Control parameter automatic adjustment system and autonomous travel assistance device

Abstract

The purpose of the present invention is to promptly realize operation starting of an autonomous travel system by efficiently performing parameter adjustment related to travel control of a moving body. To this end, a parameter updating unit (306) learns a relation between a control parameter (weighting matrix Q, R) and one evaluation index selected from among a plurality of evaluation indices (c1, c2, c3) while the moving body (vehicle 101) travels repeatedly over predetermined routes (global routes 401, 402), and also learns a relation between the control parameter and other evaluation indices other than the one evaluation index out of the plurality of evaluation indices (c1, c2, c3).

Description

Control parameter automatic adjustment system and autonomous driving support device

[0001] The present invention relates to an automatic adjustment function for control parameters, and an autonomous driving control technology utilizing the same, for realizing safe and efficient autonomous driving of moving bodies such as vehicles.

[0002] With labor shortages caused by a declining birthrate and aging population, and the expansion of the e-commerce market, labor reduction and improved work efficiency in logistics warehouses and factories have become important issues. To address these issues, the introduction of autonomous vehicles (hereinafter referred to as vehicles) that can operate without drivers and autonomous driving systems that efficiently operate multiple vehicles is progressing. For safety reasons, conventional autonomous driving systems have limited the range of movement of vehicles, and the movement and working ranges of people (such as field workers) and vehicles have generally been different. However, with the development of autonomous driving technology in the field of robotics, even when the movement paths of vehicles and people overlap, vehicles can determine the need for avoidance or detours and flexibly choose actions according to the surrounding conditions.

[0003] Autonomous vehicle driving control involves various requirements, such as the time to reach a destination, energy consumption, and tolerance for approaching obstacles. Therefore, when developing a control function, it is necessary to adjust control parameters based on these multiple requirements. Traditionally, adjustment of control parameters has been largely determined manually through trial and error, requiring a significant amount of work. On the other hand, recent advances in machine learning technology have enabled the development of technologies that automatically update and adjust control parameters based on update rules that maximize or minimize evaluation indices, such as rewards corresponding to driving results.

[0004] As a specific example, Patent Document 1 proposes a method for preparing an evaluation index that scores fuel efficiency and exhaust gas performance based on changes in the state quantity (speed, etc.) of a vehicle (automobile), and for determining control commands that improve the value of the evaluation index, as well as a vehicle control device. By adopting the configuration of Patent Document 1, it is possible to automatically and appropriately adjust the vehicle's control parameters.

[0005] JP 2020-56737 A

[0006] Here, we will take a deeper look at the requirements for autonomous vehicle driving. Taking the example of autonomous transport robots, the transport robot is required to follow a target trajectory and stop at the target position with as little error as possible. However, even in the same site environment, there may be multiple requirements for tracking accuracy relative to the target trajectory and stopping accuracy relative to the target position, depending on the driving area and driving purpose. For example, let's focus on the stopping operation of a vehicle and consider the following two cases.

[0007] (A) As a result of traffic control, the vehicle stops in the intersection area to give priority to other vehicles.

[0008] (B) The vehicle stops when loading or receiving a load onto the conveyor. Generally, when it comes to vehicle stopping control, there is a trade-off between the time to stop, traveling speed, and accuracy. In the case of (A), the vehicle only needs to avoid interfering with other vehicles crossing in front of it, and quick stopping is required over stopping accuracy. More specifically, even if the vehicle stops slightly tilted from the target stopping direction, it will not affect other vehicles crossing in front. On the other hand, in the case of (B), if the vehicle does not stop accurately at the specified position and direction, the load transported from the conveyor will not be received at the expected position, which could lead to the load falling, so stopping accuracy is prioritized over quick stopping.

[0009] As such, the prioritized requirements for autonomous vehicle driving vary depending on the target operation and driving environment. In this invention, the prioritized requirements are referred to as "driving strategies." Control parameter automatic adjustment technologies, including the technology of Patent Document 1, prepare evaluation indices for each driving strategy and perform parameter adjustment. More specifically, for the two previous examples, an evaluation indices for a driving strategy prioritizing stopping time and an evaluation indices for a driving strategy prioritizing stopping accuracy are prepared, and a predetermined number of runs are performed for both cases. While the previous example only illustrates two elements, stopping time and accuracy, there are various types of vehicle state variables and upper and lower limits for required performance. Therefore, depending on the number of driving strategies to be realized, there may be a large number of evaluation indices, which increases the number of runs required for parameter adjustment. Therefore, the configuration of Patent Document 1 is considered to have room for improvement in terms of the efficiency of parameter adjustment and the labor (time) required for adjustment.

[0010] The present invention has been made to solve the above problems, and aims to quickly realize the start of operation of an autonomous driving system by efficiently adjusting parameters related to the driving control of a moving object.

[0011] In order to achieve the above object, the present invention provides an automatic control parameter adjustment system for adjusting control parameters of a moving body, comprising: a sensor for detecting a state quantity of the moving body; a target state setting unit for setting a target state quantity of the moving body; a control command generation unit for generating a control command for the moving body using the control parameters so as to reduce the deviation of the state quantity from the target state quantity; an evaluation index design unit for designing a plurality of evaluation indexes for calculating an evaluation value for the driving result of the moving body; and a parameter update unit for learning the relationship between the plurality of evaluation indexes and the control parameters and updating the control parameters so as to improve the evaluation value calculated using the plurality of evaluation indexes, wherein the parameter update unit learns the relationship between the control parameter and one evaluation index selected from the plurality of evaluation indexes while the moving body repeatedly travels a predetermined route, and also learns the relationship between the control parameter and evaluation indexes other than the one evaluation index among the plurality of evaluation indexes.

