JP2022150281A - Learning device, trajectory generator and manipulator system - Google Patents

Abstract

To provide a learner which can generate a trajectory that can be commonly used for a plurality of work manipulators having different configurations, a trajector generator and a manipulator system.SOLUTION: A learning device performs learning by use of a drove command and a state signal of a manipulator for learning when the manipulator for learning is driven according to the drive command. The learning device comprises: standardization means that standardizes the drive command and the state signal on the basis of specification of the manipulator for learning; and a first learner which learns by use of the standardized drive command and state signal.SELECTED DRAWING: Figure 3

Description

The present invention relates to a learning device, a trajectory generator, and a manipulator system for generating a trajectory of an end effector position or posture of a working manipulator.

2. Description of the Related Art Conventionally, a work manipulator composed of at least a robot arm and an end effector has been used as a work device that automatically performs work such as pick-and-place, fitting, screw tightening, cutting, and drilling. In order to perform a desired task with this task manipulator, it is necessary to move the position and/or posture of the end effector that performs the task in a trajectory that allows the task to be performed. Therefore, a work manipulator is known that includes a trajectory generator that generates a trajectory of the position and orientation of the end effector and a tracking control system that causes the end effector to follow the generated trajectory.

For example, Patent Literature 1 describes a motion editing device that can generate a trajectory of a robot by creating a target robot model in a virtual space and manipulating the model in the virtual space.

JP 2008-254074 A

In such a trajectory generator, it is possible to generate a trajectory of the position/orientation of the end effector, which is the working part of the working manipulator in the real space, based on the manipulation of the model in the virtual space.

However, with the technique described in Patent Literature 1, when generating a trajectory for a working manipulator with a new configuration, it is necessary to build a model of the new working manipulator and operate it in virtual space.

SUMMARY OF THE INVENTION The present invention has been made in consideration of the above situation, and an object of the present invention is to provide a learning apparatus capable of generating a trajectory that can be used in common for a plurality of work manipulators having different configurations. It is to provide a generator and a manipulator system.

From the above, in the present invention, "a learning device that performs learning using a drive command and a state signal of the learning manipulator when the learning manipulator is driven in accordance with the drive command, wherein the learning device includes a drive command and a state signal of the learning manipulator. A learning device comprising standardization means for standardizing the state signal based on the specifications of the manipulator for learning, and a first learning device for learning using the standardized drive command and the state signal." It is what I did.

Further, in the present invention, "a trajectory generator for providing a commanded trajectory for the working manipulator using a driving command for the working manipulator and a state signal of the working manipulator, wherein the trajectory generator is a driving command for the working manipulator. and a state signal based on the specifications of the working manipulator; a second learner for inferring using the normalized drive command and state signal; and an output of the second learner for working denormalization means for denormalizing based on the specifications of the manipulator for work, the output of the denormalization means is used as the command trajectory of the manipulator for work, and the second learning device is configured to drive the manipulator for learning. In order to perform learning using the command and the state signal of the learning manipulator, the drive command and state signal in the learning manipulator are standardized based on the specifications of the learning manipulator, and the standardized drive command and state signal are used. A trajectory generator characterized by being a learned first learner."

Further, in the present invention, there is provided a "manipulator learning system for reflecting learning results of a learning manipulator on a plurality of types of working manipulators, wherein when a learning mechanism is driven in accordance with a drive command, a drive command and a state signal and a learning device that performs learning using a driving command and a state signal of the learning manipulator when the learning manipulator is driven, wherein the driving command and the state signal conform to the specifications of the learning manipulator a learning device comprising a normalization means for normalizing based on a first learning device for learning using the normalized drive command and state signal; The trajectory generator includes the above working manipulator, and the trajectory generator uses the driving command for the working manipulator and the state signal of the working manipulator to give the command trajectory for the working manipulator, and the driving command and the state signal for the working manipulator. a normalization means for normalizing based on the specifications of the manipulator for work; a second learning device for inferring using the normalized drive command and state signal; a denormalization means for denormalizing based on the denormalization means, the output of the denormalization means is used as the command trajectory of the working manipulator, and the second learning device is the first learning device. Manipulator learning system."

