JP5450218B2 - Robot and control system - Google Patents

Description

  The present invention relates to a control system and a robot that is one of its control objects.

  There has been proposed a method for causing a robot to behave adaptively against high-speed disturbances such as unexpected contact with an object (see Patent Document 1). According to this method, the configuration ratio or distribution between the control command value of the overall control system and the control command value of the high-speed local control system whose operation cycle is shorter than that of the overall control system is adjusted. For example, immediately after the occurrence of a high-speed disturbance, the robot is adjusted so that the ratio of the control command values of the local control system is increased, thereby realizing a reflective behavior mode of the robot with respect to the disturbance. .

  Then, with the passage of time from the occurrence of the disturbance, the ratio of the control command values of the overall control system is adjusted so as to gradually increase, so that the operation mode of the robot in the normal state without the disturbance is recovered. Here, the change curve of the composition ratio of each control command value can be changed over time.

JP 2004-167666 A

  However, depending on the form of disturbance, the behavior form of the robot may become an inappropriate behavior form in view of the purpose of the robot.

  Therefore, an object of the present invention is to provide a control system and the like that can cause the control target to act in an appropriate form in view of the action purpose of the control target with respect to an arbitrary form of disturbance.

The control system of the present invention for solving the above-described problem includes a plurality of modules that are hierarchized according to the relative length of the calculation cycle, and each of the modules is a candidate of a behavior form to be controlled. In addition to searching for action candidates, for each of the action candidates, the contribution of the main fitness to the main purpose for which the self is in charge is higher than the contribution of the sub-fit to the sub-objects of other modules. The overall fitness is evaluated in consideration of both the main purpose and the sub-object, and the computation cycle is relatively higher than the evaluation result of the overall fitness by the lower module having a relatively long computation cycle. characterized in that it is configured to control the operation of the control object the evaluation results of the overall goodness of fit in the form of preferentially reflect by short-level module The first aspect of the present invention).

  According to the control system of the first aspect of the present invention, the priority of the main purpose for which the own module is primarily responsible for each of the plurality of modules, while taking into account the sub-objects for which the other modules are primarily responsible, Candidates are searched. In this way, action candidates to be controlled can be searched in a manner in which the purpose is gently shared between different modules.

In addition, for each of the action candidates, the main purpose and the sub-purpose are determined so that the main fitness level in consideration of the main purpose for which the self is responsible is higher than the fitness level in view of the sub-object for which the other module is mainly responsible. Overall suitability is evaluated considering both objectives. Then, the behavior of the control target is reflected in a form in which the overall fitness evaluation result by the higher module having a relatively short calculation cycle is reflected in preference to the overall fitness evaluation result by the lower module having a relatively long calculation cycle. Is controlled.

  Therefore, when a disturbance occurs, the higher-order objective mainly handled by the higher-order module that responds quickly is reflected in the action form to be controlled, and the influence of this disturbance can be reduced quickly. Further, the sub-objects that are mainly handled by the subordinate modules that are slow to respond with the passage of time are also reflected in the action form of the control object, so that the action form of the control object that meets all of the plurality of purposes can be recovered. Thereby, the said control object can be made to act with a suitable form in view of the action objective of a control object with respect to the disturbance of arbitrary forms.

In the control system of the first invention, the intermediate module excluding the highest module having the shortest calculation cycle and the lowest module having the longest calculation cycle is a high-order purpose mainly handled by an upper module having a relatively shorter calculation cycle than itself. And, each sub-object whose main purpose is a sub-module whose operation cycle is relatively longer than that of the sub-object is the sub-object, and for each of the action candidates, the high-level fitness in consideration of the high-order object is the sub-object You may comprise so that the said comprehensive fitness may be evaluated so that contribution may become higher than the low-order fitness in view (2nd invention).

  According to the control system of the second invention, for each of the action candidates searched for by the intermediate module, the order of the main suitability for the main purpose for which the self is responsible, the higher suitability for the higher purpose, and the lower suitability for the lower purpose. The total fitness for all the objectives is evaluated so that the degree of contribution becomes high. As a result, the purpose of the intermediate module (intermediate purpose) is reflected in preference to the higher-order goal, and the behavior of the control target is controlled in a form in which the higher-order goal is reflected in preference to the lower-order goal. . As a result, it is possible to cause the control target to behave in an appropriate form in view of the action purpose of the control target with respect to an arbitrary form of disturbance.

In the control system of the second invention, the module in which a lower module having a relatively longer calculation cycle than the self evaluates the estimated lower fitness by estimating the lower fitness for each of the action candidates, The total fitness may be evaluated using the estimated lower fitness as the lower fitness (third invention).

  According to the control system of the third aspect of the present invention, the self-module is reflected in a form reflecting not only the purpose of the module in charge of the subordinate module but also the purpose of the module inferior to the subordinate module. Action candidates can be searched by the module. That is, the control target action candidate can be searched in such a manner that the upper module in which there are two or more lower modules having different ranks gently share the purpose of each of the two lower modules. As a result, the target to be controlled by the higher-level module is preferentially reflected for disturbances of any form, while the target for the lower-level module is also indirectly or secondarily reflected. Can act.

In the control system according to the third aspect of the invention, the module in which a lower module having a relatively longer calculation cycle than the self is evaluated by the lower module among the lower action candidates as the action candidates searched by the lower module. The lower behavior candidate having the maximum or maximum lower comprehensive fitness as the total fitness is received from the lower module as a lower behavior guideline, and the degree of approximation with the lower behavior guideline for each of the behavior candidates searched by itself May be evaluated as the estimated lower matching degree (fourth invention).

  According to the control system of the fourth aspect of the present invention, the lower action guide received from the lower module, which is the evaluation criterion of the estimated lower matching degree by the self module, is the lower module received by the lower module from the lower module. The action guidelines are reflected. Therefore, the self-module searches for action candidates in a form that reflects not only the purpose of the module in charge of the subordinate module but also the purpose of the module inferior to the subordinate module. sell. That is, the control target action candidate can be searched in such a manner that the upper module in which there are two or more lower modules having different ranks gently share the purpose of each of the two lower modules. As a result, the target to be controlled by the higher-level module is preferentially reflected for disturbances of any form, while the target for the lower-level module is also indirectly or secondarily reflected. Can act.

In the control system of the third or fourth invention, each of the first action candidates as the action candidates searched for by the first module as the highest module includes the first to Nth modules as the plurality of modules. The second estimated fitness e ₂ is evaluated by evaluating the first fitness e ₁ for the first goal as the main goal and estimating the fitness for the second goal mainly handled by the second module as the lower module. After evaluating ^, the first overall fitness f ₁ is evaluated according to the evaluation formula f ₁ = e ₁ (e ₂ ^ + 1), and the i-th module (i = 2 to 2) as the intermediate module is configured. N-1) is, for each i-th action candidates as the behavior candidate of searching itself, to evaluate the i-th fitness e _i in light to the i-th object as the main purpose, the upper eye (I-1) th adaptability in view to the (i-1) th object as e _i-1 is evaluated and the (i + 1) th estimated fitness by estimating the adaptability that light of the (i + 1) object of the subgoals e _{i + 1} ^ is evaluated, and the _i- th total fitness f _i is evaluated according to the evaluation formula f _i = e _i (e _i-1 (e _{i + 1} ^ + 1) +1). The Nth module as the lower module evaluates the Nth adaptability e _N with respect to the Nth objective as the main objective for each of the Nth action candidates as the behavioral candidates searched by itself, and the Nth module as the upper module. -1 module evaluates the N-1 fitness level e _N-1 and then evaluates the i-th total fitness level f _i according to the evaluation formula f _N = e _N (e _N-1 +1). (5th invention).

  According to the control system of the fifth invention, each of the two lower modules is included in an upper module in which two or more lower modules having different levels (the first lower module and the second lower module) exist. Action candidates to be controlled can be searched for in a manner in which the purpose of the main charge is gently shared. As a result, the target to be controlled by the higher-level module is preferentially reflected for disturbances of any form, while the target for the lower-level module is also indirectly or secondarily reflected. Can act.

In the control system of the fifth invention, the j-th module (j = 1, 2,..., N) is included in the j-th designated action candidate group among all the j-th action candidates as the action candidates searched for by itself. About j behavior candidate, while evaluating the j fit e _j for the j purpose as the main object to a constant value, said of all of the j behavior candidate, is not included in the j-th designated action candidates About j behavior candidate, which may be configured to evaluate lower than the first j adaptability e _j the predetermined value (sixth invention).

According to the control system of the sixth aspect of the present invention, among all of the j behavior candidate searched by the j-th module, for the first j action candidates included in the j-specified behavior candidate group, the j-th fitness e _j constant Evaluates to a value. Meanwhile, for the j-th action candidates are not included in the j-th specified behavior candidate group, the j fit e _j evaluates to a value lower than a predetermined value. That is, of the set can be taken as an action candidates in the set of actions candidates j-th fitness e _j takes a sufficiently high value (j-th solution space), the j-th fitness e _j is evaluated in the constant value a region (flat region), the j-th fitness e _j is a region lower than the predetermined value is formed. This configuration is based on the idea that j-th action candidates that are suitable for the j-th purpose to some extent may consider the j-th suitability to be approximately equivalent.

