JPH09222852A - Device with cognitive map - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a cognitive map serving as the spatial knowledge expressing the environment, capable of expressing the uncertain spatial knowledge, capable of being utilized for the recognition of the environment even if there is some error, and capable of flexibly coping with the change of the environment. SOLUTION: A cognition map memory means 20 stores the definition information of the components of a cognition map and the probability information showing the probability on the components. An environment cognitive means 30 recognizes the environment with the definition information and probability information of the components stored in the cognitive map memory means 20 in view of probability. The probability density function in the cognitive map memory means 20 is learned by the cognitive map learning means 50 according to the actual behavior result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device having a cognitive map which is a spatial knowledge representing an environment, and in particular, by expressing a cognitive map stochastically, a robot or the like even when the environment changes A device having a cognitive map that can be flexibly dealt with.

Humans carry out their daily lives by remembering and utilizing past experiences. These memories are diverse, including the knowledge of the language required for conversation and common sense for living in a social life. Among them, spatial knowledge that expresses the environment is called a cognitive map, and is used not only for daily activities but also for things. Also plays an extremely important role in the recognition and understanding of.

For example, a cleaning robot and a security robot,
Consider a device that performs a given task while moving in the real world. The robot makes a movement plan based on its own map and performs the work given at the destination. It is necessary for maps used for such environmental recognition to be able to obtain reasonable results without being strict and to be able to flexibly respond to changes in the environment.

[0004]

2. Description of the Related Art Conventionally, a map created by a robot has been created by a designer based on an expected action space, and an absolute map has been stored internally. Here, an absolute map is a map such as a commercial topographic map in which the latitude and longitude are accurately marked. In the case of an indoor map, the positions of walls, doors, desks, etc. are accurate. It is described in.

Generally, the map given to the robot at the time of design is not changed, and the action plan is always created based on the same map. The most basic task given to robots is movement. Whether it's a cleaning robot or a security robot, a travel plan is a must. further,
An important part of the movement plan is how to perform the specified movement at the lowest cost.

Conventionally, in order to make a movement plan, a route that can obtain the shortest cost has been obtained by thoroughly checking movable routes based on an absolute map and calculating a cost given at the time of design. However, it is difficult to calculate the shortest cost in an unknown environment in the real world within a realistic time, regardless of whether the layout or the existence of obstacles hardly changes, such as in a very limited narrow space or in a factory. It is impossible.

[0007]

A robot acting in the real world must consider an error in a sensor for recognizing an environment and a movement amount. Since the absolute map originally does not take into account the error, it has a drawback that it is difficult to use when correcting the measurement error and making a plan. Also, it is impossible to express a map based on uncertain information about a place that has been visited only once in the past.

Further, with the conventional absolute map, it is impossible to deal with a layout change or a sudden accident without changing the map given at the time of design. Also, the time required for the robot to make a plan must be sufficiently realistic.

The present invention solves the above problems, can express uncertain spatial knowledge, can be used for environment recognition even if there is some error, and can flexibly respond to environmental changes. It is intended to provide a device having a cognitive map that can be adapted.

[0010]

FIG. 1 is a diagram illustrating the principle of the present invention. In FIG. 1, 1 is a processing device including a CPU and a memory, 10 is a prior knowledge input means for inputting prior knowledge for constructing a cognitive map, 20 is a cognitive map storage means for storing a cognitive map, and 30 is a cognitive map. Environment recognition means for recognizing environment by map, 31 is means for searching shortest cost route by simulated annealing, 32 is means for searching shortest cost route by genetic algorithm, 40 is movement control based on searched movement route Reference numeral 50 denotes an execution control means for performing the above, and 50 denotes a cognitive map learning means for learning the information of the cognitive map.

The present invention relates to a device having a cognitive map which is spatial knowledge representing an environment, and has the following means. The prior knowledge input means 10 inputs the definition information and the probability information (probability density function) regarding the constituent elements stored in the cognitive map storage means 20 as prior knowledge. The cognitive map storage means 20 is a means for storing, for each constituent element of the cognitive map, definition information of the constituent element and probability information indicating the position, the existence, the cost probability or the accuracy of the information regarding the constituent element. . The environment recognition unit 30 is a unit that probabilistically recognizes the environment using the definition information and the probability information of each component stored in the recognition map storage unit 20.

