JP2016024598A - Control method of autonomous mobile apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a control method of an autonomous mobile apparatus having high reliability and capable of improving accuracy.SOLUTION: The control method is provided for an autonomous mobile apparatus 1 including movement means 3 for moving itself, a sensor 2 for acquiring environment information around itself, a control device 4 for moving the itself by controlling the movement means 3 and a storage unit 7 storing map information. In the control device 4, first estimation information indicating a self position estimated by a stochastic method by using a particle filter on the basis of the environment information and the map information is calculated by a first self position estimation section 51, second estimation information indicating a self position estimated by a matching method using line matching on the basis of the environment information and the map information is calculated by a second self position estimation section 52, and integrated information is calculated by an integration section 9 by using an integration function on the basis of the first estimation information and the second estimation information to control the movement means 3 on the basis of the integration information and move itself.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an autonomous mobile device that performs autonomous movement to a destination.

  The conventional autonomous mobile device has two estimation means for estimating the self-location, and is estimated by initializing the estimation result of the other estimation means based on the estimation result of one estimation means. A method of improving the reliability of self-position, that is, the degree of coincidence between the estimated self-position and the actual self-position has been devised (for example, see Patent Document 1).

JP 2003-166824 A

  However, in the method described in Patent Document 1 described above, when there is a large difference in the estimation results of the two estimation means, the reliability is obtained by initializing the other means with the result of one of the more reliable means. However, if the difference between the estimation results of the two estimation means is small, it may not be possible to improve the self-position estimation accuracy using the estimation results of the two estimation means. .

  Therefore, an object of the present invention is to provide a control method for an autonomous mobile device that can improve the estimation accuracy of its own position even when the estimation results of two estimation means are used.

  In order to achieve the above object, a method for controlling an autonomous mobile device according to the present invention includes a moving unit that moves itself, a sensor that acquires environmental information around the device, and a control that moves the self by controlling the moving unit. A method for controlling an autonomous mobile device comprising a device and a storage unit having map information, wherein the control device is self-estimated by a probabilistic method using a particle filter based on the environment information and the map information First estimation information indicating a position is calculated by a first self-position estimation unit, and second estimation information indicating a self-position estimated by a matching method using line matching based on the environment information and the map information is obtained as a second self Calculated by the position estimation unit, calculated by the integration unit using the integration function based on the first estimation information and the second estimation information, and moved based on the integration information And wherein the moving the self-control the stage.

  According to the present invention, it is possible to estimate the position of the self-position with high accuracy even when the estimation results of the two estimation means are used.

FIG. 1 is an overall configuration diagram of an autonomous mobile device according to Embodiment 1 of the present invention. FIG. 2 is a perspective view of the autonomous mobile device according to Embodiment 1 of the present invention. FIG. 3 is an explanatory diagram showing a case where self-location information is updated using the first estimation information in the first embodiment of the present invention. FIG. 4 is an explanatory diagram showing a case where the second estimation information is calculated in the first embodiment of the present invention. FIG. 5 is a flowchart of calculation of self-position estimation information by the integration unit according to Embodiment 1 of the present invention. FIG. 6 is an explanatory diagram of the neural network according to the first embodiment of the present invention. FIG. 7 is a flowchart of learning of the neural network according to the first embodiment of the present invention. FIG. 8 is an explanatory diagram of the steepest descent method according to Embodiment 1 of the present invention. FIG. 9 is an explanatory diagram of learning data creation using the simulator according to Embodiment 1 of the present invention. FIG. 10 is an explanatory diagram relating to a method for improving the accuracy of a particle filter according to Embodiment 4 of the present invention. FIG. 11 is an explanatory diagram of the difference between the estimated position of the particle filter and the line segment matching in the fourth embodiment of the present invention. FIG. 12 is an explanatory diagram of cluster formation in the particle filter according to the fifth embodiment of the present invention. FIG. 13 is an explanatory diagram of the trajectory of the autonomous mobile device that travels to the destination while repeating the travel route in the sixth embodiment of the present invention, such as temporarily stopping.

  Hereinafter, an autonomous mobile device according to an embodiment of the present invention will be described with reference to the drawings.

(Embodiment 1)
As shown in FIG. 1, the autonomous mobile device 1 is acquired by a sensor 2 that acquires environmental information including position information of obstacles that exist in its surrounding space, a moving means 3 that moves itself, and a sensor 2. The control apparatus 4 which controls the moving means 3 based on the environmental information (sensor information) which was performed, and the battery 8 which supplies electric power to these are provided.

