JP2021124403A - Control device and control method for automatically manipulating robot - Google Patents

Abstract

To provide a control device and control method for an automatically manipulating robot (drive robot).SOLUTION: The control device for an automatically manipulating robot comprises: a manipulation content inference unit 41 for inferring manipulation in a first cycle by a manipulation inference learning model 50 generated by reinforced training of a machine learner; an adjustment coefficient inference unit 45 for inferring an adjustment coefficient by an adjustment coefficient inference learning model 70 that adjusts the manipulation inferred by the manipulation content inference unit 41 during the first cycle on the basis of a traveling state and is generated by reinforced training of a machine learner so as to infer an adjustment coefficient; and a vehicle manipulation control unit 22 for adjusting manipulation by the adjustment coefficient and generating manipulation after adjustment during the first cycle and controlling an automatically manipulating robot 4 on the basis of the manipulation after adjustment.SELECTED DRAWING: Figure 2

Description

The present invention relates to a control device and a control method for an autopilot robot.

Generally, when manufacturing and selling vehicles such as ordinary automobiles, the fuel consumption and exhaust gas when the vehicle is driven according to a specific driving pattern (mode) specified by the country or region are measured and displayed. There is a need to.
The mode can be represented by a graph as, for example, the relationship between the time elapsed from the start of traveling and the vehicle speed to be reached at that time. This vehicle speed to be reached is sometimes referred to as a command vehicle speed in terms of a command regarding the speed to be achieved given to the vehicle.
The above tests on fuel consumption and exhaust gas are carried out by placing the vehicle on the chassis dynamometer and driving the vehicle according to the mode by the autopilot robot, so-called drive robot (registered trademark) mounted on the vehicle. Will be done.

The margin of error is specified for the command vehicle speed. If the vehicle speed deviates from the margin of error, the test becomes invalid. Therefore, the control of the autopilot robot is required to have high followability to the commanded vehicle speed. Therefore, for example, reinforcement learning technology may be applied to the control of the autopilot robot.
For example, Patent Document 1 discloses a vehicle driving simulation device, a driver model construction method, and a driver model construction program capable of constructing a driver model that performs a human-like pedal operation by reinforcement learning.
More specifically, the vehicle driving simulation device travels the vehicle model multiple times while changing the gain value of the driver model, and evaluates the changed gain value based on the reward value. The gain of the driver model is set automatically.
In the configuration of Patent Document 1, the driver model PID controls the vehicle by using the gain value determined in advance by reinforcement learning when the vehicle is actually driven.

Japanese Unexamined Patent Publication No. 2014-115168

The driving mode for driving the vehicle and measuring the characteristics, for example, the WLTC (World Harmonized Light Vehicles Test Cycle) mode, includes a wide variety of patterns of driving modes. In a device that controls a vehicle by a predetermined gain value as in Patent Document 1, the vehicle can be made to follow a commanded vehicle speed with high accuracy by flexibly responding to each of such a wide variety of patterns. Is not easy.

On the other hand, a machine learning device such as a neural network, which is constructed to input the vehicle state such as the detected vehicle speed and the command vehicle speed and output the operation of the vehicle suitable for the state, is trained by reinforcement learning. , It is conceivable to generate a learning model that infers the operation. When the vehicle is actually driven, the drive robot is controlled so as to input the state of the vehicle into the operation inference learning model and apply the operation inferred by the operation inference learning model to the vehicle.
In general, inference by a learning model generated by learning a machine learning device such as a neural network tends to require a large amount of calculation. Therefore, the inference cycle, which is the time interval of the inference time for inferring the operation by the operation inference learning model, is longer than the control cycle, which is the time interval of the control time that actually controls the drive robot, and is within one inference cycle. May include multiple control times.
In such a case, it is conceivable to apply the same operation as the operation most recently inferred by the operation inference learning model at all of the plurality of control times included in a certain inference cycle, but this is a precise control. However, high followability to the command vehicle speed cannot be expected.
Alternatively, it is also conceivable to infer operations at all of a plurality of control times included in the next inference cycle at once. However, in this case, the structure of the operation inference learning model becomes complicated because the number of inferred operations increases. Also, learning an operation inference learning model is not easy.

The problem to be solved by the present invention is an autopilot robot (drive robot) in which the structure of a learning model for inferring vehicle operation is simple, machine learning is easy, and the commanded vehicle speed can be followed with high accuracy. Is to provide a control device and a control method for the above.

The present invention employs the following means in order to solve the above problems. That is, the present invention is a control device for an automatic control robot that controls an automatic control robot mounted on a vehicle to drive the vehicle so that the vehicle travels in accordance with a specified command vehicle speed. An operation inference learning model generated by strengthening learning of a machine learning device so as to infer the operation of the vehicle such that the vehicle travels according to the commanded vehicle speed based on the traveling state of the vehicle including the commanded vehicle speed. An adjustment coefficient that adjusts the operation inferred by the operation content inference unit during the first cycle based on the operation content inference unit that infers the operation in the first cycle and the running state. With the adjustment coefficient inference learning model generated by strengthening the machine learning device so as to infer Provided is a control device for an automatic control robot, comprising a vehicle operation control unit that adjusts the operation to generate an adjusted operation and controls the automatic control robot based on the adjusted operation.

Further, the present invention is a control method for an automatic control robot, which controls an automatic control robot mounted on a vehicle to drive the vehicle so that the vehicle travels in accordance with a specified command vehicle speed. An operation inference learning model generated by reinforcement learning of a machine learning device so as to infer the operation of the vehicle such that the vehicle travels according to the commanded vehicle speed based on the traveling state of the vehicle including the commanded vehicle speed. The machine learning device is strengthened so as to infer the adjustment coefficient, which infers the operation in the first cycle and adjusts the inferred operation during the first cycle based on the running state. The adjustment coefficient is inferred by the adjustment coefficient inference learning model generated by learning, and during the first cycle, the operation is adjusted by the adjustment coefficient to generate an adjusted operation, and the adjusted operation is generated. Provided is a control method of an automatic control robot that controls the automatic control robot based on the above.

According to the present invention, a control device for an autopilot robot (drive robot), which has a simple structure of a learning model for inferring vehicle operation, easy machine learning, and can follow a commanded vehicle speed with high accuracy. A control method can be provided.

It is explanatory drawing of the test environment using the autopilot robot (drive robot) in embodiment of this invention. It is a block diagram of the control device of the autopilot robot in the said embodiment. It is a processing block diagram which shows the data flow of the said control device. It is a flowchart at the time of learning in the control method which controls the autopilot robot. It is a flowchart at the time of running control of a vehicle for performance measurement in the control method of the autopilot robot. It is a processing block diagram which shows the data flow of the control device of the autopilot robot in the 1st modification of the said Embodiment. It is a processing block diagram which shows the data flow of the control device of the autopilot robot in the 2nd modification of the said Embodiment.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
In the present embodiment, since the drive robot (registered trademark) is used as the autopilot robot, the autopilot robot will be referred to as a drive robot below.