According to the present invention, a common trajectory generator can be provided for a plurality of working manipulators having different configurations.

Problems, configurations, and effects other than those described above will be clarified by the following description of the embodiments.

The figure which shows the structural example of the manipulator system based on the Example of this invention. The figure which shows roughly the structural example of a learning apparatus. The figure which shows roughly the structural example of a trajectory generator. The figure which shows roughly another structural example of a trajectory generator. FIG. 4 schematically illustrates an example of upsampling; FIG. 4 schematically shows an example of downsampling;

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described with reference to the drawings. The present invention is not limited to the examples, and various signals and the like in the examples are examples. In addition, in the present specification and drawings, the same components or components having substantially the same functions are denoted by the same reference numerals, and overlapping descriptions are omitted.

FIG. 1 is a schematic diagram showing an example of the configuration of a manipulator system according to an embodiment of the invention.

The manipulator system 1 comprises a learning manipulator 11 , a learning device 12 and one or more working manipulators 13 . The learning manipulator 11 may be one or part of one or more working manipulators 13 .

The learning manipulator 11 is composed of a learning mechanism 111 , a learning control device 112 and an operation terminal 113 . Of these, the learning mechanism 111 includes a learning working unit 1111 that performs a predetermined work, a learning moving mechanism 1112 that arranges the learning working unit 1111 in a desired position and posture, and a learning mechanism 111 that senses the state of the learning mechanism 111 . It is composed of a sensor 1113 for learning. Here, the learning sensor 1113 includes, for example, an encoder that detects the joint angle of the learning moving mechanism 1112 and a force/torque sensor that can detect work reaction force acting on the learning working unit 1111 .

The learning control device 112 is composed of a learning drive unit 1121 , a learning state calculation unit 1122 and a learning storage unit 1123 . The learning drive unit 1121 drives the learning mechanism 111 in accordance with a drive command 201 from the operation terminal 113 operated by an operator or the like.

Based on the sensor output 202 of the learning sensor 1113, the learning state calculation unit 1122 calculates the state signal 203 of the learning mechanism 111, the target work, and the like. At this time, the learning state calculation unit 1122 calculates and outputs at least one state signal 203 that does not depend on the configuration of the learning mechanism 111 . For example, it is the position/orientation of the learning work unit 1111 and work reaction force. The position/orientation of the learning working unit 1111 can be calculated by coordinate conversion from the link length and joint angle of the learning moving mechanism 1112, and the work reaction force is the relative position/posture between the force/torque sensor and the learning working unit 1111. It can be calculated by coordinate transformation from the posture. The learning storage unit 1123 stores the drive command 201 output from the operation terminal 113 and the state signal 203 output from the learning state calculation unit 1122 .

With this configuration, the learning manipulator 11 simulates or actually performs the target work with the learning mechanism 111 based on the operation of the operation terminal 113 by an operator or the like, while the learning mechanism 111 simulates or actually performs the target work. The signal 203 is collected and stored in the learning storage unit 1123 .

In FIG. 1, the learning device 12 learns data such as the drive command 201 and the work state 203 stored in the learning storage unit 1123 through simulated or actual work by the learning manipulator 11, and performs the target work. Build a learner that generates the trajectory of the position and posture of the working part. Details of the learning device 12 will be described later with reference to FIG.

One or more working manipulators 13 are composed of a working mechanism 131 and a working control device 132 . Here, the working mechanism 131 of the working manipulator 13 may have the same configuration as the learning mechanism 111 of the learning manipulator 11, or may have different dimensions. . Conventionally, what is learned with the same model is applied to the actual device, but in the present invention, the learning result can be applied to different types of devices.

Among them, the working mechanism 131 of the working manipulator 13 is composed of a working working part 1311 , a working moving mechanism 1312 , and a working sensor 1313 , like the learning mechanism 111 . As with the learning sensor 1113, the working sensor 1313 may be, for example, an encoder or a force/torque sensor.

The work control device 132 includes a work drive unit 1321 , a work state calculation unit 1322 and a trajectory generator 1323 . The working drive unit 1321 drives the working mechanism 131 in accordance with the position/orientation command trajectory 204 of the working working unit 131 , which is output from the trajectory generator 1323 . Like the learning state calculator 1322, the work state calculator 1322 calculates and outputs the state signal 203. FIG.