  By forming the flat area, an area in which a solution having the maximum or maximum total fitness can be searched in the j-th solution space is expanded. For this reason, the upper module receives the lower action candidate having the maximum or maximum lower overall fitness from the lower module as the lower action guideline, and estimates the degree of approximation with the lower action guideline for each action candidate searched by itself. In the solution space (see the fourth aspect of the invention), an area where a solution can be searched for by the higher module in the solution space is expanded.

Further, the plurality of first j behavior candidate included in the flat region in the j-th solution space, the superiority in accordance with the height of the j-fit e _j may not be considered. For this reason, searching for a solution that increases the j-th matching degree more than necessary, and thus, excessive optimization of the operation control of the controlled object is prevented. Further, the j-th module has a maximum or maximum total fitness according to the fitness in consideration of the sub-objects in which the other modules (one or both of the j-1 module and the j + 1 module) are mainly responsible. It is sufficient to search for j action candidates. In other words, the j-th module can be concentrated on searching for a solution that increases the degree of fitness in consideration of the sub-object, not the main object.

Therefore, the j adaptability e _j reduce processing load of each module for the evaluation of, and thus, while achieving earlier stage of calculation of the calculation result, the j overall fitness f _j can be evaluated or calculated . Thereby, it is possible to promptly respond or act on the control target in an appropriate form in view of the behavioral purpose of the control target with respect to an arbitrary form of disturbance.

In the control system of a sixth aspect of the invention, the j-th module has a divergence degree from a reference point defined in a state space corresponding to the j-th objective among all the j-th action candidates being not more than a j-th threshold value. the j-th action candidates, as the j action candidates included in the j-th designated action candidates, while evaluating the j-th fitness e _j to the predetermined value, among the all of the j behavior candidate, the second About j behavior candidate that is not included in the j specified action candidates, based on the deviation degree, it is adapted to evaluate the first j adaptability e _j according continuously or stepwise decreasing function of the degree of deviation (7th invention).

According to the control system of the seventh invention, similar to the control system of the sixth aspect of the present invention, reduction in processing load for the evaluation of the j fit e _j, therefore, while achieving earlier stage of calculation of the calculation result, The jth total fitness _fj can be evaluated or calculated. Thereby, it is possible to promptly respond or act on the control target in an appropriate form in view of the behavioral purpose of the control target with respect to an arbitrary form of disturbance.

  In any one of the control systems according to the first to seventh aspects of the invention, each of the modules predicts a plurality of future states of the control target according to each of the action candidates searched for by itself, and each of the action candidates May be configured to evaluate the overall fitness based on each of the future states (eighth invention).

  According to the control system of the eighth aspect of the present invention, based on each of the future states predicted according to each of the action candidates by the self module, for each of the action candidates, the main purpose for which the self responsible for the self module is the main purpose And the overall fitness for each of the sub-objects for which the other modules are primarily responsible. Then, as described above, the behavior of the control target is controlled in a form in which the evaluation result of the overall fitness by the higher frequency higher module is reflected in preference to the evaluation result of the overall fitness by the lower frequency lower module. As a result, it is possible to cause the control target to behave in an appropriate form in view of the action purpose of the control target with respect to an arbitrary form of disturbance.

In any one of the control systems according to the first to eighth aspects, each of the modules is configured to search for the action candidate that determines a position or a position and a posture of the mobile device as the control target. The higher module having a relatively short calculation cycle may be configured to search for the action candidate over a short period of the mobile device as compared to a lower module having a relatively long calculation cycle (first). 9 invention).

According to the control system of the ninth invention, the position or the position and orientation of the moving apparatus in accordance with the appropriate purpose in accordance with the length of the computation cycle of that in each module, as a control object for a period of suitable length ( Hereinafter, a search for action candidates for determining “position etc.” as appropriate can be executed. As a result, in response to an arbitrary form of disturbance, the mobile device can be acted so that its position and the like are appropriately controlled in view of the behavioral purpose of the mobile device.

  In a control system according to a ninth aspect of the present invention, a position trajectory and a posture trajectory of a robot as the moving device having a base and a plurality of legs extending from the base are searched as the action candidates. A first module, a second module, and a third module are provided as the plurality of modules, and the first module searches for a gait over the first specified number of steps of the robot for stabilizing the posture of the robot as a first action candidate. The second module searches for a local route that defines a gait over a second specified number of steps larger than the first specified number of steps for causing the robot to avoid contact with an object as a second action candidate. As a third action candidate, a global route for the third module to cause the robot to reach a target position is configured. It is configured to search good (tenth aspect).

According to the control system of the tenth aspect of the present invention, the main purpose of the first action purpose of “stabilizing the robot's posture” is reflected in the first module having the shortest calculation cycle while reflecting the sub-object that the other module is mainly responsible for. As a result, it is possible to search for a first action candidate over a short period when a disturbance occurs. In addition, while reflecting the sub-object that the other module is mainly responsible for in the second module having an intermediate calculation period , the main purpose is the second action purpose of “making the robot avoid contact with the object”. It is possible to search for a second action candidate over a longer period than one action candidate. In addition, while reflecting the sub-object that the other module is mainly responsible for in the third module with the longest calculation cycle , the third objective “to make the robot reach the target position” is compared with the second action candidate. The third action candidate over a long period can be searched.

  One or both of the second module and the third module may correspond to the “other module” viewed from the first module. Similarly, one or both of the first module and the third module may correspond to “another module” viewed from the second module. In addition, one or both of the first module and the second module may correspond to “another module” viewed from the third module.

Then, the behavior of the control target is reflected in a form in which the overall fitness evaluation result by the higher module having a relatively short calculation cycle is reflected in preference to the overall fitness evaluation result by the lower module having a relatively long calculation cycle. Is controlled. As a result, it is possible to cause the control target to behave in an appropriate form in view of the action purpose of the control target with respect to an arbitrary form of disturbance.

  A robot of the present invention for solving the above-mentioned problems is a robot as the moving device that includes a base and a plurality of legs extending from the base and moves by the movement of the plurality of legs. And a control system according to the tenth aspect of the invention (eleventh aspect).

  According to the robot of the eleventh aspect of the invention, when a disturbance occurs, the higher-order objective mainly handled by the higher-order module that is initially responding is reflected in the action form of the robot as the control target, and the effect of the disturbance can be quickly reduced. Figured. In addition, the subordinate objectives, which are mainly handled by the subordinate modules whose response is slow over time, are also reflected in the behavioral form of the robot as the control target, so that the robot's behavioral form can be restored to suit all of the plural objectives. Thereby, the robot can behave in an appropriate form in view of the purpose of the action with respect to an arbitrary form of disturbance.

BRIEF DESCRIPTION OF THE DRAWINGS The structure explanatory drawing of the robot as one Embodiment of this invention. FIG. 2 is a configuration explanatory diagram of a robot control system. Explanatory drawing regarding the exchange of the signal between parallel modules. Explanatory drawing regarding the cooperation process of a parallel module. Explanatory drawing regarding the action search method by each module. Explanatory drawing regarding the evaluation method of the 1st fitness. Explanatory drawing regarding the evaluation method of a 2nd fitness. Explanatory drawing regarding the evaluation method of the 3rd fitness. Explanatory drawing regarding the diversity and convergence of the behavior search method. Explanatory drawing regarding the diversity of behavior search methods. Explanatory drawing regarding the diversity of behavior search methods. Explanatory drawing regarding the relationship between the behavior state of a robot and the behavior state of an inverted pendulum. Explanatory drawing of the experimental result by an inverted pendulum. Explanatory drawing regarding the action of the robot.

  An embodiment of a control system of the present invention and its control object will be described with reference to the drawings.

(Robot configuration)
First, a configuration of a robot as an embodiment of the present invention will be described.

  The robot R shown in FIG. 1 is a legged mobile robot. Like a human, the robot R extends from both sides of the base B0, the head B1 disposed above the base B0, and the top of the base B0. The left and right arm bodies B2, the hand H provided at the tip of each of the left and right arm bodies B2, and the left and right leg bodies B4 extending downward from the lower portion of the base body B0 are provided.

  The base B0 is composed of an upper part and a lower part that are connected vertically so as to be relatively rotatable about the yaw axis. The head B1 can move, such as rotating around the yaw axis with respect to the base B0.

  The arm body B2 includes a first arm body link B22 and a second arm body link B24. The base B0 and the first arm body link B22 are connected via a shoulder joint mechanism (first arm joint mechanism) B21, and the first arm body link B22 and the second arm body link B24 are connected to an elbow joint mechanism (second arm). The second arm body link B24 and the hand H are connected via a wrist joint mechanism (third arm joint mechanism) B25. The shoulder joint mechanism B21 has a degree of freedom of rotation around the roll, pitch and yaw axes, the elbow joint mechanism B23 has a degree of freedom of rotation around the pitch axis, and the wrist joint mechanism B25 is around the roll, pitch and yaw axes. It has a degree of freedom of rotation.

  The leg B4 includes a first leg link B42, a second leg link B44, and a foot B5. The base B0 and the first leg link B42 are connected via a hip joint mechanism (first leg joint mechanism) B41, and the first leg link B42 and the second leg link B44 are connected to a knee joint mechanism (second leg joint). The mechanism) is connected via B43, and the second leg link B44 and the foot B5 are connected via an ankle joint mechanism (third leg joint mechanism) B45.

