CN108656119A - A kind of control method of humanoid robot - Google Patents
A kind of control method of humanoid robot Download PDFInfo
- Publication number
- CN108656119A CN108656119A CN201810774005.8A CN201810774005A CN108656119A CN 108656119 A CN108656119 A CN 108656119A CN 201810774005 A CN201810774005 A CN 201810774005A CN 108656119 A CN108656119 A CN 108656119A
- Authority
- CN
- China
- Prior art keywords
- robot
- action
- model
- data
- gaussian process
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 19
- 230000009471 action Effects 0.000 claims abstract description 25
- 230000033001 locomotion Effects 0.000 claims abstract description 22
- 238000004540 process dynamic Methods 0.000 claims abstract description 15
- 238000000513 principal component analysis Methods 0.000 claims abstract description 14
- 230000009916 joint effect Effects 0.000 claims abstract description 4
- 230000009467 reduction Effects 0.000 claims description 19
- 238000012545 processing Methods 0.000 claims description 14
- 230000008569 process Effects 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 4
- 239000011159 matrix material Substances 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 3
- 230000006870 function Effects 0.000 description 6
- 150000001875 compounds Chemical class 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000002939 conjugate gradient method Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000036544 posture Effects 0.000 description 1
- 230000003252 repetitive effect Effects 0.000 description 1
- 238000012827 research and development Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
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
- B25J9/1605—Simulation of manipulator lay-out, design, modelling of manipulator
Landscapes
- Engineering & Computer Science (AREA)
- Automation & Control Theory (AREA)
- Robotics (AREA)
- Mechanical Engineering (AREA)
- Manipulator (AREA)
Abstract
A kind of control method of humanoid robot, includes the following steps：Step 1：The general joint action data feature that the humanoid robot is extracted using principal component analysis carries out dimension-reduction treatment to data；Step 2：Pass through the action model of Gaussian process Dynamic Model robot in lower dimensional space；Step 3：The humanoid robot is controlled by the action model of robot to be moved.Directly the action of robot learn by using Gaussian process dynamic model and effectively establishes action model, to the effective solution excessively high problem of robot motion data dimension, can learning training effectively be carried out to the action of robot by dimension-reduction treatment, the feature of robot motion is showed in lower dimensional space.
Description
Technical Field
The invention relates to the field of robots, in particular to a control method of a humanoid robot.
Background
In recent years, research and development of a humanoid robot have been actively carried out, and there are many problems to be overcome in the research of a humanoid robot motion model. For the research of human motion models, there are many researchers, and various human motion simulation systems have been established using humanoid robots. Human motion models are not directly applicable to humanoid robots.
And, with the increase of the degree of freedom of the robot, the control difficulty to the robot will increase correspondingly, this leads to the robot motion accuracy to be lower. In addition, the action amplitude of the robot is small, various interference factors cannot be avoided along with the action of the robot, and the factors increase the complexity of establishing a robot action model and the difficulty of learning.
Disclosure of Invention
Aiming at the problems, the invention provides a method for controlling the robot-like robot to move by establishing a robot action model after carrying out dimensionality reduction on high-dimensional data through a Gaussian process dynamic model. The control method can effectively simplify the complexity of establishing the robot action model, and the control of the robot is more stable and accurate.
In order to achieve the purpose, the invention adopts the following technical scheme:
a method for controlling a humanoid robot, comprising the steps of:
step 1: adopting principal component analysis to extract the characteristics of the whole body joint action data of the humanoid robot, and performing dimensionality reduction on the data;
step 2: establishing an action model of the robot in a low-dimensional space through a Gaussian process dynamic model;
and step 3: controlling the humanoid robot to move through an action model of the robot;
specifically, the dimension reduction processing in step 1 specifically comprises the following steps:
1-1: inputting a data setOutput setWhereinAn observation set representing the pose of the robot in continuous time,representing a low-dimensional data set after principal component analysis dimensionality reduction processing;
1-2: carrying out zero mean normalization processing on Y and calculating a covariance matrix；
Calculating the eigenvector U and the eigenvalue lambda according to the eigenvaluesIn the order of (1) toIs recombined intoWherein；
1-3: to the output setPerforming a calculation in which，。
Specifically, the specific steps of establishing the robot motion model through the gaussian process dynamic model in the step 2 are as follows:
2-1: input data Y, output parameters；
2-2: the initialization of the X is carried out,executing the dynamic model of the Gaussian process from i =1 to i = Iteration, wherein Iteration is the learning times of the dynamic model of the whole Gaussian process;
2-3: pairAndget the same time to distributeTaking its negative logarithmic probability to obtain
Starting from j =1 to j = D, the formula is executed
After solving W, ending the j loop;
executing the formula starting from M =1 to M = M
To obtainEnding the m-cycle after the optimal solution;
2-4: the i loop is ended.