The trajectory generator 1323 uses the learning device 121, which will be described later in FIG. 131 position/orientation instruction trajectories are generated and output. Here, the learning device 121 is a learning function configured by reflecting (implanting) the results of learning by the learning device 12 of FIG.

With this configuration, the working manipulator 13 drives the working working unit 1311 based on the command trajectory 204 generated by the trajectory generator 1323 having the learner 121 that has learned the data when the learning manipulator 11 performs the target work. By doing so, the target work can be performed.

FIG. 2 is a schematic diagram showing an example of the configuration of the learning device 12. As shown in FIG. The learning device 12 comprises a learning device 121 , a normalization calculation section 122 and a normalization calculation section 123 . Inputs to the learning device 12 are, as shown in FIG. In this example, the drive command 201 includes command trajectory data 201a at the current control time, command trajectory data 201c at the next control cycle, and final target values 201b for position and attitude.

The normalization calculation unit 122 normalizes each input (the driving command 201 and the state signal 203) based on the specifications of the learning manipulator 11. FIG. For example, if the learning manipulator 11 has a vertical operation range from ZL1 to ZH1, the input signal Z representing vertical data is normalized by applying equation (1).

Further, for example, when the stationary external force value applied to the learning working portion 1111 in the learning manipulator 11 is Fs1, the input signal F indicating the external force applied to the learning working portion 1111, such as work reaction force, is expressed by the equation (2). Standardize by applying.

Also, the normalization calculation unit 123 normalizes each input to a value between 0 and 1. FIG. For example, when the possible values of the input are from WL to WH, the input W is normalized by applying equation (3).

Here, normalization is to set the maximum and minimum values to 1.0 and 0 and set the input within this range. Therefore, it may become 1.0 or more in a state such as an overload.

The learning device 121 can be composed of, for example, a neural network. In FIG. 2, examples of inputs to the learning device 121 include a final target value 201b, command trajectory data 201a in the current control cycle, and state signal data 203. These are standardized and normalized to Enter 121.

The final target value 201b is the target value of the position/orientation of the learning working unit 1111 when the target work is completed.・It should be posture. The command trajectory data 201a in the current control cycle is command data of the position/orientation of the learning working unit 1111 including at least the current time. The state signal data 203 is data including at least the current time, and includes, for example, realization values of the position/orientation of the learning work unit 1111, work reaction force, and the like.

The output of the learning device 121 is the next cycle command trajectory generation value 210, which is the command value of the position/orientation of the learning working unit 1111 in the next cycle. The learning device 12 sets the parameters of the learning device 121 to the next cycle command trajectory generation value 210 and the position/orientation of the learning working unit 1111 in the next cycle when data is acquired in the target work using the learning manipulator 11. Learning is performed so as to reduce the error 211 with the normalized next cycle command trajectory data 209 obtained by standardizing and normalizing the trajectory points.

As a result, from the final target value 201b, the command trajectory data 201a indicating the current command and work situation, and the state signal data 203, the next cycle command indicating the command trajectory of the position/orientation of the learning working unit 1111 in the next cycle Inferential computation of trajectory generation values 210 is enabled. By using past data in addition to current data for the input signals indicating the current state such as the command trajectory data 201a and the state signal data 203, it becomes possible to learn the temporal connection of the command trajectories, and to continuously issue commands. Trajectories can be generated. Similarly, continuous instruction trajectories can also be generated by constructing the learning device 121 with a recurrent neural network, or by constructing a network configuration that enables time-series learning.

Furthermore, here, by learning using an input signal that eliminates the influence of the specifications of the learning manipulator 11 through normalization by the normalization calculation unit 122, a working manipulator having specifications such as a size different from those of the learning manipulator 11 is obtained. 13, the application of the learning device 121 becomes possible.

Furthermore, the state signal 203, which is the input signal of the learning device 121, is a state quantity that is not affected by the configuration of the learning manipulator 11, such as the position/orientation and work reaction force of the learning working unit 1111, so that the link length, etc. It is possible to apply the learning device 121 to the working manipulators 13 having different configurations.