  The hip joint mechanism B41 has a degree of freedom of rotation about the roll, the pitch, and the roll axis, the knee joint mechanism B43 has a degree of freedom of rotation about the pitch axis, and the ankle joint mechanism B45 rotates about the roll and the pitch axis. Has a degree of freedom. The hip joint mechanism B41, the knee joint mechanism B43, and the ankle joint mechanism B45 constitute a “leg joint mechanism group”. The translational and rotational degrees of freedom of each joint mechanism included in the leg joint mechanism group may be changed as appropriate. In addition, an arbitrary one of the hip joint mechanism B41, the knee joint mechanism B43, and the ankle joint mechanism B45 may be omitted, and the leg joint mechanism group may be configured by a combination of the remaining two joint mechanisms. . Further, when the leg body B4 has a second leg joint mechanism different from the knee joint, the leg joint mechanism group may be configured to include the second leg joint mechanism. An elastic material B52 as disclosed in Japanese Patent Application Laid-Open No. 2001-129774 is provided on the bottom of the foot B5 in order to alleviate the impact when landing.

The robot R is equipped with a plurality of internal state sensors S ₁ for measuring internal states such as position and posture of the robot R in the world coordinate system. An encoder (not shown) that outputs a signal corresponding to the bending angle (joint angle) of each joint mechanism of the robot R, and an inclination sensor that outputs a signal corresponding to the posture (identified by the azimuth angle and elevation angle) of the base body B0. , and, like a pressure sensor for determining the different foot B5 and implantation and lifting corresponds to the internal state sensor S _1. An imaging apparatus for recognizing the position of the robot R in the world coordinate system by capturing an image of the surroundings of the robot R and recognizing the position of the marker fixed to the world coordinate system based on the imaging coordinates. corresponding to the state sensor S _1.

  For example, a pair of left and right head cameras C1 mounted on the head B1 and capable of sensing light in various frequency bands, such as a CCD camera and an infrared camera having an imaging range in front of the robot R, can be employed as the imaging device. Further, a waist mounted on the lower front side of the base B0 and used to measure the position and orientation of the object by detecting the reflected light of the near-infrared laser light emitted toward the front lower side of the robot R. A camera (active sensor) C2 may be employed as the imaging device.

The robot R is equipped with an external state sensor S ₂ for measuring an external state such as the position of an object around the robot R. It said such an imaging device corresponds to an external state sensor S _2.

The robot R includes a control system 1 and a plurality of actuators 2 for moving each of the plurality of joint mechanisms. Each operation of the actuator 2 is controlled according to a control command output from the control system 1 according to the internal state and the external state of the robot R, so that the robot R can behave adaptively in various modes. .
(Control system configuration)
The control system 1 shown in FIG. 2 includes a control module mod0, a first module (stabilization module) mod1, a second module (obstacle avoidance module) mod2, and a third module (path generation module) mod3. It has.

(Configuration of the first module)
The first module mod1 is mainly responsible for stabilizing the posture of the robot R as a main purpose (first action purpose). The first module mod1 is configured to repeat the generation of the output signal by processing the input signal for each first operation period T _1.

(Configuration of the second module)
The second module mod2 is mainly responsible for causing the robot R to avoid contact with an object as a main purpose (second action purpose). The second module mod2 is configured to repeat the generation of the output signal by calculating the input signal every second calculation cycle T ₂ (for example, T ₂ = 2T ₁ ) longer than the first calculation cycle T ₁ .

(Configuration of the third module)
The third module mod3 is mainly responsible for moving the robot R to the target position as the main purpose (third action purpose). The third module mod3 is configured to repeat the generation of the output signal by calculating the input signal every third calculation cycle T ₃ (for example, T ₃ = 2T ₂ ) longer than the second calculation cycle T ₂ .

(Robot function)
The function of the robot R having the above configuration will be described.

(Calculation procedure by control system)
The control system 1 controls the operation of the actuator 2 and the behavior of the robot R according to the following procedure.

First, the current state of the robot R is recognized by the control module mod0. Specifically, based on the output signal from the internal state sensor S _1, the internal state such as the position in the world coordinate system of the robot R (such as landing position and ZMP position) and orientation (posture of the body B0) is recognized. Further, based on the output signal from the external state sensor S _2, the external conditions such as the position in the size and the world coordinate system of the object in the surroundings of the robot R is recognized. It should be noted that the control module mod0 may recognize part or all of the external state by inputting or transmitting data representing the external state such as the size and position of the object from the outside of the robot R to the control system 1.

"Current state" of the robot R is input from the control module mod0 each first calculation cycle T ₁ to the first module mod1 (see FIG. 3 / Arrow D _1+). Further, the “second search strategy” and the “second action guideline” obtained by executing the second action search process are input from the second module mod2 to the first module mod1 every second operation cycle T ₂ (FIG. 3). / See arrow D _2- ).

In response to the input signal, the “first action search process” is executed by the first module mod1 every first calculation cycle T ₁ . "First action guidelines" obtained by the execution of the first action search processing is input to the control module mod0 from the first module mod1 each first calculation cycle T ₁ (see FIG. 3 / Arrow D _1-).

  If there is no input signal from the second module mod2 at the start time of the current cycle of the first module mod1, the input signal from the second module mod2 at the start time of the most recent past cycle of the first module mod1 is used. One action search process is executed.

Current state, or the latest estimated status of the robot R in the nearest future in accordance with the current state is inputted from the first module mod1 the second module mod2 every second calculation cycle T ₂ (Fig. 3 / Arrow D ₂₊ reference). Furthermore, the “first search strategy (including the first action reservation)” obtained by executing the first action search process is input from the first module mod1 to the second module mod2 every second operation cycle T ₂ ( FIG. 3 / see arrow D ₂₊ ). Further, the “third search strategy” and the “third action guideline” obtained by the execution of the third action search process are input from the third module mod3 to the second module mod2 every third operation cycle T ₃ (FIG. 3). / See arrow D _3- ).

In response to the input signal, "second action search process" is performed by the second module mod2 every second calculation cycle T _2. Thus obtained "second search strategy" and "second Action Guidelines" is inputted from the second module mod2 every second calculation cycle T ₂ in the first module mod1 (see FIG. 3 / Arrow D _2-). Further, "second search strategy" and "second Action Guidelines" is also input to the second module mod2 from the third module mod3 every second calculation cycle T ₂ (see FIG. 3 / Arrow D _2+).

  If there is no input signal from the third module mod3 at the start time of the current cycle of the second module mod2, the input signal from the third module mod3 at the start time of the latest past cycle of the second module mod2 is used. 2 Behavior search processing is executed.

Current state, or the latest estimated status of the robot R in the nearest future in accordance with the current state is inputted from the second module mod2 the third module mod3 every third calculation cycle T ₃ (Fig. 3 / Arrow D ₃₊ reference). Moreover, obtained by the execution of the second action search processing "second search strategy (. Including second action reservation)" is input from the second module mod2 the third module mod3 every third calculation cycle T ₃ ( FIG. 3 / see arrow D ₃₊ ).

In response to the input signal, "third action search process" is performed by the third module mod3 every third calculation cycle T _2. As described above, the “third search strategy” and the “third action guideline” thus obtained are input from the third module mod 3 to the second module mod 2 at every third calculation cycle T ₃ (FIG. 3 / arrow D). See _3- ).

Command signal in response to the input signal (see FIG. 3 / Arrow D _1-) from the first module mod1 first every calculation cycle T ₁ is is sequentially generated by the control module mod0. The command signal is transmitted from the control module mod0 to the actuator 2, whereby the actuator 2 is driven according to the command signal. As a result, the robot R behaves according to the latest first action guideline.

(Cooperation processing of parallel modules)
The procedure of parallel processing by each module immediately after the occurrence of a disturbance will be described.

Consider a case where the “current state” affected by the disturbance of the robot R is first input from the control module mod0 to the first module mod1 at time t = t ₀ (see FIG. 4 / arrow D ₁₊ (t ₀ )). ).

First, by executing the “first action search process” by the first module mod1, the “first action guideline (= first action reservation)” is the first at time t = t ₁ (= t ₀ + T ₁ ). The data is output from the module mod1 to the control module mod0 (see FIG. 4 / arrow D ₁₋ (t ₁ )). By controlling the operation of the actuator 2 according to this output, the robot R is adapted to the first action purpose (= the robot R stabilizes the posture) in consideration of the current state affected by the disturbance. Can act like that.

At time t = t ₁ , the “first search strategy” is output from the first module mod 1 to the second module mod 2 (see FIG. 4 / arrow D ₂₊ (t ₁ )). Furthermore, by executing the “second action search process” by the second module mod2, the “second search strategy (including the second action guideline)” becomes the second module at time t = t ₀ + T ₁ + T ₂ . The signal is output from mod 2 to the first module mod 1 (see FIG. 4 / arrow D ₂₋ (t ₀ + T ₁ + T ₂ )).

