CN113110062A - Robot control system based on deep physical network - Google Patents

Robot control system based on deep physical network Download PDF

Info

Publication number
CN113110062A
CN113110062A CN202110497945.9A CN202110497945A CN113110062A CN 113110062 A CN113110062 A CN 113110062A CN 202110497945 A CN202110497945 A CN 202110497945A CN 113110062 A CN113110062 A CN 113110062A
Authority
CN
China
Prior art keywords
steps
robot
physical network
kinetic energy
mlp
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.)
Pending
Application number
CN202110497945.9A
Other languages
Chinese (zh)
Inventor
宋子豪
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.)
Hunan Taiguan Technology Co ltd
Original Assignee
Hunan Taiguan Technology Co ltd
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 Hunan Taiguan Technology Co ltd filed Critical Hunan Taiguan Technology Co ltd
Priority to CN202110497945.9A priority Critical patent/CN113110062A/en
Publication of CN113110062A publication Critical patent/CN113110062A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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
    • G05B13/042—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 in which a parameter or coefficient is automatically adjusted to optimise the performance

Landscapes

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

Abstract

The invention discloses a robot control system based on a deep physical network, which relates to the technical field of robot systems and comprises the following steps: the method comprises the steps of firstly, obtaining position and speed information, secondly, calculating kinetic energy of the position information through kinetic energy (T) GNN, calculating speed information, calculating potential energy through potential energy (V) MLP, thirdly, calculating Lagrangian quantity of the kinetic energy data, converting the kinetic energy data into current state data according to the potential energy data, and judging the current state data through a set control strategy MLP. Through system learning compensation kinematics, dynamics and disturbance model, the learning dynamics lets the nonlinear control based on neural network have physical meaning, and the learning kinematics lets the robot oneself learn the model decision to let it adapt to self condition change, the learning disturbance lets the robot learn to compensate unknown disturbance, has improved system's body perception, utilizes the priori knowledge to self structure to learn, reduces the influence that can't discern the disturbance, improves the immunity to disturbance.

