Device for finding a workpiece holding shape for a robot hand and method for doing so

Abstract

Device for searching for a workpiece holding shape (D) for a robot hand (12) for finding a holding shape suitable for a mounted state of the placed workpiece (WK) when holding the workpiece (WK) with the robot hand (12), wherein the device comprises: a holding form generation unit (43) which performs a holding form generation process to generate the holding form for holding the workpiece (WK) with the robot hand (12) based on the assembled state of the workpiece (WK); a hand control unit (44) that performs a gripping process to hold the workpiece (WK) with the robot hand (12) in the holding shape generated by the holding shape generation unit (43); a motion execution unit (44) that performs a motion execution process to cause the robot hand (12) to perform a predetermined movement after holding the workpiece (WK) with the robot hand (12); an evaluation evaluation unit (46) that performs an evaluation evaluation process to determine an evaluation score in order to evaluate the holding form after the motion execution unit (44) causes the robot hand (12) to perform the predetermined movement; and a repetition processing unit (48) which causes the holding shape generation unit (43), the hand control unit (44), the motion execution unit (44) and the evaluation evaluation unit (46) to execute the holding shape generation process, the gripping process, the motion execution process and the evaluation evaluation process respectively a predetermined number of times, wherein the holding shape generation unit (43) generates different holding shapes when the predetermined number of times are executed.

Description

[0001] The present invention relates to a device for searching a workpiece holding shape for a robot hand and a method for searching a workpiece holding shape, which can find a more suitable holding shape when holding a workpiece with a robot hand.

[0002] In recent years, the introduction of robots in various industries has been progressing to save labor and improve efficiency. A key prerequisite for this adoption is that a robot can securely hold a workpiece with a robot hand. Examples of technology for holding a workpiece with a robot hand include the technology disclosed in JP 2017-064910 A.

[0003] The machine learning device disclosed in JP 2017-064910 A is a machine learning device that learns an action of a robot for picking up a workpiece with a hand unit from a plurality of randomly placed workpieces, including a loose state. The machine learning device includes a state quantity observation unit that observes output data from a three-dimensional measuring instrument that measures at least one three-dimensional map for each of the workpieces, an action result acquisition unit that acquires a result of the robot's picking action of the workpiece with the hand unit, and a learning unit that receives an output from the state quantity observation unit and an output from the action result acquisition unit and learns the picking action of the workpiece.The state quantity observation unit further observes output data from a coordinate calculation unit that calculates the three-dimensional position of each of the workpieces based on the output of the three-dimensional measuring instrument. The learning unit (22) includes a reward calculation unit that calculates a reward based on a determination result of success or failure of the workpiece removal, which is the output of the action result acquisition unit, and a value function update unit that includes a value function that determines the value of the workpiece removal action and updates the value function according to the reward.According to JP 2017 - 064 910 A described above, the machine learning device can learn, without human intervention, the optimal action of the robot in taking out the workpiece that is randomly placed, including a loose state, and can take out the workpiece from the plurality of randomly placed workpieces with the hand unit.

[0004] For example, if the workpiece has a movable part or a rotating shaft, such as a stuffed toy, a doll, scissors, and a book, there is a risk that the orientation or shape of the workpiece may change after the robot hand grasps the workpiece. If the workpiece is slippery with low friction force points, such as a bolt, a can, and a bottle, or if the workpiece is prone to rotation or displacement, such as a box with an unbalanced center of gravity, there is a risk that the orientation of the workpiece may change after the robot hand grasps the workpiece.

[0005] If the orientation or shape of the workpiece changes after the robot hand grasps the workpiece, for example, during assembly, positioning difficulties arise, making assembly more difficult. For example, if the workpiece is placed at a predetermined location by the robot hand, the workpiece will no longer be placed in the assumed orientation due to the orientation change. Therefore, there is a risk that the robot hand may no longer be able to grasp the workpiece in the next step, or that the degree of freedom in the gripping method may be reduced.

[0006] The machine learning device disclosed in JP 2017 - 064 910 A described above can pick the workpiece with the hand unit from the plurality of randomly placed workpieces, but does not take into account the orientation change of the workpiece after gripping.

[0007] DE 10 2018 208 126 A1 describes a method for handling a workpiece using a removal tool, comprising the method steps of generating at least one parameter for the removal tool, determining at least one characteristic from the at least one parameter and at least one piece of workpiece information, creating a removal forecast taking into account the characteristic and at least one removal result from a removal result memory, evaluating the removal forecast, removing the workpiece if the removal forecast is positive, detecting the removal result, and storing the removal result in the removal result memory. The invention also relates to a machine for carrying out such a method.

[0008] DE 10 2018 006 248 A1 describes a workpiece holding device comprising a sensor that measures a plurality of randomly stacked workpieces in a three-dimensional space, a robot that holds the workpieces, a hand mounted on the robot, the hand being capable of holding workpieces, a holding position location calculation unit that calculates holding position location data of a position and a location for holding the workpieces by the robot based on an output of the sensor, a loading state improvement operation generation unit that generates data on an operation for improving a loading state of the workpieces by the robot based on an output of the sensor, and a robot control unit that controls the robot and the hand.wherein the robot control unit controls the robot and the hand to pick up the workpieces or perform a loading condition improvement operation based on an output of the holding position calculation unit and the loading condition improvement operation generation unit.

[0009] DE 10 2016 009 030 B4 describes a machine learning device that learns a working operation of a robot for picking up, by a hand unit, any one of a plurality of workpieces. The device comprises a state variable observation unit that observes a state variable containing position information for each of the plurality of workpieces. The device further comprises a working operation result retrieval unit that retrieves a result of a picking operation of the robot for picking up a workpiece by the hand unit, and a learning unit that includes a value function that describes a value of the picking operation of the robot. The device further comprises a decision unit that decides on command data for instructing the robot to perform a picking operation of another workpiece by referring to the value function and the state variable.The learning unit includes a reward calculation unit that calculates a reward based on the output of the operation result retrieval unit and a value function update unit that updates the value function according to the reward.

[0010] JP 2020 - 082 322 A describes a machine learning device comprising a data set storage part that acquires a data set for learning comprising one or more three-dimensional coordinate data concerning a robot hand when the robot hand successfully performs gripping via a gripping operation on a simulator for gripping any one of one or more workpieces, and two-dimensional acquired image data concerning an image of one or more workpieces acquired from a prescribed field angle by a two-dimensional imaging device on the simulator, from the simulator, and stores a plurality of data sets.Further, the learning device includes a learning part that learns a learning model that derives a three-dimensional coordinate of the robot hand in a real world from the two-dimensional captured image of one or more workpieces captured by the two-dimensional imaging device from a field angle equal to the prescribed field angle; and a learned model storage part that stores the learning model learned by the learning part.

[0011] JP 2018-118343 A describes a handling system with a camera and an image recognition section that receive the shape information of an industrial component. The handling system includes multiple types of end effectors that are in direct contact with the industrial component to perform a handling action, and a handling setting section that estimates a handling position, which is a contact position on the side of the industrial component, and selects the end effector that performs the handling action on the industrial component according to the shape information based on a learning content in a mechanical learning process.Further, the handling system includes a handling robot installed with the end effector to perform a transport action of the end effector and the industrial component; and a work planning section and an inverse kinematics calculation section that control the handling action of the handling robot to bring the selected end effector into contact with an estimated handling position.

[0012] The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a workpiece holding shape searching apparatus for a robot hand and a workpiece holding shape searching method that can find a way of holding the workpiece, which can reduce the orientation change of the workpiece after gripping when holding the workpiece with the robot hand.

[0013] This object is achieved by the subject matter of the independent claims. Preferred embodiments form the subject matter of the dependent claims.

[0014] A workpiece holding shape searching apparatus for a robot hand and a workpiece holding shape searching method according to the present invention executes a process of generating a holding shape for holding the workpiece with the robot hand based on a mounted state of the workpiece, holding the workpiece with the robot hand in the generated holding shape, causing the robot hand to thereafter perform a predetermined movement, and then determining an evaluation score to evaluate the holding shape a predetermined number of times while generating different holding shapes.With this process, the robot hand workpiece holding shape searching apparatus and method can determine the plurality of evaluation scores for various workpiece-mounted state holding shapes, and can therefore find the holding shape that can reduce the orientation change of the workpiece after gripping the workpiece in the mounted state when holding the workpiece with the robot hand. The above and additional objects, features, and advantages of the present invention will become apparent from the following detailed description and the accompanying drawings. Fig. 1 is a block diagram showing a configuration of a device for searching a workpiece holding shape for a robot hand in an embodiment. Fig. 2 is a diagram for describing an example of evaluation scores used in the workpiece holding shape searching apparatus. Fig. 3 is a flowchart showing an overall action of the workpiece holding shape searching apparatus regarding evaluation score calculation and machine learning. Fig. 4 is a flowchart showing an action of the workpiece holding shape searching device regarding calculation of the evaluation score. Fig. Figure 5 is a diagram describing how a workpiece is held by the robot hand. Fig. Figure 6 is a diagram describing feature points defined for the workpiece. Fig. 7 is a diagram for describing a method for calculating the evaluation score when there is no orientation change, as an example. Fig. Figure 8 is a diagram describing the procedure for calculating the evaluation score when there is an orientation change, as another example. Fig. Figure 9 is a diagram describing a variant for defining a section in which the evaluation score for the workpiece is determined. Fig. 10 is a diagram for describing the method for calculating the evaluation score when there is no orientation change in the restricted section, as an example. Fig. 11 is a diagram for describing the procedure for calculating the evaluation score when there is an orientation change in the restricted section, as another example. Fig. 12 is a diagram for describing a motion input screen in another variation for inputting and setting a predetermined motion to be executed after holding the workpiece with the robot hand. Fig. Figure 13 is a diagram describing another variation for finding a holding shape in a series of robot actions.