FIG. 1 is an explanatory diagram of a test environment using a drive robot in the embodiment. The test device 1 includes a vehicle 2, a chassis dynamometer 3, and a drive robot 4.
The vehicle 2 is provided on the floor surface. The chassis dynamometer 3 is provided below the floor surface. The vehicle 2 is positioned so that the drive wheels 2a of the vehicle 2 are placed on the chassis dynamometer 3. When the vehicle 2 travels and the drive wheels 2a rotate, the chassis dynamometer 3 rotates in the opposite direction.
The drive robot 4 is mounted on the driver's seat 2b of the vehicle 2 to drive the vehicle 2. The drive robot 4 includes an actuator 4c. The actuator 4c is provided so as to come into contact with the accelerator pedal 2c of the vehicle 2.

The drive robot 4 is controlled by the control device 11 described in detail later. The control device 11 changes and adjusts the opening degree of the accelerator pedal 2c of the vehicle 2 by controlling the actuator 4c of the drive robot 4.
The control device 11 controls the drive robot 4 so that the vehicle 2 travels according to a specified command vehicle speed. That is, the control device 11 controls the traveling of the vehicle 2 so as to follow the defined traveling pattern (mode) by changing the opening degree of the accelerator pedal 2c of the vehicle 2. More specifically, the control device 11 controls the traveling of the vehicle 2 so as to follow the commanded vehicle speed, which is the vehicle speed to be reached at each time, as time elapses from the start of traveling.

The control device 11 includes a drive robot control unit 20 and a learning unit 30.
The drive robot control unit 20 controls the drive robot 4 by generating a control signal for controlling the drive robot 4 and transmitting the control signal to the drive robot 4. The learning unit 30 performs machine learning as described later to generate an operation inference learning model, a first action value inference learning model, an adjustment coefficient inference learning model, and a second action value inference learning model. The control signal for controlling the drive robot 4 as described above is generated based on the inference results by the operation inference learning model and the adjustment coefficient inference learning model.
The drive robot control unit 20 is, for example, an information processing device such as a controller provided outside the housing of the drive robot 4. The learning unit 30 is an information processing device such as a personal computer.

FIG. 2 is a block diagram of the test device 1 and the control device 11. FIG. 3 is a processing block diagram showing a data flow of the test device 1 and the control device 11.
The test device 1 includes a vehicle condition measuring unit 5 in addition to the vehicle 2, the chassis dynamometer 3, and the drive robot 4 as described above. The vehicle condition measuring unit 5 is various measuring devices for measuring the condition of the vehicle 2 and the chassis dynamometer 3. In the present embodiment, the vehicle state measuring unit 5 detects _{the engine speed n det} , the engine temperature d _det , and the vehicle speed v _{det of the vehicle 2.} Each of these detected values is transmitted to the drive robot control unit 20 of the control device 11 described below.

The drive robot control unit 20 includes a vehicle operation control unit 22 and a drive state acquisition unit 23. The vehicle operation control unit 22 includes an operation complement unit 24. The operation complementing unit 24 includes a traveling resistance calculation unit 25, a feedback operation amount calculation unit 26, and a vehicle driving force calculation unit 27. The learning unit 30 includes a command vehicle speed generation unit 31, an inference data molding unit 32, a learning data molding unit 33, an operation learning data generation unit 34, a learning data storage unit 35, an adjustment coefficient learning data generation unit 36, and a reinforcement learning unit 40. I have. The reinforcement learning unit 40 includes an operation content inference unit 41, a first action value inference unit 42, a reward calculation unit 43, an adjustment coefficient inference unit 45, and a second action value inference unit 46. The reward calculation unit 43 includes an operation reward calculation unit 44 and an adjustment coefficient reward calculation unit 47.
Each component of the control device 11 other than the learning data storage unit 35 may be, for example, software or a program executed by the CPU in each of the above-mentioned information processing devices. Further, the learning data storage unit 35 may be realized by a storage device such as a semiconductor memory or a magnetic disk provided inside or outside each of the information processing devices.
Operation contents Each of the inference unit 41, the first action value inference unit 42, the adjustment coefficient inference unit 45, and the second action value inference unit 46 has an operation inference learning model 50, a first action value inference learning model 60, and an adjustment coefficient inference. A learning model 70 and a second action value inference learning model 80 are provided, respectively.

As will be described later, the operation inference learning model 50 of the operation content inference unit 41 infers the operation of the vehicle 2, and the adjustment coefficient inference learning model 70 of the adjustment coefficient inference unit 45 infers the adjustment coefficient of the vehicle 2. The drive robot control unit 20 controls the drive robot 4 based on these inferred operations and adjustment coefficients.
In particular, in the present embodiment, the operation complement unit 24 refers to the operation of the vehicle 2 inferred by the traveling resistance calculation unit 25, the feedback operation amount calculation unit 26, and the vehicle driving force calculation unit 27 that constitute the operation complement unit 24. The feedback system is controlled according to the inferred adjustment coefficient, and the operation actually applied to the drive robot 4 is calculated to control the drive robot 4.
Here, first, the drive robot control unit 20 will be described in detail. Hereinafter, the time interval of the inference time for inferring the operation and the adjustment coefficient in the operation content inference unit 41 and the adjustment coefficient inference unit 45 is referred to as an inference cycle (first cycle) Tnn. Further, the time interval of the control time that actually controls the drive robot 4 is referred to as a control cycle (second cycle) Tdr. In the present embodiment, the inference cycle Tnn is set to be longer than the control cycle Tdr. That is, the same value as the inference result of the operation and the adjustment coefficient at a certain time is applied at all the control times of the drive robot 4 within the time interval until the time after the next inference cycle Tnn. Each of the following operations of the drive robot control unit 20 is executed in the control cycle Tdr.

The drive state acquisition unit 23 receives the detected engine speed n _det , the detected engine temperature d _det , and the detected vehicle speed v _det of the vehicle 2 from the vehicle state measurement unit 5. These values are provided so as to be able to be referred to from each component in the vehicle operation control unit 22.
The vehicle operation control unit 22 receives a _{command vehicle speed v ref} to be obeyed from the command vehicle speed generation unit 31 of the learning unit 30, which will be described later. The vehicle driving force calculation unit 27 of the vehicle operation control unit 22 calculates the _{vehicle driving force F x} by a predetermined approximate formula based on the _{received differential value of the command vehicle speed v ref} and the weight of the vehicle 2.
Running resistance calculating unit 25, based on the detected vehicle speed v _det, it calculates the running resistance F _RL imitating the actual running on the real road surface. The traveling resistance calculation unit 25 _{transmits the traveling resistance FLL} to the chassis dynamometer 3 to generate a traveling resistance force for the traveling vehicle 2.