[0012] According to the present invention, it is possible to quickly start operation of an autonomous driving system by efficiently adjusting parameters related to the driving control of a moving object.

[0013] A conceptual diagram showing a driving environment in a warehouse or factory, which is an example of a driving environment. A conceptual diagram showing a vehicle. A simplified functional block diagram of a vehicle and an on-board controller. An example of the functional configuration of a control parameter automatic tuning system. A conceptual diagram showing a global path for the driving environment in Figure 1. A conceptual diagram showing the relationship between global paths and local paths toward achieving autonomous vehicle driving. A function expressing the relationship between control parameters and evaluation indices using Bayesian optimization. A conceptual diagram showing how functions for multiple driving strategies are updated in parameter updating using Bayesian optimization, which is a feature of the present invention. A flowchart showing the process of adjusting weight parameters for model predictive control using a control parameter automatic tuning system. An example of a circular path combining global paths used in control parameter tuning. A conceptual diagram showing how the circular path in Figure 9 is changed depending on the progress of parameter tuning.

[0014] An embodiment of the present invention will be described in detail below with reference to Figures 1 to 10. Note that in this embodiment, an explanation will be given of inter-process transport within a warehouse or factory as an example.

[0015] <Traveling Environment 100> FIG. 1 illustrates a travelling environment 100 in a warehouse or factory assumed in this embodiment. As an example, a trolley-type transport vehicle 101 (hereinafter referred to as vehicle 101) is used for transporting work. Alternatively, a tow vehicle, a forklift, or the like may be used. The travelling environment is an environment with a series of narrow paths surrounded by shelves 102 and the like. In addition, there may be loads 103 on the shelves 102, pallets 104 temporarily placed on the route, and loads on the pallets 104. In addition, an environment is assumed in which a vehicle charging unit 105, a conveyor 106 transporting loads from other areas, and the like are also present.

[0016] <Vehicle 101> A conceptual diagram of the vehicle 101 used in this embodiment is shown in Figure 2A. In this embodiment, a differential two-wheel vehicle is used as an example, but as mentioned above, the present invention is not limited to this and can handle mobile objects with various drive modes, such as omni-wheel types, forklifts, and four-wheel vehicles. The equation of motion for a differential two-wheel vehicle can be given by equation 1.

[0017]

[0018] Here, x is the x coordinate of the vehicle 101, y is the y coordinate of the vehicle 101, θ is the heading (orientation) of the vehicle 101, v is the traveling speed of the vehicle 101, and ω is the angular velocity of the vehicle 101. Note that in the control design of the vehicle 101, Equation 2 is used, which is obtained by discretizing Equation 1 with a sampling period Δt. Note that k in Equation 2 represents a processing step (time).

[0019]

[0020] <Functional Blocks of Vehicle 101 and On-board Controller 201> In order to explain the control operation of the vehicle 101, a simplified functional block diagram of the vehicle 101 and on-board controller 201 is shown in Figure 2B. The vehicle 101 is equipped with an on-board controller 201 that performs control calculations. The vehicle 101 also includes a drive unit 207 and an encoder 208. The on-board controller 201 includes a recording unit 202, on-board sensors 203, a driving map management unit 204, a self-position calculation unit 205, and a control command generation unit 206.

[0021] The recording unit 202 is a recording device for recording the results of arithmetic processing provided in the in-vehicle controller 201, and corresponds to an HDD (hard disk drive), an SSD (solid state drive), or a memory. The recording unit 202 records various control programs, parameters used in the control programs, and the like.

[0022] The on-board sensor 203 corresponds to an external sensor mounted on the vehicle 101, and in the present embodiment, as an example, a LiDAR (Light Detection and Ranging) sensor is mounted. The LiDAR is a sensor that measures the distance to an object present within the irradiation range while changing the irradiation angle of a laser beam.

[0023] The driving map management unit 204 manages a driving map (grid map) that shows obstacle information in the driving environment 100. The map is created in advance before the start of operation using SLAM (Simultaneous Localization and Mapping) with input of measurement data during driving acquired from the on-board sensor 203 (LiDAR) and vehicle travel distance information acquired from the encoder 208 (described later). Creating a grid map using SLAM is a well-known technique, so a description thereof will be omitted.

[0024] The self-position calculation unit 205 combines the measurement results obtained from the on-board sensor 203 and the encoder 208 to calculate the self-position (x, y coordinates on a two-dimensional plane) and orientation of the vehicle itself. Specifically, the self-position and orientation of the vehicle 101 are acquired using map matching technology. Map matching is a technology that takes the measurement information (point cloud set) acquired by the on-board sensor 203 and the driving map as input, and estimates the vehicle position on the driving map by utilizing the similarity between the point cloud set and the driving map. For example, there is an Adaptive Monte Carlo Localization method that utilizes a particle filter. Since the self-position estimation technology using map matching is also a publicly known technology, a detailed description will be omitted.

[0025] The control command generating unit 206 performs calculations related to the driving control of the vehicle 101. Details of the calculations will be described later.

[0026] The drive unit 207 converts the control command input from the control command generation unit 206 into a predetermined current value in order to transmit power to the drive wheels of the vehicle 101. Furthermore, an encoder 208 provided on the drive wheel acquires the amount of rotation of the drive wheel and estimates the travel distance and current speed of the vehicle. The estimated value is transmitted as a response value from the drive unit 207 to the control command generation unit 206, the self-position calculation unit 205, and the recording unit 202. The above is an overview of the functions provided in the vehicle 101 and the on-board controller 201.