Then, by executing the “first action search process” by the first module mod1, the “first action guideline” is changed from the first module mod1 to the control module at the time t = t ₂ (= t ₀ + T ₂ + 2T ₁ ). It is output to mod0 (see FIG. 4 / arrow D ₁₋ (t ₂ )). By controlling the operation of the actuator 2 in accordance with this output, the robot R takes into consideration the current state affected by the disturbance, and the robot R adds to the second action purpose (= the robot R). To avoid contact with an object).

Also, at time t = t ₀ + T ₁ + T ₂ , “second search strategy (including second action reservation)” and the second module mod 2 are output to the third module mod 3 (FIG. 4 / arrow D ₃₊ (See t ₀ + T ₁ + T ₂ )). Furthermore, by executing the “third action search process” by the third module mod3, the “third search strategy (including the third action guideline)” is the third at time t = t ₀ + T ₁ + T ₂ + T ₃ . output from 3 modules mod3 the second module mod2 (4 / arrow _{D 3- (t 0 + T 1} + T 2 + T 3) reference).

Furthermore, the “second search strategy (including the second action guideline)” is performed at time t = t ₀ + T ₁ + 2T ₂ + T ₃ by executing the “second action search process” by the second module mod2. The data is output from the two modules mod 2 to the first module mod 1 (see FIG. 4 / arrow D ₂₋ (t ₀ + T ₁ + 2T ₂ + T ₃ )).

Then, by executing the “first action search process” by the first module mod1, the “first action guideline” is changed from the first module mod1 at the time t = t ₃ (= t ₀ + 2T ₁ + 2T ₂ + T ₃ ). It is output to the control module mod0 (see FIG. 4 / arrow D ₁₋ (t ₃ )). In accordance with this output, the operation of the actuator 2 is controlled, so that the robot R takes into consideration the current state affected by the disturbance and the third action purpose in addition to the first action purpose and the second action purpose. (= Move the robot R to the target position).

By the above-described cooperative processing of the first module mod1, the second module mod2, and the third module mod3, the first action is performed in consideration of the disturbance within the time of about 2T ₁ + 2T ₂ + T ₃ from the occurrence of the disturbance. The behavior of the robot R can be controlled so as to meet all of the purpose, the second behavior purpose, and the third behavior purpose. For example, if T ₁ = T, T ₂ = 2T, and T ₃ = 4T, the behavior of the robot R is controlled within a time of about 10T from the occurrence of the disturbance so that the action of the robot R can be adapted to all of a plurality of purposes in consideration of the disturbance. Can be done.

(Contents of each action search process)
Details of the “first action search process” executed by the first module mod1, the “second action search process” executed by the second module mod2, and the “third action search process” executed by the third module mod3 explain.

(Procedure for the first action search process)
First, the most recent future state of the robot R is predicted based on the current state according to a behavior prediction model representing the behavior of the robot R (FIG. 5 / STEP 100). Thereby, for example, the landing position when the robot R advances one step from the current position is predicted.

Further, a plurality of first action candidates {a _i1 | i ₁ = 1, 2,..., N ₁ } of the robot R are searched according to the first search strategy and the second search strategy (see arrow D _2-in FIG. 3). (FIG. 5 / STEP 110). Specifically, the first action candidate a _i1 is searched according to the evolutionary search generation change method according to the evaluation result of the first comprehensive fitness f _{1 and} the evaluation result of the second comprehensive fitness f ₂ . This search method will be described later.

Thereby, for example, the landing position pos (a _i1 ; k ₁ ) (k ₁ = 1) that defines a gait extending from the predicted landing position one step ahead of the robot R to the first designated step count q ₁ (for example, q ₁ = 2). ˜q ₁ ) are searched for as the first action candidate a _i1 . The “gait” means a repeated pattern of leg movements focusing on landing timing.

Furthermore, based on each of the plurality of first action candidates a _i1 , a plurality of first future states {s _i1 | i ₁ = 1, 2,..., N ₁ } of the robot R are predicted according to the first state prediction model. (FIG. 5 / STEP 120). Thereby, for example, the ZMP (Zero Moment Point) of the robot R at the time when the robot R moves from the predicted landing position one step ahead by the first designated step number q ₁ according to the first action candidate a _i1 is in the first future state. Predicted as s _i1 .

Then, for each of the plurality of first action candidates a _i1 , the first total fitness f ₁ (a _i1 ) is evaluated according to the following procedure (FIG. 5 / STEP 130).

First, based on the first future state s _i1 , a first fitness e ₁ (a _i1 ) representing suitability in view of the first action purpose is evaluated. The first fitness e ₁ (a _i1 ) is, for example, the deviation zmp _bias (a _i1 ) of the predicted ZMP from the target ZMP and the direction from the target ZMP toward the predicted ZMP shown in FIG. Based on the target ZMP and the edge interval L _{foot of the foot} B5, the calculation is performed according to the evaluation formula (111) whose change characteristics are shown in FIG. The position of the target ZMP corresponds to a reference point defined in the state space corresponding to the first purpose. The deviation zmp _{bias of the} predicted ZMP from the target ZMP corresponds to the degree of deviation from the reference point. For example, the center position of the landing (for example, rectangular shape) of the foot B5 is defined as the target ZMP.

e ₁ (a _i1 ) = 1 (when zmp _bias (a _i1 ) ≦ αL _foot (0 <α <1)),
exp (− ((zmp _bias (a _i1 ) −αL _foot ) / L _foot ) ² ) (in the case of αL _foot <zmp _bias (a _i1 )) (111).

As is apparent from the evaluation formula (111), when the interval between the predicted ZMP of the robot R and the target ZMP is in the range of the first threshold value αL _foot or less, that is, when the posture of the robot R becomes stable with a certain degree of accuracy. In the case of the predicted first action candidate a _i1 , the first fitness e ₁ (a _i1 ) is evaluated to a constant value (= 1). On the other hand, as the predicted ZMP of the robot R exceeds the first threshold αL _foot and is farther from the target ZMP, that is, the first action candidate a _i1 is more likely to be unstable, the first match is made. The degree e ₁ (a _i1 ) is evaluated low.

Note that the first fitness e ₁ (a _i1 ) may be calculated according to the evaluation formula (112) whose change characteristics are shown in FIG.

e ₁ (a _i1 ) = exp (− (zmp _bias (a _i1 ) / L _foot ) ² ) (112).

As is clear from the evaluation formula (112), the closer the predicted ZMP of the robot R is to the target ZMP, that is, the first action candidate a _i1 that is more likely to be stable in the posture of the robot R, the first fitness level. e ₁ (a _i1 ) is highly evaluated.

Further, for each of the plurality of first action candidates a _i1 , the second fitness e ₂ (a _i1 ) is estimated based on the second action guide received from the second module mod 2 (see arrow D _2-in FIG. 3). The value (second estimated fitness) e ₂ ^ (a _i1 ) is evaluated. The second estimated fitness e ₂ ^ (a _i1 ) is, for example, the interval pos _err (a _i1 ; k ₁ ) between the landing position pos (a _i1 ; k ₁ ) of the robot R and the local route as the second action guideline. ) And the stride (maximum stride) L _{step of} the robot R, it is calculated according to the evaluation formula (120).

e ₂ ^ (a _i1 ) = Π _{k1 = 1 to q1} exp (−pos _err (a _i1 ; k ₁ ) / L _step ) (120).

As is apparent from the evaluation formula (12), it is predicted that the landing position pos (a _i1 ; k ₁ ) of the robot R is closer to the local route, that is, the first predicted that the robot R can avoid contact with the object. The second estimated fitness e ₂ ^ (a _i1 ) is evaluated higher as the action candidate a _i1 .

When the second action purpose is “to make the robot R avoid contact with the object while maintaining the relative attitude of the robot R with respect to the object”, the first action candidate a _i1 The smaller the deviation (relative attitude deviation) between the relative posture of the robot R determined by the gait relative to the local route (for example, specified by the relative orientation of the base body B0 with respect to the local route) and the target relative posture is smaller, the second estimated fitness. e ₂ ^ (a _i1 ) may be defined so as to be highly evaluated (see evaluation formula (120)). Such a second action purpose is to allow the human being as an object to visually recognize the posture of the robot R and to predict the behavior of the robot R based on this posture. This is meaningful from the viewpoint of avoiding contact.

Then, based on the first fitness e ₁ (a _i1 ) and the second estimated fitness e ₂ ^ (a _i1 ), the first overall fitness f ₁ (a _i1 ) according to the evaluation formula (141) or (142). Is evaluated.

f ₁ (a _i1 ) = e ₁ (a _i1 ) (e ₂ ^ (a _i1 ) +1) (141).

f ₁ = w ₁ e ₁ (w ₂ e ₂ ^ + 1) (142).

Here, “w ₂ ” and “w ₁ (= (w ₂ +1) ⁻¹ )” are weighting factors.

The evaluation result of the first comprehensive fitness f ₁ (a _i1 ) is a “first search strategy” having convergence and diversity for searching each of the first action candidate a _i1 and the second action candidate a _i2. Defined. The first search strategy includes “first action reservation” as the first action candidate a _i1 having the highest evaluation of the first total fitness f ₁ (a _i1 ). The first action reservation and the first action guideline may be common.