[0012] For example, when a robot or the like uses a cognitive map for searching a moving route, the shortest cost route is searched by simulated simulated annealing or genetic algorithm. Therefore, it has a shortest cost route searching means 31 by simulated annealing or a shortest cost route searching means 32 by a genetic algorithm. When evaluating the route in the search performed by these, the probability information is used to search for the optimum route according to the purpose of movement.

The execution control means 40 is means for controlling the movement of the robot based on the movement route searched by the search means of the environment recognition means 30 when the control target is a robot, for example.

The cognitive map learning means 50 comprises the execution control means 4
0 is a means for learning probability information about the constituent elements stored in the cognitive map storage means 20 based on the information obtained from the experience actually executed. If necessary, the components themselves will be learned.

In the present invention, the spatial information (cognitive map constituent elements) constituting the cognitive map is, for example, the following five elements, and each position, existence (whether or not present), and reliability are expressed probabilistically. . For paths, the cost (distance, time) required to move is probabilistically expressed.

Path: A path between two points A path is a path between two points that can be moved. For example, this includes corridors in offices, roads, railways, bridges, and the like.

Node: Area that can be entered There is an area that can be entered in the space connected to the path. For example, intersections, stations, airports, etc. correspond to this.

Landmark: landmark A landmark is a structure that serves as a landmark for movement. In one example, the corner is a node, but it is also a landmark because it is an important mark for changing the moving direction. Similarly, stations and airports are also landmarks.

Edge--A boundary between different regions An edge is a structure that cannot move beyond it. For example, indoors cannot move over walls and partitions, and outdoors, rivers and cliffs correspond to edges.

District: A region having a certain characteristic A relatively large region in which common features are found. For example, shopping malls, department stores, Kyushu, Shikoku, etc.

The position, the probability of existence, the reliability and the cost of the path are given by prior knowledge by the prior knowledge input means 10. When expressing a cognitive map of an office, since the layout by partition hardly changes, the data measured with high reliability can be used.
In addition, since new employees will join in spring, it is possible to describe the probability from common sense in daily life such as layout changes.

However, there may be situations where prior knowledge is completely absent. For example, if there is an unexpected visitor and you leave your luggage in the corridor and leave, the corridor path will be completely different from your prior knowledge. At this time, even if an action plan is created with prior knowledge, it is impossible to carry out the plan due to obstacles. The cognitive map learning means 50 learns the experience of failure at this time, and changes knowledge about the corridor. This makes it possible next time to make a plan that makes use of the obstacle information.

When the cognitive map of the present invention is used for searching for the shortest cost route, stochastic search can be efficiently performed. That is, by using simulated annealing or a genetic algorithm as the probabilistic search method, it is possible to find a route that follows the shortest cost in a more realistic time than the conventional method.

The probability of the position and the existence of the constituent elements of the cognitive map is a probability density function which takes a value of 0.5 when it exists or does not exist every other day, for example. Further, the reliability is information indicating the credibility of the definition information of the constituent elements including the probability, and when the information is obtained by the actual measurement, the reliability is high (value is 1
, And the information obtained from hearings has a relatively low reliability (value is close to 0). Here, such probability and reliability information is generically referred to as probability information, but only one of them is sufficient for implementing the present invention.

[0025]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below. The prior knowledge input means 10 shown in FIG. 1 inputs a probability density function determined by utilizing common sense and past experience of the environment desired to be represented by a map. Of course, since it is rare to have knowledge about all the constituents in advance, the probability is given only to the known ones, and the probability density function is learned to the unknown ones by the cognitive map learning means 50.

Details regarding the components of the cognitive map are described below. Path (route between two points) A path is represented by the coordinate value of the start position (from where), the coordinate value of the end position (to which point), the time required to pass the path, the distance, and the cost. Each value is represented by a probability density function. For example, with respect to time, P (time) =
If it is expressed as (10 seconds, 0.8), it means that it takes 10 seconds with a probability of 0.8. further,
The reliability is set for each path information,
It is an indicator of how reliable the information is.