[0027] In this embodiment, a method and system are proposed for efficiently adjusting control parameters set in a control command generator 206 to satisfy target performance of each driving strategy set for each driving area and target operation in a vehicle 101 operating in a driving environment 100 shown in Fig. 1. An overview of the control parameter automatic adjustment system will be described below.

[0028] <Control Parameter Automatic Adjustment System 300> FIG. 3 shows an example of the functional block configuration of the control parameter automatic adjustment system 300. The control parameter automatic adjustment system 300 in this embodiment includes an on-board controller 201 and a vehicle 101. Functional blocks not shown in FIG. 2B include a target position setting unit 301, a global route planning unit 302, a local route planning unit 303, an evaluation index design unit 304, a main evaluation index setting unit 305, a parameter updating unit 306, and a driving result management unit 307. In FIG. 3, arrowed lines indicate data flow. Solid lines indicate data transmitted at each control cycle, and dotted lines indicate data transmitted at regular intervals, such as before system operation or after the vehicle 101 reaches a predetermined position. Below, an overview of each newly added functional block in FIG. 3 will be described.

[0029] <Target position setting unit 301> The target position setting unit 301 sets the task currently being performed by the vehicle 101 or a driving target position for task execution (coordinate values ​​on the driving map) based on the driving map managed by the driving map management unit 204. Here, as an example, the target position may be input directly by an operator via an external device (such as a mobile display) that displays the vehicle status or the driving map, or, although not shown in Fig. 3, the target position may be input automatically calculated from the operating status of a task set in advance on the map distribution server side.

[0030] <Global Route Planning Unit 302> The global route planning unit 302 receives as input the driving map managed by the driving map management unit 204, the vehicle's current position calculated by the self-position calculation unit 205, and the target position set by the target position setting unit 301, and plans a global route from the current position to the target position. The global route is generated based on the driving map, which is expressed as a grid. There are many known methods for planning global routes, such as the Dijkstra algorithm and the A* (A star) algorithm, and therefore their explanations will be omitted here. The generated global route is expressed as a set of coordinate values ​​(e.g., row and column numbers) of each grid point on the driving map. Traveling in the order of the coordinate values ​​ensures arrival at the target position. Furthermore, a target driving speed is set for the coordinate values ​​of each grid point constituting the global route. There are various possible methods for setting the target driving speed, such as setting it based on the curvature of each section of the global route, the size of the area to be traveled, or the speed limit set for on-site operations. Furthermore, the global path planning method may be a method that takes into account a vehicle motion model, thereby setting a target traveling speed that takes into account the vehicle's motion characteristics. Routes 401 and 402 shown in FIG. 4 show generated global paths. The global path 401 travels linearly through an area with few surrounding obstacles, so the target traveling speed may be set higher. On the other hand, the path 402 travels through a narrow road surrounded by shelves, with turns, so the target traveling speed may be set lower than that of the global path 401. The target traveling speed may be a constant value for the entire path, or may be set individually for each grid (each coordinate value) that makes up the global path. The target traveling speed set here is one factor when considering a traveling strategy, which will be described later.

[0031] <Local path planning unit 303> The local path planning unit 303 receives the global path calculated by the global path planning unit 302 and the self-position as input, sets the grid coordinates of the global path that exists, for example, a distance D [m] ahead from the self-position as a local target position, and generates a local path to the local target position. The local path is not a coordinate value that indicates the global path, but may be a set of finer coordinate values ​​or a function that expresses a curve, and is a smoother path than the global path.

[0032] In this embodiment, a curve obtained by approximating the global path with a cubic function is used. The local path 501 shown in FIG. 5 is obtained by approximating the coordinate group of the global path 401 with a cubic function. The points on the global path 401 indicate the center positions of the grids, and their coordinate values ​​(x ｒ , y ｒ ) with the target running speed v ｒ The local path 501 can be expressed as in Equation 3, with the coordinate value set of the global path 401 as input. ｓ , y ｓ are coordinate values ​​on the global route converted into a vehicle coordinate system with the self-position of the vehicle 101 as the origin.

[0033]

[0034] <Control command generation unit 206> The control command generation unit 206 receives the local route planned by the local route planning unit 303 and the vehicle's own position as input, and generates a control command for the drive unit 207 equipped in the vehicle 101 according to a route tracking algorithm so as to pass along the local route as much as possible.

[0035] In this embodiment, the path tracking algorithm uses a framework of MPC (Model Predictive Control). MPC minimizes a preset evaluation function in a section of the evaluation function (vehicle behavior from the current time to N steps ahead) for each control period, and determines a control command to be input to the vehicle 101.

[0036] The evaluation function is the vehicle position at each time and the lateral deviation c ｅ , azimuth deviation θ ｅ , deviation v from target running speed ｅ The deviation vector e ｋ and the deviation vector e ｋ (the first term in the right-hand side Σ of Equation 4), and the control command u ｋ = (v ｋ , ω ｋ ) is constructed using a function (the second term in the right-hand side Σ of Equation 4) that evaluates the magnitude of

[0037]

[0038] In addition, T in Equation 4 is a symbol representing transposition. The matrix Q is a deviation vector e ｋ The weighting matrix (weighting parameter) related to the control command u ｋ and N is a parameter representing the evaluation interval (prediction interval). The weighting matrices Q and R are generally expressed as diagonal matrices as shown in Equation 5.

[0039]

[0040] However, it should be noted that the present invention can also be applied to cases where off-diagonal elements are included in the weighting matrices Q and R. As described above, the control command generator 206 includes a plurality of weighting parameters, and in this embodiment, these weighting parameters are successively updated and optimized in accordance with each driving strategy.

