JP3905413B2 - Robot equipment - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a robot apparatus that generates a gait using a genetic algorithm and walks with a leg.
[0002]
[Prior art]
Conventionally, when generating a gait of a robot having legs, it is necessary to determine many parameters, and many man-hours are required for simulations and experiments.
[0003]
For this reason, for example, in Japanese Patent Laid-Open No. 2001-191284, a parameter for controlling walking of a walking robot is updated by using a genetic algorithm so that the evaluation value of walking is improved, so that a walking pattern is updated. A technique for facilitating the generation work of is disclosed.
[0004]
[Problems to be solved by the invention]
However, since the gait generation of a walking robot is complicated, there are many evaluation conditions that must be satisfied in walking the robot, and the respective conditions may affect each other or may have a trade-off relationship.
[0005]
For this reason, even if an evolutionary learning method using a genetic algorithm is applied as in the prior art, the convergence of evolutionary computation is slow, resulting in a phenomenon where evolution is stagnated due to being trapped in a local solution There is a case.
[0006]
The present invention has been made in view of the above circumstances, and improves convergence of an optimal solution in gait generation using a genetic algorithm to prevent stagnation of evolution, and to obtain a fast and reliable gait pattern. It aims at providing the robot apparatus which can do.
[0007]
[Means for Solving the Problems]
To achieve the above object, the invention according to claim 1 is based on a genetic algorithm, wherein a plurality of evaluation conditions to be satisfied by a walking robot walking with a single leg or a plurality of legs are stepped from lower to higher. And calculating the optimal solution candidates that match the evaluation conditions of each stage as the evolution population, and the individual evaluation value between the lower evolution population and the higher evolution population. In response, the means for exchanging individuals, the means for determining the convergence of the optimization operation based on the top evolutionary population, and the top evolutionary population on which the optimization computation has converged, Means for updating control parameters The upper evaluation condition includes the lower evaluation condition. It is characterized by that.
[0008]
The invention described in claim 2 is characterized in that, in the invention described in claim 1, the calculation of the evolved population is executed by being distributed by a plurality of calculators.
[0009]
According to a third aspect of the present invention, in the first or second aspect of the invention, the evaluation value is calculated based on data from a simulator that models the movement of the walking robot.
[0010]
According to a fourth aspect of the present invention, in the invention according to any one of the first, second, and third aspects, the evaluation value is calculated based on actual machine data from the walking robot.
[0011]
That is, the invention according to claim 1 divides a plurality of evaluation conditions to be satisfied by the walking robot step by step from lower to higher. Include the lower evaluation condition in the upper evaluation condition. The candidate of the optimal solution suitable for the evaluation condition of each stage is calculated while exchanging individuals according to the evaluation value of the individual between the lower-level evolution population and the higher-level evolution population. When the optimization operation of the top evolutionary population is determined to converge, the control parameters of the walking robot are updated based on this top evolutionary population, thereby improving the convergence of the optimal solution and evolving To prevent stagnation.
[0012]
At this time, it is desirable that the computation of the evolved population is distributed and executed by a plurality of computing units, as in the invention described in claim 2, and the solution search can be further speeded up. In addition, the individual evaluation value is obtained from data from a simulator that models the movement of a walking robot as in the invention described in claim 3, or from the walking robot as in the invention described in claim 4. It can be calculated based on the data.
[0013]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention will be described below with reference to the drawings. 1 to 8 relate to an embodiment of the present invention, FIG. 1 is an overall configuration diagram of a robot apparatus, FIG. 2 is a block diagram of an optimization apparatus using a stepwise evolution method, and FIG. 3 is a plurality of arithmetic units. FIG. 4 is a flowchart of immigration processing between GA engines, FIG. 5 is an explanatory diagram of joint angles that give vertical movement to a four-legged walking robot, and FIG. 6 is four-legged walking. FIG. 7 is an explanatory diagram of a neural network that generates a gait pattern for one leg, and FIG. 8 is a flowchart showing the processing of the simulator.
[0014]
The robot apparatus shown in FIG. 1 has a single or a plurality of legs, a walking robot 1 that walks by driving the legs, and a parameter generator that generates and updates walking control parameters of the walking robot 1. 10, a simulator 20 that models the walking robot 1, and N evaluation units 31 in which evaluation conditions are sequentially added from lower to higher in order to realize stepwise evolution according to a plurality of evaluation conditions And a GA engine group 30 having N genetic algorithm calculation units (GA engines) 32 that execute operations in parallel.