Node (Area That Can Be Entered) A node is represented by a name such as a station or an airport, its position, and reliability regarding description. Again, like the path,
The position is expressed by probability.

Similar to the landmark node, the landmark is also represented by the name, its position, and the reliability of the description.

Edge (Boundary of Different Areas) Similar to nodes and landmarks, edges are also represented by names, their positions, and reliability regarding description.

District (area having a certain characteristic) Like the node, the landmark, and the edge, the district is also represented by the name, its position, and the reliability of the description.

FIG. 2 is a diagram for explaining the search for the shortest cost route by simulated annealing. Simulated annealing is a calculation method inspired by annealing in physics. It is a kind of sequential search improvement that repeats state transition from one initial state to generate a state sequence and searches for a solution in it. Is the law.

The shortest cost path problem is a certain constraint condition g _i
The cost f under (x) ≦ 0 (i = 0, 1, ..., M)
It can be formulated as to find the minimum value of (x). Given such a shortest cost route problem, in step S1, an initial solution x is randomly generated and 0 is set to j, which is the number of iterations. afterwards,
Repeat steps S2 to S4.

In step S2, the neighborhood S of the current solution x
A candidate solution y is randomly selected from (x). At this time, 1 is added to j and the temperature parameter c is further reduced. Temperature parameter c is a control variable corresponding to the temperature in physical annealing, in finite iterations, according to j of c _j is increased, the value of c _j is determined so as to decrease.

In step S3, if the cost of the neighborhood solution y selected in step S2 is lower than that of the current solution x, the solution y is set as the current solution x. That is, f (y) ≦ f (x)
Then, x = y. Conversely, even if the selected neighborhood solution y has a higher cost than the current solution x, the current solution x is set at a certain probability p. That is, if f (y)> f (x),
p (x → y) = exp (− (f (x) −f (y)) /
The probability of c) is x = y.

In step S4, it is determined whether or not a predetermined end condition is satisfied, and if the end condition is satisfied, the solution x obtained up to that point is regarded as the final solution and the process is ended.
If the end condition is not satisfied, the process returns to step S2 and the same process is repeated. The termination conditions are that a solution satisfying the given conditions has been obtained, that j has reached a predetermined value, that the cost improvement rate has converged within a predetermined range,
It can be arbitrarily set in advance such that the calculation time reaches a predetermined time.

In the cost calculation in step S2, the probability information about the components of the cognitive map can be used in various methods according to the purpose of route search. For example, if the destination must be reached within a certain time limit at the latest, cost calculation is performed without using information with low probability or reliability. In other words, the cost is calculated by weighting only the information with high probability and reliability. On the other hand, even if it fails in the worst case, if you want to experimentally find the route with the shortest cost in practice, information with low probability and reliability is the same as information with high probability and reliability. To calculate the cost. Furthermore, when an average route that is close to the shortest cost is to be obtained, the cost can be calculated by setting the coefficient of the evaluation function according to the probability and reliability.

FIG. 3 is a diagram for explaining the search for the shortest cost route by the genetic algorithm. As is well known, the genetic algorithm repeatedly applies genetic operations such as crossover, mutation, and selection to a population of solution candidates expressed in the form of chromosomes, and then the environment is used as a result of survival competition. This is a calculation method for finding a sub-optimal solution using the mechanism of biological evolution in which individuals with higher fitness of survive.

First, given the shortest cost path problem, the optimization terms and constraints are coded on the chromosome used in the genetic algorithm. The chromosome here is, for example, a sequence of codes x ₁ , x ₂ , ..., x representing a path or a node passing from a starting point to a destination point on a cognitive map.
Expressed as _n .

In the optimization using the genetic algorithm, the following processing is performed. First, in step S21, a group of chromosomes (solution candidates) is generated. Here, a predetermined number of chromosomes are randomly generated.

