Patent Application
20230271319
The present disclosure relates to a method of generating a learning model for transferring fluid from one container to another by controlling a robotic arm, and a method and system for weighing fluid. In particular, the present disclosure relates to a method for generating a learning model for controlling a robot arm based on a machine-learned learning model to transfer a fluid in one container to another container, and a method and system for weighing the fluid.
In recent years, there has been development of technology that allows robots to perform tasks that were previously performed by humans. One such technology uses a robot arm and machine learning to automatically perform the weighing process, for example, transferring a specified amount of liquid from one container to another. However, in the conventional configuration, there is still room for improvement in terms of machine learning methods.
The purpose of the present disclosure is to make further improvements in terms of configuration and processing with respect to machine learning methods in a method that incorporates machine learning in the control of weighing fluids, including liquids, with a robot arm.
According to an exemplary embodiment of the present disclosure, there is provided a method of generating a learning model for machine learning, comprising: acquiring a plurality of learning data (teaching data) including i) time-series information showing a posture of a robot arm, the robot arm holding a first container containing therein a targeted fluid and pouring the fluid from the first container to a second container, and ii) a weight of the second container which changes time serially; and generating a learning model based on the learning data, the learning model being given, as input thereto, only two types of information consisting of the information showing the posture of the robot arm at a first time and the weight of the weight of the second container and outputting the information showing the posture of the robot arm at a second time.
According to another aspect, there is also provided a system and a method, in which a robot arm is controlled by using AI (artificial intelligence) to transfer the fluid in a container into another container using the teacher data (learning data) generated as described above.
Therefore, the robot arm can be controlled based on the learning model generated based on the teacher data (learning data) to automate the process of transferring the target fluid from the first container to the second container. Therefore, the various benefits of such automation can be enjoyed.
In addition, the learning model is generated using only two types of input information: the time-series information about the robot arm's posture and the time-series weight of the second container. This differs from previously known methods in that it does not require an image of the first container. This simplifies the system configuration by eliminating the need for image information from the optical imaging system, and enables weighing with the same or higher accuracy and in a shorter time than conventional known methods.
Other configurations and effects other than those described above will become clear in the examples described below with the drawings.
In the accompanying drawings:
One embodiment is described below with reference to the drawings. A weighing system 1 is a system capable of performing the weighing process of pouring and transferring a specified amount of fluid from a first container 91 to a second container 92 by controlling a robot arm 10 based on machine learning. Note that the above machine learning employs, in particular, a learning method classified as supervised learning.
The weighing system 1 machine-learns the relationship between the movement of the robot arm 10 when pouring the contents of the first container 91 and the weight of the second container 92, i.e., the weight of the fluid poured into the second container 92, which is changed by the movement of the robot arm 10. Using the learning model obtained by the machine learning, the robot arm 10 is automatically operated to achieve the target weighing value specified by the user, for example.
The robot arm 10 is configured to hold the first container 91, for example by grasping it. The robot arm then transfers the held first container 91 to the vicinity of the second container 92 and tilts the first container 91 with respect to the second container 92 to pour and transfer the fluid contained in the first container 91 into the second container 92. In this way, the weighing system 1 weighs a predetermined amount of fluid contained in the first container 91.
The shape, size, color, and weight of the first and second containers 91 and 92 are not limited to any particular configuration herein. The fluid to be weighed by the weighing system 1 is not limited to liquids, but can also be powders or granules that have flowable properties. When a liquid is the object of weighing, the color and viscosity of the liquid are not limited. The amount of fluid contained in the first and second containers 91 and 92 at the time of weighing is also not particularly limited. In this specification, when referring to the weight of the first or second container 91 or 92, it shall also include the weight of the fluid contained in the first or second container 91 or 92.
The weighing system 1 is equipped with a robot arm 10, a control unit 20, a weighing instrument 30, and an information processing unit 40, as shown in
The robot arm 10 can be composed of a vertically articulated robot with six axes (joints), for example, as shown in
The arms 121-126 are connected to each other via a plurality of axes J1-J6, respectively, to allow rotation around said axes. In this case, the axes are referred to as a first axis J1, a second axis J2, a third axis J3, a fourth axis J4, a fifth axis J5, and a sixth axis J6, in the order from the base 11 side. If each axis J1(−J6) is not specified, the respective axes J1-J6 are collectively referred to simply as an axis J. Each of the axes J1-J6 can be individually driven by a servo motor, for example. In this embodiment, information about (or showing) the posture of the robot arm 10 means the state of the robot arm 10, which is composed of a set of rotation angles θn of the respective axes Jn. In this case, “n” means a positive integer corresponding to each of the axes J1-J6. For example, when the angle θ1 is used, it means the angle of the first axis J1. The n-th axis rotation angle θn is sometimes referred to herein as the n-th axis angle θn.
