admittance parameters are measured according to the execution of the track following experiment of a plurality of groups of robots, admittance parameter values under the optimal scheme are determined according to the track deviation, and the admittance parameters of the robot system are changed in real time, so that the flexibility of human-computer interaction is improved, the sensitive and quick response of the robots is realized, and the effect is better.
Example 3
In the present embodiment, a robot admittance compliance control system is provided, as shown in fig. 6, the robot admittance compliance control system includes a detection module 31, an admittance module 32;
The detection module 31 is used for detecting the operating force exerted on the robot and detecting the movement speed and acceleration of the robot;
the admittance module 32 is configured to input the operating force, the motion velocity, and the acceleration to a trained neural network model, which outputs admittance parameters, including virtual damping and virtual mass.
The robot in the embodiment can be used for scenes such as auxiliary medical treatment and art carving, is particularly suitable for auxiliary medical treatment, and the specific application field of the robot is not limited by the embodiment of the invention. In addition, the robot should be understood in a broad sense, for example: the robot may be a robotic arm capable of human-machine interaction and intelligent control.
The operation force may be an interaction force given to the robot by an operator during human-computer interaction. For example: in the process of clinical operation, the tail end of the mechanical arm is connected with one end of the sensor, the other end of the sensor is connected with the clamp, an operation tool is fixed in the clamp, an operator controls the operation tool to perform displacement in multiple directions by hands, and then the operation process is executed, at the moment, the interaction force applied to the operation tool by the operator can be collected by the sensor, namely, the operation force is the operation force.
The effect of admittance parameters (virtual mass and virtual damping) on robot kinematics is explained below by a specific control principle of admittance control.
The admittance control law proposed by the present embodiment is based on force monitored using position, desired velocity and sensors.
Because the influence of system rigidity on admittance control can be ignored, the mass and the damping of admittance parameters control the overall movement speed and the interactive performance of the robot, and an admittance model is as follows:
Figure BDA0002243014120000121
wherein F is an operating force vector measured by the sensor, m is a virtual mass parameter, c is a virtual damping parameter,
Figure BDA0002243014120000122
respectively velocity and acceleration vectors.
For the formula (1), the operation force is used as the system input quantity, the speed is used as the system output quantity to carry out Laplace change, and the obtained equation is as follows:
Figure BDA0002243014120000123
h(s) determining the value of H(s) for the admittance transfer function, the virtual damping c and the ratio m/c of the virtual mass to the virtual damping; where the virtual damping c affects the steady state value of the system response and the ratio of the virtual mass to the virtual damping m/c affects the dynamic performance of the system. It can also be said that the virtual damping c affects the steady state value of the response, while the virtual mass to virtual damping ratio m/c affects the dynamic performance of the entire robot system.
When the admittance parameters (virtual mass and virtual damping) are fixed values, the dynamic performance of the robot remains unchanged, but coordination between fine small-range operation movement and coarse large-range operation movement cannot be performed, i.e. the force required to move the robot and the fine movement cannot be coordinated; when the admittance parameter is set to a high value, a larger operation force is generally required to move the robot at a given speed or acceleration, at which point the robot performs a fine motion more easily, because the robot is less reactive and the human-machine interaction is smooth. When the admittance parameter is set to a low value, it is easier to move the robot at a high speed or acceleration, but it is more difficult to perform a fine action.
Based on the principles of the above formulas (1) and (2), the embodiment can change admittance parameters in real time according to the change of the stiffness of the human-computer interaction environment, i.e., different forces externally applied to the robot, by using the neural network model to realize the compliance control algorithm, thereby improving the flexibility of the human-computer interaction of the robot and realizing the sensitive and quick response of the robot.
Example 4
Compared with embodiment 3, the difference of this embodiment is that, as shown in fig. 6, the robot admittance compliance control system further includes a model building module 33 and a model training module 34;
The model establishing module 33 is used for establishing a neural network model;
the model training module 34 is used to train the neural network model using a genetic algorithm.
Preferably, the neural network model is a fully-connected multi-layer feedforward network, the fully-connected multi-layer feedforward network comprises a hidden layer and an output layer, the hidden layer comprises a plurality of neurons, and the output of the hidden layer is the input of the output layer.
In this embodiment, the neural network model structure adopts an adaptive system, and the adopted structure is a fully-connected multi-layer feedforward network with a single hidden layer. The fully-connected multilayer feedforward network comprises an input layer, a hidden layer and an output layer, wherein the hidden layer comprises a plurality of neurons, and the output of the hidden layer is the input of the output layer. The number of layers per layer and the number of neurons are a result of tradeoffs between simplicity, performance, and training time. Since the input layer of the neural network model receives as input the robot's speed of motion and acceleration and the operating force F, it consists of several nodes, which depend on the dimensions of the problem under consideration. In this embodiment, the neural network model structure is specifically shown in fig. 3, the hidden layer and the output layer are respectively composed of 5 neurons and 1 neuron, and the reference standard for the number of the hidden layers is selected (the performance of neurons with hidden layers is not significantly improved by adding hidden layers).