[0015] One or more embodiments of the present invention will be described below with reference to the drawings. However, the scope of the invention is not limited to the disclosed embodiments. It is noted that components with the same reference numeral in respective drawings denote the same component, and descriptions thereof will be omitted where appropriate. In this specification, when a component is referred to generically, the component is denoted by a reference numeral without a subscript, whereas when reference is made to an individual component, the component is denoted by a reference numeral with a subscript.

[0016] A workpiece holding shape search device for a robot hand in the embodiment is a device that can find a holding shape suitable for a mounted state of the workpiece placed when holding the workpiece with the robot hand. The workpiece holding shape search device includes: a holding shape generation unit that executes a holding shape generation process to generate the holding shape for holding the workpiece with the robot hand based on the mounted state of the workpiece; a hand control unit that executes a gripping process to hold the workpiece with the robot hand in the holding shape generated by the holding shape generation unit; a motion execution unit that executes a motion execution process to cause the robot hand to perform a predetermined movement after the robot hand holds the workpiece; and an evaluation evaluation unit that executes an evaluation evaluation process.to determine an evaluation score to evaluate the holding shape after the motion execution unit causes the robot hand to execute the predetermined motion; and a repetition processing unit that causes the holding shape generation unit, the hand control unit, the motion execution unit, and the evaluation evaluation unit to execute the holding shape generation process, the gripping process, the motion execution process, and the evaluation evaluation process a predetermined number of times, respectively. The holding shape generation unit generates different holding shapes upon execution of the predetermined number of times. In the present embodiment, the workpiece holding shape search device includes a machine learning unit,which performs machine learning on the holding shape based on the holding shape generated by the holding shape generation unit and the evaluation score determined by the evaluation evaluation unit. The repetition processing unit further causes the machine learning unit to perform the machine learning a predetermined number of times. Such a workpiece holding shape searching device will be described in more detail below.

[0017] Fig. 1 is a block diagram showing a configuration of a device for searching a workpiece holding shape for a robot hand in the embodiment. Fig. 2 is a diagram for describing an example of evaluation scores used in the workpiece holding shape searching apparatus.

[0018] The workpiece holding shape searching device for a robot hand (hereinafter, abbreviated as “workpiece holding shape searching device”) D in the embodiment comprises, for example, as shown in Fig. 1, a robot 1, a first workpiece detection unit 2, a second workpiece detection unit 3, a control processing unit 4, an input unit 5, an output unit 6, an interface unit (IF unit) 7, and a storage unit 8 are shown. The robot 1 is a mechanical device that is connected to the control processing unit 4 and performs predetermined work (action, movement) in response to the control of the control processing unit 4, and includes, for example, a robot body 11 and a robot hand 12. The robot body 11 is, for example, a six-axis articulated robot that is connected to the control processing unit 4 and moves in response to the control of the control processing unit 4, and has the robot hand 12 at a tip thereof.The robot hand 12 is a mechanism connected to the control processing unit 4 via the robot body 11 and capable of holding and releasing a workpiece WK in response to the control of the control processing unit 4. The robot hand 12 has, for example, as shown in FIG. Fig. 5, which will be described below, comprises a pair of first and second finger parts 121 and 122, a support part 123 which supports these first and second finger parts, and a connecting part 124 which connects the support part 123 to the tip of the robot body 11. In the Fig. In the example shown in Fig. 5, the first and second finger parts 121 and 122 are each a plate-shaped member extending in one direction. To allow at least one end of each of the first and second finger parts 121 and 122 (e.g., each tip) to be separated and connected, the support part 123 is connected to the other end of each of the first and second finger parts 121 and 122 (e.g., each base portion). The robot hand 12 can hold the workpiece WK with a predetermined gripping force by bringing one end of each of the first and second finger parts 121 and 122 close to each other, and can release the workpiece WK by separating one end of each of the first and second finger parts 121 and 122. One end of the connecting part 124 is connected to the support part 123, and the other end, although not shown, is rotatably connected to the tip of the robot body 11.The robot hand 12 can rotate with respect to the robot body 11 with the connecting part 124 as the axis of rotation.

[0019] The first workpiece detection unit 2 is a device connected to the control processing unit 4 and detects the workpiece WK held by the robot hand 12 in response to the control of the control processing unit 4. The first workpiece detection unit 2 is, for example, an image pickup device (so-called digital camera), a light detection and ranging (LiDAR), or the like. The second workpiece detection unit 3 is a device connected to the control processing unit 4 and detects the workpiece WK placed on an assembly table in response to the control of the control processing unit 4. The second workpiece detection unit 3 is analogous to the first workpiece detection unit 2 and is, for example, an image pickup device, a LiDAR, or the like.Therefore, an image pickup device, a LiDAR, or the like can be used as both the first workpiece detection unit 2 and the second workpiece detection unit 3 by making the detection direction (image pickup direction or scanning direction) changeable.

[0020] The input unit 5 is a device connected to the control processing unit 4 and inputs, for example, various commands such as a command to instruct the workpiece holding mold searching device D to start an action and various data necessary for the operation of the workpiece holding mold searching device D to the workpiece holding mold searching device D. The input unit 5 is, for example, a plurality of input switches assigned predetermined functions, a keyboard, a mouse, and the like.

[0021] The output unit 6 is a device connected to the control processing unit 4 and outputs commands and data input from the input unit 5 and an evaluation score EV determined by holding the workpiece WK and performing a predetermined movement in response to the control of the control processing unit 4. The output unit 6 is, for example, a display device such as a CRT display, a liquid crystal display (LCD), and an organic EL display, or a printing device such as a printer.

[0022] Note that the input unit 5 and the output unit 6 may comprise a touch panel. When the touch panel is included, the input unit 5 is, for example, a position input device that detects and inputs the operation position, such as a resistive film method or a capacitive method, and the output unit 6 is a display device. In this touch panel, a position input device is provided on a display surface of the display device, and one or more input content candidates that can be input are displayed on the display device. When a user touches the display position at which the input content to be input is displayed, the position is detected by the position input device, and the display content displayed at the detected position is input to the workpiece holding shape search device D as the user's operation input content.With such a touch panel, the user can intuitively understand the input process, and therefore, the device for searching a workpiece holding shape D is provided that is easy for the user to operate. The IF unit 7 is, for example, a circuit connected to the control processing unit 4 and inputting and outputting data from and to an external device in response to the control of the control processing unit 4. The IF unit 7 is, for example, an interface circuit of RS-232C, which is a serial communication system, an interface circuit using the Bluetooth standard (registered trademark), an interface circuit using the USB standard, and the like.The IF unit 7 may be, for example, a communication interface circuit for transmitting and receiving communication signals to and from an external device such as a data communication card, a communication interface circuit conforming to the IEEE 802.11 standard, and the like.

[0023] The storage unit 8 is a circuit connected to the control processing unit 4 and stores various predetermined programs and various predetermined data in response to the control of the control processing unit 4. The various predetermined programs include, for example, a control processing program and the like. The control processing program includes: a control program for controlling respective units 1 to 3 and 5 to 8 of the workpiece holding mold searching device D according to functions of each unit; a workpiece recognition program for recognizing, for example, the mounted state of the workpiece WK placed on the mounting table or the like based on a detection result of the second workpiece detection unit 3; a workpiece orientation recognition program,to detect the orientation of the workpiece WK held by the robot hand 12 based on a detection result of the first workpiece detection unit 2; a holding shape generation program for executing a holding shape generation process to generate the holding shape for holding the workpiece WK with the robot hand 12 based on the mounted state of the workpiece WK; a robot control program for executing a gripping process to hold the workpiece WK with the robot hand 12 in the holding shape generated by the holding shape generation program and executing a movement execution process to cause the robot hand 12 to perform a predetermined movement after holding the workpiece WK with the robot hand 12; an evaluation scoring program for executing an evaluation scoring process to determine the evaluation score EV to evaluate the holding shape after the robot hand 12 is controlled by the robot control program,to execute the predetermined movement; a repetition processing program for executing the holding shape generation process, the grasping process, the movement execution process, and the evaluation evaluation process a predetermined number of times; and the like. In the present embodiment, the control processing program further includes a machine learning program for executing machine learning on the holding shape based on the holding shape generated by the holding shape generation program, the predetermined movement, and the evaluation score determined by the evaluation evaluation program. The repetition processing program further causes the machine learning program to execute machine learning upon execution the predetermined number of times. The various predetermined data include data necessary for executing each program, such as score conversion information,Movement information and result information. The storage unit 8 includes, for example, a read-only memory (ROM), which is a non-volatile storage element, an electrically erasable programmable read-only memory (EEPROM), which is a rewritable non-volatile storage element, and the like. The storage unit 8 may include a relatively large-capacity hard disk drive. The storage unit 8 includes a random access memory (RAM), which is a so-called working memory for the control processing unit 4, for storing data and the like generated during the execution of the predetermined program, and the like. Then, for storing the score conversion information, movement information, and result information, respectively. The storage unit 8 functionally includes a score conversion information storage unit 81,a movement information storage unit 82 and a result information storage unit 83.,

[0024] The score conversion information storage unit 81 stores the score conversion information. The score conversion information is information used in determining the evaluation score EV for evaluating the holding shape. In the present embodiment, the evaluation score EV is determined based on the pre-movement orientation, which is the orientation of the workpiece WK before the predetermined movement is executed, and the post-movement orientation, which is the orientation of the workpiece WK after the predetermined movement is executed. More specifically, the evaluation score EV is determined based on a discrepancy amount between the post-movement orientation and the pre-movement orientation.Therefore, the score conversion information may be information that associates the discrepancy amount between the orientation before the movement and the orientation after the movement with the value of the evaluation score EV. However, in the present embodiment, since the evaluation score EV is determined by weighting with the weight WT, the score conversion information is information that associates the discrepancy amount between the orientation before the movement and the orientation after the movement with the value of the evaluation score EV before weighting. The score conversion information may be stored in the score conversion information storage unit 81 by incorporating it into the evaluation scoring program, or may be stored in the score conversion information storage unit 81 by inputting it by the user (operator) from the input unit 5. In one example, as shown in FIG. Fig. 2, when there is almost no discrepancy amount ((discrepancy amount) < 1 [mm]), the evaluation score is ev 100, and when the discrepancy amount is small (1 [mm] ≤ (discrepancy amount) < 5 [mm]), the evaluation score is ev 70, when the discrepancy amount is medium (5 [mm] ≤ (discrepancy amount) ≤ 10 [mm]), the evaluation score is ev 40, when the discrepancy amount is large (10 [mm], (discrepancy amount)), the evaluation score is ev 0, when the robot hand 12 does not grip the workpiece WK or drops it to make the detection of the discrepancy amount impossible (feature points undetectable (failure)), the evaluation score is ev -50.

