JP2019098419A - Machine learning method for learning extraction operation of powder, grain or fluid, and robot machine learning control device - Google Patents

Abstract

To provide a machine learning method for making a robot learn an operation for scooping and subdividing a defined amount only of powder, grains or fluid accommodated in bulk in a container into another container, and to provide a robot machine learning control device.SOLUTION: Provided is a machine learning method for making a robot machine learning control device learn an operation for scooping a defined amount only of powder, grains or fluid that are objects from a first container into a second container by using a multi-axial robot and extraction means disposed at a tip thereof. The machine learning method comprises: a reality learning process for making the multi-axial robot and the object perform reinforcement learning in a reality space; and a simulation learning process for making an artificial multi-axial robot which artificially simulates the multi-axial robot, and an artificial object which artificially simulates the object perform reinforcement learning in a simulation space. The artificial object includes a variation parameter comprised in the object.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a machine learning method for causing a robot to learn a motion of skimming and dividing a large volume of powder, particles or fluid contained in a container into another container by a predetermined amount, and a robot machine learning control device About.

<Description of Background Art-1> From the past, the robot was operated with the command amount calculated by calculating the three-dimensional map for each work, and the plurality of works placed randomly were taken out with the hand, and the third order for each work There is known an apparatus that performs machine learning while associating an original map and a result of retrieval operation (see Patent Document 1).

<Description of Background Art-2> Patent Document 1 observes a state quantity of a robot that takes out a workpiece from a plurality of workpieces with a hand, and a robot including output data of a three-dimensional measuring instrument that acquires a three-dimensional map for each workpiece. A state quantity observation unit, an operation result acquisition unit for acquiring the result of the extraction operation of the robot, and instruction data for instructing the robot to the extraction operation are described, and an operation amount including the instruction data in association with the state amount and the result of the extraction operation. Has a system to learn.

<Required Technology> However, unlike the work as described in Patent Document 1, the powder, the particles or the fluid have no stylistic property, and the quantitative removal operation for the powder, the particles or the fluid in an operation for partial dispensing There is a problem of being required, and a technique for solving these problems is required.

Unexamined-Japanese-Patent No. 2017-30135

<Problems of Background Art> Therefore, the present invention has been made in view of such circumstances, and an object thereof is to set powder, particles or fluid contained in a large volume in a container in another container. It is an object of the present invention to provide a machine learning method and a robot machine learning control device that cause a robot to learn an operation of taking out and dividing by the specified amount.

<Contents of claim 1> In order to achieve such an object, the present invention is grasped by the following constitution. (1) The present invention uses a multi-axis robot and a take-out means installed at the tip thereof to make a second container a predetermined amount of powder, particles or fluid as an object from the first container. A machine learning method for causing a robot machine learning control device to learn a motion of picking up, wherein the reality learning in which a reinforcement learning is performed by an object (multi-axis robot, taking means, first container, second container, and) in real space In the process and simulation space, (multi-axis robot, taking-out means, first container, pseudo-multi-axis robot simulating second container simulated, pseudo-taking-out means, first pseudo container, second pseudo Container) A simulated target that simulates an object in a simulated manner, and consists of a simulation learning process for reinforcement learning, and the simulated target has a variation parameter that the object has A machine learning method characterized.

<Contents of Claim 2> (2) In the configuration of the above (1), the present invention is characterized in that the fluctuation parameter of the powder or particles includes the repose angle of the object.

<Contents of Claim 3> (3) In the configuration of the above (2), in the configuration of the above (2), the variation parameter of the powder or particles includes a plurality of repose angles according to the temperature and / or the humidity of the object. It is characterized by

<Contents of Claim 4> (4) In the configuration of the above (1), the present invention is characterized in that the fluctuation parameter of the powder or the particles is the interparticle friction coefficient of the object.

<Contents of Claim 5> (5) In the configuration of the above (4), according to the present invention, the fluctuation parameter of the powder or the granules is a plurality of interparticle friction coefficients according to the temperature and / or humidity of the object. It is characterized by including.

<Contents of Claim 6> (6) In the configuration of the above (1), the fluctuation parameter of the fluid according to the present invention is the viscosity or the kinematic viscosity of the object.

<Contents of Claim 7> (7) The present invention is a robot machine learning control device learned by the machine learning method according to any one of the above (1) to (6).