The invention has the beneficial effects that: the Gaussian process dynamic model is adopted to directly learn the motion of the robot and effectively establish the motion model, so that the problem of overhigh dimension of motion data of the robot is effectively solved, the motion of the robot can be effectively learned and trained through dimension reduction processing, and the motion characteristics of the robot are displayed in a low-dimensional space. Meanwhile, the action of the robot at the next moment can be predicted, and the action prediction and planning of the robot are realized, so that the robot is controlled more effectively, and the self-control of the robot in an unmanned environment and under the condition of no instruction can be realized. The control method can be widely applied to the establishment of action models of industrial robots such as humanoid robots and mechanical arms, and has great significance for realizing more stable and accurate realization of various actions of the robots.
Drawings
Fig. 1 shows a humanoid robot controlled by the present invention.
FIG. 2 is a flow chart of principal component analysis dimension reduction processing according to the present invention.
Fig. 3 is a walking diagram of the humanoid robot of the present invention.
Fig. 4 is a walking model of the humanoid robot of the present invention.
Fig. 5 is a comparison graph of the restored data after the learning of the humanoid robot of the present invention and the original robot joint data.
Detailed Description
The present invention will be described in more detail with reference to the accompanying drawings and specific embodiments, but the present invention is not limited to the following embodiments, and can be modified and implemented as desired within a scope not departing from the gist of the present invention.
As shown in figure 1, the invention adopts an NAO robot to perform modeling processing after collecting the motion data. As shown in fig. 1, the NAO robot has 25 degrees of freedom of joints throughout the body, and the motor controlling the 25 joints can drive the NAO robot to perform various motions.
And (3) extracting the characteristics of the whole body joint action data of the NAO robot by adopting principal component analysis, and performing dimensionality reduction on the data. Through dimension reduction processing, learning can be simplified, and the calculation cost of the whole robot action data learning is reduced.
As shown in fig. 2, the vectorIs an observation set of the robot postures in continuous time,comprisesNA state, each stateIs as followsRepresents an instantaneous state value of the robot motion trajectory. Each state is defined asWhereinIs the dimension of each state. XIs a low-dimensional data set after principal component analysis and dimension reduction processing. At this time. Dimension reduction from D-dimensional data to 3D data can be realized by performing dimension reduction on principal component analysis (）。
The principal component analysis dimension reduction treatment comprises the following specific steps:
inputting a data setOutput set(ii) a Carrying out zero mean normalization processing on Y and calculating a covariance matrix(ii) a Calculating the eigenvector U and the eigenvalue lambda according to the eigenvaluesIn the order of (1) toIs recombined intoWherein(ii) a To the output setPerforming a calculation in which，。
After the principal component analysis and the dimension reduction processing, an action model of the robot is established in a low-dimensional space through a Gaussian process dynamic model. In the learning process of the Gaussian process dynamic model, the learning parameters are
Wherein,for the observation set and the low-dimensional state set,in order to be a hyper-parameter,is a scale parameter. At any timeThe state equation is:
(1)
(2)
is at the moment of timeThe state value of the space after the dimension reduction,it is the observed value corresponding to the time of day,andin order to be a noise, the noise is,andis formed by a non-linear basis functionAndas defined.
(3)
(4)
Here, ,,a and B follow a normal distribution.
Observation ofThe probability distribution with respect to X is as follows
(5)
Wherein,is formed byA defined core matrix. In GPDM, we use RBF kernel functions,
(6)
wherein,，in order to be a scale parameter,to control the parameters of the kernel function width,is a noise variance parameter.
(7)
Here, the number of the first and second electrodes,。
（8）
is defined by a linear kernel function and an RBF kernel function together.The part of the RBF kernel function is controlled,the linear kernel function portion is controlled.
During the learning process, some parameters are processed approximately. The method comprises the following specific steps:
。
observation ofThe X initial value is a known condition and can also be initialized by principal component analysis.
It needs to be estimated by learning. By the formulae (5) to (8), the compounds can be obtainedAndwhile being distributed as
(9)
The negative log probability is:
(10)
wherein,are constant parameters.