  As a basic operation of the autonomous mobile device 1 of the present embodiment, the update unit updates the self-position information based on the travel information, and controls based on the self-position information that is sequentially updated, to the destination. Moving.

  The control device 4 includes a storage unit 7, a first self-position estimating unit 51 that calculates first estimation information indicating the self-position by estimation, and a second self-position estimation that calculates second estimation information indicating the self-position by estimation. Unit 52, third self-position estimation unit 53, integration unit 9 that calculates integrated information indicating self-location using a plurality of pieces of estimation information such as first estimation information and second estimation information, and map information A route generation unit 6 that generates a route to the destination, a learning device 10 that learns based on learning data, and an update unit 11 that updates self-location information stored in the storage unit 7 using integrated information or the like. Prepare. The storage unit 7 stores control parameters, map information, self-location information, acquired environment information (sensor information), travel information, and the like. The first self-position estimation unit 51 calculates first estimation information using the environment information and map information from the sensor 2. The third self-position estimation unit 53 that calculates third estimation information indicating the self-position estimates the self-position as necessary when learning the integration unit 9.

  Here, the estimated information indicating the self position is information for the update unit 11 to update the self position information indicating the self position stored in the storage unit 7 by the autonomous mobile device 1. The update of the self-location information is to replace the self-location information stored in the storage unit 7 with new information using information corresponding to the self-location information included in the integrated information.

  For example, as shown in FIG. 2, the sensor 2 acquires information on an obstacle that exists in a measurement plane 2 </ b> P parallel to the traveling plane P of the autonomous mobile device 1.

  As the sensor 2, for example, a laser radar that scans in a plane can be listed. The sensor 2 (laser radar) oscillates the laser beam within a predetermined fixed angle range on a semicircular measurement plane 2P that is a scan plane ahead of the traveling direction, and detects the distance to the obstacle and the distance to the obstacle. It is a sensor that can acquire a direction (angle). For example, the upper limit of the measurable distance of the sensor 2 is 30 m, and the range in which the laser beam is oscillated is an angle range of ± 90 ° to the left and right with respect to the traveling direction of the autonomous mobile device 1. Ranging every time. For example, the sensor 2 scans intermittently under a constant control cycle, and obtains a set of distance data acquired for each scan and a set of scan angles when the distance data is acquired. The information is stored in the storage unit 7 as information (sensor information). The environmental information is information on an obstacle (in the present embodiment, a corridor wall) obtained based on the distance of the obstacle detected by the sensor 2 at every predetermined angle. The environment information may include other information, for example, discrimination information (for example, time information) for distinguishing each scan.

  Note that the sensor configuration and sensor examples described above are examples for realizing the present embodiment, and the present invention is not limited to these, and various sensors can be used.

  The moving means 3 is composed of, for example, a motor driven by the battery 8 and driving wheels. This motor is provided with an encoder for measuring the rotation speed and rotation speed of the motor as the travel information acquisition means 31. The control device 4 can acquire travel information including the travel distance, the travel direction, and the like of the autonomous mobile device 1 at predetermined time intervals by the output of the encoder as the travel information acquisition means 31. The obtained travel information is stored in the storage unit 7. The stored travel information is appropriately referred to in dead reckoning (estimated navigation) or the like.

  The travel information is information including a relative position and a relative movement direction in a coordinate system based on the reference position. The reference position is, for example, a position when the autonomous mobile device 1 starts moving. The relative position indicates a position relative to the reference position.

  The self-location information is sequentially updated by dead reckoning using travel information stored in the storage unit 7. Furthermore, the self-location information is also updated by integrated information calculated by the integration unit 9 described later. Dead reckoning is performed, for example, in a short sampling period of about 50 msec, and the updating unit 11 updates the self-location information at intervals of, for example, 50 msec based on the travel information. Update of the self-location information by the integration information by the integration unit 9 is performed with a long sampling period of, for example, about 1000 msec, and the update unit 11 updates the self-location information at intervals of, for example, 1000 msec based on the integration information. That is, in the update unit 11, the update of the self-location information based on the integrated information is an interval longer than the update interval based on the travel information.