[0015]
The GA engine group 30 is configured as an optimization device using a so-called stepwise evolution method that prevents the stagnation of evolution by gradually complicating the task given to the walking robot 1 from the initial stage of evolution. Specifically, as shown in FIG. 2, the GA engine group 30 has (N−1) individual immigration control units 33 and a convergence with respect to the N evaluation units 31 and the corresponding N GA engines 32. And a determination unit 34.
[0016]
The GA engine 32 is provided in a plurality of stages from the lower order to the upper order, and as will be described later, the evaluation conditions set by the evaluation unit 31 for each GA engine 32 are added stepwise from the lower order to the upper order. Note that the evaluation unit 31 and the GA engine 32 at each stage define the lowest level as the first level, the second level, the third level,...
[0017]
The individual immigration control unit 33 is provided between the GA engines 32 adjacent to each other, and the partial populations in the evolutionary population PGn (n = 1 to N) to be optimized are assigned to the nearest adjacent stages. Move between. In the present specification, an individual that is moved between the upper and lower evolved populations PGn is referred to as “immigrants”.
[0018]
The convergence determination unit 34 is attached only to the highest GA engine 32, and performs optimization convergence determination based on the evolved population PGN output from the highest GA engine 32. Here, the “highest level” means a target stage that can include an optimal solution as a final objective, in other words, a stage in which all conditions to be finally given are set. Therefore, the optimum solution finally obtained, that is, a solution that meets all the conditions to be considered is calculated in the highest-level GA engine 32.
[0019]
The evaluation conditions set for the GA engine 32 by the evaluation unit 31 at each stage are added stepwise from the lower level to the upper level. That is, the first-stage evaluation condition (evaluation condition set for the first-stage GA engine 32) is only a single “condition 1”, whereas the second-stage evaluation condition is the first-stage evaluation condition. “Condition 1 + 2” (“+” means “and”) is obtained by adding a new “Condition 2” to “Condition 1”. Further, the third-stage evaluation condition is “condition 1 + 2 + 3” in which a new “condition 3” is added to “condition 1 + 2” which is the immediately lower evaluation condition. That is, “evaluation condition 1 + 2 +... + N” (final evaluation condition), which is a combination of a plurality of conditions, is divided step by step, and conditions are added step by step from lower to higher. Set the evaluation conditions individually.
[0020]
Note that the condition added to each stage is not necessarily a single condition, and a plurality of conditions may be added simultaneously. For example, the first-stage evaluation condition is “condition 1”, and the second-stage evaluation condition is “condition 1 + 2 + 3”.
[0021]
Conditions that are added stepwise to each GA engine 32 are arranged in a lower order as the restraining force is stronger, and placed in a higher order as the restraining force is weaker. In other words, necessary conditions are arranged on the lower side, and sufficiently conditional conditions are arranged on the upper side. Specifically, as described below, conditions to be considered in the walking optimization related to the walking robot 1, such as “stand with legs”, “do not fall”, “walk without making footsteps as much as possible”, etc. The condition of strong restraint of “standing with legs”, which is the main premise of walking, corresponds to “Condition 1” at the bottom, and the condition “Do not fall” corresponds to “Condition 2”, The advanced condition of “walking without making footsteps as much as possible” corresponds to the highest condition “condition N” (the binding force is the weakest).
[0022]
In detail, as conditions required for walking of a robot having a single or a plurality of legs, for example, there are conditions shown in the following (1) to (9), and generally (1) → (9) Strong binding power in order.
(1) Stand with your legs (do not put your body on the ground).
(2) Do not tip over (keep body angle within limits).
(3) Move (not just stepping).
(4) Proceed in the direction you want to go.
(5) Walk faster.
(6) Walk more stably (withstand disturbances).
(7) Less energy is used.
(8) Walk with as little posture as possible.
(9) Walk with as little footsteps as possible.
