Summary of the invention
The purpose of the present invention is to provide artificial intelligence trajectory reproducing method, robot passes through running track digitization, intelligenceTechnique utilization can be reappeared.
The invention is realized in this way artificial intelligence trajectory reproducing method, robot is autonomous to learn by record target trajectoryTarget trajectory variation is practised, using high-order pose spline curve fitting optimization aim track, generates the target trajectory way of continual curvatureDiameter, robot control mechanical arm and restore the target trajectory approach.
Further, target trajectory is formed by dragging the mechanical arm of robot.
Further, the output end mounting torque sensor to slow down in the joint of the mechanical arm is passed according to the torqueSensor obtains joint moment data, carries out elasticity modeling by the joint moment data, carries out to the joint moment dataRobot records target trajectory is realized in amendment.
Further, during dragging the mechanical arm, pass through external encoder or external grating device straightenerThe position in the joint of the end and mechanical arm of tool arm.
Further, by the combination of 2D 3 D visual marker mark point and depth RGB image, target trajectory is carried outPoint cloud data, is converted to the detection template of concern by detection and in real time tracking, realizes robot records target trajectory.
Further, the step of robot autonomous learning objective track is as follows:
1) the deep learning network FCN based on self-supervision, is created；
2) data of target trajectory, the data training points comprising entire project task, are acquired；
3) it, according to the collected data, is fitted by multiple spline curve means；
4) it, is iterated using small lot gradient descent method, obtains target trajectory approach.
7, artificial intelligence trajectory reproducing method as claimed in claim 6, which is characterized in that in the step 3), lead toIt is as follows to cross the fitting of Quintic spline curve means:
Q (t)=q0+a1(t-t0)+a2(t-t0)2+a3(t-t0)3+a4(t-t0)4+a5(t-t0)5
Wherein, q (t) is target trajectory position,For speed,For acceleration, t is time, t0For initial time,t1For end time, q0For initial position, q1For end position, a1,a2,a3,a4,a5Respectively Quintic spline curve coefficients, v0ForInitial velocity, v1To terminate speed, a0 is initial acceleration, and a1 is to terminate acceleration；
Meet the following conditions:
q(t0)=q0 q(t1)=q1
Define T=t1-t0, obtain following result:
Further, it in the step 4), is iterated using X sample, wherein 1 < X < m, m is the step3) sample number being fitted in.
Compared with prior art, artificial intelligence trajectory reproducing method provided by the invention, robot to be learned by recordThe target trajectory of habit can be fitted target trajectory with the variation of autonomous learning target trajectory, be formed at robot dataThe target trajectory approach of reason in this way, robot then can control mechanical arm performance objective trajectory paths, that is, is realized to variousThe digitization of target trajectory realizes that mechanical arm uses to execute technique.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, rightThe present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, andIt is not used in the restriction present invention.
Realization of the invention is described in detail below in conjunction with specific embodiment.
The same or similar label correspond to the same or similar components in the attached drawing of the present embodiment；In description of the inventionIn, it is to be understood that if there is the orientation or positional relationship of the instructions such as term " on ", "lower", "left", "right" for based on attached drawing instituteThe orientation or positional relationship shown, is merely for convenience of description of the present invention and simplification of the description, rather than the dress of indication or suggestion meaningIt sets or element must have a particular orientation, be constructed and operated in a specific orientation, therefore describe the use of positional relationship in attached drawingLanguage only for illustration, should not be understood as the limitation to this patent, for the ordinary skill in the art, can be withThe concrete meaning of above-mentioned term is understood as the case may be.
It is preferred embodiment provided by the invention shown in referring to Fig.1.
Artificial intelligence trajectory reproducing method provided in this embodiment can make robot have the energy of study people's execution movementPower is reappeared by track artificial intelligence, will be difficult to the process experiences stated with data in the past by running track digitization, and was completedThe robot technique application high to professional skill level requirement such as spraying, welding, polishing, specific as follows:
Robot passes through record target trajectory, autonomous learning target trajectory variation, using high-order pose spline curve fittingOptimization aim track, generates the target trajectory approach of continual curvature, and robot controls mechanical arm and restores the target trajectory approach.
The artificial intelligence trajectory reproducing method of above-mentioned offer, robot, can be with by the record target trajectory to be learntThe variation of autonomous learning target trajectory, is fitted target trajectory, forms the target trajectory approach of robot dataization processing,In this way, robot then can control mechanical arm performance objective trajectory paths, that is, realize the digitization to various target trajectories,Realize that mechanical arm uses to execute technique.
In the present embodiment, the mechanical arm by dragging robot forms target trajectory, specific as follows:
In the output end mounting torque sensor that the joint of mechanical arm is slowed down, joint moment number is obtained according to torque sensorAccording to carrying out elasticity modeling, it can be achieved that high-precision dragging teaching by joint moment data, and reduce speed reducer frictional forceIt influences, joint moment data is modified, realize robot records target trajectory.
During driving machinery arm, end and the machine of mechanical arm are corrected by external encoder or external grating deviceThe position in the joint of tool arm.
Alternatively, the knot of 2D 3 D visual marker mark point and depth RGB image can also be passed through as other embodimentsIt closes, carries out the detection and tracking in real time of target trajectory, point cloud data is converted to the detection template of concern, realizes robot noteRecord target trajectory.
In the present embodiment, the step of robot autonomous learning objective track, is as follows:
1) the deep learning network FCN based on self-supervision, is created；
2) data of target trajectory, the data training points comprising entire project task, are acquired；
3) it, according to the collected data, is fitted by multiple spline curve means；
4) it, is iterated using small lot gradient descent method, obtains target trajectory approach.
In step 3), it is fitted by Quintic spline curve means as follows:
Q (t)=q0+a1(t-t0)+a2(t-t0)2+a3(t-t0)3+a4(t-t0)4+a5(t-t0)5
Wherein, q (t) is target trajectory position,For speed,For acceleration, t is time, t0For initial time,t1For end time, q0For initial position, q1For end position, a1,a2,a3,a4,a5Respectively Quintic spline curve coefficients, v0ForInitial velocity, v1To terminate speed, a0 is initial acceleration, and a1 is to terminate acceleration；
Meet the following conditions:
q(t0)=q0 q(t1)=q1
Define T=t1-t0, obtain following result:
It in step 4), is iterated using X sample, wherein 1 < X < m, m is the sample being fitted in the step 3)This number.
Small lot gradient descent method is the compromise of batch gradient descent method and stochastic gradient descent method, that is, for mSample, using X appearance come iteration, 1 < X < m can generally take X=10, certainly according to the data of sample, can adjust for weThe value of this whole X.
In order to make it easy to understand, we are unfolded using the linear regression for containing only a feature, at this point, the vacation of linear regressionIf function are as follows:
hθ(x(i))=θ1x(i)+θ0
Wherein, i=1,2 ... ..., m are expressed as sample number.
Corresponding objective function are as follows:
Batch gradient descent method is the form of most original, when referring to iteration each time, is carried out using all samplesThe update of gradient, as follows:
Local derviation is asked to above-mentioned objective function:
Wherein, i=1,2 ... ..., m are expressed as sample number, j=0, and 1, it indicates characteristic, biasing top is used herein
When each iteration, parameter is updated.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the inventionMade any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.