FIG. 3 is a schematic diagram showing an example of the configuration of the trajectory generator 1323. As shown in FIG. The trajectory generator 1323 includes a working normalization calculation unit 13231, a working normalization calculation unit 13232, a learning device 121, a delay device 13233, an inverse normalization calculation unit 13234, and an inverse normalization calculation unit 13235. Become.

The work time normalization calculation unit 13231 normalizes the final target value 201b and the state signal 203 given by the work instruction from the work terminal 133, which is an external device. Here, for example, if the target work is to pick up an article, even if the operating range of the working working part 1311 and the specification of the external force are different between the learning manipulator 11 and the working manipulator 13, the work is essentially unaffected. In this case, the working normalization calculation unit 13231 uses the specifications of the working manipulator 13 to normalize the input signal so that the working manipulator 13 specifications can be used effectively. Further, for example, for an input signal whose absolute magnitude is important in the target work, such as work reaction force in cutting work, the specification of the learning manipulator 11 is used to standardize the input signal, and other input signals is standardized using the specifications of the working manipulator 13 .

The working normalization calculation unit 13232 normalizes the normalized final target value 201b and the state signal 203 so as to have values between 0 and 1. FIG. At this time, the maximum value WH and the minimum value WL used in the normalization calculation shown in equation (3) are the same values as those of the normalization calculation section 123 in the learning device 12 .

The learning device 121 uses the learning device 121 trained by the learning device 12 . By delaying the next cycle trajectory generation value 210, which is the output of the learning device 121, by one cycle with the delay device 13233, it is input to the learning device 121 as normalized command trajectory data 212 together with the normalized final target value 201b and the state signal 203. By doing so, the next periodic trajectory generation value 210 can be inferentially calculated. Here, when past values are used in addition to the present values for the state signal 203 and the normalized command trajectory data 212, past values of the state signal input 203 and the normalized command trajectory data 212 are stored in a memory (not shown). should be saved and used.

The inverse normalization calculation unit 13234 inversely normalizes the next periodic trajectory generation value 210 (W2 in the expression) using expression (4). At this time, for WH and WL, the parameters of equation (3) used for normalization of the position/orientation of the work unit 1311 by the work normalization calculation unit 13232 are used.

The denormalization calculation section 13235 denormalizes the output signal (G2 in the formula) of the denormalization calculation section 13232 using the equation (5). At this time, for GH and GL, the parameters used for normalization of the position/orientation of the working part 1311 by the working normalization calculating part 13231 are used.

With this configuration, the working manipulator 13 can generate the position/orientation command trajectory 204 of the working working part 1311 using the learner 121 learned by the learning manipulator 11 in the target work.

For example, by configuring the manipulator system 1 as described above, when performing the target work with the working manipulator 13, the learner 121 learned in the target work with the learning manipulator 11 is used to determine the position of the working working unit 1311. • A command trajectory 204 for attitude can be generated. Here, by standardizing the input of the learning device 121, it becomes possible to apply the learning device 121 to working manipulators 13 having different specifications such as size.

In addition, by setting the state signal 203 in the learning state calculation unit 1122 and the work state calculation unit 1322 to a state quantity that is not affected by the configuration of the learning manipulator 11 and the work manipulator 13, work with different configurations such as link length can be performed. It becomes possible to apply the learning device 121 to the manipulator 13 for use.

In other words, by learning the target work with the learning manipulator 11, one or more work manipulators 13 with different specifications and configurations can perform the target work.

According to the first embodiment, even if the manipulator for learning 11 and the manipulator for work 13 are different models having different sizes, etc., the learning result of the manipulator for learning 11 can be transplanted and reflected in the manipulator for work 13. It becomes possible.

However, the problem assumed in this case is that since the sizes are different, for example, if it is assumed that the control amount in one control cycle is the same value, for example, the size of the working manipulator 13 is larger than that of the learning manipulator 11. If it is large, the time required for movement to the target position will be long. Conversely, when the size is small, the time required for movement is short, and it becomes difficult to secure positional accuracy. In Example 2, this point is further reviewed. In addition, in the second embodiment, only parts different from the manipulator system 1 described in the first embodiment are shown.