Driving state acquisition unit 23, a driving force demand _{F ref} vehicle driving force _{F x} and the running resistance _{F RL} is added value, detected engine speed _{n det,} detected engine temperature _{d det,} and the detected vehicle speed (vehicle speed) The v _date is transmitted to the inference data forming unit 32 described later.
The inference data forming unit 32 combines each of the values received from the drive state acquisition unit 23 and the command vehicle speed v _ref separately received from the command vehicle speed generation unit 31 into the operation content inference unit 41 as the running state of the vehicle 2. Send.
The operation content inference unit 41 is reinforcement-learned to infer the operation of the vehicle 2 so as to cause the vehicle 2 to travel according _{to the command vehicle speed vref based on these traveling states.} Operation content The inference unit 41 infers the operation of the vehicle 2 based on the received running state for each inference cycle Tnn. In the present embodiment, the operation target includes the accelerator pedal 2c. Therefore, the operation content inference unit 41 calculates the amount of change in the accelerator opening in the present embodiment. Strictly speaking, the amount of change in the accelerator opening is calculated by inferring the feedforward system based on the required driving force _{Ref calculated from the command vehicle speed vref.} That is, the change amount of the accelerator opening calculated by the operation content inference unit 41 is the feedforward change amount (hereinafter, referred to as the FF change amount) θ _FF .

The inference data forming unit 32 also transmits the traveling state of the vehicle 2 to the adjustment coefficient inference unit 45. The adjustment coefficient inference unit 45 infers the adjustment coefficient for adjusting the FF change amount inferred by the operation content inference unit 41, that is, the operation θ _FF during the next inference cycle Tnn, based on the traveling state. Reinforcement learning is being done. The adjustment coefficient inference unit 45 infers the adjustment coefficient of the vehicle 2 based on the received running state for each inference cycle Tnn. In the present embodiment, the adjustment coefficient includes a proportional gain Kp, an integrated gain Ki, and a differential gain Kd.

The feedback manipulated variable calculation unit 26 receives the vehicle speed error dv, which is the difference between the _{command vehicle speed v ref} and the detected vehicle speed v _det. The feedback manipulated variable calculation unit 26 also receives the inferred adjustment coefficients Kp, Ki, Kd, that is, the proportional gain Kp, the integral gain Ki, and the differential gain Kd from the adjustment coefficient inference unit 45 for each inference cycle Tnn.
The feedback manipulated variable calculation unit 26 feeds back the adjusted variable θ _{FB of the operation θ FF} , that is, the feedback of the accelerator opening degree by feedback control based on the latest inference results of the adjustment coefficients Kp, Ki, and Kd received for each inference cycle Tnn. Change amount (hereinafter referred to as FB change amount) θ _FB is calculated. In particular, in the present embodiment, the feedback control is a PID (Proportional-Differential Control) control. As described above, the feedback manipulated variable calculation unit 26 calculates the adjusted variable θ _FB with a control cycle Tdr shorter than the inference cycle Tnn.

_{The operation complement unit 24 receives the inferred operation θ FF} from the operation content inference unit 41 for each inference cycle Tnn.
_{The operation complement unit 24 adds the adjustment amount θ FB} calculated by the feedback operation amount calculation unit 26 to the latest inference result of _{the operation θ FF} received for each inference cycle Tnn, and performs the adjusted operation θ _ref , that is, Calculate the amount of change θ _{ref that is actually used.} As described above, the operation complement unit 24 calculates the _{adjusted operation θ ref with a control cycle Tdr shorter than the inference cycle Tnn.}
The operation complement unit 24 _{transmits this adjusted operation θ ref} to the drive robot 4. The drive robot 4 changes the accelerator opening degree by driving the actuator 4c based on the _{adjusted operation θ ref and operating the accelerator pedal 2c.}

Thus, the vehicle operation control unit 22, the adjustment factor Kp, Ki, and adjust the operation theta _FF by Kd to generate adjusted operation theta _ref, controls the drive robot 4 based on the adjusted operating theta _ref. The operation θ _FF and the adjustment coefficients Kp, Ki, and Kd are inferred and updated with an inference period Tnn longer than the control period Tdr.

Next, the learning unit 30 will be described.
_{As described above, the operation content inference unit 41 infers the operation θ FF} of the vehicle 2 after the time based on the traveling state at a certain time. _{In order to effectively infer the operation θ FF} of the vehicle 2, the operation content inference unit 41 is provided with a machine learning device as will be described later, and the drive robot 4 based on the _{inferred operation θ FF is provided.} The machine learning device is strengthened and learned based on the reward calculated based on the running state at the time after the operation of, and the operation inference learning model 50 is generated. When actually controlling the running of the vehicle 2 for performance measurement, the operation content inference unit 41 infers the operation θ _{FF of the} vehicle 2 by using the operation inference learning model 50 for which this learning is completed.
Further, as described above, the adjustment coefficient inference unit 45 infers the adjustment coefficients Kp, Ki, and Kd of the vehicle 2 after the time based on the traveling state at a certain time. In order to effectively infer the adjustment coefficients Kp, Ki, and Kd of the vehicle 2, the adjustment coefficient inference unit 45 is provided with a machine learning device as will be described later, and the inferred adjustment coefficients Kp, Ki, Ki are provided. , The machine learner is strengthened and learned based on the reward calculated based on the running state at the time after the operation of the drive robot 4 based on Kd, and the adjustment coefficient inference learning model 70 is generated. When the adjustment coefficient inference unit 45 actually controls the running of the vehicle 2 for performance measurement, the adjustment coefficient inference learning model 70 for which this learning is completed is used by the adjustment coefficient inference unit 45 to adjust the adjustment coefficients Kp, Ki, Kd of the vehicle 2. Infer.
That is, the control device 11 mainly includes an operation theta _FF and adjustment factor Kp during reinforcement learning, Ki, learning and Kd, operation theta _FF and adjustment factor Kp in time for travel control of the vehicle 2 for performance measurement, Ki , Kd inference is performed in two ways. In order to simplify the explanation, in the following, first, after explaining each component of the control device 11 at the time of learning _{the operation θ FF and the coordinators Kp, Ki, and Kd numbers, the operation θ at the time of measuring the performance of the vehicle 2 The behavior of each component when inferring FF} and adjustment coefficients Kp, Ki, and Kd will be described.
In FIG. 2, the transmission / reception of data related to the learning models 50 and 70 at the time of learning the learning models 50 and 70 is shown by a broken line.

First, the behavior of the components of the learning unit 30 during learning of the operation θ _FF and the adjustment coefficients Kp, Ki, and Kd will be described.
The command vehicle speed generation unit 31 holds the _{command vehicle speed v ref} generated based on the information regarding the mode. The command vehicle speed generation unit 31 _{transmits the command vehicle speed vref} to the vehicle operation control unit 22 and the inference data molding unit 32.
As described above, the vehicle operation control unit 22 controls the _{drive robot 4 based on the command vehicle speed vref} received from the command vehicle speed generation unit 31 to drive the vehicle 2. The driving state acquisition unit 23 _{collects the required driving force F ref} , the detected engine speed n _det , the detected engine temperature d _det , and the detected vehicle speed (vehicle speed) v _det , and transmits them to the inference data forming unit 32.
The inference data forming unit 32 receives the required driving force _Ref , the detected engine speed n _det , the detected engine temperature d _det , and the detected vehicle speed v _det from the driving state acquisition unit 23. Further, the inference data forming unit 32 receives the command vehicle speed _vref from the command vehicle speed generation unit 31. The inference data forming unit 32 puts these together into a running state, and after appropriately forming them, transmits them to the operation content inference unit 41 and the adjustment coefficient inference unit 45 of the reinforcement learning unit 40.