[0023]
Among these conditions, the condition with the strongest binding force, that is, the condition of “standing on the leg” in (1) is determined by the first evaluation unit 31, and the condition with the strongest binding force and the second binding force are determined. In this case, the evaluation unit 31 in the second stage simultaneously determines the strong condition of (2) and the “not toppling” condition in (2). In this way, by preparing N independent GA engines 32 for the N conditions so as to sequentially add to the evaluation unit 31 from the condition with strong binding force, the N-stage GA engine 32 Evaluation will be performed for all required conditions.
[0024]
When the GA engine group 30 is realized by a single arithmetic device such as a stand-alone single processor type computer, the single computer has N evaluation units 31 and GA engines 32, and (N-1) individual immigrants. It is responsible for all functions of the control unit 33 and the convergence determination unit 34, but in order to increase the processing speed, a plurality of operations such as a multiprocessor computer or a plurality of single processor computers connected by a network via a LAN, etc. It is also possible to implement the GA engine group 30 with the device.
[0025]
In this case, as shown in FIG. 3, each of the N computers (processors) 30-1, 30-2,..., 30-N includes one evaluation unit 31, a GA engine 32, and an individual immigrant corresponding thereto. Information transmission (individual exchange) between computers (processors) that realize the functions of the control unit 33 and perform parallel processing is performed via a network, a bus, a multi-port memory, and other interfaces. Then, the computer (processor) corresponding to the uppermost GA engine 32 has the function of the convergence determination unit 34.
[0026]
Since each GA engine 32 performs a GA operation independently of each other, it can be easily parallelized, and the processing time required for optimization is shortened as the number of computers (processors) used increases. The amount of data required for individual exchange between mutually adjacent GA engines 32 is relatively small, and the frequency of occurrence is not high. Therefore, it is not always necessary to use a high-speed dedicated line as an interface for individual exchange. Note that if the single GA engine 32 is further subdivided and parallelized using a normal distributed GA, the processing speed can be further increased.
[0027]
In the robot apparatus having the above configuration, the operation state is simulated by the simulator 20 or the walking robot 1 using the gait parameters generated by the parameter generator 10, and the result is evaluated by the evaluation unit 31 of the GA engine group 30. In other words, since the GA operation is a method of performing repetitive learning, the evaluation progresses using a simulator instead of the actual machine (walking robot 1) because of the limitation of calculation time, and the result of the best individual with the highest evaluation condition is obtained. However, it is also possible to apply a part or all of the evaluation to the actual machine in consideration of the balance between the certainty of control and the time until the control result is obtained.
[0028]
The GA engine group 30 performs calculation so that the evaluation value of the best individual in each GA engine 32 is improved based on the evaluation value from each evaluation unit 31, passes a new parameter to the parameter generator 10, and sets the parameter The gait parameters (control parameters) output from the generator 10 to the walking robot 1 are updated. Then, the evaluation of the operation state by the simulator 20 or the walking robot 1 is repeated again to gradually evolve into a gait parameter having a high evaluation value.
[0029]
This optimization calculation in the GA engine group 30 performs a solution search based on only the simplest evaluation condition 1, for example, the “stand on the leg” condition, in the lowest GA engine 32, and outputs an evolved individual group PG1. In addition, the GA engine 32 that is one level higher than this adds a new evaluation condition 2, for example, the “do not fall” condition, to the evaluation condition 1 and performs a solution search under both evaluation conditions (1 + 2), and the evolved individual The group PG2 is output.
[0030]
Similarly, as the GA engine 32 becomes higher, evaluation conditions are added in stages, and optimization under more complicated evaluation conditions is attempted. The GA engine 32 at each stage outputs a solution candidate that matches an individually set evaluation condition as an evolved population PGn. Individual individuals constituting the evolved population PGn are coded according to rules appropriately set according to optimization. The initial setting of the evolution population PGn at each stage is preferably created randomly for each GA engine 32 in order to ensure individual diversity. In addition, the number of individuals of the evolved population PGn set in the GA engine 32 of each stage may be the same, but may be set to a different number of individuals.
[0031]
On the other hand, an individual with a poor evaluation value in the evolved individual group migrates to a group with a lower evaluation condition, and an individual with a good evaluation value migrates to a group with a higher evaluation condition. That is, the individual immigration control unit 33 moves a predetermined number of individuals (immigrants) from the evolved individual group PGn to a higher-order evolved individual group PGn + 1 in the order of good relative evaluation values at a predetermined timing. (Step up) The higher-order GA engine 32 that has accepted the immigrant performs an evolution operation from an individual who has obtained a reasonable evaluation of the lower evaluation conditions. Therefore, since the space where the optimal solution will exist can be limited with the stepwise addition of the evaluation conditions, the possibility of finding the optimal solution can be increased.