Other common models can be used as the neural network model, and are not limited herein. However, in a large number of experiments, compared with other models, the control accuracy of the fully-connected multilayer feedforward network is higher, the response sensitivity and the quick response time are shorter, and the effect is better.
More specifically, the model building module 33 is further configured to set the output function of the hidden layer as:
Figure BDA0002243014120000141
where s is the sigmoid activation function, wikWeight coefficient vector for k-th neuron of hidden layer, bhT is a matrix transposition symbol;
the model building block 33 is further configured to set the output function of the output layer as:
Figure BDA0002243014120000142
wherein woAs a vector of weight coefficients of the output, boTo output the deviation, y is the admittance parameter and T is the matrix transposition symbol.
Preferably, the model training module 34 is also used to collect several sets of sample operating forces applied to the robot, as well as corresponding sample motion velocities and sample accelerations.
Under the general condition, different application scenes have different operating forces and moving spaces, an operating force range and a moving space position range can be preset in the specific application scenes, a series of data and process data processing are convenient to collect, discretization processing can be carried out on the collected continuous sample operating force in the training process, and the data range after the discretization processing extends over the whole preset operating force range.
The model training module 34 is further configured to calculate and obtain an ideal admittance parameter according to an admittance control formula by using the sample operation force, the sample movement velocity, and the sample acceleration;
the model training module 34 is further configured to set an energy function, a penalty function, and an optimal energy equation, respectively;
the model training module 34 is further configured to randomly generate a weight coefficient vector and a deviation amount of an output function of the hidden layer by using a genetic algorithm.
And randomly generating a weight coefficient vector and a deviation amount of the output function for each neuron of the hidden layer.
The model training module 34 is further configured to input the sample operation force, the sample movement velocity, and the sample acceleration as input parameters to the neural network model;
the model training module 34 is further configured to process the input parameters using an energy function, a penalty function, and an optimal energy equation to obtain sample admittance parameters.
And calculating admittance parameters of the system and forward and backward penalty functions according to the detected operating force, speed and acceleration for each neuron of the hidden layer. An energy function is calculated for each sub-neuron.
The model training module 34 is further configured to determine whether a difference between the ideal admittance parameter and the sample admittance parameter is within a preset range, and if so, the training is ended; and if not, adjusting the weight coefficient vector and the deviation amount, and returning to the step of inputting the sample operating force, the sample motion speed and the sample acceleration as input parameters to the neural network model.
Preferably, the first and second liquid crystal display panels are,
the model training module 34 is also configured to set the energy function as:
Figure BDA0002243014120000151
where p is the sample coordinate position, prefSetting the coordinate position as a preset coordinate position, and setting T as a response characteristic, wherein the response characteristic comprises at least one of response time, a steady-state value and rise time;
model training module 34 is also configured to set a penalty function:
Figure BDA0002243014120000152
wherein R isforwardAs a forward penalty function, RbackwardIs a backward penalty function, where N is the sample operating force, JfiAnd JbiRespectively a forward energy equation and a backward energy equation of the step i;
model training module 34 is also configured to set the optimal energy equation to:
F＝max(Rforward,Rbackward)。
the main body of the experiment of this example is the UR5 robot manufactured by the priority robot of denmark, weighing 18.4kg (kg), maximum load 5kg, maximum operating space 850mm (mm) of the robot. The controller was a CX2030 controller manufactured by German Kyoto company, and the sampling frequency was 1000Hz (Hz).
Firstly, the environmental rigidity interacting with the UR5 robot is obtained according to experience, the stress value (discrete value 10-200N (Newton)) of the sensor at the moment is measured, the default virtual mass value is set to be 5-10kg, the maximum value of the virtual damping is 1000Ns/m, and the minimum value is set to be 250 Ns/m. The response of the system in the ideal state is calculated from the environmental stiffness (operating force) according to equations (1) and (2) to ensure that the system is in a steady state when the admittance parameters change.
Secondly, a neural network model is constructed, and the output quantity is the trained admittance parameter because the input layer receives the motion speed, the acceleration and the operation force of the robot as input.
Training a neural network model by using a set energy function, a penalty function and an optimal energy equation in a preset operating force range and a preset moving space position range in a specific application scene, randomly generating a population consisting of individuals representing different neural network weight sets by using a genetic algorithm, respectively simulating the operation of the robot by aiming at the operating force in each discrete range, and then minimizing the energy function by using the genetic algorithm; different admittance parameters are measured according to the following track experiment of executing a plurality of groups of robots, and admittance parameter values under the optimal scheme are determined according to track deviation, so that variable admittance parameter control is realized, the following accuracy of the robots is higher, the response time of the robots is short, and the effect is better.
While specific embodiments of the invention have been described above, it will be appreciated by those skilled in the art that this is by way of example only, and that the scope of the invention is defined by the appended claims. Various changes and modifications to these embodiments may be made by those skilled in the art without departing from the spirit and scope of the invention, and these changes and modifications are within the scope of the invention.