Description

Robot control system based on deep physical network
Technical Field
The invention relates to the technical field of robot systems, in particular to a robot control system based on a deep physical network.
Background
The bionic robot mainly runs in complex and unknown environments, the control problem of a robot body is under-actuated control, the solution is more difficult, the requirements on perception and planning functions are higher, the existing kinetic equation numerical solver is difficult to meet most of under-actuated control problems, a deep neural network is often not stable enough on the robot, a framework based on a deep physical network is provided, the control stability of a complex system is improved, a DRL method adopts a black box control system based on the neural network to directly learn control equations from a large number of samples, for example, the acceleration of the system: the position of the particle, the velocity of the particle, and the control signal are used, so that the learning result is unstable, and is mainly reflected in the following aspects: most of the numerical solutions have no physical significance, do not exist in the real world, the system lacks ontology perception, the prior knowledge of the structure of the system is difficult to learn, the influence of disturbance cannot be distinguished, and the disturbance rejection capability is poor.
Disclosure of Invention
The present invention is directed to a robot control system based on a deep physical network, which solves the above-mentioned problems in the related art.
In order to achieve the purpose, the invention is realized by the following technical scheme: the robot control system based on the deep physical network comprises the following steps:
acquiring position and speed information;
secondly, the position information carries out kinetic energy calculation through kinetic energy (T) GNN, speed information carries out potential energy calculation through potential energy (V) MLP;
thirdly, performing Lagrange's quantity calculation on the kinetic energy data, converting the kinetic energy data into current state data according to the potential energy data, and judging the current state data by a set control strategy MLP;
step four, the set disturbance model BNN calculates the current state and compensates the control input;
and step five, determining the acceleration of the robot through the Lagrange quantity and the control input.
Further, the method comprises the following steps: according to the operation steps in the first step, 1, solving a forward model based on position, speed and control output:
Figure BSA0000241408650000021
Figure BSA0000241408650000022
further, the method comprises the following steps: according to the operation steps in the step one, obtaining the generalized acceleration
Figure BSA0000241408650000023
We can use Runge-Kutta (RK) numerical integration to get the future shape of the dynamic systemState of the art
Figure BSA0000241408650000024
And calculating a control variable tau by adopting an MPC algorithm.
Further, the method comprises the following steps: according to the operation steps in the second step, the robot passes p(s)t|ot，st-1，at-1) Obtaining xtIs represented by the low order of stWe can assume stX represents the coarse particle sizetAnd according to the same rule, substituting
Figure BSA0000241408650000025
Figure BSA0000241408650000026
Figure BSA0000241408650000027
Multi-layer perceptron (MLP) learns kinetic and potential energies:
Figure BSA0000241408650000028
VMLP＝MLP(q)。
further, the method comprises the following steps: according to the operation steps in the third step, the mechanical structure of the robot is unknown to the robot, the control problem is to solve the pattern recognition problem, the robot can learn the kinematic model of the self mechanical structure through a Graph Neural Network (GNN), the input of the GNN is a graph G ═ V, E describing the particle dynamics, and comprises a variable number of points (Vertices) and Edges (Edges), wherein the points are particles presumed by the system, the Edges are interparticle actions presumed by the system, and the output of the GNN is the kinematic characteristic of the system by using the GNN framework
Figure BSA0000241408650000031
I.e. the kinetic energy of the approximation:
Figure BSA0000241408650000032
we then have equations of control with kinematics:
Figure BSA0000241408650000033
Figure BSA0000241408650000034
further, the method comprises the following steps: according to the operation steps in step four, the non-conservative force is further decomposed into the influence of the mechanical output of the system on the whole system, and the system is subjected to the disturbance, such as friction, load and the like: τ ═ b (q) · a + ∈, where a ═ pi (a)t|st) For the control strategy, converting the Jacobian matrix B (q) into control input of the whole system, and various disturbances epsilon-p (epsilon | s) suffered by the systemt) Depending on the current state, the Jacobian matrix B (q) is learned by MLP, the disturbance model p (E | s) is learned by Bayesian Neural Network (BNN)t)。
Further, the method comprises the following steps: according to the operation steps in the step one, the Lagrangian mechanics is used for describing xtA dynamic system of generalized coordinates of
Figure BSA0000241408650000035
During the motion of the dynamic system, all particles are from state xtEnter the next state xt+1These particles may follow different paths from xtTo xt+1Each path having a lagrange component,
Figure BSA0000241408650000036
v (q) is kinetic energy and potential energy.
Further, the method comprises the following steps: according to the operation steps in the third step, the robot has various path selections, and the system state is changed by controlling the output, which comprises the steps of keeping the system in a certain unstable state, and possibly kicking the system out of the stable state by external force, so that the Lagrange quantity conforms to the non-conservative Euler-Lagrange equation:
Figure BSA0000241408650000041
further, the method comprisesThe method comprises the following steps: according to the operation steps in step three, the tau represents the non-conservative force in the system, so that
Figure BSA0000241408650000042
We obtained:
Figure BSA0000241408650000043
Figure BSA0000241408650000044