When the operation content inference unit 41 receives the traveling state, the operation inference learning model 50 during learning infers _{the operation θ FF} of the vehicle 2 for driving the vehicle 2 according _{to the command vehicle speed v ref.} .. Since this operation θ _FF is not updated during the inference cycle Tnn in which the operation content inference unit 41 is executing the next inference, it is continuously used for controlling the drive robot 4 during the next inference cycle Tnn. ..
In the present embodiment, the operation inference learning model 50 includes an input layer having input nodes corresponding to each of the traveling states, a plurality of intermediate layers, and an output node corresponding to _{the operation θ FF of the vehicle 2.} It is a neural network.
When the corresponding running state value is input to each of the input nodes, the calculation based on the weight is performed, and the calculation result is performed for each of the intermediate nodes of the intermediate layer provided as the next stage of the input node. Is stored. Such an operation and storage of the operation result in the intermediate node of the next stage are sequentially executed for each intermediate layer. Finally, the same calculation is performed based on the calculation result stored in the intermediate node in the middle layer of the final stage, and the result is stored in the output node as _{the operation θ FF of the vehicle 2.
The operation content inference unit 41 transmits the operation θ FF} of the vehicle 2 generated in this way to the vehicle operation control unit 22.

Similarly, when the adjustment coefficient inference unit 45 receives the traveling state, the adjustment coefficient inference unit 45 for driving the vehicle 2 according to the command vehicle speed _vref by the adjustment coefficient inference learning model 70 under learning based on the traveling state, the operation content inference unit 41. The adjustment coefficients Kp, Ki, and Kd applied to the operation θ _{FF of the vehicle 2 inferred by the above are inferred.} Since the adjustment coefficients Kp, Ki, and Kd are not updated during the inference cycle Tnn in which the adjustment coefficient inference unit 45 is executing the next inference, the adjustment coefficients Kp, Ki, and Kd continue to be controlled by the drive robot 4 during the next inference cycle Tnn. Is used.
In the present embodiment, the adjustment coefficient inference learning model 70 includes an input layer having an input node corresponding to each of the traveling states, a plurality of intermediate layers, and an output node corresponding to each of the adjustment coefficients Kp, Ki, and Kd. It is a neural network equipped with.
When the corresponding running state value is input to each of the input nodes, the calculation based on the weight is performed, and the calculation result is performed for each of the intermediate nodes of the intermediate layer provided as the next stage of the input node. Is stored. Such an operation and storage of the operation result in the intermediate node of the next stage are sequentially executed for each intermediate layer. Finally, the same operation is performed based on the operation result stored in the intermediate node in the intermediate layer of the final stage, and the result is stored in the output node as the adjustment coefficients Kp, Ki, and Kd.
The adjustment coefficient inference unit 45 transmits the adjustment coefficients Kp, Ki, and Kd thus generated to the vehicle operation control unit 22.

_{The inference of the operation θ FF} of the vehicle 2 and the adjustment coefficients Kp, Ki, and Kd in the operation content inference unit 41 and the adjustment coefficient inference unit 45 as described above is performed for each inference cycle Tnn. Each of the operation inference learning model 50 and the adjustment coefficient inference learning model 70 _{infers only the operation θ FF of the} vehicle 2 and the adjustment coefficients Kp, Ki, and Kd used during the next inference cycle Tnn in one inference. , No further inferences are made. _{Further, the operation θ FF of the} vehicle 2 and the adjustment coefficients Kp, Ki, and Kd used in the next inference cycle Tnn are derived in the next inference.
The vehicle operation control unit 22 _{receives and updates the operation θ FF} of the inferred vehicle 2 and the adjustment coefficients Kp, Ki, and Kd for each inference cycle Tnn. _{The vehicle operation control unit 22 constantly changes the running state based on the latest updated operation θ FF} of the vehicle 2 and the adjustment coefficients Kp, Ki, and Kd until the time after the next inference cycle Tnn. input and generates an adjusted operating theta _ref, controls the drive robot 4 based on the adjusted operating theta _ref.
Operational inference learning model 50 and adjustment coefficient Inference learning model 70 learning, that is, adjustment of the values of each parameter constituting the neural network by the error back propagation method and the stochastic gradient descent method is not performed at this stage, and the operation inference learning is performed. Model 50 and adjustment coefficient inference The learning model 70 _{only infers the operation θ FF of the} vehicle 2 and the adjustment coefficients Kp, Ki, and Kd. The learning of the operation inference learning model 50 and the adjustment coefficient inference learning model 70 is later performed in association with the learning of the first and second action value inference learning models 60 and 80.

As a result of controlling the drive robot 4 based on the inference results of the operation inference learning model 50 and the adjustment coefficient inference learning model 70, the running state of the vehicle 2 is changed. The drive state acquisition unit 23 acquires the changed _{running state as the next running state after the operation θ FF of the} vehicle 2 and the adjustment coefficients Kp, Ki, and Kd are applied.
The reward calculation unit 43 calculates the reward used for reinforcement learning of the operation reasoning learning model 50 and the adjustment coefficient reasoning learning model 70.
More specifically, the operation reward calculation unit 44 includes the running state, the operation θ _FF inferred by the operation inference learning model 50 corresponding to the running state, and the next running newly generated based on the _{operation θ FF.} Based on the condition, the reward is calculated by a well-designed formula. Further, the adjustment coefficient reward calculation unit 47 is newly added based on the running state, the adjustment coefficients Kp, Ki, Kd inferred by the adjustment coefficient inference learning model 70 corresponding to the running state, and the adjustment coefficients Kp, Ki, Kd. Based on the next running condition generated in, the reward is calculated by a properly designed formula.
In the present embodiment, since the control cycle Tdr is shorter than the inference cycle Tnn, the drive robot 4 is controlled a plurality of times at the control cycle Tdr interval during the inference cycle Tnn. Along with this, the reward in the present embodiment is set as an absolute value of the average value of the errors _{of the command vehicle speed v ref} and the detected vehicle speed v _{det after each of the plurality of controls.} That is, in the present embodiment, the absolute value as described above is calculated, and the closer it is to 0, the higher the reward is designed.
The first and second action value inference learning models 60 and 80, which will be described later, calculate the action value so that the smaller the reward, the higher the action value, and the operation inference learning model 50 and the adjustment coefficient inference learning model 70 calculate these action values. Reinforcement learning is performed so as to output the operation θ _FF and the adjustment coefficients Kp, Ki, and Kd that increase the value.