[0041] Deviation vector e ｋ The initial value c in ｅ (0), θ ｅ (0), v ｅ (0) is calculated from the local path and the global path. The lateral deviation is expressed as the intercept term d of the cubic function (Equation 3) expressed as the local path, as shown in Equation 6.

[0042]

[0043] The azimuth deviation is the coordinate (x ｒ , y ｒ ) the current azimuth angle θ based on the tangent direction of the global path 401 at the point 401, and the coordinate (x ｒ , y ｒ ) and the Nr-1 coordinate (x ｒ＋Ｎｒ－１ , y ｒ＋Ｎｒ－１ ) and the azimuth angle θr of the line segment connecting them (Equation 7).

[0044]

[0045] Here, the coordinates on the global route nearest to the current self-position are (x ｒ , y ｒ The speed deviation is expressed as the target running speed v ｒ＋Ｎｒand the current running speed v. In addition, in MPC, it is necessary to handle the change in vehicle behavior in the evaluation section indicated by N, and the deviation vector e ｋ Therefore, using the vehicle motion shown in Equation 2 and the local path shown in Equation 3, the changes in position deviation and heading deviation within the evaluation section N are expressed as in Equation 8 and Equation 9, respectively.

[0046]

[0047]

[0048] In addition, d / ds in Equation 8 represents the first derivative of the cubic function representing the local path. Regarding the change in speed deviation, the running speed v ｋ and the target driving speed v ｒ＋Ｎｒ (See Equation 4).

[0049] In addition, MPC can easily handle constraint conditions. For example, in the case of the vehicle 101 (differential two-wheel vehicle) used in this embodiment, the characteristics of the drive unit 207, including the vehicle 101's traveling speed v and the upper and lower limits of the turning angular velocity ω, can be incorporated as constraint conditions. This makes it possible to generate control commands that the vehicle 101 can actually achieve. MPC that takes such constraint conditions into consideration can be expressed as in Equation 10.

[0050]

[0051] In addition, [s.t.] in Equation 10 is an abbreviation of [subject to], and means that the optimization problem is solved under the conditions after [s.t.]. Also, the subscripts max and min are the control command u ｋ Each element (travel speed v ｋ , angular velocity ω ｋ ) corresponds to the upper and lower limit values. Furthermore, the constraint conditions include the addition of Equation 2, Equation 8, and Equation 9. Note that the MPC in this embodiment is configured to obtain a control command u that minimizes the evaluation function J, but it may also be configured to set a target value for the evaluation function J and obtain a control command u that is below the target value.

[0052] <Parameter Adjustment (Characteristic of the Present Invention)> In MPC, even a slight change in the weight parameters (weight matrices Q, R, etc.) described above can significantly change the actual vehicle behavior, making weight parameter design extremely important. On the other hand, as the number of weight parameters increases, the number of combinations also increases, requiring more man-hours to adjust the values. This embodiment deals with a case in which the weight parameters described above are adjusted more efficiently for the multiple driving strategies described at the beginning. Specifically, this embodiment deals with a case in which, even for the same global path, there are different situations in which the allowable maximum lateral deviation, directional deviation, and deviation of the target driving speed differ from the global path, and efficiently finds optimal weight parameters according to each situation (driving strategy).

[0053] The proposals of the present invention can also be applied to global routes that are planned each time according to the destination during actual operation of an autonomous driving system. However, in this embodiment, for the sake of simplicity, the following description will focus on global routes 401 and 402 as the target global routes.

[0054] <Evaluation index design unit 304> In recent years, a method has been devised for adjusting control parameters such as weight parameters of MPC using Bayesian optimization (hereinafter referred to as BO). BO is an algorithm that searches for parameters that maximize or minimize a specific evaluation index c based on the specific evaluation index c through repeated simulations and experiments.

[0055] In this embodiment, multiple evaluation indexes c are set based on various driving strategies, and like the evaluation function J of MPC, they must be described by some kind of mathematical formula. However, unlike the evaluation function of MPC, it is not necessary to use state quantities or control commands; for example, it is also possible to use the time it takes for the vehicle to reach a target driving position as the evaluation index c. In parameter adjustment of MPC using BO, a vector ρ that brings together the diagonal elements of the weight matrices Q and R that correspond to the adjustment parameters is given by equation 11.

[0056]

[0057] An example of the evaluation index c is the time t ｅｎｄThen, the optimal ρ that minimizes this value is ＊ An optimization problem such as the equation (12) can be given as a problem for finding

[0058]

[0059] In general, this optimization problem cannot be solved directly. In BO, several solution candidates ρ ｉ While checking the value of the evaluation index c actually obtained by giving ｉ Calculate the solution candidate ρ. The black circles in the figure represent the solution candidates ρ ｉ (i = 1...6). Using the solution candidates obtained so far, it is possible to calculate the estimated value of the function shape (solid line in the figure) and its variance (dashed line in the figure). The variance is the value of the solution candidate ρ ｉ The next solution candidate ρ is calculated by taking into account the estimated value of the function shape and the variance. ７ By performing the above process, the solution candidate ρ ｉ It is known that this method allows for extremely efficient parameter search (weight design) compared to the case of brute force verification.

[0060] The evaluation index design unit 304 in this embodiment designs a plurality of evaluation indexes c. In this embodiment, evaluation indexes corresponding to the following three driving strategies are prepared.

[0061] (1) Emphasis on Lateral Deviation An example of an evaluation index corresponding to a driving strategy that emphasizes lateral deviation is shown in Equation 13.