further, the method comprises the following steps: according to the operation steps in step three, the chain rule further expands the time derivative:
Figure BSA0000241408650000045
Figure BSA0000241408650000046
the invention provides a robot control system based on a deep physical network. The method has the following beneficial effects:
through system learning compensation kinematics, dynamics and disturbance model, the learning dynamics lets the nonlinear control based on neural network have physical meaning, and the learning kinematics lets the robot oneself learn the model decision to let it adapt to self condition change, the learning disturbance lets the robot learn to compensate unknown disturbance, has improved system's body perception, utilizes the priori knowledge to self structure to learn, reduces the influence that can't discern the disturbance, improves the immunity to disturbance.
Drawings
FIG. 1 is a control schematic diagram of a robot control system based on a deep physical network according to the present invention;
FIG. 2 is a schematic diagram of a GNN architecture of the robot control system based on a deep physical network according to the present invention;
fig. 3 is a schematic diagram of a control model of the robot control system based on the deep physical network.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments.
Examples of which are illustrated in the accompanying drawings, wherein like reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are illustrative and intended to be illustrative of the invention and are not to be construed as limiting the invention.
The invention will be further illustrated with reference to the following examples and drawings:
referring to fig. 1-3, the present invention provides a technical solution: the robot control system based on the deep physical network comprises the following steps:
acquiring position and speed information;
secondly, the position information carries out kinetic energy calculation through kinetic energy (T) GNN, speed information carries out potential energy calculation through potential energy (V) MLP;
thirdly, performing Lagrange's quantity calculation on the kinetic energy data, converting the kinetic energy data into current state data according to the potential energy data, and judging the current state data by a set control strategy MLP;
step four, the set disturbance model BNN calculates the current state and compensates the control input;
and step five, determining the acceleration of the robot through the Lagrange quantity and the control input.
Specifically, according to the operation steps in the step one, 1, solving a forward model based on position, speed and control output:
Figure BSA0000241408650000051
specifically, according to the operation steps in the step one, the generalized acceleration is obtained
Figure BSA0000241408650000052
We can use Runge-Kutta (RK) numerical integration to obtain the future state of the dynamic system
Figure BSA0000241408650000061
The control variable τ is calculated using the MPC algorithm.
Specifically, according to the operation steps in the second step, the robot passes p(s)t|ot，st-1，at-1) Obtaining xtIs expressed as stWe can assume stX represents the coarse particle sizetAnd according to the same rule, substituting
Figure BSA0000241408650000062
Figure BSA0000241408650000063
Figure BSA0000241408650000064
Multi-layer perceptron (MLP) learns kinetic and potential energies:
Figure BSA0000241408650000065
VMLP＝MLP(q)。
specifically, according to the operation steps in the third step, the mechanical structure of the robot is unknown to the robot, the control problem is to solve the pattern recognition problem first, the robot can learn the kinematic model of the self mechanical structure through a Graph Neural Network (GNN), the input of the GNN is a graph G ═ V, E describing the particle dynamics, and the graph G ═ V, E comprises a variable number of points (Vertices) and Edges (Edges), wherein the points are particles presumed by the system, the Edges are interparticle actions presumed by the system, and the output of the GNN is the kinematic feature of the system by using the GNN framework
Figure BSA0000241408650000066
I.e. the kinetic energy of the approximation:
Figure BSA0000241408650000067
we then have equations of control with kinematics:
Figure BSA0000241408650000068
Figure BSA0000241408650000069
specifically, according to the operation steps in step four, the non-conservative force is further decomposed into the influence of the mechanical output of the system on the whole system, and the disturbance to which the system is subjected, such as friction, load and the like: τ ═ b (q) · a + ∈, where a ═ pi (a)t|st) For the control strategy, converting the Jacobian matrix B (q) into control input of the whole system, and various disturbances epsilon-p (epsilon | s) suffered by the systemt) Depending on the current state, the Jacobian matrix B (q) is learned by MLP, the disturbance model p (E | s) is learned by Bayesian Neural Network (BNN)t)。
Specifically, according to the operation steps in the step one, the Lagrangian mechanics is used for describing xtA dynamic system of generalized coordinates of
Figure BSA0000241408650000071
During the motion of the dynamic system, all particles are from state xtEnter the next state xt+1These particles may follow different paths from xtTo xt+1Each path having a lagrange component,
Figure BSA0000241408650000072
v (q) is kinetic energy and potential energy.
Specifically, according to the operation steps in the third step, the robot has various path selections, and the system state is changed by controlling the output, which includes keeping the system in a certain non-steady state, and possibly kicking the system out of the steady state by external force, so that the lagrangian quantity conforms to a non-conservative euler-lagrangian equation:
Figure BSA0000241408650000073
specifically, according to the operation steps in step three, τ represents the non-conservative force in the system, such that
Figure BSA0000241408650000074
We obtained:
Figure BSA0000241408650000075
specifically, according to the operation steps in step three, the chain rule further expands the time derivative:
Figure BSA0000241408650000076
the above is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, many variations and modifications can be made without departing from the inventive concept of the present invention, which falls into the protection scope of the present invention.