[0032]
Further, the individual immigration control unit 33 moves a predetermined number of immigrants from the evolved individual group PGn to the lower evolved individual group PGn-1 in the order of poor relative evaluation values at a predetermined timing ( Step down). The lower-level GA engine 32 that has accepted immigrants retries the solution search under simpler evaluation conditions than before. This is because when it is difficult to find a solution under a complex condition, the condition is simplified, a new local solution in the vicinity of the individual is quickly found, and the individual is again moved to the upper level.
[0033]
In this way, an individual that satisfies a simpler evaluation condition is stepped up to the next higher evolution group, and optimization under more complicated evaluation conditions is performed. On the other hand, an individual that fails to satisfy the complicated evaluation condition steps down to the next lower evolutionary population, and tries optimization under a simple evaluation condition again.
[0034]
In other words, under the evaluation environment in which conditions are added step by step, optimization operations are executed step by step, and a part of the evolved population PGn is exchanged between the upper and lower sides to optimize the degree of individual optimization It is possible to automatically and autonomously perform evolutionary calculations according to the situation. At the same time, the solution search of the evolutionary population PGn on the upper side is efficiently performed, and the bias of individual genes can be avoided (maintenance of diversity). Therefore, even if it falls into a local solution, the possibility of local solution escape increases dramatically and the progress of evolution becomes more continuous, so a high-quality solution is calculated at high speed even under complicated evaluation conditions. It becomes possible.
[0035]
Here, let us consider a case in which the approach is difficult to approach a local solution (optimal solution) in a solution space with an increased degree of complexity positioned at the top (a case where there is one mountain before reaching the solution). In the present embodiment, when the step up by the control by the individual immigration control unit 33 is stuck in the evolution, the retry loop of stepping down the individual, making some corrections on the lower side, and stepping up again is the result. Composed.
[0036]
With this retry loop, if you make a slight change to the individual stepped down and step up again, the angle of the approach will change, and as a result, it is possible to increase the possibility of bypassing the mountain and reaching the solution. In other words, since the solution space drastically expands as a complex evaluation condition in which a large number of conditions are combined, the conventional simple GA is likely to fall into a local solution. The progress is continuous, and a good optimal solution can be obtained without falling into a local solution.
[0037]
Further, according to the present embodiment, there is an effect that diversity of the group can be maintained. Maintaining diversity in GA is the only source that guarantees a broad solution search, and migrating individuals created and changed on the lower side greatly contributes to maintaining diversity in the group on the upper side. In addition, by providing a condition with a stronger restraining force on the lower side than on the upper side, it is possible to efficiently search for the final optimum solution. In general, a space in which a solution that matches a certain condition exists tends to increase as the constraint force of the given condition increases. Therefore, if a lower level is added, a condition with stronger binding force is added step by step, and a solution search on the upper side is performed after considering the global solution obtained on the lower side. The search can be expected to be performed efficiently.
[0038]
Immigration processing between GA engines is executed according to the procedure shown in FIG. This process is executed independently in each individual immigration control unit 33, and will be described below by exemplifying a process related to the n-th stage GA engine 32 as an intermediate stage.
[0039]
First, in step S1, it is determined whether it is upper immigration timing, that is, whether it is a timing to step up a part of the evolved population PGn to the immediately higher level. This timing may be a fixed timing such as every 20th generation, or may be determined by a lottery with a certain probability.
[0040]
If a positive determination is made in step S1, the process proceeds to step S2. On the other hand, if a negative determination is made in step S1, the upper-level immigration process (steps S2 to S5) is skipped and the process proceeds to step S6. Note that the uppermost GA engine 32 does not have the GA engine 32 on the upper side, and therefore, the upper immigration process (that is, the immigration step-up process and the exchange immigration receiving process from the upper side) is not performed.