Claims (8)

1. A robot admittance compliance control method, characterized in that the robot admittance compliance control method comprises:
detecting an operation force applied to the robot, and detecting a movement speed and an acceleration of the robot;
inputting the operating force, the motion velocity and the acceleration to a trained neural network model, wherein the neural network model outputs admittance parameters, and the admittance parameters comprise virtual damping and virtual mass;
the step of inputting the operating force, the movement velocity and the acceleration to a trained neural network model is preceded by:
establishing the neural network model;
training the neural network model using a genetic algorithm.
2. The robot admittance compliance control method of claim 1, wherein the neural network model is a fully-connected multi-layer feedforward network, the fully-connected multi-layer feedforward network including a hidden layer and an output layer, the hidden layer including a plurality of neurons, an output of the hidden layer being an input of the output layer;
the step of establishing the neural network model comprises:
setting the output function of the hidden layer as:
Figure FDA0003611204630000011
where s is the sigmoid activation function, w ikA weight coefficient vector for the kth neuron of the ith hidden layer, wherein i is 1-3, bhT is the matrix transpose symbol for the offset,
Figure FDA0003611204630000012
the speed is shown, and F is an operating force vector measured by a sensor;
setting the output function of the output layer as:
Figure FDA0003611204630000013
wherein woAs a vector of weight coefficients of the output, boTo output the deviation, y is the admittance parameter, T is the matrix transposition sign, and h is the output function of the full link layer.
3. The robotic admittance compliance control method of claim 2, wherein the step of training the neural network model using a genetic algorithm includes:
acquiring a plurality of groups of sample operating forces applied to the robot, and corresponding sample movement speeds and sample accelerations;
calculating and obtaining ideal admittance parameters according to an admittance control formula by using the sample operating force, the sample movement speed and the sample acceleration;
respectively setting an energy function, a penalty function and an optimal energy equation;
randomly generating the weight coefficient vector and the deviation amount of the output function of the hidden layer by using a genetic algorithm;
inputting the sample operation force, the sample movement velocity, and the sample acceleration as input parameters to the neural network model;
Processing the input parameters by using the energy function, the penalty function and the optimal energy equation to obtain sample admittance parameters;
judging whether the difference value of the ideal admittance parameter and the sample admittance parameter is in a preset range, if so, finishing the training; if not, adjusting the weight coefficient vector and the deviation amount, and returning to the step of inputting the sample operation force, the sample motion speed and the sample acceleration as input parameters to the neural network model.
4. The robotic admittance compliance control method of claim 3,
setting the energy function as:
Figure FDA0003611204630000021
where p is the sample coordinate position, prefSetting the coordinate position as a preset coordinate position, and setting T as a response characteristic, wherein the response characteristic comprises at least one of response time, a steady-state value and rise time;
and/or setting the penalty function as:
Figure FDA0003611204630000022
Figure FDA0003611204630000023
wherein R isforwardAs a forward penalty function, RbackwardIs a backward penalty function, where N is the sample operating force, JfiAnd JbiRespectively a forward energy equation and a backward energy equation of the ith step;
and/or setting the optimal energy equation as follows:
F＝max(Rforward，Rbackward)。
5. the robot admittance compliance control system is characterized by comprising a detection module and an admittance module;
The detection module is used for detecting the operating force applied to the robot and detecting the movement speed and acceleration of the robot;
the admittance module is used for inputting the operation force, the movement velocity and the acceleration into a trained neural network model, and the neural network model outputs admittance parameters which comprise virtual damping and virtual mass;