Evaluation formula (141) and the contribution to As is clear from (142) the first overall fitness f ₁ (a _i1), the first fit e ₁ (a _i1) towards the second estimated relevance e ₂ ^ Higher than (a _i1 ). For example, when the first fitness e ₁ (a _i1 ) approaches 0, the first overall fitness f ₁ (a _i1 ) also becomes 0 regardless of the level of the second estimated fitness e ₂ ^ (a _i1 ). Get closer. On the other hand, even if the second estimated fitness e ₂ ^ (a _i1 ) approaches 0, the first overall fitness f ₁ (a _i1 ) approaches not the 0 but the first fitness e ₁ (a _i1 ), height of the first fit e ₁ (a _i1) is to be directly reflected in the first overall fitness f ₁ (a _i1).

Here, the second estimated relevance e ₂ ^ is a second fit e ₂ (a _i1) rather than its estimate (a _i1) is used, the second module mod2 (than the first module mod1 1 The first search strategy is defined based not only on the second behavioral objective of the lower-level module) but also on the third behavioral objective of the third module mod3 (module lower in the second rank than the first module mod1). This is to make it happen.

That is, as described above, the suitability of each of the first action candidates a _i1 for the second action purpose is evaluated using the local route as the second action guideline as a reference (see evaluation formula (120)). However, as will be described later, the second action guideline is searched by the second module mod2 in a form reflecting the third action guideline. Therefore, by calculating the second estimated fitness e ₂ ^ (a _i1 ), the fitness of the first action candidate a _i1 for the third action purpose is indirectly evaluated. Then, by evaluating the first overall fitness f ₁ (a _i1 ) based on the second estimated fitness e ₂ ^ (a _i1 ) (see the evaluation formulas (141) and (142)), the evaluation result is In the first search strategy, not only the second action purpose is directly reflected, but also the third action purpose can be reflected indirectly.

(Second action search process)
First, based on the current state, the latest future state of the robot R is predicted according to a behavior prediction model representing the behavior of the robot R (FIG. 5 / STEP 200). Thereby, for example, the landing position when the robot R advances one step from the current position is predicted. Note that the latest future state of the robot R predicted by the first module mod1 may be output from the first module mod1 to the second module mod2 (see FIG. 3 / arrow D2 ₊ ).

Further, a plurality of second action candidates {a _i2 | i ₂ = 1, 2,..., N ₂ } of the robot R are searched according to the first search strategy, the second search strategy, and the third search strategy (FIG. 5). / STEP 210). Specifically, the second action is performed according to the individual generation change method according to the evaluation result of the first overall fitness f ₁ , the evaluation result of the second overall fitness f ₂ , and the evaluation result of the third overall fitness f _3. Candidate a _i2 is searched. This search method will be described later.

Accordingly, for example, as shown in FIG. 7A, the landing position pos (a _i2 ; k) of the robot R that determines the gait over the second designated step count q ₂ (for example, q ₂ = 3 <q ₁ ). ₂ ) A local route including (k ₂ = 1 to q ₂ ) is searched as the second action candidate a _i2 . For example, the landing position one step ahead of the robot R predicted as described above is set as the starting point of the local route. The predicted landing position of the robot R at the time when the gait over the first designated number of steps q ₁ as the first action reservation (see arrow D _{2+ in} FIG. 3) starts from the landing position one step ahead of the robot R. May be the starting point of the local route. The “local route” means a route that is determined based on the position of an object as an external state and the like for the robot R to avoid contact with the object.

Further, based on each of the plurality of second action candidates a _i2 , a plurality of second future states {s _i2 | i ₂ = 1, 2,..., N ₂ } of the robot R are predicted according to the second state prediction model. (FIG. 5 / STEP 220). Thereby, for example, when the robot R moves according to the local route as the second action candidate a _i2 , the shortest distance C (a _i2 ) between the robot R and the object is predicted as the second future state s _i2 ( FIG. 7 (a)).

Then, for each of the plurality of second action candidates a _i2 , the second total fitness f ₂ (a _i2 ) is evaluated according to the following procedure (FIG. 5 / STEP 230).

First, based on the second future state s _i2 , a second fitness e ₂ (a _i2 ) representing suitability in view of the second action purpose is evaluated. The second fitness e ₂ (a _i2) is, for example, based on the shortest distance between the topical route and the object shown in Figure _{7 (a) C (a i2} ), and the size R _object of the object, FIG. 7 It is calculated according to the evaluation formula (221) whose change characteristics are shown in (b). A local path (line segment) that is infinitely or very far from the center position of the object corresponds to a reference point defined in the state space according to the second purpose. The reciprocal C ⁻¹ of the shortest distance between the local route and the center position of the object corresponds to the degree of deviation from the reference point. For example, when the contour of the object is approximated to a circle on a plane, the radius of the circle is defined as the size R _object of the _object .

e ₂ (a _i2 ) = 1 (when C ⁻¹ (a _i2 ) ≦ βR _object ⁻¹ (0 <β <1)),
exp (− (R _object (C ⁻¹ (a _i2 ) −βR _object ⁻¹ )) ² ) (in the case of βR _object ⁻¹ <C ⁻¹ (a _i2 )) (221).

As is apparent from the evaluation formula (221), when the reciprocal C ⁻¹ of the minimum distance between the local path of the robot R and the object is in the range of the second threshold value βR _object ⁻¹ or less, that is, the robot R exceeds a certain level. If the second action candidate a _i2 is predicted to be able to avoid contact with an object with accuracy, the second fitness e ₂ (a _i2 ) is evaluated to a constant value (= 1). On the other hand, as the reciprocal C ⁻¹ of the minimum distance between the local path of the robot R and the object exceeds the second threshold value βR _object ⁻¹ , that is, the second action candidate that is likely to contact the object with the robot R. a higher is _i2 second fit e ₂ (a _i2) is evaluated lower.

Note that the second fitness e ₂ (a _i2 ) may be calculated according to the evaluation formula (222) whose change characteristic is shown in FIG.

e ₂ (a _i2 ) = exp (− (R _object / C (a _i2 )) ² ) (222).

As is clear from the evaluation formula (222), the smaller the reciprocal C ⁻¹ of the minimum distance between the local path of the robot R and the object, that is, the second action that the robot R is more likely to avoid contact with the object. The second fitness e ₂ (a _i2 ) is evaluated higher as the candidate a _i2 .

When the second action purpose is “to make the robot R avoid contact with the object while maintaining the relative posture of the robot R with respect to the object”, the second action candidate a _i2 The smaller the deviation (relative posture deviation) between the relative posture of the robot with respect to the posture of the local route (for example, specified by the relative orientation of the base body B0 with respect to the local route) and the target relative posture, the second fitness e ₂ (a _i2 ) may be defined so as to be highly evaluated (see evaluation formulas (221) and (222)).

In addition, for each of the plurality of second action candidates a _i2 , the third fitness e ₃ (a _i2 ) is estimated based on the third action guideline received from the third module mod3 (see FIG. 3 / arrow D _3- ). The value (third estimated fitness) e ₃ ^ (a _i2 ) is evaluated. The third estimated fitness e ₃ ^ (a _i2 ) is, for example, the interval pos _err (a _i2 ; k ₂ ) between the landing position pos (a _i2 ; k ₂ ) of the robot R and the global route as the third action guideline. ) And the stride (maximum stride) L _{step of} the robot R, it is calculated according to the evaluation formula (230).

e ₃ ^ (a _i2 ) = Π _{k2 =} 2˜q2 exp (−pos _err (a _i2 ; k ₂ ) / L _step ) (230).

As is clear from the evaluation formula (230), it is predicted that the landing position pos (a _i2 ; k ₂ ) of the robot R is closer to the global path, that is, the robot R moves according to the global path for reaching the target position. The third estimated fitness e ₃ ^ (a _i2 ) is evaluated higher as the second action candidate a _i2 to be performed.

When the third action purpose is “to make the robot R reach the target position in the target posture”, the posture of the robot R determined according to the local route as the second action candidate a _i2 (for example, the base B0) It may be defined so that the third estimated fitness e ₃ ^ (a _i2 ) is evaluated higher as the deviation (posture deviation) between the position and the global path posture is smaller (evaluation). (Refer Formula (230)). Such a third action purpose is meaningful from the viewpoint of causing the robot R to perform a task whose degree of difficulty changes according to its posture, such as gripping an object when the robot R arrives at or after the target position. .

Based on the first fitness e ₁ (a _i2 ) evaluated by the first module mod 1 in addition to the second fitness e ₂ (a _i2 ) and the third estimated fitness e ₃ ^ (a _i2 ), The second total fitness f ₂ (a _i2 ) is evaluated according to the evaluation formula (241) or (242).

f ₂ (a _i1 ) = e ₂ (a _i2 ) (e ₁ (a _i2 ) (e ₃ ^ (a _i2 ) +1) +1) (241).

f ₂ = w ₂ e ₂ (w ₁ e ₁ (w ₃ e ₃ ^ + 1) +1) (242).

Here, “w ₁ ” “w ₃ ” and “w ₂ (= {w ₁ (w ₃ +1) +1} ⁻¹ ” are weighting factors.

The evaluation result of the second total fitness f ₂ (a _i2 ) has convergence and diversity for searching each of the first action candidate a _i1 , the second action candidate a _i2, and the third action candidate a _i3. Defined as “second search strategy”. The second search strategy includes “second action reservation” as one second action candidate a _i2 having the highest evaluation of the second total fitness f ₂ (a _i2 ). The second action reservation and the second action guideline may be common.