[0062]

[0063] In narrow or complicated areas such as area 112 in Figure 1, if the amount of lateral deviation increases with respect to the target overall path, it will lead to interference with surrounding obstacles. Therefore, the evaluation index c1 shown in equation 13 is set to a smaller maximum allowable value for the amount of lateral deviation compared to other evaluation indexes described later. There are two types of lateral deviations to be evaluated, and the average value e of the lateral deviations obtained in each control cycle is used. ａｖｅ＿ｃ and the maximum lateral deviation e determined from the results up to the end of the run ｍａｘ＿ｃHere, the maximum lateral deviation e ｍａｘ＿ｃ When this occurs, the evaluation index value (hereinafter referred to as the driving score) is designed to be significantly worse (larger). ｔ１ , W ｃ１ indicates the weight parameters of various evaluation items, which are values ​​that are set in advance. Although these weights themselves are also adjustment items, there are fewer of them compared to the number of parameters of the weight matrices Q and R that you ultimately want to adjust, and the conditions that must be achieved (required performance) can be easily expressed as mathematical formulas, making design easier than a comprehensive search for the weight matrices Q and R.

[0064] (2) Emphasis on Heading Deviation An example of an evaluation index corresponding to a driving strategy that emphasizes heading deviation is shown in Equation 14.

[0065]

[0066] When transporting a load, if the azimuth deviation becomes large and the vehicle behaves in a meandering manner relative to the target overall route, this may lead to excessive shaking of the transported load, which may cause the load to collapse or fall. Therefore, the evaluation index c2 shown in Equation 14 is designed to worsen the travel score according to the average azimuth deviation compared to the other evaluation indexes (see the third term of evaluation index c2).

[0067] (3) Emphasis on Travel Time An example of an evaluation index corresponding to a travel strategy that emphasizes travel time is shown in Equation 15.

[0068]

[0069] Even in an environment with few surrounding obstacles, such as area 111 in Figure 1, if there is material handling equipment such as conveyor 106 at the destination, if the material handling equipment does not arrive at the material handling equipment within a specified time, the load will remain on the conveyor 106, reducing the operational efficiency of the entire warehouse. Therefore, using the same design guidelines as in (1), an upper limit is set for the difference from the target travel time, and if this upper limit is exceeded (arrival is late), the travel score is designed to be significantly worse. However, lateral deviation from the target route is also evaluated, and parameters that achieve both path tracking accuracy and travel speed are determined.

[0070] The feature of the present invention is that, as described above, multiple evaluation indices are set, and when updating parameters by BO for any evaluation indices, the relationship between the input weight parameters and driving scores for other evaluation indices is simultaneously calculated, and the function shapes of the evaluation indices for other driving strategies are also updated. Details will be explained below.

[0071] <Primary Evaluation Indicator Setting Unit 305> The primary evaluation indicator setting unit 305 sets a primary evaluation indicator to be used for actual parameter update from among the multiple evaluation indicators designed by the evaluation index design unit 304. The primary evaluation indicator selection policy may be random, or the on-site operator may intentionally select the primary evaluation indicator. For example, a policy may be considered in which an evaluation indicator that is difficult to obtain a good driving score is preferentially selected as the primary evaluation indicator. This is because the more difficult an evaluation indicator is to obtain a good driving score, the more attempts it generally requires. Therefore, the mechanism of the parameter update unit 306, which will be described in detail later, increases the possibility of obtaining a good driving score for other evaluation indicators. As an example, among the various evaluation indicators shown in Equations 13, 14, and 15, the evaluation indicator c1 shown in Equation 13 is designed to lead to a deterioration in the score even with a slight lateral deviation compared to the other indicators. Therefore, it is determined to be the most difficult indicator to adjust, and in this embodiment, evaluation indicator c1 is set as the primary evaluation indicator.

[0072] <Parameter Update Unit 306> The parameter update unit 306 receives inputs of the changes in state quantities associated with the current drive recorded in the recording unit 202, the function shape of the BO, weight parameters implemented in the past, and the driving score recorded in the driving result management unit 307 (described later), and obtains weight parameters to be input for the next drive according to the BO update rules. Note that the BO update rules are also applied to evaluation indexes other than the primary evaluation index (hereinafter, "other evaluation indexes"). Figure 7 is a schematic diagram illustrating this situation. This situation is the focus of the present invention, and the function shape of the evaluation index c corresponding to each driving strategy is updated by inputting the relationship between the weight parameters and the driving score in addition to the primary evaluation index shown in Figure 7. In a conventional BO, the next parameters are generally proposed according to the estimated function shape. However, in the present invention, the weight parameters updated according to the rules for the primary evaluation index are simply applied to updating the function shapes of the other evaluation indexes. As a result, the other evaluation indexes are different from the process of obtaining optimal parameters (next parameters) estimated from the current function shape. Specifically, in the situation of FIG. 7, if the original BO is used, the other driving strategy (evaluation index c2) is ρ' ７ is output as the next parameter candidate, but here, in accordance with the function shape of the primary evaluation index (evaluation index c1), ρ ７ is set as the next parameter. Therefore, although the other driving strategies will not be updated to lead to efficient estimation of the function shape, when the other evaluation index itself becomes the primary evaluation index, the relationship between the weight parameters and the driving score has already been estimated in advance, so there is a high possibility that parameters that lead to a deterioration in the score will not need to be updated. By using the above mechanism, the number of trials (number of runs) of other driving strategies can be reduced, and the labor required for parameter adjustment can be reduced.

[0073] The weight parameters newly proposed for the primary evaluation index based on the BO update rules are used the next time the vehicle travels along a similar global route, and are transmitted to the control command generator 206 .

[0074] The above is an explanation of each functional block constituting the control parameter automatic adjustment system 300 proposed in the present invention. Hereinafter, the actual flow of parameter automatic adjustment will be explained according to the flowchart shown in FIG.