[0025] The motion information storage unit 82 stores motion information representing the predetermined motion. The predetermined motion includes, for example, a velocity motion to move the robot hand 12 in a predetermined direction at a predetermined speed for a predetermined time, and / or an acceleration motion to move the robot hand 12 in a predetermined direction with a predetermined acceleration for a predetermined time, and / or a rotation motion to rotate the robot hand 12 in a predetermined angular range at a predetermined speed (or predetermined angular velocity) or a predetermined acceleration (or predetermined angular acceleration) for a predetermined time, and / or an oscillation motion to oscillate the robot hand at a predetermined amplitude and frequency (or cycle) for a predetermined time.When the predetermined motion is velocity motion, the motion information includes a predetermined direction, a predetermined speed, and a predetermined time. When the predetermined motion is acceleration motion, the motion information includes a predetermined direction, a predetermined acceleration, and a predetermined time. When the predetermined motion is rotational motion, the motion information includes a predetermined amplitude, a predetermined speed (predetermined angular velocity), or a predetermined acceleration (predetermined angular acceleration and a predetermined time). When the predetermined motion is oscillation motion, the motion information includes a predetermined amplitude, a predetermined frequency, and a predetermined time. The motion information is appropriately set in advance and stored in the motion information storage unit 82.

[0026] The result information storage unit 83 stores the result information. The result information is a mounted state, a holding shape, and a movement (action in operation) of the workpiece, which are generated (obtained) when the robot 1 is actually operated after machine learning as described below. The result information storage unit 83 associates these pieces of information and stores the information as the result information. The holding shape is represented (defined) by the gripping position and gripping force when gripping the workpiece WK with the robot hand 12.

[0027] The control processing unit 4 in the present embodiment is a circuit for controlling respective units 1 to 3 and 5 to 8 of the device for searching a workpiece holding shape D according to functions of each unit, determining the evaluation score EV in various holding shapes, and further executing machine learning. The control processing unit 4 includes, for example, a central processing unit (CPU) and peripheral circuits thereof. By executing the control processing program, the control processing unit 4 functionally includes a control unit 41, a workpiece recognition unit 42, a holding shape generation unit 43, a robot control unit 44, a workpiece orientation recognition unit 45, an evaluation evaluation unit 46, a machine learning unit 47, and a repetition processing unit 48.

[0028] The control unit 41 controls respective units 1 to 3 and 5 to 8 of the workpiece holding shape searching device D according to functions of each unit and has the overall control over the workpiece holding shape searching device D.

[0029] The workpiece recognition unit 42 recognizes, for example, the mounted state of the workpiece WK placed on the mounting table or the like based on the detection result of the second workpiece detection unit 3. For example, when the second workpiece detection unit 3 is an image pickup device, the workpiece recognition unit 42 detects an outline of the workpiece WK and recognizes the mounted state of the workpiece WK by extracting an edge from an image of the workpiece WK captured by the image pickup device (first workpiece image). The holding mold generation unit 43 executes the holding mold generation process to generate the holding mold for holding the workpiece WK with the robot hand 12 based on the mounted state of the workpiece WK.For example, the holding shape generating unit 43 generates the holding shape (gripping position and gripping force) by setting the gripping position to grip the workpiece WK with the robot hand 12 for the mounted state of the workpiece WK within the gripping range in which the robot hand 12 can grip the workpiece (range equal to or smaller than the maximum distance of the first and second finger parts 121 and 122 in the example shown in FIG. Fig. 5), and by randomly setting the gripping force for holding the workpiece WK with the robot hand 12 within the gripping force range in which the robot hand 12 can grip the workpiece.

[0030] The robot control unit 44 controls the robot 1 to enable the robot 1 to perform predetermined work. The predetermined work includes the gripping process for holding the workpiece WK with the robot hand 12 in the holding shape generated by the holding shape generation unit 43, and the movement execution process for causing the robot hand 12 to perform a predetermined movement represented by the movement information stored in the movement information storage unit 82 after holding the workpiece WK with the robot hand 12. That is, in the present embodiment, the robot control unit 44 executes the gripping process and executes the movement execution process.Note that the robot control unit 44 corresponds to an example of a hand control unit that executes the gripping process for holding the workpiece with the robot hand in the holding shape generated by the holding shape generating unit, and also corresponds to an example of a movement executing unit that executes the movement executing process for causing the robot hand to perform a predetermined movement after holding the workpiece with the robot hand.

[0031] The workpiece orientation detection unit 45 detects the orientation of the workpiece WK held by the robot hand 12 based on the detection result of the first workpiece detection unit 2. For example, when the first workpiece detection unit 2 is an image pickup device, the workpiece orientation detection unit 45 detects the outline of the workpiece WK and detects the orientation of the workpiece WK by extracting the edge from the image of the workpiece WK captured by the image pickup device (second workpiece image).In the present embodiment, since the evaluation score EV is determined as described below based on the discrepancy amount at each of the plurality of feature points SM set for the workpiece WK, the workpiece orientation detection unit 45 detects the orientation of the workpiece WK by extracting the plurality of feature points SM from the outline of the workpiece WK and determining the position of each of the plurality of feature points SM. The plurality of feature points SM are appropriately set to be able to express the orientation of the workpiece WK. For example, the feature point is set for each movable part of the workpiece WK. Note that the first workpiece detection unit 2 and the workpiece orientation detection unit 45 correspond to an example of an orientation detection unit that detects the orientation of the workpiece held by the robot hand.In the present embodiment, the workpiece orientation detection unit 45 further determines whether the robot hand 12 grips the workpiece WK. For example, the workpiece orientation detection unit 45 can determine whether the robot hand 12 grips the workpiece WK by determining whether the workpiece WK (part or all of the workpiece WK) appears in the image of the workpiece WK captured by the image pickup device. In this case, the shape of the workpiece WK is stored in advance in the storage unit 8, the image of the workpiece WK is searched for the shape of the workpiece WK, and the shape of the workpiece WK is used to determine whether the workpiece WK appears in the image of the workpiece WK.

[0032] The evaluation evaluation unit 46 executes the evaluation evaluation process to determine the evaluation score EV for evaluating the holding shape after the robot control unit causes the robot hand 12 to execute the predetermined movement. More specifically, in the present embodiment, the evaluation evaluation unit 46 determines the evaluation score EV based on the orientation before the movement and the orientation after the movement. The orientation before the movement is the orientation of the workpiece WK detected by the workpiece orientation detection unit 45 based on the detection result of the first workpiece detection unit 2 before the execution of the predetermined movement. The orientation after the movement is the orientation of the workpiece WK detected by the workpiece orientation detection unit 45 based on the detection result of the first workpiece detection unit 2 after the execution of the predetermined movement.Specifically, the evaluation judging unit 46 determines the amount of discrepancy between the orientation after movement and the orientation before movement, and converts the determined amount of discrepancy into the evaluation score EV. In the present embodiment, the plurality of feature points SM are set for determining the evaluation score EV for the workpiece WK. The weight WT in determining the evaluation score EV is set for each of the plurality of feature points SM from the viewpoint that each feature point SM has a different degree of importance according to the work purpose of the robot 1, such as a part of the workpiece WK that is difficult to change in orientation and a part of the workpiece WK that is easy to change in orientation. Each evaluation score EV is determined for each of the plurality of feature points SM using the weight WT.Therefore, in the present embodiment, for each of the plurality of feature points SM, the evaluation evaluation unit 46 determines the discrepancy amount between the position of the feature point SM in the orientation detected by the workpiece orientation detection unit 45 before the movement of the workpiece WK and the position of the feature point SM in the orientation detected by the workpiece orientation detection unit 45 after the movement of the workpiece WK, which correspond to each other. The evaluation evaluation unit 46 converts the determined discrepancy amount of the feature point SM into the evaluation score ev before weighting using the score conversion information stored in the score conversion information storage unit 81, thereby determining a plurality of temporary evaluation scores EVt at the plurality of feature points SM.Then, for each of the plurality of feature points SM, the evaluation evaluation unit 46 determines the evaluation score EV at each of the plurality of feature points SM by multiplying the temporary evaluation score EVt at the feature point SM by the weight WT corresponding to the feature point SM. Furthermore, in the present embodiment, the evaluation evaluation unit 46 determines an average of the evaluation scores EV determined at the plurality of feature points SM as the final evaluation score EVr.