Claims (10)

1. The robot control system based on the deep physical network is characterized by comprising the following steps:
s1, acquiring position and speed information;
s2, calculating kinetic energy through kinetic energy (T) GNN by the position information, calculating speed information through the kinetic energy, and calculating potential energy through potential energy (V) MLP;
s3, performing Lagrange' S quantity calculation on the kinetic energy data, converting the kinetic energy data into current state data according to the potential energy data, and judging the current state data by a set control strategy MLP;
s4, calculating the current state and compensating the control input by the set disturbance model BNN;
and S5, determining the acceleration of the robot through the Lagrange quantity and the control input.
2. The deep physical network based robot control system according to claim 1, comprising the steps of: according to the operation steps in the first step, 1, solving a forward model based on position, speed and control output:
Figure FSA0000241408640000011
Figure FSA0000241408640000012
3. the deep physical network based robot control system according to claim 1, comprising the steps of: according to the operation steps in the step one, obtaining the generalized acceleration
Figure FSA0000241408640000013
We can use Runge-Kutta (RK) numerical integration to obtain the future state of the dynamic system
Figure FSA0000241408640000014
And calculating a control variable tau by adopting an MPC algorithm.
4. The deep physical network based robot control system according to claim 1, comprising the steps of: according to the operation steps in the second step, the robot passes p(s)t|ot，st-1，at-1) Obtaining xtIs expressed as stWe can assume stX represents the coarse particle sizetAnd according to the same rule, substituting
Figure FSA0000241408640000021
Figure FSA0000241408640000022
Figure FSA0000241408640000023
Multi-layer perceptron (MLP) learns kinetic and potential energies:
Figure FSA0000241408640000024
VMLP＝MLP(q)。
5. the deep physical network based robot control system of claim 1,the method is characterized by comprising the following steps: according to the operation steps in the third step, the mechanical structure of the robot is unknown to the robot, the control problem is to solve the pattern recognition problem, the robot can learn the kinematic model of the self mechanical structure through a Graph Neural Network (GNN), the input of the GNN is a graph G ═ V, E describing the particle dynamics, and comprises a variable number of points (Vertices) and Edges (Edges), wherein the points are particles presumed by the system, the Edges are interparticle actions presumed by the system, and the output of the GNN is the kinematic characteristic of the system by using the GNN framework
Figure FSA0000241408640000025
I.e. the kinetic energy of the approximation:
Figure FSA0000241408640000026
we then have equations of control with kinematics:
Figure FSA0000241408640000027
Figure FSA0000241408640000028
6. the deep physical network based robot control system according to claim 1, comprising the steps of: according to the operation steps in step four, the non-conservative force is further decomposed into the influence of the mechanical output of the system on the whole system, and the system is subjected to the disturbance, such as friction, load and the like: τ ═ b (q) · a + ∈, where a ═ pi (a)t|st) For the control strategy, converting the Jacobian matrix B (q) into control input of the whole system, and various disturbances epsilon-p (epsilon | s) suffered by the systemt) Depending on the current state, the Jacobian matrix B (q) is learned by MLP, the disturbance model p (E | s) is learned by Bayesian Neural Network (BNN)t)。
7. The deep physical network based robot of claim 1A control system, comprising the steps of: according to the operation steps in the step one, the Lagrangian mechanics is used for describing xtA dynamic system of generalized coordinates of
Figure FSA0000241408640000031
During the motion of the dynamic system, all particles are from state xtEnter the next state xt+1These particles may follow different paths from xtTo xt+1Each path has a lagrange component, ≡ T-V,
Figure FSA0000241408640000037
v (q) is kinetic energy and potential energy.
8. The deep physical network based robot control system according to claim 1, comprising the steps of: according to the operation steps in the third step, the robot has various path selections, and the system state is changed by controlling the output, which comprises the steps of keeping the system in a certain unstable state, and possibly kicking the system out of the stable state by external force, so that the Lagrange quantity conforms to the non-conservative Euler-Lagrange equation:
Figure FSA0000241408640000032
9. the deep physical network based robot control system according to claim 1, comprising the steps of: according to the operation steps in step three, the tau represents the non-conservative force in the system, so that
Figure FSA0000241408640000033
We obtained:
Figure FSA0000241408640000034
Figure FSA0000241408640000035
10. the deep physical network based robot control system according to claim 1, comprising the steps of: according to the operation steps in step three, the chain rule further expands the time derivative:
Figure FSA0000241408640000036
CN202110497945.9A 2021-05-08 2021-05-08 Robot control system based on deep physical network Pending CN113110062A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110497945.9A CN113110062A (en) 2021-05-08 2021-05-08 Robot control system based on deep physical network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110497945.9A CN113110062A (en) 2021-05-08 2021-05-08 Robot control system based on deep physical network

Publications (1)

Publication Number Publication Date
CN113110062A true CN113110062A (en) 2021-07-13

Family

ID=76721319

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110497945.9A Pending CN113110062A (en) 2021-05-08 2021-05-08 Robot control system based on deep physical network

Country Status (1)