In the crossover in step S22, for a plurality of chromosomes stochastically selected from the group of chromosomes, an operation of combining a part of the chromosomes with a part of the other chromosomes is executed. Further, in the mutation in step S23, an operation of replacing a part of the gene of the chromosome stochastically selected from the population of chromosomes with another gene is performed. In the selection in step S24, the fitness (cost) is calculated for each chromosome, and the probabilistic selection of chromosomes according to the fitness is performed. Specifically, for example, by using the roulette method or the like, a process of replicating a chromosome with high fitness (low cost) and deleting a chromosome with low fitness (high cost) is executed. By the above generational change, the group of solution candidates is improved.

In step S25, it is determined whether or not a predetermined end condition is satisfied, and if the end condition is satisfied, the optimum solution is processed as the final solution among the solution candidate groups obtained up to that point. To finish. If the end condition is not satisfied, the process returns to step S22 and the same process is repeated. The termination condition can be arbitrarily selected, such as repeating the genetic operation for a prescribed number of generations, the computation time has reached a prescribed time, or that there are solution candidates that satisfy the prescribed condition in the group of chromosomes. It is possible.

In the calculation of the fitness in step S24, it is possible to use the probability information about the components of the cognitive map in various ways depending on the purpose of the route search, as in the case of the simulated annealing described above. It is the same.

For example, the evaluation function used in the route search by simulated annealing can be generalized by the following equation. Evaluation = α (sum of time) + β (sum of distance) + γ (energy consumption) In this formula, α, β, and γ are coefficients for each term. For example, if both β and γ are zero, Only the sum of time is evaluated, and α = 0.5, β = 0.5,
If γ = 0, the evaluation will be made with equal emphasis on time and distance. Since the determination of these coefficients largely depends on the problem and environment, it is desirable to determine them by experiments or the like according to each case.

The fitness in the genetic algorithm can be defined similarly.

[0045]

FIG. 4 shows an example of an office layout for explaining an embodiment of the present invention, and FIG. 5 shows an example of a cognitive map describing the office layout of FIG.

An office layout as shown in FIG. 4 is handled as the environment targeted for the cognitive map. The coordinate value (x, y) indicating the position here is from (0, 0) to (10) in the areas partitioned by the partitions P1 to P4.
It is expressed in relative coordinates up to 0,100).

The corridors A1 and A2 shown in FIG. 4 are treated as paths on the cognitive map, and are expressed as shown in FIG. 5A, for example. For example, corridor A1 has coordinates (45,15)
To (85,15).
The probability that this corridor A1 exists is 0.9. The existence probability of 0.9 means that if a large baggage is placed on the corridor A1, the corridor A1 cannot be used as a pass because the corridor A1 cannot be used as a pass.
The time required for the robot to pass through the corridor A1 is 10 seconds with a probability of 0.8. The length of the corridor A1 is 4 m with a probability of 0.9. Costs such as energy consumption are
It is 20 with a probability of 0.7. Further, it is given as prior knowledge that the reliability of these pieces of information is 0.8. The definition information of the corridor A2 has the same meaning.

The corner Co is a node, and as shown in FIG. 5B, the position is defined as (90, 90), its probability is 0.9, and its reliability is 0.8. As landmarks, a trash can Ga, a desk De, and a door Do are defined as shown in FIG. 5C. As the edges, partitions P1 to P6 are defined as shown in FIG. The district is an office, and is defined as shown in FIG. As the position coordinates, if there is a range such as a region, the coordinate value of a specific point such as the center of gravity is used. Besides, size information and the like may be held.

Using the cognitive map as shown in FIG. 5, consider the problem of moving from the trash Ga to the door Do. Recycle bin Ga
The evaluation function f (x) in the path search from the door to the door Do is
For simplicity, the time required to move each path is added. In addition to the explicitly defined paths such as the corridors A1 and A2, there are practical paths such as from one landmark to another landmark. For the convenience of explanation, the time required to move each path in this movement problem is given as shown in FIG. For example, the time required to move from the door Do to the desk De is 80 seconds. Also, the constraint conditions are shown in FIG.
As shown in (b), f (x) <150.