[0033] The machine learning unit 47 performs machine learning on the holding shape based on the holding shape generated by the holding shape generation unit 43, the predetermined movement, and the evaluation score EV (final evaluation score EVr in the present embodiment) determined by the evaluation evaluation unit 46. In the present embodiment, reinforcement learning, known as Q-learning, is used for machine learning. Reinforcement learning maximizes future value rather than current reward. Q-learning is generally represented by the following formula 1. [Formula 1] ↵Q(s,a)←Q(s,a)+α(R(s,a)+γ maxQ(s',a')−Q(s,a)) S: State (predetermined movement) A: Action (holding form (gripping position, gripping force)) R: Reward (final rating score) a: Learning rate (0 < α ≤ 1, ex 0.1) Y: discount rate (0 < y ≤ 1, ex 0.9)

[0034] Here, Q(s, a) is a value when an action a is performed in a state s (Q value). s is a state at time t, and s' is a state at time t + 1. a is an action at time t, and a' is an action at time t + 1. State s transitions to state s' through action a. R(s, a) is a reward obtained by the state transition (reward obtained when action a is performed in state s). maxQ(s', a'), where a' is appended to the lower portion, is the currently estimated maximum value. α is a learning rate (0 < α ≤ 1) and γ is an anticipation rate (0 < γ ≤ 1).

[0035] In the present embodiment, Q-learning is used by assigning the predetermined movement to state s, assigning the holding form (grip position and grip force) to action a, and assigning the final evaluation score EVr to reward R. The learning rate α and the relaxation rate γ are appropriately set. For example, the learning rate α is set to 0, 1, 0.2, or the like, and the relaxation rate γ is set to 0.8, 0.9, or the like.

[0036] In the machine learning unit 47, a machine learning model is generated by machine learning to output the holding shape for the mounted state of the workpiece WK and the predetermined movement.

[0037] The repetition processing unit 48 executes the holding shape generation process, the grasping process, the movement execution process, and the evaluation evaluation process a predetermined number of times (number of times of evaluation, number of times of machine learning). In the present embodiment, the repetition processing unit 48 further causes the machine learning unit 47 to execute machine learning when executing the predetermined number of times. The holding shape generation unit 43 generates different holding shapes when executing the predetermined number of times. In the present embodiment, as described above, since the holding shape is randomly generated, different holding shapes are generated when executing the predetermined number of times.It should be noted that since the holding shape is randomly generated, it is possible that the same holding shape is randomly generated, which can be ignored by increasing the number of times the processes are repeatedly executed.

[0038] The control processing unit 4, the input unit 5, the output unit 6, the IF unit 7, and the storage unit 8 in the workpiece holding shape searching device D may be configured using, for example, a tower or desktop computer.

[0039] Next, the action of the device for searching a workpiece holding shape D is described. Fig. 3 is a flowchart showing an overall action of the workpiece holding shape searching apparatus regarding evaluation score calculation and machine learning. Fig. 4 is a flowchart showing the action of the workpiece holding shape searching device regarding calculation of the evaluation score. Fig. Figure 5 is a diagram describing how a workpiece is held by the robot hand. Fig. Figure 6 is a diagram describing feature points defined for the workpiece. Fig. 6A shows the workpiece WK and Fig. Figure 6B shows the first to sixth six feature points SM1 to SM6, which are used for the Fig. 6A shown workpiece WK are specified. Fig. 7 is a diagram for describing a method for calculating the evaluation score when there is no orientation change, as an example. Fig. Figure 7A shows six feature points SM1b to SM6b in the orientation before the movement on the left side of the paper and shows six feature points SM1a to SM6a in the orientation after the movement on the right side of the paper. Fig. 7B shows each temporary evaluation score (temporary evaluation score) and each evaluation score in the first to sixth feature points SM1 to SM6 and the final evaluation score of the workpiece WK in the first holding mold. Fig. Figure 8 is a diagram describing the procedure for calculating the evaluation score when there is an orientation change, as another example. Fig. Figure 8A shows six feature points SM1b to SM6b in the orientation before the movement on the left side of the paper and shows six feature points SM1a to SM6a in the orientation after the movement on the right side of the paper. Fig. 8B shows each temporary evaluation score (temporary evaluation score) and each evaluation score in the first to sixth feature points SM1 to SM6 and the final evaluation score of the workpiece WK in the second holding mold.

[0040] When the power is turned on, the workpiece holding shape searching device D with the above-described configuration performs the necessary initialization of each unit and starts the unit's operation. In the workpiece holding shape searching device D, by executing the control processing program, the control processing unit 4 functionally includes the control unit 41, the workpiece recognition unit 42, the holding shape generation unit 43, the robot control unit 44, the workpiece orientation detection unit 45, the evaluation evaluation unit 46, the machine learning unit 47, and the repetition processing unit 48.

[0041] Note that, for simplicity, the predetermined motion is a speed motion for moving the robot hand 12 in a predetermined direction at a predetermined speed for a predetermined time, and is stored in the motion information storage unit 82 as the motion information. The range of the holding shape generated by the holding shape generation unit 43 in association with the predetermined motion (gripping range and gripping force range) can be specified (limited). The number of times the repetition processing unit 48 repeats the processes is also stored in the storage unit 8.

[0042] First, the workpiece WK is placed in a predetermined mounted state on the mounting table or the like. For example, the workpiece WK is placed in the predetermined mounted state by the user (operator). Alternatively, for example, another robot is prepared, and the workpiece WK is placed in the predetermined mounted state by the other robot. Therefore, the second workpiece detection unit 3 and the workpiece recognition unit 42 can be omitted in the process of searching for a holding shape of the workpiece WK and machine learning. Note that the predetermined mounted state is stored in advance in the storage unit 8.

[0043] Afterwards, Fig. 3, the workpiece holding shape searching device D executes the holding shape generation process for generating the holding shape for holding the workpiece WK with the robot hand 12 based on the mounted state of the workpiece WK (S1) with the holding shape generation unit 43 of the control processing unit 4. More specifically, the holding shape generation unit 43 randomly generates the holding shape of the workpiece WK (gripping position and gripping force) for the mounted state of the workpiece WK. In one example, the workpiece WK is a humanoid doll that is in Fig. 6A is shown.

[0044] Next, with the robot control unit 44 of the control processing unit 4, the workpiece holding shape searching device D executes the gripping process for holding the workpiece WK with the robot hand 12 in the holding shape generated by the holding shape generating unit 43 in process S1 (S2). For example, as shown in Fig. 5, the workpiece WK of the dummy is gripped by the robot hand 12 on a body of the dummy. Next, the device for searching a workpiece holding form D detects the workpiece WK gripped in process S2 (S3) with the first workpiece detection unit 2. For example, the image of the workpiece WK is acquired by the first workpiece detection unit 2 for the workpiece WK of the Fig. 6A.

[0045] Next, with the workpiece orientation detection unit 45 of the control processing unit 4, the workpiece holding shape search device D determines whether the robot hand 12 is gripping the workpiece WK in process S2 (S4). Specifically, the workpiece orientation detection unit 45 determines whether the robot hand 12 is gripping the workpiece WK based on the detection result of the workpiece WK before the movement in process S3. Specifically, when the workpiece WK is detected before the movement in process S3, in this example, when the workpiece WK (part or all of the workpiece WK) appears in the image of the workpiece WK, the workpiece orientation detection unit 45 determines that the robot hand 12 is gripping the workpiece WK (Yes), and the workpiece holding shape search device D then executes process S5.On the other hand, when the workpiece WK is not detected before the movement at process S3, in this example, when the workpiece WK does not appear in the image of the workpiece WK, the evaluation unit 46 determines that the robot hand 12 does not grip the workpiece WK (No), and the workpiece holding shape searching device D then executes process S9.

[0046] In the process S5, the workpiece holding shape searching device D executes the movement execution process with the robot control unit 44 of the control processing unit 4 to cause the robot hand 12 to execute the predetermined movement represented by the movement information stored in the movement information storage unit 82.

[0047] Next, with the first workpiece detection unit 2, the device for searching a workpiece holding shape D detects the workpiece WK after the execution of the predetermined movement at process S5 analogously to process S3 (S6).

[0048] Next, the workpiece holding shape search device D determines, with the workpiece orientation detection unit 45, whether the robot hand 12 is gripping the workpiece WK, analogously to process S4 (S7). As a result of this determination, if the robot hand 12 is gripping the workpiece WK (Yes), the workpiece holding shape search device D then executes process S8. On the other hand, if the robot hand 12 is not gripping the workpiece WK (No), the workpiece holding shape search device D then executes process S9 as a result of the determination.

[0049] In the process S8, the workpiece holding shape searching device D determines the evaluation score EV when the grip is successful (final evaluation score EVr in the present embodiment) with the evaluation evaluation unit 46 of the control processing unit 4, and then executes the process S10.

[0050] More precisely, extracted as in Fig. 4, the workpiece holding die search device D, firstly, includes the workpiece orientation detection unit 45, which determines the pre-movement orientation of the workpiece WK based on the detection result of the first workpiece detection unit 2 (S21) detected in process S3. Specifically, by extracting the edge from the image of the workpiece WK generated in process S3, the workpiece orientation detection unit 45 detects the outline of the workpiece WK, extracts the plurality of feature points SM from the extracted outline of the workpiece WK, and detects the orientation of the workpiece WK by determining the position of each of the plurality of feature points SM.