Country Link
CN (1) CN113110062A (en)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH10187654A (en) * 1996-12-27 1998-07-21 Canon Inc Motion simulation method for virtual object
CN103331756A (en) * 2013-06-04 2013-10-02 浙江工业大学 Mechanical arm motion control method
CN108319144A (en) * 2018-02-21 2018-07-24 湘潭大学 A kind of robotic tracking control method and system
CN110488608A (en) * 2019-08-14 2019-11-22 深圳市烨嘉为技术有限公司 For controling the intelligent kinetic parameter identifying approach and module of integral control system
CN110597051A (en) * 2019-09-24 2019-12-20 南京理工大学 Stewart stable platform control method based on RBF neural network
US20210089275A1 (en) * 2019-09-25 2021-03-25 Siemens Aktiengesellschaft Physics Informed Neural Network for Learning Non-Euclidean Dynamics in Electro-Mechanical Systems for Synthesizing Energy-Based Controllers

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH10187654A (en) * 1996-12-27 1998-07-21 Canon Inc Motion simulation method for virtual object
CN103331756A (en) * 2013-06-04 2013-10-02 浙江工业大学 Mechanical arm motion control method
CN108319144A (en) * 2018-02-21 2018-07-24 湘潭大学 A kind of robotic tracking control method and system
CN110488608A (en) * 2019-08-14 2019-11-22 深圳市烨嘉为技术有限公司 For controling the intelligent kinetic parameter identifying approach and module of integral control system
CN110597051A (en) * 2019-09-24 2019-12-20 南京理工大学 Stewart stable platform control method based on RBF neural network
US20210089275A1 (en) * 2019-09-25 2021-03-25 Siemens Aktiengesellschaft Physics Informed Neural Network for Learning Non-Euclidean Dynamics in Electro-Mechanical Systems for Synthesizing Energy-Based Controllers

Similar Documents

Publication Publication Date Title
Liang et al. Three‐dimensional trajectory tracking of an underactuated AUV based on fuzzy dynamic surface control
CN110275436B (en) RBF neural network self-adaptive control method of multi-single-arm manipulator
de Sousa et al. Adaptive control for mobile robot using wavelet networks
Singh et al. Stability analysis of robust adaptive hybrid position/force controller for robot manipulators using neural network with uncertainties
KR101706367B1 (en) Neural network-based fault-tolerant control method of underactuated autonomous vehicle
Qi et al. Stable indirect adaptive control based on discrete-time T–S fuzzy model
Nguyen et al. Adaptive chattering free neural network based sliding mode control for trajectory tracking of redundant parallel manipulators
Hassanein et al. Model-based adaptive control system for autonomous underwater vehicles
Duan et al. Fuzzy observer‐based tracking control of an underactuated underwater vehicle with linear velocity estimation
Xu et al. Robot trajectory tracking control using learning from demonstration method
Karami et al. Adaptive neural observer-based nonsingular terminal sliding mode controller design for a class of nonlinear systems
Yu et al. Finite-time adaptive fuzzy backstepping control for quadrotor UAV with stochastic disturbance
CN108972546A (en) A kind of robot constant force curved surface tracking method based on intensified learning
Elhaki et al. Robust amplitude-limited interval type-3 neuro-fuzzy controller for robot manipulators with prescribed performance by output feedback
Zhao et al. Experimental evaluation of human motion prediction toward safe and efficient human robot collaboration
Li et al. A learning-based model predictive control scheme and its application in biped locomotion
Ngo et al. A novel self-organizing fuzzy cerebellar model articulation controller based overlapping Gaussian membership function for controlling robotic system
Çallar et al. Hybrid learning of time-series inverse dynamics models for locally isotropic robot motion
Zhang et al. Adaptive control of discrete-time nonlinear systems using ITF-ORVFL
Rincon et al. Expressive states with a robot arm using adaptive fuzzy and robust predictive controllers
Jäntsch et al. Adaptive neural network dynamic surface control: An evaluation on the musculoskeletal robot anthrob
CN113110062A (en) Robot control system based on deep physical network
Yousuf Robust output-feedback formation control design for nonholonomic mobile robot (nmrs)
Yu et al. Robust visual servoing of robot manipulators with neuro compensation
CN117377560A (en) Object handling in case of collision avoidance using complementary constraints

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
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20210713

WD01 Invention patent application deemed withdrawn after publication