According to the present invention, there is provided a machine learning method for causing a robot to learn an operation of skimming and dividing a large amount of powder, particles or fluid contained in a container into another container by a predetermined amount, and robot machine learning A control device can be provided.

It is a block diagram showing a notional composition of a machine learning method concerning an embodiment of the present invention.

<Description of Embodiment-1> Hereinafter, a mode for carrying out the present invention (hereinafter referred to as "embodiment") will be described in detail with reference to the attached drawings. The same components are denoted by the same reference symbols throughout the description of the embodiments.

<Description of Embodiment-2> First, a machine learning method according to the present invention will be described based on FIG.

<Objects> In the present invention, powder, granules or fluid means powders such as sugar, salt, organic acid salt, powder spices, granules of grains, beans, minced vegetables, etc., water, oil, vinegar And egg yolk, soy sauce, and other fluids, and viscous materials such as miso and eggplant which have viscosity are also included in the fluid. Since these powders, granules or fluids are not routine, when prepared at the time of food production, they need to be quantitatively measured and prepared. However, in the case of producing a small amount of various kinds of food, since the material and physical properties are diverse, it is difficult to automate the weighing, and often the human weighs.

<Multi-Axis Robot> The multi-axis robot in the real space B is not limited to a human type but refers to industrial robots such as a vertical articulated robot, a horizontal articulated robot, an orthogonal robot, and a parallel link robot. In particular, a low-output vertical articulated robot with six or more axes does not pose a danger even when operated in a human work area, and human work while maintaining the conventional work environment in a large variety of food production sites etc. Is preferable because it is possible to substitute In addition, a multi-axis robot is preferably one compatible with ROS (Robot Operating System). If it is a multi-axis robot compatible with ROS, it will be easy to change to models of different manufacturers. Multi-axis robots that do not support ROS are handled in the same way as ROS-compatible multi-axis robots by incorporating a driver layer including coordinate axis transformation matrix, motion speed transformation matrix, and motion angle transformation matrix into the robot machine learning controller described later. It becomes possible.

<Extraction Means> An extraction means for an object such as powder, particles or fluid is provided at the tip of the multi-axis robot in the present invention. The takeout means may be any one that can scoop out powder, particles, or fluid such as a spoon, a scissors and a cup. Although the taking-out means does not have a precise measuring function such as a load cell but stopping operation and weighing, it may have an instantaneous counting function of the actuator load of the multi-axis robot. By having the function of instantaneously counting the actuator loads of the multi-axis robot, it is possible to perform feedback control in consideration of the amount of the object which the extraction means tried to pick up. Further, the takeout means preferably has a quantity control function. The quantity control function is a mechanism that changes the volume and depth by changing the depth and shape of the spoon or cup with an actuator, and the mechanism that changes the area of the shield from which the object is taken out. . By having the quantity control function, it is possible to arbitrarily change the amount taken out once. Moreover, if it is a spoon or a cup, scraping operation | movement can be enabled and the quantitative property of the amount taken out once can be improved.

<Container> The container for containing the powder, the particles or the fluid in the present invention may be an existing container such as a vat, a bucket, a pail, a drum, or a bag. The first container is not particularly limited, but it is more preferable that the bottom of the container is inclined. By being inclined, the powder, the particles or the fluid are easily collected at the deepest portion, and the learning time of the operation of the takeout means can be shortened. Moreover, even if it is a container whose bottom part is a flat, it is also possible to incline the bottom part of a container by inclining the container itself beforehand without installing horizontally. A weighing device is installed immediately below the second container to obtain a measurement result 3b.

<Determined Amount> The defined amount in the present invention refers to the amount described in the recipe in the manufacture of food and the like. The takeout means may take out a predetermined amount once or take out a plurality of times to take out a predetermined amount as a total takeoff.

<Robot Machine Learning Control Device> In the robot machine learning control device according to the present invention, different mechanisms operate in the learning process and the driving process. The learning process also includes a real learning process and a simulation learning process described later. The robot machine learning control apparatus according to the present invention in the learning process operates to generate an operation sequence for a multi-axis robot (in the real learning process) or a pseudo multi-axis robot (in the simulation learning process) within a defined range. It is more preferable that the difference between the sequence generation mechanism, the surveying mechanism for obtaining the actual measurement result (in the real learning process) or the calculation survey result described later (in the simulation learning process), and the difference between the survey result and the determined amount be smaller. It consists of a reward feedback mechanism that feeds back as a reward. The robot machine learning control apparatus according to the present invention in the driving process comprises at least an operation sequence generating mechanism, but the measuring mechanism and the reward feedback mechanism are continuously operated in the driving process to obtain the measurement result in the real space B. You may learn at the same time.