[0041]
In step S2, a partial individual group to be stepped up to the upper side in the n-th stage evolution individual group PGn, that is, an immigrant individual to the upper side is determined. Specifically, migrants who evaluate the individual individuals that make up the evolved population based on a predetermined individual selection method, and basically step up individuals with good relative evaluation within this evolved population. Choose as. The number of immigrants selected is constant, and the total number of individuals in the evolved population PGn does not change during the evolution calculation. It is reasonable to determine the number of immigrants as a percentage of the total number of individuals.
[0042]
There are several general methods for selecting individuals with good evaluation values in GA, and immigration to the upper side is determined based on appropriate methods. These methods are not a method of simply determining only those with good evaluation values from ranks, but are based on probabilistic selection from the viewpoint of maintaining the genetic diversity of a group. Typical examples include roulette type selection, rank type selection, tournament type selection, and the like. In standard GA, these methods are used for selection of a parent individual when creating a next generation individual, but in this embodiment, a selection method equivalent to that is adopted.
[0043]
In step S3, individual data relating to immigration to the higher-order evolution individual group PGn + 1 is transmitted to the upper (n + 1) stage. In other words, the immigrant determined at the n-th stage is transmitted to the (n + 1) -th stage, and the n-th stage is in a state of waiting for exchange individual data reception (step S4). When the nth stage receives the exchange individual data from the (n + 1) th stage, the received individual data is set in the evolution individual group PGn in the nth stage (step S5).
[0044]
In step S6 subsequent to step S5, it is determined whether or not migrant individual data from the (n-1) th stage in which the nth stage is the lower side is received. When the migrant individual data is received, the process proceeds to step S7, and the low-level immigration process (steps S7 to S9) is performed. Regarding the lowest-order GA engine 32, since the GA engine 32 does not exist on the lower side, the lower-level immigration process (that is, the immigration step-down process and the exchange immigration reception process from the lower side) is not performed.
[0045]
In this low-level immigration process, for example, considering an individual that satisfies the two evaluation conditions of “standing with legs” and “not falling”, this individual added the condition of “move”, which is one higher level condition. Despite having immigrated to be evaluated by the GA engine 32 on the upper side, if the evaluation value is in a bad state in the population, this individual will not be deceived, but once again the original two Immigrants are moved to the lower side of the evaluation condition and nearby solutions are searched. Thus, by becoming an individual with a good evaluation value under the two evaluation conditions, it becomes possible to challenge the optimization with the GA engine 32 to which the higher evaluation conditions are added again.
[0046]
Specifically, in step S7, an individual to be moved to the lower side in the evolved population PGn, that is, an immigrant individual to the lower side is determined. This immigration to the lower side evaluates the individual individuals that make up the evolved population based on a predetermined individual selection method, and basically steps the individuals with poor relative evaluation within this evolved population. Choose as an immigrant to bring down. The number of immigrants to the lower side is constant, and the total number of individuals in the evolution population PGn does not change during the evolution calculation. It is reasonable to determine the number of immigrants as a percentage of the total number of individuals. As a method for determining immigration to the lower side, well-known methods such as reverse roulette method selection, reverse rank method selection, and reverse tournament type selection can be used, and the selection rate is reversed with respect to the evaluation value (poor evaluation value is bad). Use it so that the selectivity is higher).
[0047]
In step S8, the n-th stage transmits individual data relating to immigration to the lower-level evolution population PGn-1 in the lower (n-1) -th stage. In step S9, instead of deleting some individuals as immigrants, the received individual data from the lower side is set in the n-th stage evolution individual group PGn, and the next generation evolution calculation is performed.
[0048]
Although the above-described immigration process has been described with respect to an example in which individual movement is performed between the most recently evolved individual groups PGn, PGn + i (or PGn, PGn-1) adjacent to each other, for example, jumping such as two jumps. It is also possible to apply to general individual movement. Further, when the processing speed is prioritized over the improvement of the accuracy of the optimum solution by the retry loop described above, only the step up may be performed alone without performing the step down. In that case, there is no point in carrying out the step-down alone, and it is necessary to carry out the step-up together.
[0049]
For the above generation evolution calculation, the convergence determination unit 34 located at the top always observes the evolution population PGN calculated by the top GA engine 32, and the final evaluation condition “condition 1 + 2 ++. It is determined whether or not optimization satisfying + N ”has been performed. Specifically, based on an evaluation function appropriately set so that an individual that satisfies this “evaluation condition 1 + 2 +... + N” has a higher fitness, the individuals that make up the highest-level evolution population PGN are selected. Evaluate individually.