the robot admittance compliance control system comprises a model establishing module and a model training module;
the model establishing module is used for establishing the neural network model;
the model training module is used for training the neural network model by utilizing a genetic algorithm.
6. The robotic admittance compliance control system of claim 5, wherein the neural network model is a fully-connected multi-layer feedforward network including a hidden layer and an output layer, the hidden layer including a plurality of neurons, an output of the hidden layer being an input of the output layer;
the model building module is further configured to set an output function of the hidden layer as:
Figure FDA0003611204630000031
where s is the sigmoid activation function, wikA weight coefficient vector for the k-th neuron of the i-th hidden layer, b hT is the matrix transpose symbol for the offset,
Figure FDA0003611204630000032
the speed is shown, and F is an operating force vector measured by a sensor;
and further configured to set the output function of the output layer to:
Figure FDA0003611204630000033
wherein woAs a vector of weight coefficients of the output, boTo output the deviation, y is the admittance parameter, T is the matrix transposition sign, and h is the output function of the full link layer.
7. The robotic admittance compliance control system of claim 6, wherein the model training module is further configured to collect sets of sample operating forces exerted on the robot, and corresponding sample movement velocities and sample accelerations;
the model training module is also used for calculating and obtaining ideal admittance parameters according to an admittance control formula by utilizing the sample operating force, the sample motion speed and the sample acceleration;
the model training module is also used for respectively setting an energy function, a penalty function and an optimal energy equation;
the model training module is further used for randomly generating the weight coefficient vector and the deviation amount of the output function of the hidden layer by utilizing a genetic algorithm;
the model training module is further used for inputting the sample operation force, the sample motion speed and the sample acceleration as input parameters to the neural network model;
The model training module is further used for processing the input parameters by using the energy function, the penalty function and the optimal energy equation to obtain sample admittance parameters;
the model training module is further used for judging whether the difference value between the ideal admittance parameter and the sample admittance parameter is within a preset range, and if so, the training is finished; if not, adjusting the weight coefficient vector and the deviation amount, and returning to the step of inputting the sample operation force, the sample motion speed and the sample acceleration as input parameters to the neural network model.
8. The robotic admittance compliance control system of claim 7,
the model training module is further configured to set the energy function to:
Figure FDA0003611204630000041
where p is the sample coordinate position, prefSetting the coordinate position as a preset coordinate position, and setting T as a response characteristic, wherein the response characteristic comprises at least one of response time, a steady-state value and rise time;
and/or the model training module is further configured to set the penalty function as:
Figure FDA0003611204630000042
Figure FDA0003611204630000043
wherein R isforwardAs a forward penalty function, RbackwardIs a backward penalty function, where N is the sample operating force, JfiAnd JbiRespectively a forward energy equation and a backward energy equation of the step i;
And/or the model training module is further configured to set the optimal energy equation as:
F＝max(Rforward，Rbackward)。
CN201911006791.8A 2019-10-22 2019-10-22 Robot admittance compliance control method and system Active CN110597072B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911006791.8A CN110597072B (en) 2019-10-22 2019-10-22 Robot admittance compliance control method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911006791.8A CN110597072B (en) 2019-10-22 2019-10-22 Robot admittance compliance control method and system