FIG. 4 is a schematic diagram showing an example of the configuration of a trajectory generator according to the second embodiment. The trajectory generator 1323 includes a work normalization calculation unit 13231, a work normalization calculation unit 13232, a learning device 121, a delay device 13233, an inverse normalization calculation unit 13234, an inverse normalization calculation unit 13235, It consists of a sampler A13236 and a sampler B13237.

Compared with the first embodiment, the configuration is such that samplers 13236 and 13237 are added before and after the learning device 121 . By applying the samplers 13236 and 13237, the operation cycle of the computer in the learning device 121 portion can be set to TC with respect to the control operation cycle TB of the working manipulator 13 .

In FIG. 4, the trajectory generator 1323 first performs the inference calculation by the learning device 121 in the inference calculation cycle TC different from the work control cycle TB which is the control calculation cycle of the work manipulator 13 .

When TB has a longer period than TC, the sampler A 13236 converts the period TB signal into period TC by upsampling, and the sampler B 13237 converts the period TC signal into period TB by downsampling.

When TC has a period longer than TB, the sampler A 13236 converts the period TB signal into period TC by downsampling, and the sampler B 13237 converts the period TC signal into period TB by upsampling.

For example, when the data recording cycle of the learning manipulator 11 to the learning storage unit 1123 is TA, the operating range is RA, the maximum speed is VA, and the operating range of the working manipulator 13 is RB, and the maximum speed is VB, By operating the learner 121 with the inference period TC derived from the equation (6), the command trajectory 204 generated by the trajectory generator 1323 in the working manipulator 13 becomes a trajectory of the maximum speed VB or less. That is, by using a plurality of cycles for the trajectory generator 1323, generation of trajectories that cannot be followed due to normalization or different control cycles can be prevented.

FIG. 5 is a schematic diagram illustrating an example of an upsampling method. In the sampler A13236 or sampler B13237, if the period of the input signal 301 is T1 and the period of the output signal 302 is T2, upsampling can be performed by the method shown in FIG. 5, for example, when T1 is longer than T2.

Interpolation characteristics 303 are obtained by interpolating from time (k−1)T1 to time kT1 at which the next input signal 301 enters. At the time iT2 when the output signal 302 is generated, the value of the interpolated characteristic 303 is used as the output signal 302 .

This makes it possible to generate an output signal 302 having a period T2 shorter than the period T1 of the input signal 301. FIG.

Here, in FIG. 5, the interpolated characteristic 303 is generated by interpolating the input signal 301 at the immediately preceding time (k−1) T1 with a zero-order function, but this is not the only option. Further past input signals 301 such as time (k-2)T1 may be used to interpolate with a linear function or higher, or a spline function or the like may be used.

FIG. 6 is a schematic diagram illustrating an example of a downsampling method. In the sampler A13236 or sampler B13237, if the period of the input signal 301 is T1 and the period of the output signal 302 is T2, down-sampling can be performed by the method shown in FIG. 6, for example, when T2 is longer than T1.

Interpolation characteristics 303 are obtained by interpolating from time (k−1)T1 to time kT1 at which the next input signal 301 enters. At the time iT2 when the output signal 302 is generated, the value of the interpolated characteristic 303 is used as the output signal 302 .

This makes it possible to generate an output signal 302 having a period T2 longer than the period T1 of the input signal 301. FIG.

Here, in FIG. 6, the interpolated characteristic 303 is generated by interpolating the input signal 301 at the immediately preceding time (k−1)T1 with a 0th-order function, but this is not the only option. Further past input signals 301 such as time (k-2)T1 may be used to interpolate with a linear function or higher, or a spline function or the like may be used.