  The first self-position estimating unit 51 uses the environmental information (sensor information) and the map information to calculate first estimation information indicating the self-position based on a probabilistic method. Here, the probabilistic method is a method using, for example, Bayesian theory. In the present embodiment, a particle filter is used as a probabilistic method in the first self-position estimation unit 51. In calculating the first estimation information using the particle filter, the first estimation information is expressed in a state in which sample sets called particles are distributed. The first estimation information is calculated at a predetermined time interval by estimating, updating, resampling, and the like of these particles. That is, the first self-position estimation unit 51 estimates the position of the autonomous mobile device 1 using a particle filter, and calculates first estimation information.

  FIG. 3 is an explanatory diagram showing a case where self-position information is updated using the first estimation information calculated by the first self-position estimation unit in the embodiment of the present invention. The self-location information 101 of the autonomous mobile device 1 before update shown in FIG. 3 is updated based on, for example, travel information, and is indicated by a two-dot chain line in FIG. The environment information in the self-location information 101 before the update is represented as environment information 201 indicated by a two-dot chain line in FIG. On the other hand, the solid straight line shown in FIG. 3 is information indicating the wall of the corridor included in the map information, and is the corresponding map information 301. As described above, there is a possibility that the environment information 201 from the sensor 2 based on the self-location information 101 before the update based on the first estimation information and the corresponding map information 301 are deviated. Therefore, in the present embodiment, a plurality of particles 102 are calculated using the particle filter by the first self-position estimating unit 51, and the estimation result using the particles 102 is used, which is indicated by a solid line in FIG. The first estimation information 111 is calculated to improve the self-position estimation accuracy. For example, the first self-position estimation unit 51 calculates first estimation information including an X coordinate value, a Y coordinate value, a value indicating a direction, a covariance value, and the like using an average value of the plurality of particles 102.

  The first estimation information may be a difference from the self-location information before update.

  The second self-position estimation unit 52 calculates the second estimation information based on a method different from that of the first self-position estimation unit 51 using the environment information and the map information from the sensor 2. The second self-position estimation unit 52 of the present embodiment uses a line segment extracted from the environmental information obtained from the sensor 2 and performs line matching of the extracted line segment with the line segment included in the map information. Then, the second estimation information is calculated. That is, the line matching in the present embodiment is an example of a matching method for comparing and matching a line acquired from environmental information from the sensor 2 and a line included in the map information.

  FIG. 4 is an explanatory diagram showing a case where the second estimation information is calculated by the second self-position estimation unit in the embodiment of the present invention. The self-location information 101 of the autonomous mobile device 1 before update shown in FIG. 4 is updated based on, for example, travel information, and is indicated by a two-dot chain line in FIG. FIG. 4 shows a line segment 401 extracted from the self-position information 101 before update and a corresponding line segment 501 extracted from the map information. The second self-position estimation unit 52 compares the line segment 401 based on the environmental information indicated by the two-dot chain line segment in FIG. 4 with the corresponding line segment 501 extracted from the map information, If the corresponding line segment 501 is within a certain distance and within a certain angle, matching is performed, and second estimation information is calculated. The second estimation information is information including an X coordinate value, a Y coordinate value, a value indicating a direction, and a matching rate. The matching rate is a value indicating how much the line segment 401 and the corresponding line segment 501 match. By using the line matching by the second self-position estimation unit 52, the second estimation information 151 indicated by the solid line in FIG. 4 is calculated, and the self-position estimation accuracy is improved.

  Note that the second estimation information may be a difference from the self-location information before update.

  With reference to FIG. 5, a calculation method of integrated information by the integration unit 9 using the first estimation information and the second estimation information in the embodiment of the present invention will be described. As described above, the calculation of the integrated information by the integration unit 9 is performed with a period longer than the update period of the self-position information based on dead reckoning, and the self-position information is updated with the obtained integrated information.

  The control device 4 performs the following processing for each relatively long sampling period while the autonomous mobile device 1 is moving.

  First, first estimation information is calculated by the first self-position estimation unit 51 (A1). Next, second estimation information is calculated by the second self-position estimation unit 52 (A2). Next, the integration unit 9 uses the first estimated information and the second estimated information as input values and integrates the self-position estimated using the integration function calculated using the learning data learned by the learning device 10. Information is calculated (A3). The update unit 11 updates the self-location information indicating the current self-location of the autonomous mobile device 1 with the integrated information obtained by the integration unit 9.

  In this embodiment, a neural network is used as the integration function. As is well known, the neural network has learning ability and can be expressed by a non-linear function, and is used in many fields such as control, prediction and diagnosis.