According to the formulae (5) and (7), the compounds are obtained
(11)
(12)
By minimizing the negative logarithmic probability of the formula (10) by the formulas (11) and (12), it can be estimatedThe value of (c). And (5) iterating for M times by adopting a scaling conjugate gradient method to obtain an optimal solution.
The specific algorithm of the gaussian process dynamic model is as follows:
input data Y, output parameters(ii) a Initializing X using principal component analysis dimension reduction processing, initializingThe gaussian process dynamic model is executed from i =1 to i = Iteration, where Iteration is the number of times the entire gaussian process dynamic model is learned.
Equation (10) is executed starting from j =1 to j = D, and the cycle of j is ended after W is solved.
Equation (10) is calculated from M =1 to M = MAnd ending the m-loop after the optimal solution.
The i loop is ended.
After the robot action model is established, the NAO robot can be controlled to move through the action model. As shown in fig. 3, the NAO robot is controlled to walk like a human, and data of 25 joints of the whole body in the walking process of the robot are collected simultaneously. Each frame has 25 joints of motion data, which represents the state of the robot at a certain moment in the walking process. And randomly recording data of 1000 continuous frames in the walking of the robot, and learning and modeling the walking action of the robot by using a Gaussian process dynamic model. The whole learning takes 5 minutes, and the robot walking action model is effectively learned and established.
As shown in fig. 4, the robot continuously performs repetitive motions during walking, so the models obtained by learning are also overlapped. In the 3D space, the walking motion of the robot can be visually and clearly shown. And the action state of the robot at the next moment can be predicted by using the model obtained by learning along with the change of time.
As shown in fig. 5, the NAO robot 25 joint data were restored to verify the validity of the created motion model. The restored data is highly consistent with the original data, the data stability is enhanced, and the robot can walk more stably by using the learned data. In this case, a joint (e.g., a LEIbowRoll joint) is randomly selected and visualized with the joint data, as can be seen in fig. 5, the fitting of the original joint data to the learned restored joint data is very high, which also demonstrates the effectiveness of the motion model created using the gaussian process dynamics model.
The present invention has been described in detail in order to enable those skilled in the art to understand the invention and to practice it, and it is not intended to limit the scope of the invention, and all equivalent changes and modifications made according to the spirit of the present invention should be covered by the present invention.
Claims (1)
1. A method for controlling a humanoid robot, characterized by: the method comprises the following steps:
step 1: adopting principal component analysis to extract the whole body joint action data characteristics of the humanoid robot,
performing dimensionality reduction on the data;
step 2: establishing an action model of the robot in a low-dimensional space through a Gaussian process dynamic model;
and step 3: controlling the humanoid robot to move through an action model of the robot;
the dimension reduction treatment in the step 1 comprises the following specific steps:
1-1: inputting a data setOutput setWhereinAn observation set representing the pose of the robot in continuous time,representing a low-dimensional data set after principal component analysis dimensionality reduction processing;
1-2: carrying out zero mean normalization processing on Y and calculating a covariance matrix；
Calculating the eigenvector U and the eigenvalue lambda according to the eigenvaluesIn the order of (1) toIs recombined intoWherein；
1-3: to the output setPerforming a calculation in which，；
The specific steps of establishing the robot motion model through the Gaussian process dynamic model in the step 1 are as follows:
2-1: input data Y, output parameters；
2-2: initializing X using principal component analysis dimension reduction processing, initializingExecuting the dynamic model of the Gaussian process from i =1 to i = Iteration, wherein Iteration is the learning times of the dynamic model of the whole Gaussian process;
2-3: pairAndget the same time to distributeTaking its negative logarithmic probability to obtain
Starting from j =1 to j = D, the formula is executed
After solving W, ending the j loop;
executing the formula starting from M =1 to M = M
To obtainEnding the m-cycle after the optimal solution;
2-4: the i loop is ended.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810774005.8A CN108656119A (en) | 2018-07-15 | 2018-07-15 | A kind of control method of humanoid robot |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810774005.8A CN108656119A (en) | 2018-07-15 | 2018-07-15 | A kind of control method of humanoid robot |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108656119A true CN108656119A (en) | 2018-10-16 |