As is clear from the evaluation formulas (241) and (242), the degree of contribution to the second total fitness f ₂ (a _i2 ) is higher than the first fitness e _{1 in} the second fitness e ₂ (a _i2 ). It is higher than each of (a _i2 ) and the third estimated fitness e ₃ ^ (a _i2 ). For example, when the second fitness e ₂ (a _i2 ) approaches 0, the second synthesis is independent of the first fitness e ₁ (a _i2 ) and the third estimated fitness e ₃ ^ (a _i2 ). The fitness f ₂ (a _i2 ) also approaches zero. On the other hand, even when the first fitness e ₁ (a _i2 ) and the third estimated fitness e ₃ ^ (a _i2 ) approach 0, the second total fitness f ₂ (a _i2 ) is not 0 but the second It approaches fit e ₂ (a _i2), so that the level of the second adaptability e ₂ (a _i2) is directly reflected in the second overall fitness f ₂ (a _i2).

Further, the contribution to the second total fitness f ₂ (a _i2 ) is higher in the first fitness e ₁ (a _i2 ) than in the third estimated fitness e ₃ ^ (a _i2 ). For example, when the first fitness e ₁ (a _i2 ) approaches 0, the second total fitness f ₂ (a _i2 ) decreases regardless of the level of the third estimated fitness e ₃ ^ (a _i2 ). . On the other hand, even if the third estimated fitness e ₃ ^ (a _i3 ) approaches 0, the second total fitness f ₂ (a _i2 ) approaches e ₂ (a _i2 ) · e ₁ (a _i1 ) The level of the first goodness of fit e ₁ (a _i1 ) can be reflected in the third estimated goodness of fit e ₃ ^ (a _i2 ).

Here, the third estimation fitness e ₃ ^ a third fit e ₃ (a _i2) instead estimates its (a _i2) is used, the third module mod3 (than the second module mod2 1 Not only the second action purpose possessed by the lower rank module) but also the fourth module mod4 (the module lower by the second rank than the second module mod2) (if there is something that does not exist in this embodiment) This is because the second search strategy is defined based on the fourth action purpose (for example, the purpose of moving the robot R to the target position and then moving it to the second target position).

That is, as described above, the suitability of each second action candidate a _i2 for the third action purpose is evaluated using the global route as the third action guideline as a reference (see evaluation formula (230)). However, the third action guideline can be searched by the third module mod3 in a form that reflects the fourth action guideline. For this reason, by calculating the third estimated fitness e ₃ ^ (a _i2 ), the fitness of the second action candidate a _i2 for the fourth action purpose is indirectly evaluated. Then, the second total fitness f ₁ (a _i1 ) is evaluated based on the third estimated fitness e ₃ ^ (a _i2 ) (see the evaluation formulas (241) and (242)), and as an evaluation result, In the second search strategy, not only the third action purpose is directly reflected but also the fourth action purpose can be reflected indirectly.

(Third action search process)
First, based on the current state, the most recent future state of the robot R is predicted according to a behavior prediction model representing the behavior of the robot R (FIG. 5 / STEP 300). Thereby, for example, the landing position when the robot R advances one step from the current position is predicted. Note that the latest future state of the robot R predicted by the first module mod1 or the second module mod2 may be output from the second module mod2 to the third module mod3 (see FIG. 3 / arrow _{D3 +} ).

Further, a plurality of third action candidates {a _i3 | i ₃ = 1, 2,..., N ₃ } of the robot R are searched according to the second search strategy and the third search strategy (FIG. 5 / STEP 310). Specifically, the third action candidate a _i3 is searched according to the individual generation change method according to the evaluation result of the second comprehensive fitness f _{2 and} the evaluation result of the third comprehensive fitness f ₃ . This method will be described later.

Thereby, for example, as shown in FIG. 8A, a global route that defines the gait of the robot R toward the target position is searched for as the third action candidate a _i3 . For example, the landing position one step ahead of the robot R predicted as described above is set as the starting point of the global route. Note that the end point of the local route as the second action reservation (see FIG. 3 / arrow D ₃₊ ) is the start point of the global route. The target position of the robot R may be input or transmitted from the outside of the robot R to the control system, or may be calculated by the control system 1 based on the recognition result of the external state.

Further, based on each of the plurality of third action candidates a _i1 , a plurality of third future states {s _i3 | i ₃ = 1, 2,..., N ₃ } of the robot R are predicted according to the third state prediction model. (FIG. 5 / STEP 320). Thereby, for example, the movement distance L (a _i3 ) when the robot R moves from the predicted landing position one step ahead according to the global route is predicted as the third future state s _i3 .

Then, for each of the plurality of third action candidates a _i3 , the third total fitness f ₃ (a _i3 ) is evaluated according to the following procedure (FIG. 5 / STEP 330).

First, based on the third future state s _i3 , a third fitness e ₃ (a _i3 ) representing suitability in view of the third action purpose is evaluated. The third fitness e ₃ (a _i3 ) is, for example, the predicted movement distance L (a _i3 ) of the robot R shown in FIG. 8A and the linear distance L _dest between the starting point of the global path and the target position. Based on the above, it is calculated according to the evaluation formula (331) whose change characteristic is shown in FIG. The straight line distance L _dest corresponds to a reference point defined in the state space according to the third purpose. The deviation between the predicted movement distance L of the robot R and the straight line distance L _dest corresponds to the degree of deviation from the reference point.

e ₃ (a _i3 ) = 1 (when L _dest ≦ L (a _i3 ) ≦ γL _dest (1 <γ)),
γL _dest / L (a _i3 ) (when γL _dest <L (a _i3 )) (331).

As is apparent from the evaluation formula (331), when the predicted movement distance L of the robot R is in the range equal to or smaller than the third threshold γL _dest , that is, the movement distance until the robot R reaches the target position is predicted to be somewhat short. In the case of the third action candidate a _i3 , the third fitness e ₃ (a _i3 ) is evaluated to a constant value (= 1). On the other hand, as the predicted movement distance L of the robot R increases beyond the third threshold γL _dest , that is, the movement distance until the robot R reaches the target position is likely to be longer than a certain level. the more in _i3 third goodness of fit e ₃ (a _i3) is evaluated lower.

Note that the third fitness e ₃ (a _i3 ) may be calculated according to the evaluation formula (332) whose change characteristic is shown in FIG.

e ₃ (a _i3 ) = L _dest / L (a _i3 ) (332).

As is apparent from the evaluation formula (332), the third action candidate a _i3 predicted that the movement distance L until the robot R linearly moves toward the target position, that is, the robot R reaches the target position, is short. The higher the third matching degree e ₃ (a _i3 ), the higher the evaluation.

When the third action purpose is “to make the robot R reach the target position with the target posture”, the posture of the robot R at the target position determined according to the global route as the third action candidate a _i3 (for example, It may be defined such that the third fitness e ₃ (a _i3 ) is evaluated higher as the deviation (posture deviation) between the target body posture and the target posture is smaller (evaluation formula). (Refer to (331) (332)).

Based on the second fitness e ₂ (a _i3 ) evaluated by the second module mod 2 in addition to the third fitness e ₃ (a _i3 ), the third fitness e ₃ (a _i3 ) is calculated according to the evaluation formula (341) or (342). The overall fitness f ₃ (a _i3 ) is evaluated.

f ₃ (a _i3 ) = e ₃ (a _i3 ) (e ₂ (a _i3 ) +1) (341).

f ₃ = w ₃ e ₃ (w ₂ e ₂ +1) (342).

Here, “w ₂ ” and “w ₃ (= (w ₂ +1) ⁻¹ )” are weighting factors.

The evaluation result of the third comprehensive fitness f ₃ (a _i3 ) is a “third search strategy” having convergence and diversity for searching each of the second action candidate a _i2 and the third action candidate a _i3. Defined.

As apparent from the evaluation formulas (341) and (342), the degree of contribution to the third total fitness f ₃ (a _i3 ) is greater in the third fitness e ₃ (a _i3 ) than in the second fitness e ₂ (a _i3 ). higher than a _i3 ). For example, when the third fitness e ₃ (a _i3 ) approaches 0, the third overall fitness f ₃ (a _i3 ) also approaches 0 regardless of the level of the second fitness e ₂ (a _i3 ). On the other hand, even if the second fitness e ₂ (a _i3 ) approaches 0, the third overall fitness f ₃ (a _i3 ) approaches not the 0 but the third fitness e ₃ (a _i3 ), and the third fitness The level of the degree e ₃ (a _i3 ) is directly reflected in the third total fitness f ₃ (a _i3 ).

(Search method for action candidates according to the search strategy)
A method for searching for action candidates will be described. As a search method, a method having “convergence” and “diversity” is employed. By a method having “convergence”, an action candidate whose evaluation is higher in the vicinity of an action candidate having a high previous evaluation in the search range (meaning a range in a space defined by the action candidate) is obtained. Can be searched. By this method having “diversity”, action candidates whose evaluation is higher this time can be searched in a wide range without being overly concerned with the previous evaluation.