[0051] For example, for the workpiece WK the Fig. 6A shown doll as in Fig. 6B shows a left and a right end of the head as the first and second feature points SM1 and SM2, respectively, a left and a right finger as the third and fourth feature points SM3 and SM4, respectively, and a left and a right toe as the fifth and sixth feature points SM5 and SM6, respectively. As shown on the left side of the paper, Fig. 7A and Fig. As shown in Fig. 8A, the position of each of the first to sixth six feature points SM1b to SM6b is determined by the workpiece orientation detection unit 45 from the image of the workpiece as the orientation before the movement. It should be noted that in Fig. 7, Fig. 8, Fig. 10 and Fig. 11 a superscript b is further added to the feature point SM of the orientation before the movement and a superscript a is further added to the feature point SM of the orientation after the movement.

[0052] Next, with the workpiece orientation detection unit 45, the workpiece holding shape searching device D extracts the orientation after the movement of the workpiece WK based on the detection result of the first workpiece detection unit 2 detected in process S5, analogously to S21 (S22).

[0053] For example, as on the right side of the paper, Fig. 7A and Fig. 8A, the position of each of the first to sixth six feature points SM1a to SM6a is determined by the workpiece orientation detection unit 45 from the image of the workpiece WK as the orientation after the movement. Fig. 7A shows the orientation before the movement and the orientation after the movement when the robot hand 12 holds the workpiece WK in the first holding shape and performs the speed movement. Fig. 8A shows the orientation before the movement and the orientation after the movement when the robot hand 12 holds the workpiece WK in the second holding shape, which is different from the first holding shape, and performs the speed movement. The position of the feature point SM is represented by, for example, the pixel position. The first workpiece detection unit 2 generates the image of the workpiece WK at the same detection position (image pickup position) between the generation of the image of the workpiece WK for detecting the orientation before the movement (when executing process S3) and the generation of the image of the workpiece WK for detecting the orientation after the movement (when executing process S5).

[0054] Next, the workpiece holding shape search device D determines the discrepancy amount of the post-movement orientation from the pre-movement orientation with the evaluation unit 46 (S23). Specifically, for each of the plurality of feature points SM, the evaluation unit 46 determines the discrepancy amount between the position of the feature point SMb determined in process S21 in the orientation before the workpiece WK moves and the position of the feature point SMa determined in process S22 in the orientation after the workpiece WK moves, which correspond to each other.

[0055] Next, the workpiece holding shape searching device D determines the evaluation score EV based on the discrepancy amount determined in process S23 (S24) with the evaluation evaluation unit 46. Specifically, for each of the plurality of feature points SM, the evaluation evaluation unit 46 determines the temporary evaluation score EVt of the feature point SM by converting the discrepancy amount of the feature point SM into the evaluation score ev before weighting using the score conversion information stored in the score conversion information storage unit 81, and determines the evaluation score EV at each of the plurality of feature points SM by multiplying the temporary evaluation score EVt of the feature point SM by the weight WT corresponding to the feature point SM.

[0056] Then, with the evaluation judging unit 46, the workpiece holding shape searching device D determines the average of each evaluation score EV determined for each of the plurality of feature points SM as the final evaluation score EVr (S25) and ends process S8.

[0057] In the Fig. For example, in the example shown in Figure 7A, the first to sixth six feature points SM1 to SM6 each have no discrepancy between the orientation after the movement and the orientation before the movement. Fig. 2 is as shown in Fig. 7B shows each of the temporary evaluation scores (temporary evaluation scores) EVt1 to EVt6 at the first to sixth feature points SM1 to SM6 100, respectively. In the Fig. 7B and Fig. In the examples shown in FIG. 8B, the first and second weights WT1 and WT2 at the first and second feature points SM1 and SM2 are set to 1.0, respectively, and each of the third to sixth weights WT3 to WT6 at the third to sixth feature points SM3 to SM6 is set to 0.7. Therefore, the evaluation scores EV1 and EV2 at the first and second feature points SM1 and SM2 are determined as 100 (= 100 × 1.0), respectively, and each of the evaluation scores EV3 to EV6 at the third to sixth feature points SM3 to SM6 is determined as 70 (= 100 × 0.7). Therefore, the final evaluation score EVr in the first holding form is determined as 80 (= (100 + 100 + 70 + 70 + 70 + 70) / 6).

[0058] In the Fig. In the example shown in Figure 8A, however, each of the first to sixth six feature points SM1 to SM6 has a discrepancy between the orientation after the movement and the orientation before the movement. Fig. 2 is as shown in Fig. 8B, each of the temporary evaluation scores (temporary evaluation scores) EVt1 and EVt2 at the first and second feature points SM1 and SM2, respectively, is 70, the temporary evaluation score EVt3 at the third feature point SM3 is 40, the temporary evaluation score EVt4 at the fourth feature point SM4 is 70, the temporary evaluation score EVt5 at the fifth feature point SM5 is 40, and the temporary evaluation score EVt6 at the sixth feature point SM6 is zero. The evaluation scores EV1 and EV2 at the first and second feature points SM1 and SM2 are 70 (= 70 × 1.0), the evaluation score EV3 at the third feature point SM3 is 28 (= 40 × 0.7), the evaluation score EV4 at the fourth feature point SM4 is 49 (= 70 × 0.7), the evaluation score EV5 at the fifth feature point SM5 is 28 (= 40 × 0.7), and the evaluation score EV6 at the sixth feature point SM6 is 0 (= 0 × 0.7).Therefore, the final evaluation score EVr in the second holding form is determined as 40 (≈ (70 + 70 + 28 + 49 + 28 + 0) / 6).

[0059] When comparing the first holding shape and the second holding shape, the final evaluation score EVr;80 in the first holding shape is higher (larger) than the final evaluation score EVr;40 in the second holding shape. Therefore, the final evaluation score EVr;80 in the first holding shape and the final evaluation score EVr;40 in the second holding shape are output from the output unit 6, and by referring to the scores, the user can determine that the first holding shape is a holding shape that can reduce the orientation change of the workpiece after gripping more than the second holding shape.

[0060] Since the workpiece holding shape searching device D determines the final evaluation score EVr for various holding shapes as described below, the final evaluation score EVr for each of the various holding shapes is output from the output unit 6, and by referring to the score, the user can find the holding shape which can reduce the orientation change of the workpiece after gripping.

[0061] With reference to Fig. 3, in process S9, the workpiece holding shape searching device D determines the evaluation score with the evaluation unit 46 when the gripping fails as in process S8, and then executes process S10. Fig. 2 and Fig. 6, the evaluation scoring unit 46 determines the evaluation scores EV1 to EV6 at the first to sixth feature points SM1 to SM6, respectively, by setting each of the temporary evaluation scores EVt1 to EVt6 at the first to sixth feature points SM1 to SM6, respectively, to -50 and multiplying the temporary evaluation scores EVt1 to EVt6 by the weights WT1 to WT6 of the first to sixth feature points SM1 to SM6, respectively, and determines the final evaluation score EVr;-40 (= (-50 - 50 - 35 - 35 - 35 - 35) / 6) in the holding form in the case of failure by determining the average thereof.

[0062] In process S10, with the machine learning unit 47 of the control processing unit 4, the workpiece holding shape searching device D performs machine learning by reinforcement learning of Q-learning based on the holding shape generated by the holding shape generating unit 43 in process S1, the predetermined motion, and the evaluation score EV (final evaluation score EVr in the present embodiment) determined by the evaluation evaluation unit 46 in process S8 or process S9. Specifically, the machine learning unit 47 performs machine learning by assigning the predetermined motion to the state s of the Q-learning formula 1, assigning the holding shape to the action a, and assigning the final evaluation score EVr to the reward R.

[0063] Next, the workpiece holding shape search device D determines, with the repetition processing unit 48 of the control processing unit 4, whether execution of the predetermined number of times is completed (S11). As a result of this determination, if the execution of the predetermined number of times is not completed (No) to repeat the holding shape generation process, the gripping process, the movement execution process, and the evaluation evaluation process, the repetition processing unit 48 returns the process to process S1. As a result of the determination, if the execution of the predetermined number of times is completed (Yes), the repetition processing unit 48 terminates this process.

[0064] In each such process, various holding shapes are evaluated, and machine learning is performed for a speed movement. As described above, since the workpiece holding shape search device D is configured to output the final evaluation score EVr in each of the various holding shapes to the output unit 6, the user can find the holding shape that can reduce the orientation change of the workpiece after gripping by referring to the final evaluation score EVr in each of the various holding shapes.

[0065] Then, as required for various speed movements, moving in various directions at various speeds for various times, the workpiece holding shape search device D actually evaluates various holding shapes and performs machine learning. Similarly, as required for various acceleration movements, moving in various directions at various accelerations for various times, the workpiece holding shape search device D evaluates various holding shapes and performs machine learning.Similarly, as required for various rotational movements to rotate in various angular ranges at various speeds (or various angular velocities) or various accelerations (or various angular accelerations) for various times, the device for searching a workpiece holding shape D evaluates various holding shapes and performs machine learning. Similarly, as required for various oscillatory movements to oscillate in various amplitudes at various frequencies for various times, the device for searching a workpiece holding shape D evaluates various holding shapes and performs machine learning.

[0066] In the process after machine learning, when the movement (action) of the robot 1 is specified and the workpiece WK is placed on the mounting table or the like, the mounted state of the workpiece WK is detected by the second workpiece detection unit 3 and the workpiece recognition unit 42, the holding shape for the mounted state of the workpiece WK and the specified movement of the robot 1 is determined by the machine learning model, the robot hand 12 is controlled to hold the workpiece in the determined holding shape, and the specified movement of the robot 1 is executed.

[0067] As described above, the workpiece holding shape searching device D in the embodiment and the workpiece holding shape searching method implemented therein can determine the final evaluation score EVr in the present embodiment by executing the holding shape generation process, the gripping process, the movement execution process, and the evaluation evaluation process a predetermined number of times, based on the plurality of evaluation scores EV for various holding shapes for the mounted state of the workpiece WK. Therefore, by comparing each evaluation score EVr, the user can find the holding shape that can reduce the orientation change of the workpiece WK after gripping for the mounted state of the workpiece WK when holding the workpiece WK with the robot hand 12.