The operation reward calculation unit 44 uses the learning data forming unit 33 to obtain the running state, the operation θ _FF inferred corresponding to the running state, the next running state newly generated based on the operation θ _{FF, and the calculated reward.} Send to. The learning data forming unit 33 appropriately forms these and stores them in the learning data storage unit 35. These data are used for learning the first action value inference learning model 60, which will be described later.
Further, the adjustment coefficient inferring unit 45 includes a traveling state, an adjustment coefficient Kp, Ki, Kd inferred corresponding to the traveling state, a next traveling state newly generated based on the adjustment coefficient Kp, Ki, Kd, and an adjustment coefficient inferring unit 45. The calculated reward is transmitted to the learning data forming unit 33. The learning data forming unit 33 appropriately forms these and stores them in the learning data storage unit 35. These data are used for learning the second action value inference learning model 80, which will be described later.
In this way, _{the inference of the operation θ FF} and the adjustment coefficients Kp, Ki, and Kd, the acquisition of the next running state corresponding to the inference result, and the calculation of the reward are the first and second action value inference learning models. It is repeated until sufficient data is accumulated for learning 60 and 80.

When a sufficient amount of running data for learning the first action value inference learning model 60 is accumulated in the learning data storage unit 35, the first action value inference learning unit 42 learns the first action value inference learning model 60. The first action value inference learning model 60 becomes a learned model in which appropriate learning parameters are learned, which is used as a program module that is a part of artificial intelligence software by learning a machine learning device.
As a whole, the reinforcement learning unit 40 _{calculates the action value indicating how appropriate the operation θ FF} inferred by the operation inference learning model 50 is, and the operation inference learning model 50 increases the action value θ. Reinforcement learning is performed so as to output _FF. The action value is expressed as a function designed so that the larger the reward, the higher the action value, with the running state and the operation θ _{FF for it as arguments.} In the present embodiment, the calculation of this function is performed by the first action value inference learning model 60 as a function approximation device designed to output the action value by inputting the _{running state and the operation θ FF.}

The operation learning data generation unit 34 forms the learning data in the learning data storage unit 35 and transmits it to the first action value inference unit 42.
The first action value inference unit 42 receives the formed learning data and causes the first action value inference learning model 60 to be machine-learned.
In the present embodiment, the first action value inference learning model 60 corresponds to an input layer having input nodes corresponding to each of _{the running state and the operation θ FF} , a plurality of intermediate layers, and the action value related to the _{operation θ FF.} It is a neural network with an output node to do. Since the first action value inference learning model 60 is realized by a neural network having the same structure as the operation inference learning model 50, detailed structural explanation is omitted.

The operation reward calculation unit 44 reduces the error of the TD (Temporal Difference) error, that is, the error _{between the action value before performing the control based on the operation θ FF} and the action value after the control, and sets an appropriate value as the action value. Is output, the values of each parameter constituting the neural network, such as the weight and bias values, are adjusted by the error back propagation method and the stochastic gradient descent method. In this way, the first action value inference learning model 60 is trained so _{that the operation θ FF} inferred by the current operation inference learning model 50 can be appropriately evaluated.

Similarly, when a sufficient amount of running data for learning the adjustment coefficient inference learning model 70 is accumulated in the learning data storage unit 35, the adjustment coefficient inference unit 45 learns the second action value inference learning model 80. .. The second action value inference learning model 80 becomes a learned model in which appropriate learning parameters are learned, which is used as a program module that is a part of artificial intelligence software by learning a machine learning device.
As a whole, the reinforcement learning unit 40 calculates the action value indicating how appropriate the adjustment coefficients Kp, Ki, and Kd inferred by the adjustment coefficient inference learning model 70 are, and the adjustment coefficient inference learning model 70 determines this action value. Reinforcement learning is performed so as to output the adjustment coefficients Kp, Ki, and Kd that increase. The action value is expressed as a function designed so that the larger the reward, the higher the action value, with the running state and the adjustment coefficients Kp, Ki, and Kd for the running state as arguments. In the present embodiment, the second action value inference learning model 80 as a function approximator is designed to output the action value by inputting the running state and the adjustment coefficients Kp, Ki, and Kd for the calculation of this function. To do.

The adjustment coefficient learning data generation unit 36 forms the learning data in the learning data storage unit 35 and transmits it to the second action value inference unit 46.
The second action value inference unit 46 receives the formed learning data and causes the second action value inference learning model 80 to be machine-learned.
In the present embodiment, the second action value inference learning model 80 includes an input layer having input nodes corresponding to the running state and the adjustment coefficients Kp, Ki, and Kd, a plurality of intermediate layers, and an adjustment coefficient Kp. It is a neural network provided with output nodes corresponding to the action values related to Ki and Kd. Since the second action value inference learning model 80 is realized by a neural network having the same structure as the adjustment coefficient inference learning model 70, detailed structural explanation is omitted.

The adjustment coefficient inference unit 45 reduces the TD error, that is, the error between the action value before performing control based on the adjustment coefficients Kp, Ki, and Kd and the action value after control, and sets an appropriate value as the action value. Is output, the values of each parameter constituting the neural network, such as the weight and bias values, are adjusted by the error back propagation method and the stochastic gradient descent method. In this way, the second action value inference learning model 80 is trained so that the adjustment coefficients Kp, Ki, and Kd inferred by the current adjustment coefficient inference learning model 70 can be appropriately evaluated.

As the learning of the first and second action value inference learning models 60 and 80 progresses, each of the first and second action value inference learning models 60 and 80 outputs a more appropriate action value value. That is, since the action value values output by each of the first and second action value inference learning models 60 and 80 are different from those before learning, the operation θ _FF and the adjustment coefficient Kp that increase the action value accordingly. It is necessary to update each of the operation inference learning model 50 and the adjustment coefficient inference learning model 70 designed to output, Ki, and Kd. Therefore, the operation content inference unit 41 and the adjustment coefficient inference unit 45 learn the operation inference learning model 50 and the adjustment coefficient inference learning model 70.
Specifically, each of the operation content inference unit 41 and the adjustment coefficient inference unit 45 uses, for example, a negative value of the action value as a loss function, and makes it as small as possible, that is, an operation θ _{FF that increases the action value.} And adjustment coefficients Kp, Ki, Kd are output, and the values of each parameter that composes the neural network, such as the weight and bias values, are adjusted by the error back propagation method and the stochastic gradient descent method, and the operation inference learning is performed. Each of the model 50 and the adjustment coefficient inference learning model 70 is trained.
When each of the operation inference learning model 50 and the adjustment coefficient inference learning model 70 is learned and updated, the output operation θ _FF and the adjustment coefficients Kp, Ki, and Kd change. Based on this, the first and second behavioral value inference learning models 60 and 80 are learned.
In this way, the learning unit 30 repeats learning between the operation inference learning model 50 and the adjustment coefficient inference learning model 70, and the first and second action value inference learning models 60 and 80, thereby causing the learning model 50, Reinforcement learning of 60, 70, 80.
The learning unit 30 executes this reinforcement learning until a predetermined learning end criterion is satisfied.

_{Next, in the case of inferring the operation θ FF} and the adjustment coefficients Kp, Ki, and Kd when measuring the performance of the vehicle 2, that is, after the reinforcement learning of the operation inference learning model 50 and the adjustment coefficient inference learning model 70 is completed. The behavior of each component of the learning unit 30 will be described.