[0050]
If satisfactory optimization is not completed in a certain generation K, generational evolution calculation with individual immigration is continued. If optimization is satisfied, the highest-level calculation result is output to the parameter generator 10. At the same time, each GA engine 32 is instructed to end the GA operation. As a result, the entire optimization calculation is completed in the GA engine group 30.
[0051]
If it is desired to sequentially add (insert) new evaluation conditions after looking at the optimization result, it is sufficient to add one new GA engine 32 to which the evaluation conditions are added. The addition (insertion) position of the GA engine 32 is determined based on the strength of the binding force related to the newly added (insertion) evaluation condition. That is, it is preferable to add the GA engine 32 on the lower side if the binding force of the condition to be added is strong, and to add the GA engine 32 on the upper side if the binding force of the condition to be added is weak. As a result, it is possible to realize a system having excellent convenience that matches an actual solution search approach.
[0052]
When the optimization calculation in the GA engine group 30 is completed, the parameter generator 10 generates and updates the control parameter output from the individual evolved in the GA engine group 30 to the simulator 20 or the walking robot 1.
[0053]
In the present embodiment, a case will be described in which the walking robot 1 is a quadruped walking robot and its gait pattern is generated. As shown in FIGS. 5 and 6, the walking robot 1 includes four leg portions 3 a, 3 b, 3 c, and 3 d extending from the body portion 2, and detects a servo mechanism and posture for driving each leg portion. A detection unit is provided. Then, by inputting the instruction values for the joint angles θ2 and θ3 of the legs that give the vertical movement and the joint angle θ1 that gives the movement in the front-rear and left-right directions, the joint of each leg is driven by the servo mechanism, Walk.
[0054]
The parameter generator 10 is composed of a three-layer feedforward neural network, and uses a sigmoid function with each individual gene as a coupling coefficient between neurons. The coupling coefficient of this neural network is adjusted by the stepwise evolution method, and an optimal gait pattern is generated. Since walking is a periodic motion, a sin function and a cosine function with an elapsed time t and a walking cycle w as arguments are input to the neural network for one leg shown in FIG. The joint angle instruction values θ1, θ2, and θ3 are output.
[0055]
When the joint angle instruction value from the parameter generator 10 is output to the walking robot 1, the walking robot 1 drives each leg with an appropriate phase difference based on each joint instruction value by the servo mechanism to walk. The actual joint angles θ1, θ2, θ3, posture data, and the like are output to the evaluation unit 31 of the GA engine group 30.
[0056]
On the other hand, the simulator 20 is a model simulating an actual walking robot 1. When the joint angle instruction value from the parameter generator 10 is input, the robot 20 performs calculations such as the posture of the robot, the walking distance, and the fall determination. The result is output to the evaluation unit 31 of the GA engine group 30.
[0057]
FIG. 8 shows a processing procedure in the simulator 20. In this processing, first, a joint instruction value is received from the parameter generator 10 in step S 11. In step S12, the position of each leg is calculated from the robot model based on the joint instruction value. In step S13, the overall posture is calculated from the position of each leg.
[0058]
In the subsequent step S14, it is determined whether or not the body angle has exceeded the set limit from the previously calculated position and posture of the leg of the robot. If it is determined that the vehicle has fallen, the process jumps to step S17. If it is determined that the vehicle has not fallen, the stability is calculated from the position and posture of each leg in step S15, and the gait is determined in step S16. It is determined whether one cycle of the pattern is completed. If the simulation for one cycle is not completed, the process returns to step S11 and the same process is continued.
[0059]
On the other hand, when the fall determination in step S14 or the simulation end determination for one cycle in step S16 is made, in step S17, the total movement distance and energy used of the robot are calculated. In step S18, data is passed to the evaluation unit 31, and one cycle of processing is completed.
[0060]
Evaluation conditions for evaluating the gait pattern in the evaluation unit 31 include conditions of strong restraint such as “stand on the leg (do not put the body on the ground)”, “do not fall”, and “use energy with relatively weak restraint” There are various conditions, such as “There are few”, “Walk as much as possible,” and “Walk as little as possible”. Evaluation conditions are individually assigned to the GA engines 32 shown in FIG. 2 so that the condition with the strong binding force is the lower order.