[0068] Since the workpiece holding shape searching apparatus D and the workpiece holding shape searching method include the machine learning unit 47 and perform machine learning on the machine learning unit 47, when holding the workpiece WK with the robot hand 12 for the mounted state of the workpiece WK, the holding shape that can reduce the orientation change of the workpiece WK after gripping can be output from the machine learning unit 47 after machine learning.

[0069] The workpiece holding shape searching apparatus D and the workpiece holding shape searching method can determine the evaluation score by the amount of orientation change between the orientation after the movement and the orientation before the movement.

[0070] The workpiece holding shape searching apparatus D and the workpiece holding shape searching method set the weight WT for each of the plurality of feature points SM set for the workpiece WK, and determine each evaluation score EV using the weight WT for each of the plurality of feature points SM. Therefore, by making the weight WT for the feature point SM that hardly allows orientation change larger than the weight WT for the feature point SM that easily allows orientation change, the holding shape that can reduce the orientation change of the workpiece WK after gripping can be found while taking into account the permissibility of the orientation change.

[0071] The workpiece holding shape searching apparatus D and the workpiece holding shape searching method can determine the overall evaluation score for the plurality of feature points SM by determining the final evaluation score EVr.

[0072] Note that in the embodiment, by referring to and comparing each final evaluation score EVr in different holding shapes, the user finds the holding shape that can reduce the orientation change of the workpiece after gripping. However, a configuration may be used to search for the holding shape that can reduce the orientation change of the workpiece after the workpiece holding shape search device D grips the workpiece. In this case, the workpiece holding shape search device D further functionally includes a holding shape search unit in the control processing unit 4.The holding shape searching unit compares the plurality of evaluation scores in the plurality of holding shapes determined by executing the predetermined number of times, compares the plurality of final evaluation scores EVr in the above description, and searches the holding shape which is the highest evaluation score (highest final evaluation score EVr) among the plurality of holding shapes.

[0073] In the embodiment, the discrepancy amount of the orientation after movement from the orientation before movement is determined for the entire workpiece WK, and the evaluation score is determined. However, the discrepancy amount between the orientation after movement and the orientation before movement for a part of the workpiece WK may be determined, and the evaluation score may be determined. In such a workpiece holding shape searching device D, the target area for determining the evaluation score is input to the input unit 5, and the evaluation evaluation unit 46 determines the evaluation score in the target area input by the input unit 5. That is, the evaluation evaluation unit 46 determines the evaluation score based on the discrepancy amount between the orientation before movement and the orientation after movement at the feature points within the target area input by the input unit 5.With this operation, the orientation change of the workpiece WK can be evaluated by focusing on a part of the workpiece WK, and the orientation change in other parts of the workpiece WK except the part described above can be neglected.

[0074] Fig. 9 is a diagram for describing a variant for defining a section of the workpiece for determining the evaluation score. Fig. Figure 9A is a diagram for describing the restricted section and shows the orientation before the movement of the workpiece WK. Fig. Figure 9B shows the orientation after the movement of the workpiece WK when there is no change in the orientation in the restricted section with respect to the Fig. 9A is present before the movement. Fig. Figure 9C shows the orientation after the movement of the workpiece WK, when there is no change in the orientation in the restricted section with respect to the Fig. 9A is present before the movement.

[0075] Fig. 10 is a diagram for describing the method for calculating the evaluation score when there is no orientation change in the restricted section, as an example. Fig. 10A shows six feature points SM1b to SM6b in the Fig. 9A shown orientation before movement on the left side of the paper and shows six feature points SM1a to SM6a in the Fig. 9B shown alignment after movement on the right side of the paper. Fig. 10B shows the temporary evaluation score (temporary evaluation score) and the evaluation score in each of the first to sixth feature points SM1 to SM6 and the final evaluation score of the workpiece WK in the third holding mold. Fig. 11 is a diagram for describing the procedure for calculating the evaluation score when there is an orientation change in the restricted section, as another example. Fig. 11A shows six feature points SM1b to SM6b in the Fig. 9A shown orientation before movement on the left side of the paper and shows six feature points SM1a to SM6a in the Fig. 9C shown alignment after movement on the right side of the paper. Fig. 11B shows the temporary evaluation score (temporary evaluation score) and the evaluation score in each of the first to sixth feature points SM1 to SM6 and the final evaluation score of the workpiece WK in the fourth holding form, which is different from the third holding form.

[0076] For example, when the first workpiece detection unit 2 is an image pickup device, after the execution of the process S3 and before the execution of the process S4, the device for searching a workpiece holding shape D outputs the image of the workpiece WK generated by the image pickup device as the first workpiece detection unit 2 to the output unit 6 having the control processing unit 4. As shown in Fig. As shown in Fig. 9A, the user refers to the image of the workpiece WK and inputs the target area AR in the workpiece WK from the input unit 5. The workpiece holding shape searching device D receives and detects the input of the target area AR in the workpiece WK with the control processing unit 4. In the Fig. In the example shown in Fig. 9A, the head of the workpiece WK is input as the target area AR, and the first and second feature points SM1 and SM2 corresponding to the input target area AR are set from the first to sixth feature points SM1 to SM6 of the workpiece WK.

[0077] Then, in process S8, process S21 and process S22 are executed as described above. For example, for the Fig. 9A shown workpiece WK at process S21 as on the left side of the paper from Fig. 10A and Fig. 11A, the position of each of the first to sixth six feature points SM1b to SM6b is determined by the workpiece orientation detection unit 45 from the image of the workpiece WK as the orientation before the movement. In process S22, for the Fig. 9B shown workpiece WK as shown on the right side of the paper in Fig. 10A, the position of each of the first to sixth six feature points SM1a to SM6a is determined by the workpiece orientation detection unit 45 from the image of the workpiece WK as the orientation after the movement. For the Fig. 9C shown workpiece WK as shown on the right side of the paper in Fig. As shown in Fig. 11A, the position of each of the first to sixth six feature points SM1a to SM6a is determined by the workpiece orientation detection unit 45 from the image of the workpiece WK as the orientation after the movement.

[0078] It should be noted that in the above description, the position of each of all feature points SM in the workpiece WK is determined at process S21 and process S22, respectively, but only the position of each of the first and second feature points SM1 and SM2 corresponding to the target area AR can be determined.

[0079] In process S23 of process S8, the workpiece holding shape search device D determines, with the evaluation unit 46, the discrepancy amount between the orientation after the movement and the orientation before the movement of the workpiece WK in the target area AR. More specifically, for each of the plurality of feature points SM corresponding to the target area AR, the evaluation unit 46 determines the discrepancy amount between the position of the feature point SMb in the orientation before the movement of the workpiece WK determined in process S21 and the position of the feature point SMa in the orientation after the movement of the workpiece WK determined in process S22, which correspond to each other.

[0080] Next, in process S24 of process S8, the workpiece holding shape searching device D determines, with the evaluation judging unit 46, the evaluation score EV from the discrepancy amount determined in process S23. More specifically, for each of the plurality of feature points SM corresponding to the target area AR, the evaluation judging unit 46 determines the temporary evaluation score EVt of the feature point SM by converting the discrepancy amount of the feature point SM into the evaluation score ev before weighting using the score conversion information stored in the score conversion information storage unit 81, and determines the evaluation score EV in each of the plurality of feature points SM corresponding to the target area AR by multiplying the temporary evaluation score EVt at the feature point SM by the weight WT corresponding to the feature point SM.

[0081] In process S25 of process S8, the workpiece holding shape searching device D determines, with the evaluation evaluation unit 46, the average of each evaluation score EV determined for each of the plurality of feature points SM corresponding to the target area AR as the final evaluation score EVr. Fig. 10A, for example, the first and second two feature points SM1 and SM2 corresponding to the Fig. 9A and Fig. 9B, there is no discrepancy between the alignment after the movement and the alignment before the movement. In the Fig. 2 is as shown in Fig. 10B, each of the temporary evaluation scores (temporary evaluation scores) EVt1 and EVt2 at the first and second feature points SM1 and SM2, respectively, is 100. Each of the evaluation scores EV1 and EV2 at the first and second feature points SM1 and SM2, respectively, multiplied by the first and second weights WT1 and WT2 of the first and second feature points SM1 and SM2, respectively, is determined as 100 (= 100 × 1.0), and the final evaluation score EVr in the third holding form is determined as 100 (= (100 + 100) / 2). In the Fig. In the example shown in Figure 11A, however, each of the first and second two feature points SM1 and SM2, which correspond to the Fig. 9A and Fig. 9B, a discrepancy between the alignment after the movement and the alignment before the movement. In the Fig. 2 shown value conversion information are as in Fig. 11B, the temporary evaluation scores (temporary evaluation scores) EVt1 and EVt2 at the first and second feature points SM1 and SM2 are 40 and 0, respectively. The evaluation scores EV1 and EV2 at the first and second feature points SM1 and SM2, multiplied by the first and second weights WT1 and WT2 of the first and second feature points SM1 and SM2, are determined to be 40 (= 40 × 1.0) and 0 (= 0 × 1.0), respectively. The final evaluation score EVr in the fourth holding shape is determined to be 20 (= (40 + 0) / 2).