Here, when pouring the fluid in the first container 91 into another container, if one tries to tilt the first container 91 by rotating only the sixth axis J6, the position of the spout of the first container 91, i.e., the part from which the fluid flows out of the first container 91, will move in the vertical and horizontal directions. Therefore, when a person weighs, he or she usually tilts the container by lifting the bottom part of the container using the area around the container spout as a fulcrum. Hence, in this embodiment, the axis Jn used for information on the posture of the robot arm 10 includes not only the axis affecting the rotation angle of the first container 91, in this case, the sixth axis J6, but also all axes affecting the position of the spout of the first container 91, that is, the position of the fluid outflow. This allows the weighing system 1 to learn and model on human behavior more accurately.
The sixth arm 126 serves as an end portion of the robot arm 10 and is configured, for example, in the shape of a flange. A tool section 13 is removably attached to the tip of the sixth arm 126. The tool section 13 is referred to as a chuck or gripper, for example. In this case, the tool section 13 can hold the first container 91. The robot arm 10 has servo motors for driving each axis J1(−J6), encoders for detecting the rotation speed and position of each of the axes J1-J6, and brakes for stopping the motion of each of the axes J1-J6, although details are not shown.
The control unit 20 is a so-called robot controller and has the function of controlling the operation of the robot arm (manipulator) 10. The robot arm 10 and the control unit 20 are configured to communicate with each other by wired or wireless means. The control unit 20 may be wired or wirelessly intercommunicatable to other external devices, such as a personal computer or a portable terminal such as a smart phone. The control unit 20 may be built into the robot arm 10, or it may be realized by a server or other device to control the robot arm 10 remotely. The control unit 20 may be configured with the same or common hardware as the information processing unit 40.
The control unit 20 can be composed of, for example, the CPU 21, a memory 22, a driver 23, and a position detection unit 24, as shown in
The position detection unit 24 is composed of, for example, an encoder or the like provided for each axis (joint) J1-J6, and can detect the rotation angle θn of each axis J1-J6, i.e., the rotation angle of each motor. The control unit 20 drives the respective motors by feedback control, for example, based on the position of each axis J1-J6 detected by the position detection unit 24. The control unit 20 transmits the position of each axis J1-J6, i.e., the rotation angle of each motor, detected by the position detection unit 24 to the information processing unit 40, along with the acquired time.
The weighing instrument 30 can be composed of, for example, an electronic scale. The second container 92 is placed on the weighing instrument 30. The weighing instrument 30 measures the weight of the second container 92 at predetermined intervals and transmits the measurement results, along with the time of acquisition, to the information processing unit 40.
As shown in
The hardware configuration of the information processing unit 40 can comprise a CPU 51, a main memory 52, an auxiliary memory 53, and an interface 54, as shown in
The CPU 51 stores a plurality of learning data 64 for machine learning generated by the learning data generator 401 and a learning model 65 generated by the learning model generator 402 in the main memory 52 or the auxiliary memory 53 according to the program. In this embodiment, the CPU 51 can store the plurality of learning data 64 in the main memory 52 or the auxiliary memory 53 as learning data group 640, even though they are compiled for use by the computer. The learning data 64 included in the learning data group 640 shown in
The auxiliary memory 53 includes a tangible, non-transitory computer reading recoding medium. Examples of auxiliary memory 53 include HDD (Hard Disk Drive), SSD (Solid State Drive), magnetic disks, optical disks, CD-ROM (Compact Disc Read Only Memory), DVD-ROM (Digital Versatile Disc Read Only Memory), and semiconductor memory. The auxiliary memory 53 may be internal media directly connected to the bus of the computer comprising the information processing unit 40, or it may be external media connected to the information processing unit 40 via the interface 54 or a communication line. When each of the programs 61, 62, and 63 is delivered to information processing unit 40 via a communication line, the information processing unit 40 that receives the delivery decompresses and executes the program in its main memory 52 to realize the above devices 401, 402, and 403.