[0075] <Driving Environment During Automatic Parameter Adjustment> In explaining the flow of the automatic parameter adjustment process, this embodiment focuses on the global routes 401 and 402 in Figure 4. The accuracy of the self-position calculation unit 205 varies slightly depending on the shape of surrounding obstacles, and global routes are also generated for each destination during field operation. Therefore, ideally, it would be desirable to adjust the weighting parameters for each global route generated during field operation, as shown in Figures 1 and 4. However, particularly in the initial stages of weighting parameter adjustment, parameters that lead to excessive meandering or lateral deviation may be attempted. This could lead to interference with surrounding obstacles when traveling through narrow spaces between shelves 102 within the area 112 in Figure 1. Therefore, as an example, a mechanism could be considered in which an on-board sensor (LiDAR) mounted on the vehicle 101 measures the distance to surrounding obstacles, and the control command generation unit 206 generates control commands to avoid or decelerate as necessary. On the other hand, in such cases, the design of evaluation indices for avoidance and deceleration operations must also be considered. Therefore, in order to focus on the main points of the present invention, in this embodiment, a separate travel area 901 is prepared in which a circular route 902 can be formed by connecting the global routes 401 and 402, as shown in Figure 9. The dotted line portion of the circular route 902 indicates a buffer route 903, and since the results while traveling on the buffer route 903 are not used for evaluation and low speeds are acceptable, travel that follows the center of the route as much as possible is achieved. Note that the black diamonds and black circles connecting the routes in the figure indicate the travel start position and target value position (end position), respectively.

[0076] <Flow of Parameter Automatic Adjustment Processing> When the control parameter automatic adjustment system 300 starts processing, it acquires the self-position of the vehicle 101 (process 801 in FIG. 8 ). Here, the processing described in the self-position calculation unit 205 is executed.

[0077] Next, a travel target position of the vehicle 101 is set (process 802). Here, the process described in the target position setting unit 301 is executed. In this embodiment, since the circular route 902 (including the general routes 401 and 402) shown in FIG. 9 has already been prepared, for example, the end position of the general route 401 is set as the target position.

[0078] Next, a global route to the set travel target position is generated (process 803). The content of this process is as described in the processing overview of the global route planning unit 302. However, in this embodiment, since the global route 401 has already been prepared, a set of coordinate values ​​of the global route 401 is set.

[0079] Next, the updated parameters are acquired (process 804). In this embodiment, the driving result management unit 307 manages the driving results (function shapes expressing the relationship between the weight parameters and the driving score) of each of the global routes 401 and 402. Here, the weight parameters updated based on the driving results of the global route 401 are transmitted to the control command generation unit 206 via the parameter update unit 306 (parameter update by BO). Note that, when driving for the first time, if there are no past driving results, the parameter update unit 306 may use parameters randomly selected from the search space of various weight parameters set by the evaluation index design unit 304.

[0080] Next, evaluation indexes are set (process 805). Here, various evaluation indexes (Equations 13, 14, and 15) designed by the evaluation index design unit 304 and the main evaluation index (Equation 13) set by the main evaluation index setting unit 305 are set.

[0081] After this, actual autonomous driving begins. First, a local route is generated (process 806). Here, the process described in the local route planning unit 303 is executed. Next, the vehicle's own position is acquired (process 807). The own position is acquired here in order to acquire changes in the vehicle's own position while the vehicle is traveling and to calculate control commands.

[0082] Next, a path tracking control process is performed (process 808). Here, the process described in the control command generation unit 206 is executed. Note that the various weight parameters of the MPC used in this process use the values ​​acquired in process 804.

[0083] Next, various state quantities during vehicle travel are recorded (process 809). Here, values ​​necessary for calculating various evaluation indices, such as lateral deviation from the local route, azimuth deviation, deviation from the target travel speed, and elapsed time from the start of travel, are acquired. The acquired values ​​are recorded in the recording unit 202.

[0084] Next, it is determined whether the vehicle has reached the travel target position (process 810). Here, for example, if the self-position acquired in process 807 is within a radius of 0.1 m from the coordinate values ​​indicating the travel target position set in process 802, a stop command (a speed command of 0 m / s) is sent to the vehicle, travel is terminated, and the process proceeds to process 811. On the other hand, if the travel target position has not been reached, the process returns to process 806, and processes 806, 807, 808, and 809 are repeatedly executed until the target position is reached.

[0085] In the next process 811, the values ​​recorded in process 809 are input and the evaluation values ​​of the multiple evaluation indexes are obtained. After that, the parameters to be tried in the next run are obtained only for the primary evaluation index in accordance with the BO update rules. Note that for indexes other than the primary evaluation index, only the function shapes are updated in accordance with the BO rules.

[0086] Next, it is confirmed whether the weight parameter adjustment is complete (process 812). Specifically, it is determined whether the weight parameter adjustment of the primary evaluation index for the global route 401 is complete. Several determination conditions are possible. For example, the evaluation index design unit 304 may set a target score for each evaluation index, and if a score equal to or less than (or in some cases equal to or greater than) the target score is obtained in this trial, it may be determined that the adjustment is complete. Alternatively, the average and variance of the driving scores for the past several runs managed by the driving result management unit 307 may be obtained, and it may be determined that the adjustment is complete if the average is equal to or less than the target score and the variance is relatively small. Alternatively, the evaluation index design unit 304 may simply set a maximum number of trials, and if the number of trials is reached, it may be determined that the adjustment is complete. If it is determined that the adjustment of the global route 401 is complete, the process proceeds to process 813.