<Operation Sequence Generation Mechanism> The operation sequence generation mechanism in the present invention comprises at least a limited random number generator and learning data to be described later. In addition, as described later, when a plurality of variable parameters are stored according to the environmental conditions, environmental condition collectors are also included. The limited random number generator is used when generating an operation sequence, and in the stage where the number of learnings is still small, increases the proportion in the operation sequence to prevent learning data from falling into a suboptimal solution, At the stage where the number has increased, the proportion in the operation sequence is reduced to prevent an increase in the number of times of learning unnecessarily.

<Reality learning process> The real learning process in the present invention is composed of an operation sequence generation mechanism that generates an operation sequence for a multi-axis robot in the real space B, a surveying mechanism of measured amount results, and a reward feedback mechanism. The learning results of the process are accumulated in the learning data.

<Pseudo multi-axis robot and pseudo take-out means> The pseudo multi-axis robot in the present invention is a solid model in which the multi-axis robot in the real space B is arranged in the simulation space A, and the pseudo take-out means is the same in the real space B It is a solid model in which the extracting means is arranged in the simulation space A, and it is modeled so that there is no difference in the operation of the robot between the real space B and the simulation space A for one operation sequence.

<Pseudo Object> The pseudo object in the present invention refers to a particle simulation model in which an object in the real space B is arranged in the simulation space A, and the behavior of the object by having a fluctuation parameter of the object in the real space B And the imitation of the simulated object so that it can be approximated. The pseudo object is placed in a solid model (pseudo first vessel) in which the first vessel in the real space B is placed in the simulation space A, and the simulation of the simulation space A is performed by the operation of the pseudo multiaxial robot and the pseudo extraction means. It is taken out and divided into a solid model (a pseudo second container) in which the second container in the real space B is arranged in the simulation space A.

<Variation parameter> The variation parameter in the present invention is a physical parameter of the object in the real space B, and in the case where the object is powder or granules, includes at least an angle of repose or an interparticle friction coefficient. In the case of a fluid, it includes at least a viscosity or a kinematic viscosity. In addition, as for the variation parameter, the difference between the behavior of the object and the behavior of the pseudo object is smaller when the parameter different depending on the environmental conditions is used, that is, the simulation accuracy is higher. Environmental conditions refer to at least temperature or humidity.

  <Repose angle> The repose angle in the present invention means the maximum angle of the slope which maintains stability without collapse when the object is piled up continuously from a conveyor or the like when the object is powder or grain. . Although simulation is possible by a particle simulation model in which the interparticle friction coefficient described later is taken into consideration, in the particle simulation in which the interparticle friction coefficient is not taken into account, the repose angle is not generated and the operation of scooping powder and particles is within the simulation space A. Can not be learned by the robot machine learning controller. The angle of repose may also differ depending on the environmental conditions, and it is preferable to have a plurality of angles of repose depending on temperature and / or humidity as a fluctuation parameter. It becomes possible to make a machine learning control device learn.

<Simulation Learning Process> The simulation learning process in the present invention comprises an operation sequence generation mechanism for generating an operation sequence for the pseudo multi-axis robot, a survey mechanism for operation survey results, and a reward feedback mechanism in the simulation space A. The learning results of the simulation learning process are accumulated in the learning data. In the learning in the reality learning process, it is necessary to prepare and learn an object, a container, a multi-axis robot, and the like, and in many cases, it also requires manpower to make the learning. However, when the simulation learning process is used, it becomes easy to reset the position information of the pseudo target after one learning, and the robot machine learning control device can continuously learn in the simulation space A.

<Calculation result of calculation> With the calculation result of calculation in the present invention, the pseudo target placed in the pseudo first container was scooped out and divided into the second pseudo container using the pseudo multi-axis robot and the pseudo take-out means. It shows the result of calculating the amount. The simulated object is a particle simulation model, and the other simulated multi-axis robots, simulated extraction means, simulated first container, and simulated second container are solid models. Arithmetic survey results can be obtained by computing the motion of the pseudo target, which is a particle simulation model, using a surveying mechanism according to the motion of the solid model.