  An evolutionary search method can be adopted as a search method for action candidates having convergence and diversity.

  Specifically, first, among a plurality of parent individuals shown in FIG. 9A (corresponding to action candidate a. Expressed by “●”), the parent surrounded by a circle. An individual is selected. The selection probability is higher as the fitness f (a) (evaluated according to the evaluation curve indicated by the solid line) is higher.

  As shown in FIG. 9B, the parent individual having a higher fitness f (a) has a relatively small noise in the parent individual (the noise becomes smaller as the fitness f (a) is higher). As a result of the addition, child individuals (represented by “◯”) are generated so as to be distributed with a high probability in a relatively narrow range around the parent individual. On the other hand, a parent individual having a lower fitness f (a) has a relatively large noise added to the parent individual (noise increases as the fitness f (a) is lower). It is generated so as to be distributed with a high probability in a relatively wide range around the. In addition, although not shown, individuals are generated irregularly. The occupancy rate of irregular individuals with respect to all individuals is set to about 0.30, for example.

  By adopting the generation change method as described above, the individual concentrates in a region where the fitness f (a) is large as the generation progresses. As a result, the action candidate a having high suitability for the purpose of which the module is mainly responsible is obtained adaptively.

Further, in order to explain the diversity of search methods, two modules A and B are considered. The module A is configured to evaluate the fitness f _A (a) indicating the suitability of the action candidate a in view of the purpose, set a search strategy according to the evaluation result, and output it to the module B. . The module B is configured to receive a search strategy from the module A and search for an action candidate a having a high fitness f _B (a) indicating suitability in view of the purpose in accordance with the search strategy.

First, let us consider a case where “one action candidate a ₁ ” whose fitness f _A (a) has the maximum value is set as the search strategy by the module A as shown in the upper side of FIG. In this case, for example, FIG. 10 (a) starting from the behavior candidate a ₁ as search strategy by module B as shown in the lower, fitness f _B (a) behavioral candidate a ₂ showing the maximum value _- is searched. However, this solution may not be an optimal solution with the highest fitness f _B (a), but only a local solution.

Therefore, as shown in the upper side of FIG. 10B, “peripheral range of _one action candidate a ₁ ” in which the fitness f _A (a) shows the maximum value (or maximum value) by the module A is used as the search strategy. Is set. In response to this, for example, as shown in the lower side of FIG. 10B, the action candidate a ₂₊ (≠ that the fitness f _B (a) becomes the highest in the range as the search strategy by the module B. a _2- ) is searched. As a result of expanding the search range in this way, the possibility of obtaining an optimal solution can be increased.

Further, consider the case where “one action candidate a ₁ ” having the highest fitness f _A (a) is set as the search strategy by module A as shown in the upper side of FIG. In this case, for example, as shown in the lower side of FIG. 11A, the fitness f _B (a) has a maximum value in a single range around one action candidate a ₁ as a search strategy by module B. The candidate action a ₂₋ shown is searched. However, this solution may not be an optimal solution with the highest fitness f _B (a), but only a local solution.

Therefore, as shown in the upper side of FIG. 11B, module A sets “plural behavior candidates a ₁₋ and a ₁₊ ” whose fitness f _A (a) has a maximum value as a search strategy. . In response to this, for example, as shown in the lower side of FIG. 11B, the fitness f _B (in the surrounding range of the plurality of action candidates a ₁₋ and a ₁₊ as the search strategy by the module B A candidate for action a ₂₊ (≠ a ₂₋ ) that maximizes a) is searched. As a result of expanding the search range in this way, the possibility of obtaining an optimal solution can be increased.

(Example)
The behavior state of the robot R expressed by the ZMP and the gravity center position (represented by “◎”) of the robot R shown below each of FIGS. 12 (a) to 12 (c) are modeled on the behavior state of the inverted pendulum expressed by the landing position and the gravity center position (represented by “◎”) of the inverted pendulum shown on the upper side of each Can think.

Therefore, the experiment was performed by moving the inverted pendulum while the upper part of the inverted pendulum was suspended from above. First, as shown in FIG. 13A, the inverted pendulum starts to move from its starting position (time t = 0), and then disturbance (force received by contact with an object) is applied to the inverted pendulum. (Time t = t ₀ ). In response to this, the inverted pendulum moves along a new path while stabilizing (time t = t ₁ ), further moves while avoiding contact with the object (time t = t ₂ ), and reaches the target position. Arrives (time t = t ₃ ).

At this time, as shown in FIG. 13B, calculation for searching for a discrete landing position as a first action candidate using an individual filter was executed. The individual is represented by individual points constituting a cloud-like lump above the ground with a grid, and the height from the ground represents the size of the first overall index f ₁ . In this case, the point where the points aligned along the ridge of the cloud are vertically projected on the ground represents the landing position as the first action reservation or the first action guideline.

Since the behavior state of the robot R is modeled as the behavior state of the inverted pendulum as described above, the behavior of the robot R can be controlled using the model. As a result, as shown in FIG. 14, the robot R can be operated in the same manner as the inverted pendulum. That is, first, the robot R starts to move from its starting position (time t = 0), and then a disturbance is given to the robot R (time t = t ₀ ). In response to this, the robot R moves along a new path while stabilizing (time t = t ₁ ), further moves while avoiding contact with the object (time t = t ₂ ), and reaches the target position. (Time t = t ₃ ).

(Operational effect of the robot R of the present invention and its control system)
According to the robot R that performs the above function, while giving priority to the main purpose of the module itself (i = 1, 2, 3), the sub-objects of the other modules are mainly responsible. In addition, the action candidate of the robot R to be controlled is searched (see FIG. 5 / STEP110, STEP210, and STEP310). In this way, action candidates to be controlled can be searched in a manner in which the purpose is gently shared between different modules.

  Further, a plurality of future states of the control target corresponding to the searched action candidate are predicted following the future state of the control target predicted by the upper module (at least the control module mod0) (FIG. 5 / STEP120, STEP220, STEP320). reference).

  Based on each future state, the high-order higher-level module that is mainly responsible for the high-frequency higher-order module takes precedence over the lower-frequency lower-level module that is mainly responsible for the main purpose and sub-purposes that the main person is responsible for. The degree is evaluated (see FIG. 5 / STEP 130, STEP 230, and STEP 330). As a result, the behavior of the robot R is controlled in a manner in which the higher-order object mainly handled by the high-frequency upper module is reflected in preference to the lower-purpose object mainly handled by the low-frequency lower module.

For this reason, when a disturbance occurs (refer to FIG. 4 / time t = t ₀ ), the higher-order purpose mainly handled by the higher-order module that is initially responding is reflected in the action form of the controlled object, and the influence of this disturbance is quickly affected. Reduction is achieved (see FIG. 4 / time t = t ₁ ). Furthermore, the sub-objects that are mainly handled by the subordinate modules whose response is slow over time are also reflected in the action form of the control object, and the action form of the control object that is suitable for all of the plurality of objectives can be recovered (FIG. 4 / reference time _{t = t 2, t = t} 3). Thereby, the robot R can behave in an appropriate form in view of the purpose of the action with respect to an arbitrary form of disturbance.

In particular, of all the j-th action candidates searched for by the j-th module (j = 1, 2, 3 in the above embodiment), the degree of deviation from the reference point defined according to the j-th purpose in the state space There the j behavior candidate a _jk is within the j threshold, i.e., for the j behavior candidate a _jk included in the j-specified behavior candidate group, the j-th fitness e _j evaluates to a constant value "1" (Refer to said Formula (111) (221) (331), FIG.6 (b), FIG.7 (b), FIG.8 (b)). This constant value is not limited to “1”, and an arbitrary value can be adopted. Meanwhile, for the j-th action candidate a _jk not included in the j-th specified behavior candidate group, the j fit e _j evaluates to a value lower than a predetermined value. That is, of the set can be taken as an action candidates in the set of actions candidates j-th fitness e _j takes a sufficiently high value (j-th solution space), the j-th fitness e _j is evaluated in the constant value a region (flat region), the j-th fitness e _j is a region lower than the predetermined value is formed. This configuration is based on the idea that the j-th action candidate a _jk that fits the j-th purpose to some extent may consider the j-th matching degree to be approximately equivalent.

For the first solution space defined by the ZMP _bias that defines the first action candidate, as shown in FIG. 6B, the range of 0 ≦ ZMP _bias ≦ αL _foot is a constant value of the first fitness e _1. The range of αL _foot <ZMP _bias is defined as a slope region where the first fitness e ₁ gradually decreases from the constant value “1”.

For the second space defined by C ⁻¹ that defines the second action candidate, the range of 0 ≦ C ⁻¹ ≦ β (R _object ) ⁻¹ is the second fit as shown in FIG. 7B. degrees e ₂ is defined as a flat area is a constant value "1", beta slope (R _object) ^-1 <C range ^-1 second fit e ₂ gradually decreases from the constant value "1" It is defined as an area.

For the third space defined by L defining the third action candidate, as shown in FIG. 8 (b), the range of L _dest ≦ L ≦ γL _dest is the constant value “1” for the third fitness e _3. ”And the range of γL _dest <L is defined as a slope region where the third fitness e ₃ gradually decreases from the constant value“ 1 ”.