[0087] On the other hand, if the adjustment is not completed, the process returns to process 801. In the present embodiment, automatic parameter adjustment is performed along the circular route 902 in FIG. 9 , so the vehicle 101 should currently be at the end point of the global route 401. After this, the vehicle 101 will enter the buffer route 903, so the parameter adjustment is temporarily stopped, and the autonomous driving process except for process 809 is performed to move to the driving start position of the global route 402. After arriving at the driving start position of the global route 402, processes 801 to 812 are performed again.

[0088] In process 813, it is confirmed whether parameter adjustment for all driving strategies has been completed. Specifically, it is confirmed whether an end determination has been made for all evaluation indexes set by the evaluation index design unit 304. If adjustment of all evaluation indexes has been completed, the process proceeds to process 814.

[0089] On the other hand, if adjustment has not been completed for all evaluation indexes, the process proceeds to process 815. In process 815, the primary evaluation index is changed (processing by the primary evaluation index setting unit 305 is performed). If there is only one evaluation index left for which adjustment has not been completed, the relevant evaluation index is changed. If there are multiple evaluation indexes, the change policy is as described in the explanation of the primary evaluation index setting unit 305, and one possible policy is to switch to an evaluation index for which it is difficult to obtain a good score. Furthermore, if the adjustment termination condition for another driving strategy is met by chance, the adjustment of that driving strategy is also determined to be completed.

[0090] Next, in process 814, it is confirmed whether the adjustment of the weight parameters for all global paths has been completed. If it has been completed, all processes of the control parameter automatic adjustment system 300 are completed. If there are global paths for which adjustment has not been completed, the process proceeds to process 816.

[0091] In process 816, the layout of the global route is changed (however, this is optional). Specifically, in the case of the circular route 902 shown in Figure 9, if adjustments have been completed for all evaluation indexes of the global route 402, for example, a circular route 1001 consisting only of the global route 401 may be set as shown in Figure 10. This allows parameter adjustments for the remaining evaluation indexes to be completed in a shorter time. However, since the above-mentioned change in the layout of the global route may not be possible in the actual operating environment shown in Figures 1 and 4 due to the influence of obstacles in the environment, process 816 is not necessarily executable and should be determined depending on the on-site situation.

[0092] The above is the content of the autonomous driving system that utilizes the control parameter automatic adjustment system 300 proposed in the present invention.

[0093] <Other Use Cases> In this embodiment, as an example, multiple driving strategies are established for the path following function and corresponding evaluation indices are designed, but the use case is not limited to this and can be applied to, for example, the stopping control problem described at the beginning of the specification. The BO exemplified in this embodiment is known as a metaheuristic technique that does not depend on a specific calculation problem in algorithms for combinatorial optimization problems. As long as the desired evaluation indices can be expressed as mathematical expressions and there is a mechanism for measuring state quantities from the vehicle that can obtain the values ​​of those evaluation indices, it can be applied to any control problem or algorithm, regardless of the MPC framework exemplified in this embodiment.

[0094] In this embodiment, the evaluation index is designed with a focus on the arrival time to the destination and the lateral deviation from the target route, but the evaluation index is not limited to this and may be designed with a focus on, for example, the energy consumption of the vehicle. In this case, it is possible to add a term indicating the energy consumption, specifically, the integral value of the speed and acceleration, to the evaluation functions shown in Equations 13, 14, and 15.

[0095] Furthermore, the control parameter automatic adjustment system 300 described in this embodiment aims to expedite the start of operation of the autonomous driving system by repeatedly running the vehicle before the autonomous driving system begins operation. However, the application method of the present invention is not limited to this, and may be implemented, for example, during operation of the autonomous driving system. Specifically, in the case of the inter-process transport work described in this embodiment, it is conceivable that the vehicle will transport loads of various weights and shapes. These load changes often affect the vehicle's dynamic characteristics, leading to changes in the driving score of the evaluation index. Therefore, for example, if the approximate weight of the transported load can be measured, the relationship between the weight of the transported load and changes in the driving score of the evaluation index may be learned using BO.

[0096] <Effects of this embodiment> In this embodiment, the problem of setting multiple driving strategies and finding control parameters that satisfy each driving strategy in autonomous driving control of a vehicle is addressed. As an invention proposal, a technology is proposed in which, when learning the relationship between the evaluation index (driving score) of a main driving strategy and the control parameters, the relationship between the evaluation index and control parameters of other driving strategies is also learned based on the update rules for the evaluation index of the main driving strategy.

[0097] As a result, when the evaluation index of the other driving strategy itself becomes the primary evaluation index, the parameters that lead to a deterioration in the evaluation index can be learned in advance, which likely reduces the number of runs (trials) required for parameter adjustment. This enables efficient parameter adjustment related to the driving control of the vehicle (moving body), leading to a faster start of operation of the autonomous driving system.

[0098] In this embodiment, a control parameter automatic adjustment system 300 that adjusts control parameters (weight matrices Q, R) of a moving body (vehicle 101) includes sensors (on-board sensor 203, encoder 208) that detect state quantities of the moving body, a target state setting unit (target position setting unit 301, global path planning unit 302) that sets target state quantities of the moving body, and a deviation e of the state quantities from the target state quantities. ｋ The control command u of the moving body is calculated using the control parameters so that ｋ an evaluation index design unit 304 that designs a plurality of evaluation indexes c1, c2, c3 for calculating an evaluation value (driving score) for the driving result of the mobile body; and a parameter update unit 306 that learns the relationship (function shape) between the plurality of evaluation indexes c1, c2, c3 and the control parameters and updates the control parameters so as to improve the evaluation value calculated using the plurality of evaluation indexes c1, c2, c3, and while the mobile body repeatedly travels a predetermined route (overall routes 401, 402), the parameter update unit 306 learns the relationship between the control parameters and one evaluation index selected from the plurality of evaluation indexes c1, c2, c3, and also learns the relationship between the control parameters and other evaluation indexes other than the one evaluation index selected from the plurality of evaluation indexes c1, c2, c3.