The realization of each of the devices 401, 402, 403 is not limited to the combination of hardware and program described above, but may be realized by hardware alone, such as integrated circuits with programs implemented in the devices 401, 402, 403, or some functions may be realized by dedicated hardware and some by a combination of information processing unit 40 and programs.
The learning data generator 401 has the function of generating learning data (teacher data) for use in the learning model generator 402. The learning data generator 401 has a state data acquisition unit (first and second state data acquisition unit) 41 and a learning data generation processing unit 42. The state data acquisition unit 41 is capable of executing the state data acquisition process. The state data acquisition process includes a process that functions to acquire state data of the robot arm 10 and the second container 92 when the sample operation of the weighing process is executed by the robot arm 10 as learning data, or teacher data.
Therefore, the weighing system 1 and the weighing method performed therein are implemented in a state where the robot arm is controlled to perform the task of transferring the target fluid from one container to another. In particular, a weighing system and weighing method based on AI (artificial intelligence) technology using a learning model generated by machine learning, called supervised learning, are provided.
The state data acquisition unit 41 successively acquires the state data of the robot arm 10 and the second container 92 at the current time from the control unit 20 and the weighing instrument 30 while the robot arm 10 is being made to perform the sampling operation. The sample operation includes a series of movements of the robot arm 10 in which the robot arm tilts the first container 91 to pour a predetermined amount of fluid into the second container 92, and then returns the tilt of the first container 91 to the previous stage. The sample operation may include the operation from the point at which the robot arm 10 starts grasping the first container 91.
The state data of robot arm 10 to be acquired from the control unit 20 includes the angle θn(t) of each axis J1 to J6 of robot arm 10 at the current time (t) and the current value In(t) of the motor of each of the axes J1 to J6 at the current time (t). In this embodiment, “n” in the current value In means a positive integer corresponding to each of the axes J1-J6, as in the case of angle θn. For example, if the current value I1 is used, it means the current value of the motor corresponding to the first axis 31.
In this specification, the current value In of the motor of the nth axis is sometimes referred to as the nth axis current value In. Since the current value In of each of the axes J1-J6 varies with the load acting on the motor of each of the axes J1-J6, it can be treated as information about the load acting on each of the axes J1-J6, i.e., the load amount. When the robot arm 10 holds the first container 91, the load acting on each of the axes J1-J6 varies according to the weight of the first container 91, so that the current value In of each of the axes J1-J6 can be treated as information about the weight of the first container 91.
The learning data generation processing unit 42 is capable of performing learning data generation processing. The learning data generation process generates learning data from the state data acquired by the state data acquisition unit 41. The learning data generation process generates learning data by mapping the state data of the robot arm 10, i.e., the angle θn(t) of each of the axes J1-J6 at each time and the current In(t) of the motor of each of the axes J1-J6 at each time, and the weight w(t) of the second container 92 at each time at the time of acquisition.
That is, the data for learning is, for example, the time-series data of the angle θn(t) of each of the axes J1-J6, the current value In(t) of the motor of each of the axes J1-J6, and the weight w(t) of the second container 92, as an example shown in
The learning data generator 401, for example, generates learning data based on one sample operation by executing a series of processes shown in
In step S11, the learning data generator 401 first starts the execution of the sample weighing operation by operating the robot arm 10 via the control unit 20. Next, the learning data generator 401 acquires the current state data of the robot arm 10 and the weighing instrument 30 by the function of the state data acquisition unit 41 in step S12. Then, the learning data generator 401 stores the acquired state data in the main memory 52 or auxiliary memory 53 with the acquisition time in step S13 to accumulate them chronologically.
Next, in step S14, the learning data generator 401 determines whether the sample operation has been completed. If the sample operation has not been completed (NO in step S14), the learning data generator 401 repeats steps S12-S14 until the sample operation is completed. The learning data generator 401 can determine that the sample operation has ended, for example, when a predetermined period of time has elapsed since the robot arm 10 stopped or when an input operation indicating the end is received from the operator. When the sample operation is completed (YES in step S14), the learning data generator 401 moves the process to step S15, where the learning data is generated by the function of the learning data generation processing unit 42. The learning data generator 401 then terminates the series of processes (i.e., the processes are ended).
The learning model generator 402 has the function of generating the learning model 65 illustrated in
In this embodiment, as shown in
The learning model 65 has as its output the rotation angle θn(t+1) of each axis J1(−J6) of the robot arm 10 at the second time (t+1). The second time (t+1) is a time later than the first time (t), and the interval between the first time (t) and the second time (t+1) can be set arbitrarily in consideration of data volume and accuracy.