[0061]
For example, in the lower space when the evaluation condition “standing with legs” is assigned to the lowest position (the space formed by the adjustment parameter), several local solutions “standing with legs” are calculated by evolution by GA. Although many local solutions exist in the lower space and may be in positions that are not related to each other, "stand on the leg + do not fall" is evaluated by selecting individuals in the vicinity of this local solution. Immigrate to the higher level as a condition. On the upper side, the condition of “standing with legs” can be evolved from an individual who has obtained a reasonable evaluation, thus increasing the possibility of finding the optimal solution.
[0062]
Also, from the upper side, individuals with low evaluation are accepted as immigrants, and the solution search is performed again only with the simple condition of “standing with legs”. The reason for this is that, as described above, when it is difficult to find a solution under a complex condition, a new local solution existing in the vicinity of the individual can be quickly found by unifying the conditions and simplified, and the individual is again placed in the higher rank. It is for moving.
[0063]
In this way, even when there are many evaluation conditions to be satisfied as a walking robot and they have complicated relationships with each other, by connecting the optimization operations by GA step by step and performing evolutionary individual exchange with each other, Evolutionary calculations can be performed automatically and autonomously according to the degree of individual optimization, and an optimal gait can be generated.
[0064]
In addition, it is possible to efficiently search for a gait solution that realizes a more desirable gait, and to avoid bias of individual genes. Some evaluation conditions reduce the likelihood that the gait will be trapped in the local solution, and the degree of evolution will be more continuous.
[0065]
【The invention's effect】
As described above, according to the present invention, the plurality of evaluation conditions to be satisfied by the walking robot are divided stepwise from the lower level to the higher level. Include the lower evaluation condition in the upper evaluation condition. , While performing individual exchange between the lower-level evolution population and the higher-level evolution population according to the individual evaluation value, each candidate of the optimal solution that matches the evaluation condition of each stage is calculated, and the highest When the optimization operation of the evolutionary population of the robot is determined to converge, the control parameters of the walking robot are updated based on this top-level evolutionary population, so the convergence of the optimal solution in gait generation is improved and It is possible to prevent stagnation and obtain a fast and reliable gait pattern.
[Brief description of the drawings]
FIG. 1 is an overall configuration diagram of a robot apparatus.
FIG. 2 is a block diagram of an optimization apparatus using a stepwise evolution method.
FIG. 3 is a block diagram of an optimization device using a plurality of arithmetic devices.
FIG. 4 is a flowchart of immigration processing between GA engines.
FIG. 5 is an explanatory diagram of joint angles that give vertical movement to a quadruped walking robot.
FIG. 6 is an explanatory diagram of a joint angle that gives a four-legged walking type robot a forward / backward / left / right movement.
FIG. 7 is an explanatory diagram of a neural network that generates a gait pattern for one leg.
FIG. 8 is a flowchart showing simulator processing.
[Explanation of symbols]
1 Walking robot
3a, 3b, 3c, 3d Leg
10 Parameter generator
20 Simulator
30 GA engine group
31 Evaluation Department
32 GA engine
33 Individual Immigration Control Department
34 Convergence determination unit
PGn evolution population

Claims (4)

Based on the genetic algorithm, multiple evaluation conditions to be satisfied by a walking robot walking with single or multiple legs are divided step by step from lower to higher, and an optimal solution that matches the evaluation conditions of each stage Means for computing each candidate as an evolutionary population,
Means for exchanging individuals according to the evaluation value of the individual between the lower-level evolution population and the higher-level evolution population,
Means for determining the convergence of the optimization operation based on the top-level evolution population,
A means for updating the control parameters of the walking robot based on the top evolutionary population in which the optimization operation has converged ,
The robot apparatus characterized in that the higher-order evaluation condition includes the lower-order evaluation condition .   The robot apparatus according to claim 1, wherein the computation of the evolved population is executed by a plurality of computing units in a distributed manner.   The robot apparatus according to claim 1, wherein the evaluation value is calculated based on data from a simulator that models the movement of the walking robot.   4. The robot apparatus according to claim 1, wherein the evaluation value is calculated based on actual machine data from the walking robot.

Images