The command vehicle speed generation unit 31 _{transmits the command vehicle speed vref} to the drive robot control unit 20 and the inference data molding unit 32.
The drive robot control unit 20 transmits the required driving force _Ref , the detected engine speed n _det , the detected engine temperature d _det , and the detected vehicle speed (vehicle speed) v _det to the inference data forming unit 32.
The inference data forming unit 32 receives the required driving force F _ref , the detected engine rotation number n _det , the detected engine temperature d _det , the detected vehicle speed (vehicle speed) v _det , and the commanded vehicle speed v _ref as the traveling state, and appropriately forms the data. After that, it is transmitted to the operation content inference unit 41 and the adjustment coefficient inference unit 45 of the reinforcement learning unit 40.
Operation content inference unit 41 receives the running condition, based on this, the operation inference learning model 50 completion of the learning, during the next inference cycle Tnn, for running the vehicle in accordance with a command vehicle speed v _ref, vehicle Infer the operation θ _FF of 2.
Similarly, adjustment coefficient inference unit 45 receives the running condition, based on this, the adjustment coefficient inference learning model 70 in the training, during the next inference cycle Tnn, for running the vehicle in accordance with a command vehicle speed v _ref The adjustment coefficients Kp, Ki, and Kd applied to _{the operation θ FF} of the vehicle 2 inferred by the operation content inference unit 41 are inferred.

The feedback manipulated variable calculation unit 26 is based on the inferred adjustment coefficients Kp, Ki, Kd, that is, the proportional gain Kp, the integral gain Ki, and the differential gain Kd, and is controlled by PID at a control cycle Tdr interval shorter than the inference cycle Tnn. , The adjustment amount θ _FB is calculated. The adjustment coefficients Kp, Ki, and Kd used in this calculation are inferred and updated by the adjustment coefficient inference unit 45 at intervals of inference period Tnn longer than the control period Tdr.
_{The operation complement unit 24 receives the adjustment amount θ FB} from the feedback operation amount calculation unit 26, _{and based on the inferred operation θ FF} of the vehicle 2, the _{operation complement unit 24 receives the adjustment amount θ ref} at a control cycle Tdr interval shorter than the inference cycle Tnn. To calculate. _{The operation θ FF of the} vehicle 2 used in this calculation is inferred and updated by the operation content inference unit 41 at an inference cycle Tnn interval longer than the control cycle Tdr.
The operation complement unit 24 _{transmits this adjusted operation θ ref} to the drive robot 4. The drive robot 4 changes the accelerator opening degree by driving the actuator 4c based on the _{adjusted operation θ ref and operating the accelerator pedal 2c.}

Next, a method of controlling the drive robot 4 by the control device 11 of the drive robot 4 will be described with reference to FIGS. 1 to 3 and 4 and 5. FIG. 4 is a flowchart at the time of learning in the control method of the drive robot 4. FIG. 5 is a flowchart in the control method of the drive robot 4 when the vehicle 2 is controlled to travel for performance measurement.
First, the operation at the time of learning will be described with reference to FIG.

When learning is started (step S1), each parameter of each learning model 50, 60, 70, 80 and the like is initialized (step S3).
After that, the traveling data of the vehicle 2 is collected (step S5). More specifically, the control device 11 has not yet completed learning, the operation inference learning model 50 in the middle of learning, the operation θ _{FF of the} vehicle 2 inferred by the adjustment coefficient inference learning model 70, and the adjustment coefficients Kp, Ki, By controlling the running of the vehicle 2 by Kd, the running data is accumulated.

When sufficient running data is accumulated in the learning data storage unit 35, the operation inference learning model 50 and the adjustment coefficient inference learning model 70 are strengthened and learned by using the learning data storage unit 35, and the learning models 50 and 70 are updated (step S7). ..
When the update of the operation inference learning model 50 and the adjustment coefficient inference learning model 70 is completed, it is determined whether or not the learning of the operation inference learning model 50 and the adjustment coefficient inference learning model 70 is completed (step S9).
If it is determined that the learning has not been completed (No in step S9), the process proceeds to step S5. That is, the control device 11 further collects the traveling data, and repeats the update of the operation inference learning model 50 and the adjustment coefficient inference learning model 70 using the travel data.
When it is determined that the learning is completed (Yes in step S9), the learning process is ended (step S11).

Next, using FIG. 5, when actually _{inferring the operation θ FF} of the vehicle 2 and the adjustment coefficients Kp, Ki, and Kd when actually measuring the performance of the vehicle 2, that is, the operation inference learning model 50 and the adjustment coefficient inference learning. The operation when the vehicle 2 is controlled to travel after the reinforcement learning of the model 70 is completed will be described.

When the vehicle 2 starts traveling (step S51), the traveling environment is initially set, and the control device 11 observes the traveling state at this point as the initial state (step S53).
The inference data forming unit 32 appropriately forms the running state, and then transmits the operation content inference unit 41 and the adjustment coefficient inference unit 45 of the reinforcement learning unit 40.
When the operation content inference unit 41 receives the traveling state, the operation content inference unit 41 infers _{the operation θ FF} of the vehicle 2 for driving the _{vehicle according to the command vehicle speed v ref} during the next inference cycle Tnn based on the traveling state.
Similarly, adjustment coefficient inference unit 45 receives the running condition, based on this, during the next inference cycle Tnn, for running the vehicle in accordance with a command vehicle speed v _ref, inferred by the operation content inference unit 41 The adjustment coefficients Kp, Ki, and Kd applied to the operation θ _{FF of the vehicle 2 are inferred (step S55).}

_{The feedback manipulated variable calculation unit 26 calculates the adjusted variable θ FB} by PID control based on the inferred adjustment coefficients Kp, Ki, and Kd at a control cycle Tdr interval shorter than the inference cycle Tnn.
_{The operation complement unit 24 receives the adjustment amount θ FB} from the feedback operation amount calculation unit 26 at a control cycle Tdr interval shorter than the inference cycle Tnn, and based on the inferred operation θ _FF of the vehicle 2, the operation complement unit 24 after adjustment θ _ref. To calculate.
The operation complement unit 24 _{transmits this adjusted operation θ ref} to the drive robot 4. The drive robot 4 changes the accelerator opening degree by driving the actuator 4c based on the _{adjusted operation θ ref and operating the accelerator pedal 2c.}
Then, the drive state acquisition unit 23 acquires the running state of the vehicle 2 after the operation again in the same manner as in step S53 (step S57).
The drive state acquisition unit 23 transmits the running state of the vehicle 2 after the operation to the learning unit 30.

The control device 11 determines whether or not the running of the vehicle 2 has been completed (step S59).
If it is determined that the running has not been completed (No in step S59), the process proceeds to step S55. _{That is, the control device 11 repeats the operation θ FF} based on the traveling state acquired in step S57, the inference of the adjustment coefficients Kp, Ki, and Kd, and the observation of the further traveling state.
When it is determined that the running is completed (Yes in step S59), the running process is finished (step S61).

Next, the effects of the control device and the control method of the drive robot 4 will be described.