[0082] When comparing the third holding shape and the fourth holding shape, the final evaluation score EVr; 100 in the third holding shape is higher (larger) than the final evaluation score EVr; 20 in the fourth holding shape. Therefore, the final evaluation score EVr; 100 in the third holding shape and the final evaluation score EVr; 20 in the fourth holding shape are output from the output unit 6, and by referring to the scores, the user can determine that the third holding shape is a holding shape that can reduce the orientation change of the workpiece after gripping more than the fourth holding shape.

[0083] In the embodiment, the discrepancy amount between the orientation before movement and the orientation after movement was determined from the discrepancy amount in the position of the feature point SM of the workpiece WK, but this method is not limitative. For example, the discrepancy amount can be determined by detecting the workpiece WK as point group data with a LiDAR or the like, and comparing the point group representing the workpiece WK in the orientation before movement with the point group representing the workpiece WK in the orientation after movement to determine a match ((discrepancy amount) = (match)). In this case, the match is assigned to the evaluation score ev before weighting, and the match is converted to the evaluation score ev before weighting.

[0084] In the embodiment, the movement information was appropriately set in advance and stored in the movement information storage unit 82, but the device for searching a workpiece holding shape D may be configured to allow user input. In such a device for searching a workpiece holding shape D, the predetermined movement is input into the input unit 5, and the inputted predetermined movement is stored in the movement information storage unit 82 as the movement information. When each of the Fig. 3 and Fig. 4, the predetermined movement inputted from the input unit 5 is executed by the robot control unit 44, which is an example of a movement execution unit.

[0085] Fig. 12 is a diagram for describing a motion input screen in another variation for inputting and setting the predetermined motion to be executed after holding the workpiece with the robot hand.

[0086] When entering the predetermined movement from the input unit 5, for example, the Fig. 12 is displayed on the output unit 6, and the user uses the motion input screen 9 to input the predetermined motion from the input unit 5.

[0087] This motion input screen 9 is provided with an input area for each type of motion, and includes a first input area 91-1 for inputting the speed motion, a second input area 91-2 for inputting the acceleration motion, a third input area 91-3 for inputting the rotation motion, and a fourth input area 91-4 for inputting the oscillation motion. The first input area 91-1 includes a first checkbox 92-1 for determining whether to input the speed motion and a first content setting area 93-1 for inputting the content of the speed motion. The first content setting area 93-1 includes four radio buttons 931-1 to 931-4 for setting the speed to fast speed, medium speed, slow speed, and low speed, respectively.User-defined, a speed input field 932 for entering a numerical value when the speed is set to user-defined, and a time input field 933 for entering the time of the speed movement. A default value is appropriately set for high speed, medium speed, and low speed, respectively. When inputting the speed movement as the predetermined movement, the user activates the first check box 92-1 using the input unit 5, and when setting the speed to the medium speed, the user selects the radio button 931-2 for the medium speed using the input unit 5 and enters a numerical value in the time input field 933 using the input unit 5.With this operation, the speed motion is input as the predetermined motion, and the content is input and stored in the motion information storage unit 82 as the motion information. Similarly, the second input area 91-2 includes a second checkbox 92-2 for specifying whether to input the acceleration motion, and a second content setting area 93-2 for inputting the content of the acceleration motion. The third input area 91-3 includes a third checkbox 92-3 for specifying whether to input the rotation motion, and a third content setting area 93-3 for inputting the content of the rotation motion. The fourth input area 91-4 includes a fourth checkbox 92-4 for specifying whether to input the oscillation motion, and a fourth content setting area 93-4 for inputting the content of the oscillation motion.

[0088] In the embodiment, after the machine learning, the machine learning unit 47 may perform machine learning using the result obtained from the operation of the robot 1.

[0089] For example, in the device for searching a workpiece holding shape D during the operation after machine learning, the machine learning unit 47 may further perform machine learning at a learning rate lower than a learning rate during machine learning. In this case, to detect the pre-movement orientation before the operation action, the workpiece WK is detected by the first workpiece detection unit 2, and to detect the post-movement orientation after the operation action, the workpiece WK is detected by the first workpiece detection unit 2.Such a device for searching a workpiece holding shape D can reflect the result during the operation to the machine learning unit 47 and can reduce the impact on the machine learning unit 47 by reducing the learning rate even when a relatively unfavorable holding shape is executed, and therefore, deterioration of reducing the orientation change of the workpiece after gripping can be suppressed.

[0090] For example, in the device for searching a workpiece holding shape D, during the machine learning process, instead of the holding shape generating unit 43, the machine learning unit 47 generates the holding shape based on the mounted state of the workpiece WK. During the machine learning process, each time the robot hand 12 holds the workpiece WK, the result information storage unit 83 stores, in association with each other, as result information, the holding shape generated by the machine learning unit 47, the movement (action) in the process, and the evaluation score EV determined by the evaluation judging unit 46 (final evaluation score EVr in the above description).At a predetermined time after the machine learning (for example, every week, every month, or the like), machine learning is performed again using the holding shape generated by the machine learning unit 47 and stored in the result information storage unit 83 in association with each other, the movement (action) in the operation, and the evaluation score EV determined by the evaluation evaluation unit 46 (final evaluation score EVr in the above description). In this case, likewise, to detect the orientation before the movement before the action of the operation, the workpiece WK is detected by the first workpiece detection unit 2, and to detect the orientation after the movement after the action of the operation, the workpiece WK is detected by the first workpiece detection unit 2.Such a device for searching a workpiece holding shape D can perform machine learning jointly on the result obtained through the process. Specifically, a relatively unfavorable holding shape can be excluded from the result, and machine learning can be performed on the result that excludes the relatively unfavorable holding shape.

[0091] In the embodiment, the predetermined movement may be, for example, a combination of the plurality of movements having different contents in one type of movement, or may be, for example, a plurality of combinations of the speed movement, the acceleration movement, the rotation movement, and the vibration movement, or may be, for example, a combination of these described above.

[0092] Fig. Figure 13 is a diagram describing another variant for finding a holding shape in a series of robot actions. For example, as in Fig.13, the predetermined movement includes a first movement PT1 for ascending at a first speed for a first time after gripping the workpiece WK in the gripping shape (gripping position, gripping force) generated by the gripping shape generating unit 43, a second movement PT2 for moving at a first acceleration for a second time from the mounting position of the workpiece WK to a conveyor CV, a third movement PT3 for descending at a second speed for a third time, and a fourth movement PT4 for moving at a second acceleration for a fourth time from the conveyor CV to the mounting position of the workpiece WK. By setting the predetermined movement in this way, it is possible to evaluate the gripping shape of the workpiece WK at the transition between processes.

[0093] This specification discloses various aspects of the technology as described above, and the main technology thereof is summarized below.

[0094] A device for searching a holding shape for a robot hand according to one aspect is a device that can find a holding shape suitable for a mounted state of a placed workpiece when holding the workpiece with the robot hand, and includes: a holding shape generation unit that executes a holding shape generation process to generate the holding shape for holding the workpiece with the robot hand based on the mounted state of the workpiece; a hand control unit that executes a gripping process to hold the workpiece with the robot hand in the holding shape generated by the holding shape generation unit; a movement execution unit that executes a movement execution process to cause the robot hand to perform a predetermined movement after holding the workpiece with the robot hand; an evaluation evaluation unit that executes an evaluation evaluation process to determine an evaluation score,to evaluate the holding shape after the movement execution unit causes the robot hand to execute the predetermined movement; and a repetition processing unit that causes the holding shape generation unit, the hand control unit, the movement execution unit, and the evaluation evaluation unit to execute the holding shape generation process, the gripping process, the movement execution process, and the evaluation evaluation process, respectively, a predetermined number of times. The holding shape generation unit generates different holding shapes when executing the predetermined number of times. Preferably, in the device for searching a holding shape for a robot hand, the holding shape is represented (defined) by a gripping position and gripping force for gripping the workpiece with the robot hand. Preferably, in the device for searching a holding shape for a robot hand, the predetermined movement includes a speed movement,to move the robot hand in a predetermined direction at a predetermined speed for a predetermined time, and / or an acceleration movement to move the robot hand in a predetermined direction at a predetermined acceleration for a predetermined time, and / or a rotational movement to rotate the robot hand in a predetermined angular range at a predetermined speed (or predetermined angular velocity) or a predetermined acceleration (or predetermined angular acceleration) for a predetermined time, and / or an oscillation movement to oscillate the robot hand in a predetermined amplitude range at a predetermined frequency (or cycle) for a predetermined time.

[0095] Such a robot hand holding shape search device can determine the plurality of evaluation scores for various holding shapes for the mounted state of the workpiece by executing the holding shape generation process, the gripping process, the movement execution process, and the evaluation evaluation process a predetermined number of times. Therefore, by comparing each evaluation score, it is possible to find the holding shape that can reduce the orientation change of the workpiece after gripping for the mounted state of the workpiece when holding the workpiece with the robot hand.

[0096] In another aspect, the apparatus for searching a holding shape for a robot hand further includes a machine learning unit that performs machine learning on the holding shape based on the holding shape generated by the holding shape generation unit, the predetermined motion, and the evaluation score determined by the evaluation evaluation unit. The repetition processing unit further causes the machine learning unit to perform the machine learning upon execution the predetermined number of times. Preferably, in the apparatus for searching a holding shape for a robot hand, the machine learning is reinforcement learning. Preferably, the reinforcement learning is Q-learning.

[0097] Since such a device for searching a workpiece holding shape for a robot hand has the machine learning unit and performs machine learning on the machine learning unit, when holding the workpiece with the robot hand for the mounted state of the workpiece, the holding shape which can reduce the orientation change of the workpiece after gripping can be output from the machine learning unit after machine learning.