FIG. 7 shows an example of searching by simulated annealing. First, an initial solution x is randomly given. In this example, as shown in FIG. 7A, an initial solution such as x = {trash can Ga → corridor A1 → corner Co → corridor A1 → trash can Ga → desk De → door Do} is obtained. From the time shown in FIG. 6, f (x) = 320> 15
It is 0, and this initial solution does not satisfy the constraint condition. Therefore, the number j of repetitions is set to 0, and the process proceeds to the next step.

As shown in FIG. 7B, candidate solutions y near x are randomly generated. In this example, y = {trash can Ga → corridor A1 → corner Co → corridor A2 → door D
o}. f (y) = 200. Set j = 1.

As shown in FIG. 7C, f (y) and f (y)
Compare with (x). As a result of comparison, f (y) <f
Since (x) is an improvement in the solution, the solution is replaced with x = y. However, this solution is f
Since (x) <150 is not satisfied, the process proceeds to the next step.

Next, as shown in FIG. 7D, candidate solutions y near x are randomly generated. In this example, y =
It becomes {trash can Ga → desk De → door Do}. f
(Y) = 120. Set j = 2.

As shown in FIG. 7 (e), f (y) and f
Compare with (x). As a result of comparison, f (y) <f
Since (x) is an improvement in the solution, the solution is replaced with x = y. Examining the constraints, f (x) = 12
Since 0 satisfies the condition of f (x) <150, x = {trash can Ga → desk De → door Do} is set as a route to end the process.

FIG. 8 shows an example of search by the genetic algorithm. First, as shown in FIG. 8A, a group of chromosomes (solution candidates) is generated. Actually, a group of many chromosomes is used, but here, for simplification, three individuals are used. The numbers 1 to 3 are assigned to these three individuals. The initial population of generated chromosomes is as follows.

No1 = {trash can Ga → corridor A1 → corner Co → corridor A1 → trash can Ga → desk De → door Do} No2 = {trash can Ga → desk De → trash can Ga → corridor A1
→ Corner Co → Corridor A2 → Door Do} No3 = {trash can Ga → corridor A1 → corner Co → corridor A2 → corner Co → corridor A2 → door Do} Each evaluation value (reciprocal of fitness) is f (No1) =
320, f (No2) = 280, f (No3) = 220
It is.

Next, a crossover operation, which is one of gene manipulations, is randomly performed. In this example,
As shown in (b), crossover is performed between No2 and No3 chromosomes. The results of the crossover are as follows.

No1 = {trash can Ga → corridor A1 → corner Co → corridor A1 → trash can Ga → desk De → door Do} No2 = {trash can Ga → desk De → trash can Ga → corridor A1
→ Corner Co → Corridor A2 → Corner Co → Corridor A2 →
Door Do} No3 = {trash can Ga → corridor A1 → corner Co → corridor A2 → door Do} Each evaluation value is f (No1) = 320, f (No)
2) = 300 and f (No3) = 200.

Next, a mutation operation, which is one of gene manipulations, is randomly performed. In this example,
As shown in (c), No1 is mutated. Desk No, which is the No1 gene, was mutated in the corridor A2, and the chromosome population is as follows.

No1 = {trash can Ga → corridor A1 → corner Co → corridor A1 → trash can Ga → corridor A2 → door Do} No2 = {trash can Ga → desk De → trash can Ga → corridor A1
→ Corner Co → Corridor A2 → Corner Co → Corridor A2 →
Door Do} No3 = {trash can Ga → corridor A1 → corner Co → corridor A2 → door Do} The respective evaluation values are f (No1) = 410, f (No)
2) = 300 and f (No3) = 200.

In the following selection operation shown in FIG. 8 (d), the roulette method or the like is used to perform probabilistic multiplication and lethality operations according to the evaluation values so that those with small evaluation values survive. In FIG. 8 (d), the No3 chromosome has proliferated and has replaced the disappeared No1 chromosome.

The above operations (b) to (d) are repeated until a solution for a predetermined number of generations or a constraint condition is obtained. For example, as shown in FIG. 8E, it is assumed that the solution of No1 = {trash can Ga → desk De → door Do} is obtained. This evaluation value is f (No1) = 120, and the constraint condition f
Since (No1) <150 is satisfied, this is the solution to be obtained.