The control device 11 of the drive robot 4 of the present embodiment controls the drive robot (automatic control robot) 4 mounted on the vehicle 2 to drive the vehicle 2 so that the vehicle 2 travels according to a specified command vehicle speed _vref. The operation θ of the drive robot 4 that causes the vehicle 2 to travel according _{to the command vehicle speed v ref} based on the traveling state of the vehicle 2 including the vehicle _{speed v date} and the command vehicle speed v _{ref. The operation content inference unit 41 that infers the operation θ FF} in the inference cycle (first cycle) Tnn by the operation inference learning model 50 generated by strengthening the machine learning device so as to infer the FF, and the running state. Based on the above, the operation θ _FF inferred by the operation content inference unit 41 is adjusted during the inference period Tnn. _{Adjustment coefficient inference The operation θ FF} is adjusted by the adjustment coefficients Kp, Ki, and Kd between the adjustment coefficient inference unit 45 that infers the adjustment coefficients Kp, Ki, and Kd by the adjustment coefficient inference learning model 70 and the inference cycle Tnn, and the operation after adjustment. It _{includes a vehicle operation control unit 22 that generates θ ref} and controls the drive robot 4 based on the adjusted operation θ _ref.
Further, the control method of the drive robot 4 of the present embodiment is such that the drive robot (automatic control robot) 4 mounted on the vehicle 2 and traveling the vehicle 2 travels according to the command vehicle speed _{vref specified by the vehicle 2.} A control method for the drive robot 4 to be controlled, which is an operation θ of the vehicle 2 such that the vehicle 2 is driven according _{to the command vehicle speed v ref} based on the traveling state of the vehicle 2 including the vehicle _{speed v det} and the command vehicle speed v _{ref. The operation θ FF} is inferred by the inference cycle (first cycle) Tnn by the operation inference learning model 50 generated by strengthening the machine learning device so as to infer the _{FF, and is inferred based on the running state.} The adjustment coefficient Kp, is adjusted by the adjustment coefficient inference learning model 70 generated by strengthening the machine learning device so as to infer the adjustment coefficients Kp, Ki, and Kd that adjust the operation θ _{FF during the inference period Tnn.} ki, infer Kd, between inference cycles Tnn, the adjustment factor Kp, Ki, and adjust the operation theta _FF by Kd to generate adjusted operation theta _ref, drive the robot 4 based after the adjustment operation theta _ref Control.
According to the above configuration, the operation reasoning learning model 50 is a vehicle 2 that causes the vehicle 2 to travel according _{to the command vehicle speed v ref} based on the traveling state of the vehicle 2 including the vehicle _{speed v date} and the command vehicle speed v _ref. Reinforcement learning is done to infer the operation θ _{FF of.} Therefore, at least operations inference learning model 50 to infer period Tnn every a cycle of inference operations theta _FF of the vehicle 2, the vehicle 2 the command vehicle speed v _ref in accurately operation theta _FF vehicle 2, such as to follow the output Will be done.
Here, the operation inference learning model 50 as described above tends to have a large amount of calculation. Therefore, the inference cycle Tnn is longer than the control cycle Tdr of the drive robot 4, and a plurality of control times are included in one inference cycle Tnn. Therefore, the operation θ _{FF of the} vehicle 2 is not output so as to correspond to each of the control times individually. In such a case, if the same operation θ _FF of the vehicle 2 is applied to each of the plurality of control times, precise control cannot be performed and the followability to the command vehicle speed cannot be improved.
On the other hand, in the present embodiment, the adjustments that are strengthened and learned to infer the adjustment coefficients Kp, Ki, and Kd that adjust the _{inferred operation θ FF} during the inference cycle Tnn based on the running state. The coefficient inference learning model 70 infers the adjustment coefficients Kp, Ki, and Kd. That is, at each control time included in the inference cycle Tnn, the operation θ _FF is adjusted at any time by the adjustment coefficients Kp, Ki, and Kd, and the drive robot 4 is controlled. Accordingly, supplemented sampled difference inference cycle Tnn the control period Tdr is, during a certain time, even operating theta _FF is not newly inferred, during which can be used while adjusting the operation theta _FF. Therefore, the ability to follow the command vehicle speed is improved.
_{Further, since the same operation θ FF} is adjusted and used at a plurality of control times of the drive robot 4 included in the inference cycle Tnn, the operation inference learning model 50 uses a plurality of operations θ _{FF in one inference.} There is no need to infer. As a result, the structure of the operation inference learning model 50 can be simplified, and the operation inference learning model 50 can be easily machine-learned.

Further, the inference cycle Tnn is set longer than the control cycle (second cycle) Tdr that controls the drive robot 4, and the adjustment coefficient inference learning model 70 also infers the adjustment coefficients Kp, Ki, and Kd for each inference cycle Tnn. , Each of the operation inference learning model 50 and the adjustment coefficient inference learning model 70 _{infers only the operation θ FF} and the adjustment coefficients Kp, Ki, and Kd of the vehicle 2 used during the next inference cycle Tnn in one inference. Then, the vehicle operation control unit 22 uses the latest operation θ _FF of the vehicle 2 and the adjustment coefficients Kp, Ki, and Kd to generate the _{adjusted operation θ ref until the next inference is made.}
Further, the adjustment coefficients Kp, Ki, and Kd include the proportional gain Kp, the integral gain Ki, and the differential gain Kd, and the vehicle operation control unit 22 operates θ by feedback control based on the adjustment coefficients Kp, Ki, and Kd. The adjustment amount θ _{FB of FF} is calculated, and the operation θ _{FF is adjusted based on the adjustment amount θ FB} to generate the adjusted operation θ _ref.
Further, _{the target of the operation θ FF} includes the accelerator pedal 2c.
According to the above configuration, the control device 11 of the drive robot 4 can be appropriately realized.

[First Modified Example of Embodiment]
Next, a modified example of the control device 11 and the control method of the drive robot 4 shown as the above embodiment will be described with reference to FIG. FIG. 6 is a processing block diagram showing a data flow of the control device of the drive robot in this modified example.
The control device of the drive robot 4 in this modification is the integration buffer in which the feedback operation amount calculation unit 26A of the vehicle operation control unit is accumulated by the integration term of PID control with the control device 11 of the drive robot 4 of the above embodiment. The difference is that i_buff is calculated and transmitted to the adjustment coefficient inference unit 45A.
Along with this, the adjustment coefficient inference learning model provided in the adjustment coefficient inference unit 45A has a configuration in which the input layer includes an input node corresponding to the integration buffer i_buff in addition to the input node corresponding to the running state. ing. As a result, the adjustment coefficient inference learning model infers the adjustment coefficient based on the running state and the integration buffer i_buff.

Needless to say, this first modification has the same effect as that of the embodiment described above.
In the configuration of this modification, the integration buffer i_buff used in the feedback manipulation amount calculation unit 26A located at the subsequent stage of the adjustment coefficient inference learning model, in which the adjustment coefficient which is the inference result of the adjustment coefficient inference learning model is used, is It is the input of the adjustment coefficient inference learning model. Therefore, the accuracy of the adjustment coefficient is improved as compared with the above embodiment.