Family
ID=63789081
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810774005.8A Pending CN108656119A (en) | 2018-07-15 | 2018-07-15 | A kind of control method of humanoid robot |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108656119A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109732609A (en) * | 2019-02-21 | 2019-05-10 | 电子科技大学中山学院 | Redundant degree of freedom mechanical arm motion planning method and device |
| CN111571588A (en) * | 2020-05-15 | 2020-08-25 | 深圳国信泰富科技有限公司 | Robot whole-body action planning method and system |
| CN112605973A (en) * | 2020-11-19 | 2021-04-06 | 广东省科学院智能制造研究所 | Robot motor skill learning method and system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2011032207A1 (en) * | 2009-09-15 | 2011-03-24 | The University Of Sydney | A method and system for multiple dataset gaussian process modeling |
| CN102122391A (en) * | 2010-12-13 | 2011-07-13 | 中国人民解放军国防科学技术大学 | Automatic partitioning method for motion capture data |
| CN106227037A (en) * | 2016-07-29 | 2016-12-14 | 宁波工程学院 | A kind of humanoid robot action planning parameter optimization method based on FPGA |
-
2018
- 2018-07-15 CN CN201810774005.8A patent/CN108656119A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2011032207A1 (en) * | 2009-09-15 | 2011-03-24 | The University Of Sydney | A method and system for multiple dataset gaussian process modeling |
| CN102122391A (en) * | 2010-12-13 | 2011-07-13 | 中国人民解放军国防科学技术大学 | Automatic partitioning method for motion capture data |
| CN106227037A (en) * | 2016-07-29 | 2016-12-14 | 宁波工程学院 | A kind of humanoid robot action planning parameter optimization method based on FPGA |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109732609A (en) * | 2019-02-21 | 2019-05-10 | 电子科技大学中山学院 | Redundant degree of freedom mechanical arm motion planning method and device |
| CN111571588A (en) * | 2020-05-15 | 2020-08-25 | 深圳国信泰富科技有限公司 | Robot whole-body action planning method and system |
| CN111571588B (en) * | 2020-05-15 | 2021-05-18 | 深圳国信泰富科技有限公司 | Robot whole-body action planning method and system |
| CN112605973A (en) * | 2020-11-19 | 2021-04-06 | 广东省科学院智能制造研究所 | Robot motor skill learning method and system |
| WO2022105635A1 (en) * | 2020-11-19 | 2022-05-27 | 广东省科学院智能制造研究所 | Robot motor-skill learning method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Chebotar et al. | Path integral guided policy search | |
| Juang et al. | Wall-following control of a hexapod robot using a data-driven fuzzy controller learned through differential evolution | |
| Ficuciello et al. | Synergy-based policy improvement with path integrals for anthropomorphic hands | |
| CN108656119A (en) | A kind of control method of humanoid robot | |
| Park et al. | Inverse optimal control for humanoid locomotion | |
| Castillo et al. | Reinforcement learning meets hybrid zero dynamics: A case study for rabbit | |
| Singla et al. | Realizing learned quadruped locomotion behaviors through kinematic motion primitives | |
| Bischoff et al. | Policy search for learning robot control using sparse data | |
| Zhai et al. | An intelligent control system for robot massaging with uncertain skin characteristics | |
| Jandaghi et al. | Motion dynamics modeling and fault detection of a soft trunk robot | |
| Ma et al. | An efficient robot precision assembly skill learning framework based on several demonstrations | |
| Dong et al. | Passive bimanual skills learning from demonstration with motion graph attention networks | |
| Chao et al. | Learning robotic hand-eye coordination through a developmental constraint driven approach | |
| CN117733870A (en) | Mechanical arm target capturing method, device and medium based on track prediction control | |
| Kuo et al. | Fuzzy deep deterministic policy gradient-based motion controller for humanoid robot | |
| Koropouli et al. | Generalization of Force Control Policies from Demonstrations for Constrained Robotic Motion Tasks: A Regression-Based Approach | |
| Beltran-Hernandez et al. | Learning to grasp with primitive shaped object policies | |
| Ting et al. | Locally Weighted Regression for Control. | |
| Mosbach et al. | Learning generalizable tool use with non-rigid grasp-pose registration | |
| Matsubara et al. | Spatiotemporal synchronization of biped walking patterns with multiple external inputs by style–phase adaptation | |
| Jia et al. | Learning-based feedback controller for deformable object manipulation | |
| Ariki et al. | Latent Kullback-Leibler control for dynamic imitation learning of whole-body behaviors in humanoid robots | |
| Pavlichenko et al. | Flexible-joint manipulator trajectory tracking with learned two-stage model employing one-step future prediction | |
| Ruud | Reinforcement learning with the TIAGo research robot: manipulator arm control with actor-critic reinforcement learning | |
| Luo et al. | Human-agent joint learning for efficient robot manipulation skill acquisition |
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: 20181016 |
|
| WD01 | Invention patent application deemed withdrawn after publication |