  By forming the flat area, an area in which a solution having the maximum or maximum total fitness can be searched in the j-th solution space is expanded. For this reason, the upper module receives the lower action candidate having the maximum or maximum lower overall fitness from the lower module as the lower action guideline, and estimates the degree of approximation with the lower action guideline for each action candidate searched by itself. In the solution space, a region where a solution can be searched for by the higher-level module is expanded (see FIGS. 10 and 11).

Further, the plurality of first j behavior candidate included in the flat region in the j-th solution space, the superiority in accordance with the height of the j-fit e _j may not be considered. For this reason, searching for a solution that increases the j-th matching degree more than necessary, and thus, excessive optimization of the operation control of the controlled object is prevented. Further, the j-th module has a maximum or maximum total fitness according to the fitness in consideration of the sub-objects in which the other modules (one or both of the j-1 module and the j + 1 module) are mainly responsible. It is sufficient to search for j action candidates. In other words, the j-th module can be concentrated on searching for a solution that increases the degree of fitness in consideration of the sub-object, not the main object.

Thus, reduction in processing load of each module modj for evaluation of the j fit e _j, therefore, while achieving earlier stage of calculation of the calculation result, the j overall fitness f _j is evaluated or calculated (See the above formulas (141), (142), (241), (242), (341), and (342)). Thereby, this robot R can be made to respond or act quickly with respect to the disturbance of arbitrary forms in an appropriate form in view of the action purpose of the robot R as a control object.

(Other embodiments of the present invention)
In addition to the robot R (see FIG. 1), the control target may be any device equipped with an actuator, such as a vehicle.

  In the embodiment, the number of behavior search modules is three, but may be two, or may be four or more.

When only the first module mod1 and the second module mod2 are included in the control system 1, the second total fitness f ₂ (a _i2 ) is the second fitness e ₂ (a _i2 ) and the first fitness e ₁ (a Based on _i2 ), the evaluation is made according to the evaluation formula (244).

f ₂ (a _i1 ) = e ₂ (a _i2 ) (e ₁ (a _i2 ) +1) (244).

In addition, when the fourth module mod4 (having a longer calculation cycle) than the third module mod3 is further included as a component of the control system 1, the third total fitness f ₃ (a _i3 ) is Evaluation is performed according to an evaluation formula (343) similar to the evaluation formula (241) of the fitness f ₂ (a _i2 ).

f ₃ (a _i3 ) = e ₃ (a _i3 ) (e ₂ (a _i3 ) (e ₄ ^ (a _i3 ) +1) +1) (343).

In this case, the fourth total fitness f ₄ (a _i4 ) is evaluated by the fourth module mod 4 according to the same evaluation formula as the evaluation formula (34) of the third total fitness f ₃ (a _i3 ).

Further, when four behavior search modules are included as components of the control system 1, the second fitness e ₂ (a _i2 ), the first fitness e ₁ (a _i2 ), and the third estimated fitness e ₃ ^ in addition to (a _i2), further based on the fourth estimated relevance e ₄ ^ (a _i2), second overall fitness f ₂ (a _i2) may also be evaluated in accordance with equation (244).

f ₂ (a _i2 ) = e ₂ (a _i2 )
× (e ₁ (a _i2 ) (e ₃ ^ (a _i2 ) (e ₄ ^ (a _i2 ) +1) +1) +1) (244).

  Even when five or more action search modules are included in the control system 1 as constituent elements, the total fitness of the intermediate modules can be calculated according to the same concept.

DESCRIPTION OF SYMBOLS 1 ... Control system, 2 ... Actuator, mod1, 1st module, mod2 ... 2nd module, mod3 ... 3rd module, R ... Robot (control object).

Claims (11)

It has multiple modules that are layered according to the relative length of the calculation cycle ,
Each of the modules searches for a plurality of action candidates that are candidates for the action form to be controlled, and each module has a contribution degree of the main fitness for the main purpose that the main person is responsible for for each of the action candidates. Configured to evaluate the overall fitness in view of both the primary objective and the secondary objective so as to be higher than the contribution of the secondary fitness to the secondary objective;
The evaluation result of the control target is reflected in a form that preferentially reflects the evaluation result of the comprehensive fitness by the higher module having a relatively short calculation cycle rather than the evaluation result of the comprehensive fitness by the lower module having a relatively long calculation cycle. A control system configured to control operation. The control system according to claim 1,
The intermediate module excluding the highest module with the shortest calculation cycle and the lowest module with the longest calculation cycle is the higher-level objective mainly handled by the higher module whose calculation cycle is relatively shorter than itself, and the calculation cycle is lower than the self. For each of the action candidates, the higher fitness level in view of the higher-level goal contributes more than the lower-level fitness level in consideration of the lower-level goal, with each of the lower-level goals mainly handled by a relatively long lower-level module as the sub-purpose. A control system configured to evaluate the total fitness so as to increase the degree. The control system according to claim 2, wherein
The module in which a lower module having a relatively longer calculation cycle than the self evaluates the estimated lower fitness by estimating the lower fitness for each of the action candidates, and determines the estimated lower fitness as the lower A control system configured to evaluate the overall fitness as fitness. The control system according to claim 3, wherein
The module in which a lower module having a relatively longer calculation cycle than the self exists as the overall fitness evaluated by the lower module among the lower action candidates as the action candidates searched by the lower module. The lower action candidate having the maximum or maximum lower overall fitness is received from the lower module as a lower action guideline, and the degree of approximation with the lower action guideline is evaluated as the estimated lower suitability for each of the action candidates searched by itself. A control system characterized by being configured to do so. The control system according to claim 3 or 4,
The first to Nth modules as the plurality of modules,
The first module as the highest module evaluates the first suitability e ₁ for the first purpose as the main purpose for each of the first action candidates as the action candidates searched by itself, After evaluating the second estimated fitness e ₂ ^ by estimating the fitness of the second module for the second purpose, the first overall fitness according to the evaluation formula f ₁ = e ₁ (e ₂ ^ + 1) configured to evaluate f ₁ ,
The i-th suitability in consideration of the i-th purpose as the main purpose for each of the i-th behavior candidates as the behavior candidates searched by the i-th module (i = 2 to N-1) as the intermediate module. evaluating e _i , evaluating the i−1th fitness e _i−1 in view of the i−1th objective as the higher purpose, and estimating the fitness in view of the i + 1th objective as the lower purpose. After evaluating the i + 1-th estimated fitness level e _{i + 1} ^, the _i- th total fitness level f _i is calculated according to the evaluation formula f _i = e _i (e _i-1 (e _{i + 1} ^ + 1) +1). Configured to evaluate,
The Nth module as the lowest module evaluates the Nth adaptability e _N for the Nth objective as the main objective for each of the Nth action candidates as the behavioral candidates searched for by itself, After evaluating the N-1 matching degree e _N-1 of the N-1th module, the i-th total matching degree f _i is evaluated according to the evaluation formula f _N = e _N (e _N-1 +1). A control system configured as described above. The control system according to claim 5, wherein
The j-th module (j = 1, 2,..., N) determines the main behavior of the j-th behavior candidate included in the j-th designated behavior candidate group among all the j-th behavior candidates as the behavior candidates searched by itself. while evaluating the j fit e _j for the j object of purposes to a constant value, said of all of the j behavior candidate, for the j action candidates are not included in the j-th specified behavior candidate group, the second control system characterized by being configured to the j adaptability e _j ratings lower than the predetermined value. The control system according to claim 6, wherein
The j-th module selects a j-th action candidate having a deviation degree from a reference point defined in a state space corresponding to the j-th purpose among the all j-th action candidates, being a j-th threshold value or less. as a j action candidates included in the j specified action candidates, while evaluating the j-th fitness e _j to the predetermined value, said of all of the j behavior candidate, included in the j-th designated action candidates for no j-th behavior candidate, control system, characterized in that the basis of the deviation degree, is adapted to evaluate the first j adaptability e _j according continuously or stepwise decreasing function of the degree of deviation . In the control system according to any one of claims 1 to 7,
Each of the modules predicts a plurality of future states of the control target according to each of the behavior candidates searched for by itself, and evaluates the overall fitness for each of the behavior candidates based on each of the future states. A control system characterized by being configured to do so. In the control system according to any one of claims 1 to 8,
Each of the modules is configured to search for the action candidate that determines the position or position and posture of the mobile device as the control target,
The higher module having a relatively short calculation cycle is configured to search for the action candidate over a short period of the mobile device as compared to a lower module having a relatively long calculation cycle. Control system. The control system according to claim 9, wherein
A first module configured to search, as the action candidate, a position trajectory or a position trajectory and a posture trajectory of a robot as the moving device having a base and a plurality of legs extending from the base; 2 modules and a third module as the plurality of modules,
The first module is configured to search for a gait over a first designated number of steps of the robot for stabilizing the posture of the robot as a first action candidate;
The second module is configured to search as a second action candidate a local route that defines a gait over a second specified number of steps greater than the first specified number of steps for causing the robot to avoid contact with an object;
The control system, wherein the third module is configured to search as a third action candidate a global route for causing the robot to reach a target position. A robot as the moving device that includes a base body and a plurality of legs extending from the base body and moves by an operation of the plurality of legs;
A robot comprising the control system according to claim 10.

Images