[0099] According to this embodiment configured as described above, parameter adjustment related to the travel control of a moving object can be efficiently performed, thereby enabling the autonomous travel system to be quickly put into operation.

[0100] Furthermore, the control parameter automatic adjustment system 300 in this embodiment includes a recording unit 202 that records the time-series changes in the state quantities and the control parameters. This makes it possible to perform offline parameter adjustment targeting a separately considered driving strategy (evaluation index) using the time-series changes in the state quantities of the vehicle 101 and the control parameters recorded in the recording unit 202.

[0101] In this embodiment, the evaluation index design unit 304 sets a target value (target score) for each of the multiple evaluation indexes c1, c2, and c3, and the parameter update unit 306 finishes updating the control parameters for those evaluation indexes c1, c2, and c3 that have reached the target value. This makes it possible to efficiently adjust the parameters for the multiple evaluation indexes c1, c2, and c3.

[0102] Furthermore, the control parameter automatic adjustment system 300 in this embodiment includes a main evaluation index setting unit 305 that sets, as a main evaluation index, an evaluation index from among the multiple evaluation indexes c1, c2, and c3 that is expected to result in the greatest number of runs of the specified route required to complete the update of the control parameters, based on the function shapes, constraints, and target values ​​of the multiple evaluation indexes c1, c2, and c3, and the parameter update unit 306 prioritizes updating the control parameters targeted at the main evaluation index. This reduces the number of trials (runs) targeted at other driving strategies, thereby making it possible to reduce the man-hours required for parameter adjustment.

[0103] Furthermore, the control command generating unit 206 in this embodiment calculates the deviation e of the state quantity from the target state quantity. ｋ and the control command u ｋ and deviation e ｋ and the control command u ｋ The control command u is set so that the value of the evaluation function J defined by the weighting parameters (weighting matrices Q, R) forｋ The parameter update unit 306 updates the weight parameters as the control parameters. This allows for efficient parameter adjustment related to the travel control of the moving object by the MPC, thereby enabling the autonomous travel system to be quickly put into operation.

[0104] The control parameter automatic adjustment system 300 in this embodiment may be provided in an autonomous driving assistance device that realizes functions such as an advanced driver assistance system (ADAS).

[0105] It should be noted that the present invention is not limited to the above-described embodiment, and includes various modifications. For example, the above-described embodiment is shown to clearly explain the present invention, and the present invention is not necessarily limited to an embodiment having all of the described configurations.

[0106] 100... Driving environment, 101... Vehicle (moving object), 102... Shelf, 103... Load, 104... Pallet, 105... Charging unit, 106... Conveyor, 111, 112... Area, 201... On-board controller, 202... Recording unit, 203... On-board sensor, 204... Driving map management unit, 205... Self-position calculation unit, 206... Control command generation unit, 207... Drive unit, 208... Encoder (sensor), 300... Control parameter automatic adjustment system , 301...target position setting unit (target state setting unit), 302...global route planning unit (target state setting unit), 303...local route planning unit, 304...evaluation index design unit, 305...main evaluation index setting unit, 306...parameter update unit, 307...driving result management unit, 401, 402...global route, 501...local route, 801...processing, 801 to 816...processing, 901...driving area, 902...loop route, 903...buffer route, 1001...loop route.

Claims

1. A control parameter automatic adjustment system for adjusting control parameters of a moving body, comprising: a sensor for detecting a state quantity of the moving body; a target state setting unit for setting a target state quantity of the moving body; a control command generation unit for generating a control command for the moving body using the control parameters so as to reduce a deviation of the state quantity from the target state quantity; an evaluation index design unit for designing a plurality of evaluation indexes for calculating an evaluation value for the traveling result of the moving body; and a parameter updating unit for learning the relationship between the plurality of evaluation indexes and the control parameters, and updating the control parameters so as to improve the evaluation value calculated using the plurality of evaluation indexes, wherein the parameter updating unit learns the relationship between the control parameters and one evaluation index selected from the plurality of evaluation indexes while the moving body repeatedly travels a predetermined route, and learns the relationship between the control parameters and other evaluation indexes among the plurality of evaluation indexes other than the one evaluation index.

2. The control parameter automatic tuning system according to claim 1, further comprising a recording unit for recording the time series changes in the state quantities and the control parameters.

3. A control parameter automatic adjustment system as described in claim 1, characterized in that: the evaluation index design unit sets a target value for each of the plurality of evaluation indexes; and the parameter update unit terminates the update of the control parameters for an evaluation index among the plurality of evaluation indexes that has reached the target value.

4. A control parameter automatic adjustment system as described in claim 1, further comprising a primary evaluation index setting unit which sets, from among the multiple evaluation indexes, an evaluation index which is expected to result in the greatest number of times that the specified route will be driven before the update of the control parameters is completed, as a primary evaluation index, based on the function shapes of the multiple evaluation indexes, each constraint condition, and each target value, and the parameter update unit prioritizes updating of the control parameters targeted at the primary evaluation index.

5. A control parameter automatic adjustment system as described in claim 1, characterized in that the control command generation unit determines the control command so as to improve the value of an evaluation function defined by the deviation of the state quantity from the target state quantity, the control command, and a weighting parameter for the deviation and the control command, and the parameter update unit updates the weighting parameter as the control parameter.

6. An autonomous driving assistance device comprising the control parameter automatic adjustment system according to claim 1.

Images