<Learning Data> The learning data in the present invention is a function consisting of a multistage neural network that returns an operation sequence to a predetermined amount and a variation parameter. In the learning process, the learning data is subjected to backpropagation so as to reduce the difference between the obtained survey result and the fixed amount for the fixed amount, the fluctuation parameter, and the motion sequence, the variables in the function are We will correct (learn). In addition, the learning data when not passing through the learning process at all are generated by random numbers.

<Effect of Embodiment-1> The effect of the embodiment described above will be described. In the simulation space A of the embodiment, according to the motion sequence generated from the motion sequence generation mechanism, the pseudo multi-axis robot and the pseudo pick-up means can operate to pick up the pseudo object. The surveying mechanism fetches the calculated survey result and the learning feedback is corrected by the reward feedback mechanism according to the difference between the calculated survey result and the determined amount. As described above, by repeating the learning process in the simulation space A, learning data can be efficiently corrected without human intervention.

<Effect of Embodiment 2> In the real space B of the embodiment, the multi-axis robot and the extraction means operate in accordance with the operation sequence generated from the operation sequence generation mechanism, and the object can be skimmed out. The measured amount result obtained by the surveying mechanism is obtained, and the learning data is corrected by the reward feedback mechanism according to the difference between the measured amount result and the determined amount. As described above, by repeating the learning process in the real space B, it is possible to correct minute learning data that could not be corrected in the simulation space A.

<Advantage-3 of the embodiment> In the simulation space A of the embodiment, the fluctuation parameter of the object in the real space B is given to the pseudo object, thereby reducing the difference between the result of the operation survey and the result of the actual measurement amount. It is possible to correct learning data close to the real learning process in the process.

<Effects of the Embodiment-4> The learning data in the embodiment corrects the learning data in the simulation learning process and the real learning process to increase the number of repetitions of the learning process while reducing the time for human intervention. It is possible to make the robot machine learning control device learn efficiently the action of scooping powder, particles or fluid by a predetermined amount.

<Structure of Modified Example-1> In the above-described embodiment, the repose angle and / or the inter-particle friction coefficient may be included as a variable parameter of the powder or particles as the object. In such a case, a particle simulation model close to the real space becomes possible, and the learning efficiency is further enhanced.

<Configuration-2 of Modified Example> Further, in the above-described embodiment, a viscosity or a kinematic viscosity may be included as a variable parameter of the fluid that is the target. In such a case, a particle simulation model close to the real space becomes possible even with an object stuck to the taking-out means such as miso, for example, and the learning efficiency is further enhanced.

<Configuration-3 of Modification> In the above-described embodiment, a plurality of fluctuation parameters according to environmental conditions such as temperature and / or humidity may be included. By efficiently learning for various environmental conditions, it is possible to generate an operation sequence according to environmental conditions, even for powders and granules that differ in character depending on temperature and humidity like sugar. It becomes.

Although the present invention has been described above using the embodiment, it goes without saying that the technical scope of the present invention is not limited to the scope described in the above embodiment. It is apparent to those skilled in the art that various changes or modifications can be added to the above embodiment. It is also apparent from the scope of the claims that the embodiments added with such alterations or improvements can be included in the technical scope of the present invention.

Claims (7)

An operation of skimming a predetermined amount of powder, particles or fluid as an object from a first container into a second container using a multi-axis robot and a take-out means installed at the tip of the multi-axis robot A machine learning method for causing a robot machine learning control device to learn a robot, comprising: a reality learning process for performing reinforcement learning with the object in a real space; and a pseudo object simulating the object in a simulation space, A machine learning method comprising: a simulation learning process of reinforcement learning, wherein the pseudo target has a variation parameter of the target. The machine learning method according to claim 1,
A machine learning method characterized in that the fluctuation parameter of the powder or particles includes an angle of repose of the object. The machine learning method according to claim 2,
A machine learning method characterized in that the fluctuation parameter of the powder or particles includes a plurality of repose angles according to the temperature and / or humidity of the object. The machine learning method according to claim 1,
A machine learning method characterized in that the fluctuation parameter of the powder or particles is an interparticle friction coefficient of the object. The machine learning method according to claim 4,
A machine learning method characterized in that the fluctuation parameter of the powder or particles includes a plurality of interparticle friction coefficients according to the temperature and / or humidity of the object. The machine learning method according to claim 1,
A machine learning method characterized in that the fluctuation parameter of the fluid is the viscosity or the kinematic viscosity of the object. A robot machine learning control device learned by the machine learning method according to any one of claims 1 to 6.

Images