The learning model 65 can use a neural network capable of processing time-series data, such as an RNN (Recurrent Neural Network), MTRNN (Multi Timescale RNN), LSTM (Long Short Term Memory), ARIMA (Auto Regressive Integrated Moving Average), or one-dimensional CNN (Convolutional Neural Network), for example.
The learning model generator 402 has a learning data acquisition unit 43 and a learning processing unit 44, as shown in
When the robot arm 10 holds the first container 91 containing the target fluid and pours the target fluid from the first container 91 into the second container 92, that is, when performing the weighing process, an inference unit 403 has the ability to infer and output information about the posture of the robot arm 10 using the learning model 65. The inference unit 403 has the state data acquisition unit 41 and an inference processing unit 45. In this embodiment, the state data acquisition unit 41 is shared by the learning data generator 401 and the inference unit 403, but it may be configured not to be shared.
The inference processing unit 45 can input the information θn(t) about the posture of the robot arm 10 at the first time (t) and the weight w(t) of the second container 92 to the learning model 65 and output the information θ(t+1) about the posture of the robot arm at the second time (t+1). In this embodiment, in addition to the information θn(t) about the posture of the robot arm 10 at the first time (t) and the weight w (t) of the second container 92, the inference processing inputs to the learning model 65 the information about the force of the robot arm 10, that is, the current In (t) of the motor of each axis J1 to J6. The inference processing unit 45 passes the output information θn(t+1) about the attitude at the second time (t+1) to the control unit 20, as shown in
In the weighing process, the weighing instrument measures the weight w(t) of the second container 92 at the current time (t) and successively passes the measurement results to the inference unit 403, and the control unit 20 acquires the angles θn(t) of each of the axes J1 to J6 of the robot arm 10 at the current time (t) and successively passes them to the inference unit 403. That is, the inference unit 403 acquires the state data of the robot arm 10 and the second container 92 at the current time (t) from the control unit 20 and the weighing instrument 30. The control unit 20 controls the motion of the robot arm 10 so that the angle θn(t) of each of the axes J1 to J6 of the robot arm 10 at the next time (t+1) output from the inference unit 403.
The weighing system 1 can perform the weighing process based on the flow illustrated in
Next, the weighing system 1 performs the operation of pouring the target amount of fluid into the second container 92 by tilting the first container 91, i.e., weighing, by repeating steps S23 through S26. In step S23, the weighing system 1 obtains the state data of the robot arm 10 and the second container 92 at the current time (t) by the function of the inference unit 403. Next, in step S24, the weighing system 1 calculates information about the posture of the robot arm 10 at the next time (t+1), i.e., each axis angle θn(t+1), by the function of the inference unit 403.
Next, in step S25, the weighing system 1 operates the robot arm 10 so that each axis angle θn(t+1) is calculated in step S24 by the function of the control unit 20. Then, in step S26, the weighing system 1 determines whether the weighing has been completed, and if not (NO in step S26), the process returns to step S23, and if completed (YES in step S26), the process moves to step S27. Whether the weighing has been completed can be judged based on, for example, the fact that the tilt of the first container 91 has returned to the initial state, i.e., the state in step S22, or that no weight change has occurred in the second container 92 for a predetermined period. The weighing system 1 then returns the first container 91 to the original container storage area 931, and so on, thereby completing the series of weighing processes.
According to the embodiment described above, the method of generating the learning model 65 first acquires the learning data 64. The learning data 64 has a plurality of learning data including the time-series information On about the posture of the robot arm 10 when holding the first container 91 containing the target fluid and pouring the target fluid from the first container 91 into the second container 92, and the time-series weight w of the second container 92. Next, using the learning data contained in the learning data 64, a learning model 65 is generated with the information θn(t) about the posture of the robot arm 10 at the first time (t) and the weight w(t) of the second container 92 as inputs thereto and the information θn(t+1) about the posture of the robot arm 10 at the second time (t+1) as output therefrom.
The method of generating learning data of this embodiment includes the process of acquiring time-series information On about the posture of the robot arm when holding the first container 91 containing the target fluid and pouring the target fluid from the first container 91 into the second container 92, the process of acquiring the time-series weight w of the second container 92, and the process of performing the correspondence between the time-series information On about the posture of the robot arm 10 and the time-series weight w of the second container 92.