1: Manipulator system 11: Learning manipulator 111: Learning mechanism 1111: Learning working unit 1112: Learning moving mechanism 1113: Learning sensor 112: Learning control device 1121: Learning driving unit 1122: Learning state calculation unit 1123: storage unit for learning 113: operation terminal 12: learning device 121: learning device 122: normalization calculation unit 123: normalization calculation unit 13: working manipulator 131: working mechanism 1311: working working part 1312: working Moving mechanism 1313: Learning sensor 132: Work control device 1321: Work drive unit 1322: Work state calculation unit 1323: Trajectory generator 13231: Work normalization calculation unit 13232: Work normalization calculation unit 13233: Delay Unit 13234: Inverse normalization calculation unit 13235: Inverse normalization calculation unit 133: Work terminal 201: Drive command 202: Sensor output 203: State signal 204: Commanded trajectory 201b: Final target value 201a: Commanded trajectory data 208: Next cycle command Trajectory data 209: Normalized next cycle command trajectory data 201c: Next cycle command trajectory generation value 211: Error 212: Normalized command trajectory data 301: Interpolation characteristics

Claims (9)

A learning device that performs learning using a drive command and a state signal of the learning manipulator when the learning manipulator is driven in accordance with the drive command,
The learning device includes normalization means for normalizing the driving command and the state signal based on the specifications of the manipulator for learning, and a first learning device for learning using the normalized driving command and the state signal. A learning device comprising: The learning device according to claim 1,
The learning device, wherein the drive command includes command trajectory data in the current control cycle and the next control cycle, and final target values of position and attitude. The learning device according to claim 2,
The first learning device receives the command trajectory data in the current control cycle, the final target value, and the state signal as inputs, and performs learning that optimizes the difference between the learning result and the command trajectory data in the next control cycle. A learning device characterized by executing The learning device according to any one of claims 1 to 3,
A learning device comprising normalization means for converting the output of the normalization means into a value within a predetermined numerical range, and using the output as the input of the first learning device. A trajectory generator that provides a command trajectory for the working manipulator using a drive command for the working manipulator and a state signal of the working manipulator,
The trajectory generator includes standardization means for standardizing the drive command and the state signal for the work manipulator based on the specifications of the work manipulator, and a trajectory generator for reasoning using the standardized drive command and the state signal. 2 learners and denormalization means for denormalizing the output of the second learner based on the specifications of the working manipulator, wherein the output of the denormalization means is converted to the command of the working manipulator. Orbital and
In order to perform learning using the driving command and the state signal of the learning manipulator, the second learning device transmits the driving command and the state signal of the learning manipulator based on the specifications of the learning manipulator. A trajectory generator, wherein the trajectory generator is a first learning device that is normalized and learns using the normalized drive command and the state signal. A trajectory generator according to claim 5, comprising:
The drive command includes final target values of position and orientation,
The second learning device is characterized in that the command trajectory data in the next control cycle, which is the output thereof, is delayed by one cycle, and input together with the final target value and the state signal as command trajectory data in the current control cycle for learning. trajectory generator. A trajectory generator according to claim 5 or claim 6,
The second learning device includes normalization means for converting the output of the normalization means into a value within a predetermined numerical range, and denormalization means for denormalizing the output of the second learning device, A trajectory generator, wherein the output of denormalization means is used as the input of said denormalization means. A trajectory generator according to any one of claims 5 to 7,
A trajectory generator, wherein the trajectory generator is composed of a computer, and a control cycle in the second learning device and a control cycle in portions other than the second learning device are different. A manipulator system for reflecting learning results of a learning manipulator on a plurality of types of working manipulators,
a learning manipulator that stores the driving command and the state signal when the learning mechanism is driven according to the driving command;
A learning device that performs learning using a drive command and a state signal of the learning manipulator when the learning manipulator is driven, wherein the drive command and the state signal are standardized based on the specifications of the learning manipulator. a learning device comprising normalization means and a first learning device that learns using the normalized drive command and the state signal;
one or more working manipulators driven based on commanded trajectories from a trajectory generator;
The trajectory generator provides a command trajectory for the work manipulator using a drive command for the work manipulator and a state signal of the work manipulator, and generates the drive command and the state signal for the work manipulator according to the specifications of the work manipulator. a second learning device for reasoning using the standardized drive command and the state signal; and an output of the second learning device based on the specifications of the working manipulator. denormalization means for performing denormalization, the output of the denormalization means is used as the command trajectory of the working manipulator, and the second learning device is the first learning device. and manipulator system.

Images