[Second variant of the embodiment]
Next, a further modification of the control device and control method of the drive robot 4 shown as the first modification will be described with reference to FIG. 7. FIG. 7 is a processing block diagram showing a data flow of the control device of the drive robot in this modified example.
The control device of the drive robot 4 in this modification is different from the control device of the drive robot 4 in the first modification, in which the adjustment coefficient inference learning model is integrated with the operation inference learning model in the operation content inference unit 41B and operated. The difference is that the inference learning model and the adjustment coefficient inference learning model are realized as one learning model.

That is, the learning model provided in the operation content inference unit 41B in this modification is an operation of an input layer having input nodes corresponding to each of the traveling state and the integration buffer i_buff, a plurality of intermediate layers, and the vehicle 2. It is a neural network provided with output nodes corresponding to each of θ _{FF and adjustment coefficients Kp, Ki, and Kd.}
Along with this, the behavioral value inference learning model used for reinforcement learning of this learning model _{has a large reward by inputting the running state, the operation θ FF of the} vehicle 2 with respect to this, and the adjustment coefficients Kp, Ki, and Kd. It is a learning model as a function approximator designed to output a moderately high behavioral value.

In such a configuration, the operation content inference unit 41B outputs the adjustment coefficients Kp, Ki, and Kd, which are transmitted to the feedback operation amount calculation unit 26A.
Further, the integration buffer i_buff output by the feedback manipulated variable calculation unit 26A is transmitted to the operation content inference unit 41B and input to the learning model.

Needless to say, this second modification has the same effect as that of the embodiment described above.
In the configuration of this modification, the number of learning models is reduced, so that it can be implemented even in a smaller resource environment.

The control device and control method for the drive robot of the present invention are not limited to the above-described embodiment and each modification described with reference to the drawings, and various other modifications are included in the technical scope thereof. Conceivable.
For example, in the above embodiment, the operation amount of the accelerator pedal is output as the operation of the vehicle, but in addition to this, other operations such as the brake pedal may be output.
Further, in the above embodiment, the learning of the operation inference learning model 50 and the adjustment coefficient inference learning model 70 and the learning of the first action value inference learning model 60 and the second action value inference learning model 80 are repeated. .. However, the order in which these learning models 50, 60, 70, and 80 are trained is not limited to this as long as they are trained with sufficient accuracy. For example, after repeating the learning of the operation inference learning model 50 and the first action value inference learning model 60 to complete these learnings, the learning of the adjustment coefficient inference learning model 70 and the second action value inference learning model 80 is repeated. These learnings may be completed.
In addition to this, as long as the gist of the present invention is not deviated, the configurations given in the above-described embodiment and each modification can be selected or changed to other configurations as appropriate.

1 Test equipment 2 Vehicle 2c Accelerator pedal 3 Chassis dynamometer 4 Drive robot (autopilot robot)
11 Control device 20 Drive robot control unit 22 Vehicle operation control unit 23 Drive state acquisition unit 24 Operation complement unit 25 Travel resistance calculation unit 26, 26A Feedback operation amount calculation unit 27 Vehicle drive force calculation unit 30 Learning unit 31 Command vehicle speed generation unit 35 Learning data storage unit 40 Enhanced learning unit 41, 41B Operation content reasoning unit 42 First action value inference unit 43 Reward calculation unit 45, 45A Adjustment coefficient inference unit 46 Second action value inference unit 50 Operation inference learning model 60 First action value Inference learning model 70 Adjustment coefficient Inference learning model 80 Second action value Inference learning model θ _FF Feed forward Change amount (operation)
θ _FB feedback change amount (adjustment amount)
Operation after θ _ref adjustment Kp proportional gain (adjustment coefficient)
Ki integrated gain (adjustment coefficient)
Kd derivative gain (adjustment coefficient)
i_buff Integral buffer v _det detection Vehicle speed (vehicle speed)
v _ref command vehicle speed

Claims (7)

An autopilot robot control device that controls an autopilot robot mounted on a vehicle to drive the vehicle so that the vehicle travels according to a specified command vehicle speed.
An operation generated by reinforcement learning of a machine learning device so as to infer an operation of the vehicle such that the vehicle travels according to the commanded vehicle speed based on the traveling state of the vehicle including the vehicle speed and the commanded vehicle speed. An operation content inference unit that infers the operation in the first cycle using an inference learning model, and an operation content inference unit.
Adjustment generated by strengthening learning of a machine learning device so as to infer an adjustment coefficient that adjusts the operation inferred by the operation content inference unit during the first period based on the running state. An adjustment coefficient inference unit that infers the adjustment coefficient using a coefficient inference learning model, and an adjustment coefficient inference unit.
During the first cycle, the vehicle operation control unit adjusts the operation according to the adjustment coefficient to generate the adjusted operation, and controls the autopilot robot based on the adjusted operation.
The control device of the autopilot robot. The first cycle is set longer than the second cycle for controlling the autopilot robot.
The adjustment coefficient inference learning model also infers the adjustment coefficient for each first cycle.
Each of the operation inference learning model and the adjustment coefficient inference learning model infers only the operation of the vehicle and the adjustment coefficient used during the next first period in one inference.
The autopilot robot according to claim 1, wherein the vehicle operation control unit uses the latest operation of the vehicle and the adjustment coefficient to generate the adjusted operation until the next inference is made. Control device. The adjustment coefficients include proportional gain, integral gain, and derivative gain.
The vehicle operation control unit calculates the adjustment amount of the operation by feedback control based on the adjustment coefficient, adjusts the operation based on the adjustment amount, and generates the adjusted operation. Alternatively, the control device for the autopilot robot according to 2. The vehicle operation control unit calculates the integration buffer and
The control device for an autopilot robot according to claim 3, wherein the adjustment coefficient inference learning model infers the adjustment coefficient based on the traveling state and the integration buffer. The control device for an autopilot robot according to any one of claims 1 to 4, wherein the operation inference learning model and the adjustment coefficient inference learning model are realized as one learning model. The control device for an autopilot robot according to any one of claims 1 to 5, wherein the operation target includes an accelerator pedal. A control method for an autopilot robot that controls an autopilot robot mounted on a vehicle to drive the vehicle so that the vehicle travels according to a specified command vehicle speed.
An operation generated by reinforcement learning of a machine learning device so as to infer an operation of the vehicle that causes the vehicle to travel according to the commanded vehicle speed based on the traveling state of the vehicle including the vehicle speed and the commanded vehicle speed. The operation is inferred in the first cycle by the inference learning model, and the operation is inferred in the first cycle.
An adjustment coefficient inference learning model generated by reinforcement learning of a machine learning device so as to infer an adjustment coefficient that adjusts the inferred operation during the first period based on the running state. Infer the adjustment coefficient and
A control method for an autopilot robot, which adjusts the operation according to the adjustment coefficient to generate an adjusted operation during the first cycle, and controls the autopilot robot based on the adjusted operation.

Images