[0098] In another aspect, the apparatus for searching a holding shape for a robot hand further comprises an orientation detection unit that detects an orientation of the workpiece held by the robot hand, wherein the evaluation evaluation unit determines the evaluation score based on a pre-movement orientation, which is the orientation of the workpiece detected by the orientation detection unit before the execution of the predetermined movement, and a post-movement orientation, which is the orientation of the workpiece detected by the orientation detection unit after the predetermined movement is executed.

[0099] Such a device for searching a workpiece holding shape for a robot hand can determine the evaluation score by the amount of orientation change between the orientation after the movement and the orientation before the movement.

[0100] In another aspect, the apparatus for searching a holding shape for a robot hand further comprises an input unit that inputs the predetermined movement, wherein the movement execution unit executes the predetermined movement input by the input unit.

[0101] Since such a device for searching a workpiece holding shape for a robot hand further includes the input unit, the user can specify the predetermined movement. In particular, the user can specify the predetermined movement taking into account the situation in which the robot hand is operated.

[0102] In another aspect, in the apparatus for searching a holding shape for a robot hand, a plurality of feature points are set for determining the evaluation score for the workpiece, a weight in determining the evaluation score is set for each of the plurality of feature points, and the evaluation evaluation unit uses the weight for each of the plurality of feature points to determine the evaluation score.

[0103] Such a workpiece holding shape search device for a robot hand sets the weight for each of the plurality of feature points specified for the workpiece and determines each evaluation score using the weight for each of the plurality of feature points. Therefore, by making the weight of the feature point that allows for orientation change with difficulty greater than the weight of the feature point that allows for orientation change easily, the holding shape that can reduce the orientation change of the workpiece after gripping can be found while taking into account the permissibility of the orientation change.

[0104] In another aspect, in the apparatus for searching a holding shape for a robot hand, the evaluation evaluation unit further determines an average of the evaluation scores determined for each of the plurality of feature points as a final evaluation score.

[0105] Such a device for searching a workpiece holding shape for a robot hand can determine overall evaluation scores for the plurality of feature points.

[0106] In another aspect, in the apparatus for searching a holding shape for a robot hand, the machine learning is reinforcement learning of Q-learning, and even during an operation after the machine learning, the machine learning unit executes the machine learning at a learning rate lower than a learning rate during the machine learning.

[0107] Such a device for searching a workpiece holding shape for a robot hand can reflect the result during the operation to the machine learning unit and can reduce the impact on the machine learning unit by reducing the learning rate even when a relatively unfavorable holding shape is executed, and therefore, deterioration of reducing the orientation change of the workpiece after gripping can be suppressed.

[0108] In another aspect, in the apparatus for searching a holding shape for a robot hand during a process after machine learning, the machine learning unit, instead of the holding shape generating unit, generates the holding shape based on the mounted state of the workpiece, the apparatus further comprising a result information storage unit that stores the holding shape generated by the machine learning unit, the movement in the process, and the evaluation score determined by the evaluation scoring unit in association with each other during the process after machine learning each time the workpiece is held with the robot hand, and the machine learning unit performs the machine learning at a predetermined time after the machine learning using the holding shape generated by the machine learning unit and stored in the result information storage unit in association with each other,the movement in the operation and the evaluation rating determined by the evaluation unit.

[0109] Such a device for searching a workpiece holding shape for a robot hand can perform machine learning jointly on the result obtained through the process. Specifically, a relatively unfavorable holding shape can be excluded from the result, and machine learning can be performed on the result that excludes the relatively unfavorable holding shape.

[0110] A method for searching a holding shape for a robot hand according to one aspect is a method for finding a holding shape suitable for a mounted state of a placed workpiece when holding the workpiece with the robot hand. The method includes: a holding shape generation step of generating the holding shape for holding the workpiece with the robot hand based on the mounted state of the workpiece; a hand control step of holding the workpiece with the robot hand in the holding shape generated in the holding shape generation step; a movement execution step of causing the robot hand to perform a predetermined movement after holding the workpiece with the robot hand; and an evaluation score step of determining an evaluation score to evaluate the holding shape after causing the robot hand to perform the predetermined movement in the movement execution step.The holding shape generation step, the hand control step, the movement execution step, and the evaluation evaluation step are executed a predetermined number of times, and the holding shape generation step includes generating different holding shapes when executed the predetermined number of times. Such a method for searching a holding shape for a robot hand can determine the plurality of evaluation scores for different holding shapes for the mounted state of the workpiece by executing the holding shape generation step, the hand control step, the movement execution step, and the evaluation evaluation step a predetermined number of times. Therefore, by comparing each evaluation score, the user can find the holding shape that can reduce the orientation change of the workpiece after gripping for the mounted state of the workpiece when holding the workpiece with the robot hand.

Claims

[1] Device for searching a workpiece holding shape (D) for a robot hand (12) for finding a holding shape suitable for a mounted state of the placed workpiece (WK) when holding the workpiece (WK) with the robot hand (12), the device comprising: a holding mold generating unit (43) that executes a holding mold generating process to generate the holding mold for holding the workpiece (WK) with the robot hand (12) based on the mounted state of the workpiece (WK); a hand control unit (44) which carries out a gripping process to hold the workpiece (WK) with the robot hand (12) in the holding shape generated by the holding shape generating unit (43); a movement execution unit (44) that executes a movement execution process to cause the robot hand (12) to perform a predetermined movement after holding the workpiece (WK) with the robot hand (12); an evaluation evaluation unit (46) that executes an evaluation evaluation process to determine an evaluation score to evaluate the holding shape after the movement execution unit (44) causes the robot hand (12) to execute the predetermined movement; and a repetition processing unit (48) which causes the holding shape generation unit (43), the hand control unit (44), the movement execution unit (44) and the evaluation evaluation unit (46) to execute the holding shape generation process, the gripping process, the movement execution process and the evaluation evaluation process a predetermined number of times, respectively, wherein the holding shape generating unit (43) generates different holding shapes when executing the predetermined number of times. [2] A device for searching a workpiece holding shape (D) for a robot hand (12) according to claim 1, further comprising a machine learning unit (47) that performs machine learning on the holding shape based on the holding shape generated by the holding shape generating unit (43), the predetermined movement, and the evaluation score determined by the evaluation evaluation unit (46), wherein the repetition processing unit (48) further causes the machine learning unit (47) to execute the machine learning upon execution of the predetermined number of times. [3] Apparatus for searching a workpiece holding shape (D) for a robot hand (12) according to claim 1 or 2, further comprising an orientation detection unit (2, 45) which detects an orientation of the workpiece (WK) held by the robot hand (12), wherein the evaluation evaluation unit (46) determines the evaluation score based on a pre-movement orientation, which is the orientation of the workpiece (WK) detected by the orientation detection unit (2, 45) before the execution of the predetermined movement, and a post-movement orientation, which is the orientation of the workpiece (WK) detected by the orientation detection unit (2, 45) after the predetermined movement is executed. [4] Apparatus for searching a workpiece holding shape (D) for a robot hand (12) according to one of claims 1 to 3, further comprising an input unit (5) which inputs the predetermined movement, wherein the movement execution unit (44) executes the predetermined movement input by the input unit (5). [5] Device for searching a workpiece holding shape (D) for a robot hand (12) according to one of claims 1 to 4, wherein a plurality of feature points are defined to determine the evaluation score for the workpiece (WK), a weight is set for each of the plurality of feature points in determining the evaluation score, and the evaluation scoring unit (46) uses the weight for each of the plurality of feature points to determine the evaluation score. [6] Apparatus for searching a workpiece holding shape (D) for a robot hand (12) according to claim 5, wherein the evaluation evaluation unit (46) further determines an average of the evaluation score determined for each of the plurality of feature points as a final evaluation score. [7] Device for searching a workpiece holding shape (D) for a robot hand (12) according to one of claims 2 to 6, wherein machine learning is reinforcement learning of Q-learning, and even during a post-machine learning operation, the machine learning unit (47) performs the machine learning at a learning rate that is lower than a learning rate during the machine learning. [8] Device for searching a workpiece holding shape (D) for a robot hand (12) according to one of claims 2 to 6, wherein during a process after machine learning, instead of the holding mold generating unit (43), the machine learning unit (47) generates the holding mold based on the mounted state of the workpiece (WK), the device further comprises a result information storage unit (83) which stores the holding shape generated by the machine learning unit (47), the movement during the process and the evaluation score determined by the evaluation evaluation unit (46) in association with each other during the process after the machine learning each time the workpiece (WK) is held with the robot hand (12), and the machine learning unit (47) executes the machine learning again at a predetermined time after the machine learning using the holding shape generated by the machine learning unit (47) and stored in the result information storage unit (83) in association with one another, the movement in the process, and the evaluation score determined by the evaluation evaluation unit (46). [9] A method for searching a workpiece holding shape (D) for a robot hand (12) for finding a holding shape suitable for a mounted state of the placed workpiece (WK) when holding the workpiece (WK) with the robot hand (12), the method comprising: a holding mold generating step (S1) of generating the holding mold for holding the workpiece (WK) with the robot hand (12) based on the mounted state of the workpiece (WK); a hand control step (S2) of holding the workpiece (WK) with the robot hand (12) in the holding mold generated in the holding mold generating step (S1); a movement execution step (S5) of causing the robot hand (12) to execute a predetermined movement after holding the workpiece (WK) with the robot hand (12); and an evaluation evaluation step (S8, S9) of determining an evaluation score to evaluate the holding shape after the robot hand (12) is caused to execute the predetermined movement in the movement execution step (S5), wherein the holding shape generating step (S1), the manual control step (S2), the movement execution step (S5) and the evaluation evaluation step (S8, S9) are executed a predetermined number of times (S11), and the holding shape generating step (S1) comprises generating different holding shapes when executing the predetermined number of times.

Images