When the robot arm 10 holds the first container 91 containing the target fluid and pours the target fluid from the first container 91 into the second container 92, The inference method in this embodiment acquires information θn(t) about the posture of the robot arm 10 at the first time (t) and the weight w(t) of the second container 92 at the first time t, and let the information θn(t) about the posture of the robot arm 10 at the first time (t) and the weight w(t) of the second container 92 as inputs. The information θn(t) about the posture of the robot arm 10 at the first time (t) and the weight w(t) of the second container 92 obtained is input to the learning model 65, which is trained using the learning data 64 with the information On(t+1) about the posture of the robot arm 10 at the second time (t+1) as the output and the information θn(t+1) about the posture of the robot arm 10 at the second time (t+1) is output.
The weighing method and weighing system in this embodiment is a method and system in which a robot arm 10 holds a first container 91 containing therein the target fluid and pours the target fluid from the first container 91 into a second container 92. The weighing method and weighing system includes the step (or means) of acquiring information θn(t) about the posture of the robot arm 10 at the first time (t), the step (or means) of acquiring the weight of the second container 92 at the first time (t), the step (or means) of inputting the information θn(t) about the posture of the robot arm and the weight w(t) of the second container 92, the information θn(t+1) about the posture of the robot arm and the weight w(t) of the second container 92 acquired at the fir time (t), the step (means) of outputting (t+1) the information θn(t+1) about the posture of the robot arm 10 at a second time t(+1), and the step (means) of controlling movements of the robot arm 10 based on the information On(t+1) about the posture of the robot arm 10 at the second time (t+1). The learning model 65 is learned using the learning data 64 with the information θn(t) about the posture of the robot arm 10 at the first time (t) and the weight w(t) of the second container 92 as inputs thereto and the information θn(t+1) about the posture of the robot arm at the second time (t+1) as output therefrom.
If the weighing process were to be performed without using machine learning, depending on the viscosity of the fluid contained in the first container 91, the amount of fluid previously contained in the first and second containers 92, and the parameters of the target weighing value, the operation of the robot arm 10 must be changed, and the control program of the robot arm must also be changed. In this case, it is necessary to change the motion of the robot arm 10 based on the difference between the target weighing value and the value of the weighing instrument 30, but it is extremely difficult to prepare a control program for all patterns because the amount of input data and decision items is too large.
In contrast, according to this embodiment, since the robot arm 10 is operated based on the learning model 65 obtained by machine learning, there is no need to prepare numerous control programs to respond to the viscosity of the fluid contained in the first container 91, the amount of fluid previously contained in the first and second containers 91 and 92, and the target weighing value parameters. As a result, the configuration for weighing fluids using the robotic arm can flexibly respond to changes in factors such as the shape of the first container 91 and the type of fluid.
In addition, this embodiment provides the following improvements over previously known weighing methods that use machine learning. First, the inventor has found that, using only two types of information about the posture of the robot arm 10 holding the first container 91 and the weight of the second container 92 as the input data for opportunity learning for weighing, without using three pieces of information, which are the angle of the first container 91, the image or video of the first container 91, and the weight of the second container 92, allows weighing with the same or better accuracy and duration than the conventional configuration which uses the three types of information described above.
Then, since the weighing system 1 and weighing method of this embodiment does not use the image or video, etc. of the first container 91 for machine learning of the motion of the robot arm 10 and control of the motion, it can be a simple configuration that does not include cameras or other optical devices. In other words, no more than three types of information, including images, are needed to generate a learning model. The benefits of reducing the type of information are enormous. Therefore, the weighing system 1 can reduce the data volume to be processed by the information processing unit 40 by not using the image or video, etc. of the first container 91 for controlling the robot arm 10 and machine learning, thereby increasing the speed of the weighing process.
Furthermore, the weighing system 1 and the weighing method use the information On about the posture of the robot arm 10 for machine learning of the movement of the robot arm 10 and control of the movement. According to this method, the weighing process can be performed with high accuracy and at high speed without using the image or video of the first container 91 during the weighing process.
The learning data 64 also contains information about the load acting on the robot arm 10, in this case, the current value In of the motor of each of the axes J1-J6. According to this method, the learning model 65 can include an element of the load acting on each of the axes J1-J6 of the robot arm 10, i.e., the weight of the first container 91. This further improves the accuracy of the learning model 65, and as a result, the weighing process can be executed with even higher accuracy and speed.
The above-described embodiments are not limited to the embodiments described above and in the drawings, but can be modified as appropriate to the extent not to depart from the gist of the invention.