  The structure, input value, and output value of the neural network in this embodiment will be described with reference to FIG. Many structures have been devised for the neural network, but in this embodiment, a three-layer neural network in which two layers of neuron elements having a sigmoid function (intermediate layer, output layer) are used is used. As shown in FIG. 6, values included in the first estimation information and the second estimation information are used as input values. The output value is integrated information. The number of units in the intermediate layer is, for example, 10. The number of units in the output layer is the same as the output value, and when the output value is an X coordinate value, a Y coordinate value, and a value indicating the direction of the autonomous mobile device 1, the number is three. The layers are connected by weights, and these weights are obtained by learning of a neural network by the learning device 10 described later. As an input value in the first estimation information, an X coordinate value, a Y coordinate value, a value indicating a direction, and a covariance value included in the first estimation information are used. As an input value in the second estimation information, the second estimation information is used. X coordinate value, Y coordinate value, orientation value, and matching rate are used.

  In addition, in the calculation of the first estimation information by the first self-position estimation unit 51, when the calculation result by the particle filter has not converged, the integration information is not calculated by the integration unit 9, and the second estimation information is obtained. The self-location information can also be updated based on this. This means that after the particle filter distributes the particles, it takes time for the particles to gather (converge) in one place, and during this time, the results from the particle filter may contain large errors. This is to prevent being there. Whether the calculation by the particle filter has converged is determined by whether the variance of the first estimation information is less than a certain value, or the difference between the minimum value and the maximum value of the first estimation information is less than a certain value (for example, the X coordinate It is determined whether or not the difference between the values is within 2 m and the difference between the Y coordinate values is within 2 m.

  Next, a method for learning a neural network by the learning device 10 will be described. In order to calculate highly accurate integrated information for updating the self-location information by the integration unit 9 including the neural network, learning of the neural network is necessary. Learning means a calculation for determining the weight between each layer. As shown in FIG. 6, the weight is a parameter that associates each unit between the layers. For learning, learning data of a combination of an input value and an output value is used. In the present embodiment, a method of once learning a neural network using data acquired by a simulator, and then re-learning the neural network using data acquired by running the autonomous mobile device 1 in an actual environment. Used.

  FIG. 7 shows a flowchart of learning of the neural network. First, data necessary for learning is acquired using a simulator described later, and the neural network learns the data as learning data (B1). Generally, there are no optimal parameters for the initial values of weights, learning coefficients, etc. when learning a neural network, and it is necessary to set trial and error for the model to which the neural network is applied. It is said. Also in this embodiment, in learning using a simulator, initial values of weights, learning coefficients, etc. are set by trial and error, and learning errors (input value and output value errors) fall within a certain range. Until now, learning is performed while changing the learning parameters. Next, the neural network is learned using the data acquired by the operation of the autonomous mobile device 1 (B2). Here, the simulator is used because the time during which an operation test can generally be performed is often short, so the time for actually learning the autonomous mobile device 1 to acquire the learning data and learning This is a countermeasure for shortening the amount of data acquired. Specifically, in the present embodiment, in the neural network learning using the data acquired by the operation of the autonomous mobile device 1 that is a real machine, the weight of the neural network that is the result of learning using the simulator is initially set. As the learning parameter, the same learning value is used as the learning parameter, and the learning result using the simulator is used as the initial value. Thereby, it is possible to shorten the learning time, the learning of the neural network using the data acquired by the actual machine operation can finely adjust the weight according to the actual traveling site, and the autonomous mobile device When 1 is introduced, a highly accurate learning result can be generated in a short time.

  A neural network learning method using learning data will be described. In the hierarchical neural network described above, learning is performed by using the backpropagation method (error back propagation method) using the learning data, and the weighting coefficient determined by adjusting the internal coupling state (weight) is stored in the storage unit 7. To do. The core of the learning algorithm based on the back-propagation method is to minimize the error between the input value and the output value based on the steepest descent method. The problem of searching for the minimum solution by the steepest descent method is that a global minimum solution cannot be obtained and the learning result falls into obtaining a local optimum solution. FIG. 8 is an example showing the relationship between the error between the input value and the output value with respect to the weight. Depending on the setting of the initial value of the weight and the setting of the learning coefficient, the global optimum solution shown in FIG. 8 cannot be obtained, resulting in obtaining the local optimum solution, and a large error between the input value and the output value remains. There is a possibility. Therefore, in this embodiment, it is possible to acquire a global optimum solution instead of a local optimum solution by learning the neural network by trial and error using data obtained by a simulator.

  In the present embodiment, after the neural network is once learned using the data acquired by the simulator, the neural network is learned again using the data acquired by running the autonomous mobile device 1 in an actual environment. In the case where there are few obstacles of complicated shape in the traveling environment of the autonomous mobile device 1, learning by data acquired by the simulator and data acquired by traveling the autonomous mobile device 1 in the actual environment are used. If there is no significant difference between the learning and the learning, it is possible to use only one of the learnings.

  An outline of the operation of the simulator will be described with reference to FIG. The simulator 600 simulates the movement of the autonomous mobile device 1 in a virtual environment on a computer. The operation by the simulator 600 includes an autonomous mobile simulator 601 and a virtual autonomous mobile device 602. The virtual autonomous mobile device 602 does not actually have the moving means 3 and the sensor 2 in the autonomous mobile device 1 and realizes these functions by the autonomous mobile simulator 601 on a computer. The autonomous movement simulator 601 simulates the operations of the sensor 2, the moving unit 3, and the storage unit 7. More specifically, the autonomous mobile simulator 601 uses the same control value as the command value to the moving means 3 acquired from the virtual autonomous mobile device 602 according to the control parameter in the virtual environment. ) To calculate virtual self-position information 611 indicating the position and orientation of the virtual autonomous mobile device 1 in the virtual environment. The autonomous mobile simulator 601 uses the virtual self-position information 611 and the map information 612 to obtain a virtual environment that will be obtained from the sensor 2 acquired when the autonomous mobile device 1 exists at a position corresponding to the virtual self-position information 611. Information 613 is calculated and the result is transmitted to the virtual autonomous mobile device 602. The virtual autonomous mobile device 602 calculates virtual self position information 611 based on the virtual environment information 613 acquired from the autonomous movement simulator 601. The virtual autonomous mobile device 602 moves to the destination in the virtual environment by performing the simulation until it reaches the destination. At this time, the first estimation information acquired by the first self-position estimation unit 51, the second estimation information acquired by the second self-position estimation unit 52, and the virtual autonomous mobile device 602 from the autonomous movement simulator 601 at a constant frequency. The acquired virtual self-location information 611 is stored in the storage unit. The first estimated information and the second estimated information are used as input values for learning data, and the virtual self-position information 611 is used as third estimated information as output values for learning data. This is based on the idea that the virtual self position information 611 corresponds to the self position of the autonomous movement simulator 601 and is the true value of the self position. In the acquisition of the learning data, various map information is set (for example, a virtual obstacle 622 is arranged), or the map information of the autonomous mobile simulator 601 and the virtual autonomous mobile device 602 is shifted. Thus, by setting various situations that can occur in the actual environment, it is possible to acquire learning data that can be used in various environments.

  As described above, by using the simulator 600, it is possible to acquire learning data on a computer without actually running the autonomous mobile device 1.

  Next, a method for acquiring learning data by a moving operation in the real environment of the autonomous mobile device 1 will be described. The first estimation information calculated by the first self-position estimation unit 51 and the second self-position estimation unit 52 calculated by the first self-position estimation unit 51 are stored in a certain frequency until the autonomous mobile device 1 travels in a real environment and arrives at the destination. The two estimation information is used as an input value of learning data. As the output value of the learning data, estimated information having higher reliability than the first estimated information and the second estimated information is used. In the present embodiment, in order to acquire the third estimation information that is more reliable than the first estimation information and the second estimation information, for example, landmarks arranged in the real environment when the driving test is performed. It is used to calculate the third estimation information. The third estimation information is calculated using the third self-position estimation unit 53. As a landmark to be arranged in the actual environment, for example, a reflector can be listed. By arranging the reflecting plate in the real environment, for example, when a laser radar is used as the sensor 2, environmental information indicating the direction in which the reflecting plate exists and the distance to the reflecting plate are accurately obtained and reflected on the reflecting plate. From the intensity of the beam when the received beam is received, it can be detected that the beam is reflected from the reflector, and can be distinguished from other environmental information. When a plurality of reflector positions can be acquired, the third estimation information of the autonomous mobile device 1 can be acquired with high accuracy by the principle of triangulation. Then, the neural network learning is performed using the third estimated information as the output data of the learning data as the highly accurate information indicating the position of itself. After the learning is completed, the landmark is removed, and thereafter, integrated information is calculated using the learned integration unit 9. By doing in this way, since the autonomous mobile device 1 can perform learning based on environmental information acquired by traveling in an actual environment, learning suitable for the actual traveling environment can be performed. .

  It should be noted that it is desirable to install many landmarks at locations where self-location information is to be updated more accurately. The location where the self-location information is to be updated more accurately is, for example, the movement start point, the arrival point (destination) of the autonomous mobile device 1, the intersection where the passage intersects, the vicinity where the door is installed, the autonomous For example, in the vicinity of an elevator where the mobile device 1 may be boarded. For example, installing many reflectors means installing reflectors at intervals of 2 to 3 m. By doing so, the arrival accuracy of the autonomous mobile device 1 can be improved, and the door portion and the like can be reliably passed. Further, every time the sensor 2 acquires reflection from a new landmark, learning by the learning device 10 is performed.

  In the above embodiment, the method using the particle filter as the first self-position estimation unit 51 and the line matching as the second self-position estimation unit 52 has been described. Without limitation, various estimation means such as a Kalman filter and an ICP algorithm (Iterative Closest Point Algorithm) can be used, and three or more of them can be combined.

  In the present embodiment, the method based on the triangulation using the landmark is described as the third estimation information calculation method of the third self-position estimation unit 53. The estimation information calculation method is not limited to this means. For example, feature parts such as edges of obstacles that are permanently arranged are extracted from the environment information, and third estimation information is obtained by triangulation using these feature parts. Various methods such as the calculation of can be adopted.

(Embodiment 2)
In the second embodiment, a plurality of integration functions having different learning results of the integration unit 9 are held, and the integration function used for calculation of integration information by the integration unit 9 can be changed depending on the area where the autonomous mobile device 1 travels. It is a feature. Other than this, since it is the same as that of the first embodiment described above, the description is omitted.

  More specifically, in the present embodiment, map information is divided into a plurality of areas in advance, learning data is prepared for each area, learning of an integration function such as the above-described neural network is performed for each area, An integrated function with different learning results is created.

  In this way, since a plurality of integrated functions that may have different learning results depending on the environment can be set for each area, it is possible to travel based on more accurate self-position information. The setting of the area can be realized more simply by dividing the map information into a lattice shape (for example, a square of 5 m square) and making each lattice one area.

(Embodiment 3)
In the third embodiment, neural network learning using learning data acquired using a simulator and neural network learning using learning data acquired by running the autonomous mobile device 1 in an actual environment that is an actual site. However, the present invention is characterized in that a difference is provided in the limitation of the correction range. Other than this, since it is the same as that of the first embodiment described above, the description is omitted.

  In neural network learning using learning data obtained using a simulator, the initial value of the weight is a randomly generated value, so there can be a large difference between the weight of the learning result and the initial value weight. It is necessary to set a large correction range. On the other hand, in the learning of the neural network using the learning data acquired by running the autonomous mobile device 1 in the real environment, the learning result (weight) of the neural network using the learning data already acquired using the simulator is an initial value. Since learning is started, it is not desirable that the weight value be large. Therefore, it is necessary to set the weight correction range to be small. In this way, by changing the range of correction between the learning of the neural network using the learning data acquired by running the actual site and the learning result of the neural network using the learning data acquired using the simulator, It is possible to prevent the position accuracy from deteriorating.

(Embodiment 4)
In the fourth embodiment, the first estimation information calculated by the first self-position estimation unit 51 using a particle filter and the second estimation information calculated by the second self-position estimation unit 52 using line segment matching are used as a neural network. In the method in which the integration unit 9 integrates and calculates integrated information using a network, the accuracy of the self-location information updated based on the integrated information is increased by increasing the estimation accuracy of the particle filter. Other than this, since it is the same as that of the first embodiment described above, the description is omitted.

  The particle filter includes a process of adding a noise to the obtained set of particles to obtain a new set of particles after the resampling process in the process of estimating, updating, and resampling the particles. Therefore, when the noise is increased, the particle distribution is expanded, and when the noise is decreased, the particle distribution is decreased. Therefore, if the noise is increased, the robustness of the position estimation is increased but the accuracy is decreased. Conversely, if the noise is decreased, the robustness of the position estimation is decreased, but the accuracy is increased.

  The side marked with (a) in FIG. 10 shows how the particles 102 spread when the noise is set large, and the side marked with (b) in FIG. 10 shows how the particles spread when the noise is set small. Is shown. Thus, by reducing the noise, the dispersion value of the particles is reduced.

  The conditions for setting the noise small will be described with reference to FIG. In the present embodiment, when the difference between the first estimation information 701 calculated by the particle filter and the second estimation information 702 calculated by line segment matching is small, the noise of the particle filter is reduced. More specifically, when the second estimation information 702 that is an estimation result of line segment matching is included in the particle spread 703 (the state on the (a) side in FIG. 11), noise is reduced. On the other hand, when the second estimation information 702 that is the estimation result of the line segment matching is not included in the particle spread 703 (the state on the (b) side in FIG. 11), the noise is not changed. With this configuration, the accuracy of the first estimation information using the particle filter is improved, and as a result, the accuracy of the integrated information using the neural network is improved, and the self-location information updated using the integrated information Accuracy can be improved.

(Embodiment 5)
In the fifth embodiment, the first estimation information calculated by the first self-position estimation unit 51 using a particle filter and the second estimation information calculated by the second self-position estimation unit 52 using line segment matching are used as a neural network. In the integrated information calculation method integrated by the integration unit 9 using the network, the first estimation information is calculated and the integration information is calculated by the integration unit 9 even when the particle distribution by the particle filter is not converged. It is a method to do. Other than this, since it is the same as that of the first embodiment described above, the description is omitted.

  As described above, the particle filter requires time until the particles gather in one place (the particle distribution converges). The first estimation information calculated using the particle distribution that has not yet converged may include a large error. On the other hand, it has been found that in the process until the particle distribution converges, the particles gather in several places to form clusters. In the present embodiment, high-precision first estimation information is calculated using clusters even in a state where the particle distribution has not converged. In the initial state, the particle filter generally distributes the particles 102 in the environment (shown on the (a) side in FIG. 12), and then may form a cluster 800 in the calculation process (see FIG. 12). 12 (shown on the (b) side). The cluster 800 is a set of particles 102 in which the particles 102 exist within a certain distance. In this state, a cluster 800 that is close (for example, closest) to the second estimation information 702 is extracted from the plurality of clusters 800, and first estimation information 711 is calculated based on the particles 102 in the extracted cluster 800. Then, the integration information is calculated by the integration unit 9 using the obtained first estimation information 711 and second estimation information 702.

  In this way, even if the distribution of the particles 102 has not yet converged, the integration information can be calculated by the integration unit 9 by using the second estimation information and the first estimation information calculated using the cluster. This makes it possible to reduce the time during which the integration unit 9 cannot use the first estimation information that is the result of the particle filter.

(Embodiment 6)
The sixth embodiment is a method of acquiring learning data by a movement operation in the real environment of the autonomous mobile device 1 in a predetermined time zone. Since other than this is the same as in the first embodiment, the description is omitted. Generally, when the autonomous mobile device 1 is moving normally every day, it is difficult to derive the third estimation information with higher accuracy than the first estimation information and the second estimation information. Therefore, in the present embodiment, a learning data acquisition time (for example, a time zone such as midnight when there is no person) is set, and the autonomous mobile device 1 is moved during the learning data acquisition time to acquire third estimation information. ing.

  In the present embodiment, the third estimation information is calculated by the first self-position estimation unit 51 and the second self-position estimation unit 52. Specifically, in the learning data acquisition time (time when there is no person, etc.), the autonomous mobile device 1 is moved at a low speed or intermittently repeatedly stopped. If the autonomous mobile device 1 moves continuously at a low speed, the autonomous mobile device 1 moves intermittently at a constant time interval (or a constant distance interval based on travel information). While the vehicle is stopped, environmental information is acquired by the sensor 2, based on the environmental information, the first self-position estimation unit 51 calculates first estimation information, and the second self-position estimation unit 52 calculates second estimation information. By doing so, the environment information can be obtained more accurately than when the autonomous mobile device 1 travels at a normal speed (for example, a set allowable maximum speed of 1 m / s), and the calculation time has a margin. The first estimation information and the second estimation information can be calculated, and the third estimation information can be derived based on these. FIG. 13 shows a state where the vehicle travels to the destination while acquiring environmental information at the information acquisition position S set at predetermined intervals on the route. The constant interval is, for example, 2 m or 10 seconds.

  At this time, when the matching rate of line segment matching is high, the second estimation information that is the result of line segment matching may be used as the third estimation information. Further, when the matching rate of line segment matching is low and the dispersion value of the particle filter is small, the first estimation information of the particle filter may be used as the third estimation information.

  Simultaneously with the above implementation, the first self-position estimation unit 51 uses the virtual environment information reflecting the position information of the virtual obstacle in addition to the environment information acquired during the movement of the autonomous mobile device 1. The first estimation information is calculated by the second self-position estimation unit 52, the second estimation information is calculated by the second self-position estimation unit 52, and this data is used as the input value of the learning data. The above-described third estimation information is used as the output value of the learning data. By doing this, in the environment where there is no obstacle such as a person, the first estimation information and the second estimation information can be calculated in a state simulating the situation where there is an obstacle such as a person. Learning data close to the situation can be obtained.

  It is to be noted that a virtual obstacle can use a preset locus of the obstacle position, and can also use a record of the obstacle position locus acquired when the autonomous mobile device 1 normally travels. Also good.

  The autonomous mobile device according to the present invention can be used when it is necessary to travel in an environment where various obstacles exist, such as in a hospital.

DESCRIPTION OF SYMBOLS 1 Autonomous mobile device 2 Sensor 3 Moving means 4 Control apparatus 6 Path | route production | generation part 7 Memory | storage part 8 Battery 9 Integration part 10 Learning device 11 Update part 31 Travel information acquisition means 51 1st self-position estimation part 52 2nd self-position estimation part 53 Third self-position estimation unit

Claims (13)

Control of an autonomous mobile device comprising a moving means for moving the self, a sensor for acquiring environmental information around the self, a control device for controlling the moving means to move the self, and a storage unit having map information A method,
The control device calculates first estimation information indicating a self-position estimated by a probabilistic method using a particle filter based on the environment information and the map information in a first self-position estimation unit, and the environment information and the Second estimation information indicating a self-position estimated by a matching method using line matching based on map information is calculated by a second self-position estimation unit, and an integrated function is calculated based on the first estimation information and the second estimation information. Using the integrated unit to calculate integrated information, and based on the integrated information to control the moving means to move itself,
Control method of autonomous mobile device. As the integration function, using a neural network,
The method for controlling an autonomous mobile device according to claim 1. As the neural network, a three-layer neural network in which two layers of neuron elements having a sigmoid function are stacked is used.
The method for controlling an autonomous mobile device according to claim 2. As an input value of the neural network in the first estimation information, an X coordinate value, a Y coordinate value, a value indicating a direction, and a covariance value of the first estimation information are used, and the input of the neural network in the second estimation information As the value, the X coordinate value, the Y coordinate value, the value indicating the direction, and the matching rate of the second estimation information are used.
The method for controlling an autonomous mobile device according to claim 2 or 3. The integration function is calculated by learning by a learning device using the first estimation information and the second estimation information as input values.
The control method of the autonomous mobile apparatus of any one of Claims 1-4. Learning with a neural network for each of a plurality of areas set in the map information,
The method for controlling an autonomous mobile device according to claim 5. The learning by the learning device is derived by a movement operation of a virtual autonomous mobile device in the simulator environment,
The method for controlling an autonomous mobile device according to claim 5 or 6. The learning by the learning device is performed by executing a moving operation that moves to the destination while operating itself at a lower speed than the actual operation in the real environment.
The method for controlling an autonomous mobile device according to claim 5 or 6. Learning by the learning device is performed using landmarks temporarily arranged in the real environment.
The method for controlling an autonomous mobile device according to claim 5 or 6. The probabilistic method using the particle filter includes a process of adding noise to the obtained set of particles after the resampling process to obtain a new set of particles, and the first estimation information and the If the difference of the second estimation information is within a certain range, reduce the magnitude of the noise,
The control method of the autonomous mobile apparatus of any one of Claims 1-9. The learning device uses only the second estimation information as an input value in a state where the particle distribution calculated using the particle filter has not converged,
The method for controlling an autonomous mobile device according to claim 5 or 6. The first self-position estimation unit is closest to the second estimation information when the particles form a plurality of clusters in a state where the particle distribution calculated using the particle filter has not converged. The first estimation information is calculated using particles in the cluster.
The method for controlling an autonomous mobile device according to claim 5 or 6. The learning device uses the second estimation information as a learning result when the matching rate calculated by the line segment matching is high, and the particle variance calculated using the particle filter is low while the matching rate is low. In this case, the first estimation information is used as a learning result.
The method for controlling an autonomous mobile device according to claim 5 or 6.

Images