FIG. 9 shows a learning example according to the embodiment of the present invention. When the shortest route {trash can Ga → desk De → door Do} is obtained by the search shown in FIG. 7 or 8, the robot is actually moved. It can be seen that when the robot actually moves, there is a discrepancy between the robot knowledge used to obtain the evaluation function and the actual experience. Learning is to reflect the difference in robot knowledge.

For example, assume that the robot has knowledge of the cognitive map as shown in FIG. 5 at some point. Here, when actually moving along the route of {trash can Ga → desk De → door Do}, the accurate positions of the trash can Ga, the desk De, and the door Do can be measured, and the reliability thereof increases. For example, before learning, the knowledge as shown in FIG.
It changes as shown in (b).

[0065]

As described above, according to the present invention, it is possible to express a cognitive map that is closer to a human way of thinking than in the prior art, and it is possible to flexibly respond to changes in the environment. become able to. Furthermore, it is possible to find a route that conforms to the shortest cost in a more realistic time as compared with the conventional technique.

[Brief description of drawings]

FIG. 1 is a diagram illustrating the principle of the present invention.

FIG. 2 is a diagram illustrating searching for a shortest cost route by simulated annealing.

FIG. 3 is a diagram illustrating a search for a shortest cost route by a genetic algorithm.

FIG. 4 is a diagram showing an example of an office layout.

FIG. 5 is a diagram showing an example of a cognitive map describing an office layout.

FIG. 6 is a diagram showing an example of conditions for explaining an embodiment of the present invention.

FIG. 7 is a diagram showing a search example by simulated annealing.

FIG. 8 is a diagram showing a search example by a genetic algorithm.

FIG. 9 is a diagram showing a learning example according to the embodiment of the present invention.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 Processor 10 Prior knowledge input means 20 Cognitive map storage means 30 Environment recognition means 31 Shortest cost route search means by simulated annealing 32 Shortest cost route search means by genetic algorithm 40 Execution control means 50 Cognitive map learning means

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Minoru Sekiguchi 1015 Kamiodanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture, Fujitsu Limited (72) Inventor, Hirohisa Naito 1015, Kamedotachu, Nakahara-ku, Kawasaki, Kanagawa Prefecture, Fujitsu Limited

Claims (6)

[Claims] 1. In a device having a cognitive map that expresses spatial knowledge used by a computer to recognize an environment, for each component of the cognitive map, the definition information of that component and the certainty regarding that component are shown. A cognitive map storage means for storing the probability information; and an environment recognition means for probabilistically recognizing the environment by using the definition information and the probability information of each component stored in the cognitive map storage means. A device with a characteristic cognitive map. 2. The device having the cognitive map according to claim 1, wherein a path indicating a path between two points and a space connected to the path are provided as components of the cognitive map stored in the cognitive map storage means. An apparatus having a cognitive map, which has at least a node indicating an area that can be entered in and an edge indicating a boundary of different areas. 3. The apparatus having a cognitive map according to claim 1 or 2, further comprising an input means for providing probability information regarding components stored in the cognitive map storage means by prior knowledge. A device with a cognitive map. 4. The device having the cognitive map according to claim 1, claim 2 or claim 3, wherein the recognition is performed based on information obtained from matters experienced based on the result of recognition of the environment by the environment recognition means. An apparatus having a cognitive map, comprising cognitive map learning means for learning probability information about components stored in a map storage means. 5. The apparatus having the cognitive map according to claim 1, claim 2, claim 3 or claim 4, wherein the environment cognitive means is the shortest from a certain starting point to an arriving point by simulated annealing. An apparatus having a cognitive map, which is a means for obtaining a cost route or a route satisfying a predetermined constraint condition. 6. A device having a cognitive map according to claim 1, claim 2, claim 3 or claim 4, wherein the environment recognition means uses a genetic algorithm to obtain the shortest cost from a certain starting point to an arriving point. A device having a cognitive map, which is a means for obtaining a route or a route satisfying a predetermined constraint condition.