Publications (2)

Publication Number Publication Date
CN110597072A CN110597072A (en) 2019-12-20
CN110597072B true CN110597072B (en) 2022-06-10

Family

ID=68851286

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911006791.8A Active CN110597072B (en) 2019-10-22 2019-10-22 Robot admittance compliance control method and system

Country Status (1)

Country Link
CN (1) CN110597072B (en)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111168677B (en) * 2020-01-08 2022-09-16 山东理工大学 Stability control method for humanoid flexible arm system
CN111290269B (en) * 2020-02-11 2022-04-08 西北工业大学深圳研究院 Self-adaptive compliance stable control method of space robot
CN112171670B (en) * 2020-09-22 2024-01-23 北京石油化工学院 Rehabilitation robot isotonic movement control method and system
JP2022117612A (en) * 2021-02-01 2022-08-12 セイコーエプソン株式会社 Operation parameter adjustment method, operation parameter adjustment program, and robot system
CN115502966B (en) * 2021-06-22 2024-04-26 腾讯科技（深圳）有限公司 Variable admittance control method for robot
CN114406983A (en) * 2021-12-06 2022-04-29 中国科学院深圳先进技术研究院 Adaptive admittance control method and related device for lower limb exoskeleton robot
CN114589702B (en) * 2022-04-24 2024-05-07 合肥工业大学 Collaborative robot dragging method based on dynamic parameter identification and admittance control
CN114932557B (en) * 2022-06-24 2023-07-28 合肥工业大学 Self-adaptive admittance control method based on energy consumption under kinematic constraint

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108227506A (en) * 2018-02-08 2018-06-29 常州普旺科技有限公司 A kind of robot admittance control system based on adaptive optimization method
CN109249394A (en) * 2018-09-27 2019-01-22 上海电气集团股份有限公司 Robot control method and system based on admittance control algorithms

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3752119A4 (en) * 2018-02-17 2021-12-08 President And Fellows Of Harvard College Wearable devices for protecting against musculoskeletal injuries and enhancing performance

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108227506A (en) * 2018-02-08 2018-06-29 常州普旺科技有限公司 A kind of robot admittance control system based on adaptive optimization method
CN109249394A (en) * 2018-09-27 2019-01-22 上海电气集团股份有限公司 Robot control method and system based on admittance control algorithms

Also Published As

Publication number Publication date
CN110597072A (en) 2019-12-20

Similar Documents

Publication Publication Date Title
CN110597072B (en) Robot admittance compliance control method and system
CN111281743B (en) Self-adaptive flexible control method for exoskeleton robot for upper limb rehabilitation
Zollo et al. Biomechatronic design and control of an anthropomorphic artificial hand for prosthetic and robotic applications
Miyamoto et al. A kendama learning robot based on bi-directional theory
CN106945043B (en) Multi-arm cooperative control system of master-slave teleoperation surgical robot
EP1974869A1 (en) Apparatus and method for generating and controlling the motion of a robot
CN111522243A (en) Robust iterative learning control strategy for five-degree-of-freedom upper limb exoskeleton system
CN112247962B (en) Man-machine game control method and system for upper limb wearable robot
CN112904728A (en) Mechanical arm sliding mode control trajectory tracking method based on improved approach law
CN113681543A (en) Mechanical arm zero-force control method based on model prediction
CN112782977B (en) Motor compliance control method based on BP neural network and extended observer
CN112894821A (en) Current method based collaborative robot dragging teaching control method, device and equipment
Yamawaki et al. Iterative learning of variable impedance control for human-robot cooperation
JP2558599B2 (en) Robot controller
CN113500597B (en) Multi-end teleoperation sliding mode impedance control method based on force translation mechanism
JPH07319507A (en) Robot controller
Aksman et al. Force estimation based compliance control of harmonically driven manipulators
JP3362379B2 (en) Method of detecting force in force control device
CN113031442A (en) Modularized mechanical arm dispersed robust fault-tolerant control method and system
JPH0635525A (en) Robot arm control method
Morozova et al. Robotic glove prototype: Development and simulation
Ohashi et al. Nonlinear disturbance compensation for haptic device
Hirzinger et al. Towards a new generation of multisensory light-weight robots with learning capabilities
Safavi et al. A novel MPC approach to optimize force feedback for human-robot shared control
CN115781695B (en) Touch sensor module, device and mobile operation robot control method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant