JP2002346957A - Movement trajectory control method for bipedal walking robot - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a movement trajectory control method for a bipedal walking robot that can generate a pattern in the middle of walking, generate a real-time walking pattern and realize follow-up motion after a human by clarifying a parameter with a small effect on walking in compensatory trajectory generation and remarkably shortening compensatory trajectory generating time. SOLUTION: The movement trajectory control method for a bipedal walking robot computes a moment-compensatory trajectory of each joint for a trunk region and a waist region separately, decides moment-compensatory trajectory computing algorithm considering cooperation between the trunk region and the waist region, and uses it to prepare a walking pattern on a real-time basis in the middle of walking.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for controlling a motion trajectory of a bipedal walking robot, and in particular, to set a parameter having little effect on walking in generating a compensation trajectory of the robot, thereby greatly shortening a motion compensation trajectory generation time. The present invention relates to a motion trajectory control method for a bipedal walking robot that enables generation of a pattern during walking and generates a real-time walking pattern.

[0002]

2. Description of the Related Art A bipedal walking robot, especially a humano robot, is mainly located near a human living environment. A very flexible adaptation of the decision is required.

According to recent research and development of this type of robot, a motion applied to such an external environment has been studied in the form of a motion following a human using a six-axis force sensor. In this method, information from the six-axis force sensor is used as external environment information, and unit patterns created in advance are sequentially output in accordance with the external environment information, thereby adapting to the environment, that is, following the environment.

[0004] As another example, a simple conversation in Japanese, recognition of an object by two eyes and a walking pattern, measurement of distance, grasping and moving of an object with both hands having tactile movement and biped walking. A humanoid intelligent robot capable of performing an electronic organ has been developed.Furthermore, a keyboard instrument capable of focusing on artistic activities as a place for robot activities, recognizing music scores with eyes, and playing an electronic organ with both hands and feet. Performance robots have also been developed. This robot can perform simple conversational exchanges with humans, or recognize the pitch of a human singing voice, and even perform an accompaniment of an organ according to the pitch.

[0005]

A human being adapted to adapt to and follow the environment by sequentially outputting unit patterns prepared in advance in accordance with the information from the above-described six-axis force sensor as external environment information. Although robots are revolutionary in adapting to the external environment, they can only move in a predetermined pattern, and a huge amount of unit patterns are required to realize various movements . In addition, there is a problem of restriction on the memory capacity, and there is also a problem that it takes too much time to generate a compensation trajectory.

Therefore, the present invention clarifies parameters that have little effect on walking in generating a compensated trajectory, and can greatly reduce the time required to generate a compensated trajectory, thereby enabling pattern generation during walking and realizing a motion following a person. An object of the present invention is to provide a motion trajectory control method for a bipedal walking robot that can be performed.

[0007]

According to the present invention, a humanoid bipedal walking robot recognizes walking pattern change information from a visual sensor and / or a voice input sensor, and follows the walking pattern change information in accordance with the walking pattern change information. And a motion trajectory control method for a bipedal walking robot, which is capable of changing in real time.

Further, according to the present invention, a walking pattern is created in real time during walking by calculating a trajectory by dividing a moment compensation amount into each part of the body by using a moment compensation trajectory calculation algorithm by coordination of each body part. 2
A motion trajectory control method for a walking robot is provided.

According to the present invention, there is provided a motion trajectory control method for a bipedal walking robot, in which a computer is separately used for controlling each body part and generating a pattern when creating the moment compensation trajectory calculation algorithm.

Further, according to the present invention, the robot main body is provided with a visual sensor and / or a voice input sensor based on a microphone sound used for designating a toe trajectory necessary for creating a walking pattern, and stereo vision or monocular is provided by the visual sensor. To determine a stride for generating a toe trajectory, the voice input sensor reacts only to a preset command, and selects a preset stride in association with a word by a voice input word. And a method for controlling the motion trajectory of a bipedal walking robot.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, a preferred embodiment of the present invention will be described with reference to the drawings.

First, a description will be given of a method of preparing the waist / trunk cooperative moment compensation trajectory calculation algorithm according to the present invention. Here, the safety of the biped humanoid robot during the whole body movement (safety of the posture state)
Is determined from a mechanical viewpoint using a ZMP stability determination criterion. Hereinafter, the ZMP and the ZMP stability determination standard will be described.

Consider the conditions necessary for a biped humanoid robot to maintain a stable support state during whole body movement. The fact that the biped humanoid robot cannot maintain a stable support state, that is, falls down, means that the support polygon (the area formed by the road surface and the ground contact point formed by the contact point (3 or more points) between the sole and the road surface) It is considered that the rotational movement is about a straight line on the road surface passing through a certain side or point of the largest convex polygon). That is, it is considered that an outward moment acts as viewed from the supporting polygon. Conversely,
As shown in FIG. 1, if no outward moment is generated around all sides and points on the supporting polygon A and only an inward moment M is generated, the biped humanoid robot is in a stable supporting state. Can be maintained. F is the resultant of the forces acting on each point.

Considering the state of support of the biped humanoid robot at this time, since the grounding points on the sole do not float, the force due to gravity and inertia force acting on the road surface from the biped humanoid robot at all the grounding points will cause the ground surface to fall. In the pushing direction, the point P where these resultant forces act must be within the supporting polygon A of the sole. The moment due to the resultant force around this point P is clearly zero, and this point is
(Zero moment point). In other words, ZMP is a point at which the moment due to gravity and inertial force of each part of the biped humanoid robot becomes zero on the road surface. Also, if at any moment during the movement, the ZMP exists in the supporting polygon and the force acting on the road surface from the biped humanoid robot is in the direction of pushing the road surface, then 2
A foot humanoid robot can continue to exercise stably. " This is referred to as a ZMP stability determination standard.

In the moment compensation trajectory calculation algorithm, this ZMP stability determination standard is used for the stability determination of the walking pattern in the pitch axis and roll axis directions on the ZMP during the whole body movement of the biped humanoid robot. Also,
As in the conventional case, the moment about the yaw axis on the ZMP is also compensated in order to stabilize the movement of the biped humanoid robot around the yaw axis.

Next, the moment compensation trajectory calculation algorithm will be described.

The outline of the present algorithm consists of the following five points.

(A) Modeling of a humanoid robot (b) Derivation of the ZMP equation in the model (a) (c) Approximate modeling of a humanoid robot (d) Distributing the compensated moment in the approximate model to the waist and trunk (e) Calculation of exact solution of moment compensation trajectory that satisfies the above b by iterative calculation using approximation model of c above (1) Modeling of humanoid robot Figure 2 shows the coordinate system of each mass point of the humanoid robot and the force acting on it FIG. 5 is a diagram showing a vector display. First, the following conditions are set for the humanoid robot, the coordinate system, and the road surface.

(A) A humanoid robot is composed of a set of mass points (FIG. 2).

(B) The road surface is sufficiently hard and does not move or displace even if any force or moment acts.

(C) A plane including the X axis (coincident with the front direction of the humanoid robot) and the Y axis coincides with the road surface,
An orthogonal coordinate system O (right-handed system) having an axis perpendicular to the road surface as the Z axis is set as shown in FIG.

(D) The contact state between the sole of the humanoid robot and the road surface is a set of point contact points.

(E) At the contact point between the walking system and the road surface,
The friction coefficient for rotation on the road surface (around the X, Y and Z axes) is ignored.

(F) The propulsive force of the humanoid robot when walking is within a range in which translational (X and Y axis) sliding does not occur at the contact point.

(2) Calculation of ZMP Equation When a moment balance equation for an arbitrary point P in the absolute coordinate system (O-XYZ) is derived based on the above assumptions and settings, the following equation is obtained.

[0026]

(Equation 1)

In equation (2.1), by setting the point P as ZMP, the following ZMP equation is obtained from T = 0.

[0028]

(Equation 2)

Further, in order to consider the relative movement of each part,
A motion coordinate system (O'-X'Y'Z ') fixed near the waist of the biped humanoid robot shown in FIG. 2 is set. When the ZMP equation is expressed using this coordinate system, the following equation is obtained.

[0030]

(Equation 3)

The trajectories of the trunk and waist are calculated so as to satisfy the ZMP equation.

(3) Approximate Modeling of Humanoid Robot In this algorithm, an approximate solution of the moment compensation trajectory is calculated and substituted into the strict model, and the exact solution of the moment compensation trajectory is repeatedly calculated by feeding back and accumulating the moment error. It has gained. The approximate modeling process is as follows.

(A) The external force is not considered in the approximate model.

(B) Parts other than the lower limbs such as the waist, trunk, upper limbs, and head are assumed to be a four-mass point approximate model (FIG. 3).

(C) The rotation of the motion coordinate system is not taken into account for the purpose of linear and decoupling (however, a part of the trunk yaw axis is considered as long as it can be made linear and decoupling).

(D) Linear and decoupling are assumed as not sharing the motion of the Z axis.

First, if the external force is not considered in the approximation model, the equation (2.3) becomes as follows.

[0038]

(Equation 4)

As shown in FIG. 3, the approximate modeling of the upper limb mass is performed by concentrating all masses of the upper limb on the shoulder, and the trunk having three masses.
The waist is an approximate model having one mass point. Here, when the moment generated by the movement of the lower limb, which is the moment to be compensated, is set to M, and the four-mass point approximation model is performed, the equation (2.4) becomes

[0040]

(Equation 5)

Equation (2.5) interferes with each other in terms of apparent forces generated by rotation of the motion coordinate system.
Therefore, in order to make these differential equations non-interfering, it is first assumed that no moment is generated by the apparent force, that is, assuming that the motion coordinate system is not rotating, equation (2.5) Are the expressions (2.6) to (2.
It becomes like 8). As for the yaw axis, there is a term in which the rotation of the motion coordinate system is linear and does not affect the decoupling (the term can be transferred to the right side as a known moment), and therefore, the calculation is performed for this term also in the approximate model.

Pitch axis;

(Equation 6) Roll axis;

(Equation 7) Yaw axis;

(Equation 8)

Equations (2.6) to (2.8) share a movement in the Z-axis direction, and since the lower limb and the trunk are connected via a rotary joint, the equations are non-linear and interference. is there. Therefore, the humanoid robot assumes that the waist height is constant during exercise, and that the trunk arm (the arm from the trunk mid-mass point to both shoulder mass points) rotates in the horizontal plane, and that the trunk mass point does not move in the Z-axis direction. Assume, linear, decoupling.

If the rotation angle θ _y of the yaw axis actuator and the length of the arm vector r _T ′ of the trunk arm are R, the following three equations are obtained.

A pitch axis;

(Equation 9) Roll axis;

(Equation 10) Yaw axis;

[Equation 11]

In the equations (2.9 to (2.11)), the known term on the left side is moved to the right side, and the equations (2.12) to (2.12)
In (2.14), the right-hand side is rewritten as -M '=-[M
_x ', M _y', the repositioning and M _z '] ^T, we obtain the following three equations.

Pitch axis;

(Equation 12) Roll axis;

(Equation 13) Yaw axis;

[Equation 14]

(4) Introduction of Relational Expression for Coordination of Waist and Trunk and Calculation of Approximate Solution In order to obtain an approximate solution of the moment compensation trajectories x _T ′, x _W ′, y _T ′ and y _W ′, An approximate solution is analytically obtained using (2.12) to (2.14). However, since a total of four approximations of x ′ and y ′ of the waist and the trunk are unknown, Equation (2.
Since sufficient equations do not exist with only the two equations (12) and (2.13), an approximate solution of the moment compensation trajectory with respect to the pitch and the roll axis cannot be specified.

Therefore, in the present algorithm, the compensated moment in the approximation model (that is, equation (2.12),
By distributing (right side of (2.13)) to the waist and the trunk, respectively, as shown in equations (2.15) and (2.16), four unknown moment compensation approximations x _T ′ and x _W ′ are obtained. ,
y _T ', y _W' make the four differential equations against. Thus, an approximate solution can be uniquely specified.

A pitch axis;

(Equation 15) Roll axis;

(Equation 16) Yaw axis;

[Equation 17]

Here, in the equation (2.15), M
_{Since yTRUNK} 'and _MyWAIST ' can be calculated from the lower limb approximation and the ZMP approximation, they become known functions. Further, assuming steady walking, each participant of the humanoid robot and the ZMP are expressed in the motion coordinate system (O'-X'Y'Z '). ), There is a periodic relative motion, so that _MyTRUNK ′ and M
_yWAIST 'is a periodic function. Therefore, the left side of Expression (2.15) is also a periodic function. Therefore,
The right-hand sides of the two equations of the formula (2.15) are expanded into Fourier series, respectively, and x _T ′ and x _W ′ are respectively represented by the Fourier series whose coefficients are unknown, and the left-hand sides of the two equations of the equation (2.15) , And the Fourier coefficients of both sides are compared to obtain the Fourier coefficients of x _T ′ and x _W ′. Then, by applying an inverse Fourier series expansion to this, an approximate solution of the moment compensation trajectories x _T ′ and x _W ′ around the pitch axis can be obtained. Further, the moment compensation trajectories y _T ′ and y _W ′ about the roll axis are obtained in the same manner.

[0052] In this walking control method, since respect to the yaw axis moment on the ZMP is assumed that compensates only movements of the trunk, the trunk yaw axis track theta _y is moment compensation orbital yaw axis It is calculated by the Fourier coefficient comparison method using the equation (2.14). However, the term relating to the offset in the Fourier coefficient becomes indefinite. This is because there is no term that determines the initial value like the gravity term in the pitch and roll axis orbits. Therefore, with respect to the offset term, a method of setting the swing width of the arm in the steady portion to be within the movable angle range is adopted.

(5) Calculation of Exact Solution of Moment Compensation Trajectory by Iterative Calculation In order to obtain an exact solution of the moment compensation trajectory of the biped humanoid robot, an iterative calculation method is used. Specifically, the approximate solution obtained as described above is substituted into the equation (2.3), which is an equation of the strict model, and is compared with the right side thereof, thereby obtaining a moment error e _M = [e _Mx , e _My ,
e _Mz ] ^T is calculated, the sign of the error is inverted and stored in the right side of the equations (2.15) to (2.17), the calculation is executed again, and the allowable value e _M with the moment error is obtained. = [E
_Mx , e _My , e _Mz ] An exact solution is obtained by repeatedly performing the calculation until it becomes ^T or less.

However, in this state, an extremely large number of calculations are required to obtain an exact solution. Therefore, in the present algorithm, the accumulated moment error E _n = [E _{Mx (n)} , E
_{My (n)} , E _{Mz (n)} ] ^T is reduced by using the regularity of convergence and reducing the calculation time by performing the repetitive calculation while estimating the limit value using equation (2.18). Aim. Note that e _(n)
Indicates a moment error calculated after the n-th repetitive calculation.

[0055]

(Equation 18)

In the above, the solution method in the steady walking has been described. However, if the postures at the start and end of the motion are the same for a general motion, a series of motions including the start and the end of the motion are defined as one motion cycle. By setting the period of the front and rear stop postures to be sufficiently long, the moment compensation trajectory can be calculated by exactly the same algorithm.

FIG. 4 is a flowchart showing the above-mentioned "coordinated waist / trunk moment compensation trajectory calculation algorithm of the moment compensation amount distribution type capable of changing directions" of the biped humanoid robot.

Next, the creation of a walking pattern will be described.

In a biped humanoid robot, when realizing the whole body movement, time series data (herein, referred to as a walking pattern) of each joint angle outputted to the biped humanoid robot is described above. The procedure created using the moment compensation trajectory calculation algorithm based on the coordination of the above includes the following five steps.

(1) Setting of Toe Trajectory (and Hand Trajectory, Trunk Trajectory If Necessary) Representative points in the time series as the toe position and posture (and the toe position and posture) of the biped humanoid robot are In consideration of the continuity up to the position, velocity, and acceleration in each coordinate axis direction, interpolation is performed using a fifth-order or sixth-order polynomial to obtain a toe trajectory pattern.

(2) Initial setting lower limb trajectory First, an initial setting trajectory of the lower back is specified based on the above-mentioned foot trajectory. It should be noted that the coordinates of the waist are changed by the later moment compensation calculation, except when the waist mass point is not used for the moment compensation. Next, the lower limb link trajectory is calculated by inverse kinematics calculation based on the waist and the toe trajectory.

(3) Setting of the set ZMP trajectory The ZMP is set in a stable area formed by the sole. X, Y
Both the coordinates are set almost at the center of the stable region, and finally, the whole is smoothed by passing through a low-pass filter to obtain a set ZMP orbit pattern.

(4) Calculation of moment compensation trajectory First, parameters for moment compensation are set, then moment compensation trajectory calculation is performed, and the moment compensation trajectory of the trunk and waist is determined. In general, the upper limb mass position is determined by forward kinematics calculation according to the trajectory of the shoulder.
When specifying the hand trajectory, the upper limb link trajectory is calculated by inverse kinematics calculation according to the trajectory of the shoulder.

(5) Creation of Walking Pattern and File Output Based on the lower limb, upper limb, waist and trunk trajectory obtained as described above, the actuator angle of each joint is calculated by inverse kinematics calculation. A walking pattern is obtained by taking time series data of actuator angles for all joints, and the output of the walking pattern is performed in a CSV (Comma Separate Value) format which is a table of values separated by commas.

Next, the near real-time pattern generation method will be described.

In this case, the quasi-real-time pattern generation is basically performed by cutting an arbitrary step from a pattern generated for each walking cycle and connecting the same, similarly to the pattern switching performed so far. It creates a pattern.

First, considering each state of the pattern, the pattern according to the method described with respect to the waist / trunk cooperative moment compensation trajectory calculation algorithm can be broadly divided into three states as shown in FIG.

(1) Stationary state This is a stationary state before the start of walking and after the end of walking. From the stationary state before the start of walking to the stationary state after the end of walking is 1
A compensation trajectory is generated as a walking cycle. Further, since the compensation trajectory is obtained by the FFT, it is necessary to take at least one step before and after the stationary phase, and make the compensation trajectory a periodic function for each walking cycle.

(2) Steady State Among the walks, all walks except the transient state described later are in the steady state.

(3) Transient state This is two steps at the start of walking and at the end of walking. These are between the stationary state and the steady state, and draw a different step trajectory from the steady state even with the same step length and walking direction.

As described above, when one walking cycle is less than 5 steps, the walking is ended in a transient state, and when the number of walking steps is 5 steps or more, all walking is performed in a steady state except for two steps before and after. Becomes As a unit pattern to realize continuous walking requires at least one step of steady state,
In this method, a pattern of five steps in one walking cycle is set as a basic pattern.

The flow of quasi-real-time pattern generation follows the following procedure. FIG. 6 shows the flow.

(1) Designation of Toe Trajectory The waist trunk cooperative moment compensation trajectory generation algorithm, which is the basis of this method, can generate a compensation trajectory for an arbitrary toe trajectory and waist trajectory. In this method, the waist trajectory is specified in advance, and the foot trajectory (step length and direction) is arbitrarily specified for each step.

At the time of generation, it is known that by fixing the trunk trajectory, the trajectory of the gait trajectory in the time axis direction can be minimized, and the same method is used for this method.

(2) Grasp of Walking State In this method, it is assumed that everything from the start to the end of walking is generated in real time. Therefore, it is first necessary to grasp whether the current state is a stationary state, a transient state, or a steady state. As the compensation, the present state is grasped by the flag using the fact that the walking transitions from stationary to transient to stationary to transient to stationary.

(3) Grasp of standing position In this method, since the steady state is one step, only one foot pattern can be generated during steady walking. To cope with this, the first foot to be swung at the start of walking is changed according to the state of the standing leg, and the left and right are alternately generated.

(4) Generation of Pattern A compensated trajectory of all five steps is generated by using the waist trunk-coordinated moment compensated trajectory calculation method described above.

(5) Preservation of Toe Trajectory When a compensated trajectory is generated, not only the present but also past and future toe trajectories are required. For the future, if the spread in the time axis direction described in the above (1) designation of the toe trajectory is within the current one step, it can be appropriately taken, so the step length is calculated as 0.0 (m). are doing. However, since the past cannot be ignored from the viewpoint of the connection of the trajectories, the toe trajectories for the past three steps are stored and used for the next calculation.

(6) Outputting a pattern Among the generated patterns, an appropriate one step is designated according to the walking state obtained by the designation of the toe trajectory in the above (1), and is output as a unit pattern.

Next, shortening of the pattern generation time will be described.

The basis of generation of the near real time pattern is as follows.
Pattern generation is performed for each step using the algorithm described for the waist / trunk cooperative moment compensation trajectory calculation algorithm, and these are output continuously.

However, the current pattern generation time requires about 1.8 seconds per pattern, and can cope only with low-speed walking of about 2.0 seconds or more per step as it is. Therefore, the parameters related to the pattern generation time are taken out of the parameter groups specified in the pattern generation, and the parameters are reduced to the minimum necessary for walking, thereby shortening the pattern generation time and reducing the number of steps by 1.0 step. Be able to cope with walking for about a second.

However, since it is expected that most of the parameters for shortening the generation time will affect the stability of walking by decreasing them, it is possible to walk by these parameters by conducting a walking experiment in advance. Let's find the minimum value.

(1) Reducing the Number of All Phases Among the parameters to be set, the one that can be considered first as related to the generation time is the number of all phases. This is because the toe trajectory, waist trajectory, compensation trajectory, and the like generated in the above-described waist / trunk cooperative moment compensation trajectory calculation algorithm are all treated as time-series data with one walking cycle as the total number of phases. This is because the calculation is performed for all phases.

Here, when the number of all phases is reduced, the result of a simulation of how much the generation time changes is shown in Table 1.

[0086]

[Table 1]

Thus, it can be seen that the reduction in the number of all phases leads to a reduction in the generation time almost in proportion.

However, since the reduction of the number of all phases is the reduction of the time-series data itself, the number of steps and time in one walking cycle are reduced, and in order to avoid this, the data itself is coarsened, and the interval between the two is reduced by interpolation. You have to get it.

Here, as described in the waist / trunk cooperative moment compensation trajectory calculation algorithm described above, since a pattern of all five steps is used as a basis, at least seven steps including two steps before and after the stationary phase are secured. Must. Therefore, it can be seen that the number of phases per step that can be secured for the total number of phases is as shown in Table 1. In order to achieve the target of walking for 1.0 second per step, it is necessary to set the total number of phases to 256 phases, since it is required to be about 0.5 seconds per pattern in consideration of the transfer time of pattern data. It is assumed that the number of phases per step is 32 phases, which is half the current number.

Here, since a method is employed in which a given pattern file is linearly interpolated by the designated number of interpolation points and output to each joint, the fact that the number of phases per step is reduced means that the number of phases per step is reduced by the sixth-order function. This indicates that the generated time-series data is coarse, which causes a difference between the generated data and the data output in the actual machine. Since this takes a trajectory different from the expected trajectory of the actual machine, it is inevitable that it will affect walking stability.

Therefore, it was examined by a preliminary walking experiment how much the number of phases per one step affects walking.
The ZMP trajectory at that time is shown in FIGS.

Looking at the ZMP trajectory, it can be seen that the state of vibration is different over the entire walking cycle, but it cannot be said that this is directly linked to the instability of walking immediately. That is, if the number of phases is reduced to this extent, it is considered that the influence on walking itself is small.

(2) Decrease in the number of repeated calculations The compensation trajectory is calculated repeatedly until the error moment when the approximation generated by the four-mass point approximation model is applied to the strict model falls below a certain allowable value. The calculation time can be reduced by reducing the number. However, a decrease in the number of repetitive calculations increases an error moment, which also causes an error with an actual machine. Therefore, the effect of increasing error moment on walking was investigated by walking experiments.

Since the specified error moment is an allowable value, if the number of calculations is the same, the actually generated error moment is the same regardless of the allowable value. Therefore, the allowable error moment was appropriately specified, and the number of calculations was arbitrarily set, so that the comparison based on the number of repetitive calculations was performed. Table 2 shows the correspondence between the allowable error moment and the number of calculations at that time. 9 and 10 show the ZMP trajectory at that time.

[0095]

[Table 2]

Looking at the ZMP approximation, it can be seen that the trajectory is almost unchanged until the number of calculations is about four. This is probably because the response delay of each joint and the impact at the landing point are much larger than the ZMP trajectory error at which the error moment occurs. However, if the number of calculations is less than three, the trajectory clearly changes, and if the number of calculations is one, it becomes impossible to walk. Therefore, here, the allowable error moment is set to 0.5 (Nm) as a value at which the number of calculations is four or more.

(3) Reduction in the number of mass points in the model As described above with respect to the reduction in the number of repetitive calculations in (2), the compensation trajectory generated by the four mass point model is applied to the strict model, It is calculated repeatedly until the moment falls within a certain tolerance. On this occasion,
In the strict model, an actual machine is modeled by a number of mass points, a moment generated by each mass point with respect to a given orbit is calculated, and an error moment with respect to a set ZMP is calculated. In other words, reducing the moment calculation in this exact model leads to shortening of the compensation trajectory calculation. However, a decrease in the number of mass points in the strict model may still cause an error between the strict model and the actual machine.

FIG. 11 shows the current mass point. As described above, the current mass point is not limited to one for each link, and it can be seen that modeling is performed by a plurality of mass points depending on the location. Therefore, by reducing this to about one for each link, the model calculation time in the strict model is reduced. At this time, the integration of the mass points of the models is set so that the model in which each mass point is generated has the same value as much as possible even in the reference link.

Next, a visual and voice input system according to the present invention will be described.

In the method of the present invention, a visual sensor and / or a voice input sensor using a microphone sound is provided in the robot main body for designating a toe trajectory required for creating a walking pattern, and stereo vision or monocular is provided by the visual sensor. To determine a stride for generating a toe trajectory, the voice input sensor responds only to a command set in advance, and selects a stride set in advance in association with a word by a voice input word. ing. Therefore, as long as the stride information can be given, the input can be made with an arbitrary input device. Here, a voice input by a visual sensor and a microphone is mainly used for designating a toe trajectory necessary for creating a walking pattern. Less than,
The relationship between sensor input and toe trajectory determination for each of the visual information input system and the voice input system will be described. FIG. 12 shows the configuration of these systems.

(1) Visual Information Input System The flow from the input of visual information to the determination of the toe trajectory is as follows.

(A) First, before the start of walking, the tracking vision is initialized and the tracking points are specified. The specification of a tracking point can be performed by using a mouse of a personal computer. Here, the follower's hands are set as tracking points.

(B) The step length is determined from the visual sensor to generate the toe trajectory. In this case, it is necessary to recognize the front, rear, left and right directions by a visual sensor in order to follow the front, rear, left and right. As is already known, it is necessary to perform stereo vision in order to recognize the front-back direction, and the humanoid robot according to the present invention also has a system compatible with stereo vision. However, for the sake of simplicity, stereo vision is not performed here, and follow-up movement in the front-back, left-right directions is enabled by gesture recognition by a single eye.

The gesture recognition is pseudo-stereoscopic viewing by a single eye so as to be able to cope with tracking by stereoscopic viewing.

In this method, two points are tracked from one camera, and these points are regarded as tracking points of the left and right cameras, respectively, and are regarded as a movement amount in the front-back direction according to a change in the distance between the two points. is there.

First, the amount of movement of the left and right tracking points from the initial value is calculated.

DV _R = V _R -V _R0 dV _L = V _L -V _L0 (3.1) The difference between the movement amounts is multiplied by a gain to obtain the stride amount in the front-back direction.

[0108] _{F f = G f (dV R} -dV L) ...... (3.2) F s = G s (dV R -dV L) ...... (3.3) Further, the left and right tracking point in the lateral direction The average of the movement amount is taken.

[0109] Here, V _R0, V _L0; initial value V _R of the tracking point, V _L; current value dV _R of the tracking point, dV _L; the difference between the initial value and the present value F _f; longitudinal direction stride F _s; lateral direction stride G _f; a lateral direction stride gain; longitudinal direction stride gain G _s.

(C) A toe base and a pattern are generated using the determined stride length. The timing of pattern generation is
The timer constantly monitors the generation permission flag from the master side, and when the permission is issued, it is performed every step.

(2) Voice Input System Voice input uses a grammar control using a text file, and does not respond to any command other than a preset command. When the generated word is recognized, a stride previously set in association with the word is selected, and pattern generation is performed. Here, as an example, before, after, right,
Only 8 words of left, right turn, left turn, spot and stop are registered, but by editing the grammar control text file and setting the stride corresponding to it,
It is possible to correspond to various walks. Table 3 shows the stride for each word in this embodiment.

[0112]

[Table 3]

In the table, "None" means that the current stride is not changed.

Tables 4 and 5 collectively describe the symbols and notations of the mathematical formulas used above.

[0115]

[Table 4]

[0116]

[Table 5]

[0117]

As described above, according to the present invention, a parameter which has little effect on walking in generation of a compensating trajectory of a robot is clarified, and a time for generating a compensating trajectory is greatly reduced, so that a pattern setting during walking can be performed. It was possible, and it was possible to follow the movement of humans while setting patterns in real time. In addition, in a biped humanoid robot, a computer for joint control and a computer for pattern management are provided side by side and connected by DP-RAM to realize continuous walking without being restricted by the memory capacity limitation. In addition, by using the pattern management computer, it is possible to perform a process that normally takes more time than the control cycle during walking. In addition, various effects are provided, such as easy setting of the toe trajectory by visual or audio.

[Brief description of the drawings]

FIG. 1 is a diagram showing a resultant force and a moment acting on a supporting polygon in a stable supporting state.

FIG. 2 is a diagram showing a coordinate system of each mass point of the humanoid robot and a vector display of a force acting on the coordinate system.

FIG. 3 is a diagram showing an approximate model of four mass points.

FIG. 4 is a diagram illustrating a flowchart of a “moment compensation trajectory calculation algorithm for cooperating waist and trunk with a moment compensation amount distribution type capable of changing directions” of a biped humanoid robot.

FIG. 5 is a view showing a walking pattern of a lower limb portion of the biped humanoid robot.

FIG. 6 is a diagram showing a flowchart of a near real-time pattern generation procedure.

FIG. 7 is a diagram showing a Y-ZMP trajectory for describing an influence of the number of phases per step on walking.

FIG. 8 is a diagram showing an X-ZMP trajectory for explaining an influence of the number of phases per step on walking.

FIG. 9 is a diagram showing a Y-ZMP trajectory for explaining an effect of the number of phases per step on walking.

FIG. 10 is a diagram showing an X-ZMP trajectory for explaining an influence of the number of phases per step on walking.

FIG. 11 is a diagram showing a weight arrangement diagram of each mass point of the biped humanoid robot.

FIG. 12 is a diagram showing a flow of an entire system for real-time pattern generation according to the present invention.

[Explanation of symbols]

X Roll axis Y Pitch axis Z Yaw axis O-XYZ An absolute coordinate system O'-X 'in which a plane including the X axis and the Y axis coincides with the road surface and the origin is fixed to the road surface with the axis perpendicular to the road surface as the Z axis. Y'Z 'Motion coordinate system whose origin is fixed near the waist of the humanoid robot

 ──────────────────────────────────────────────────続 き Continuing from the front page (72) Inventor Shunpei 3-4-1 Okubo, Shinjuku-ku, Tokyo Waseda University Faculty of Science and Technology (72) Inventor Motohiro Sato Kamiya Building 4-20-3 Ishihara, Sumida-ku, Tokyo No. 301 (72) Inventor Yusuke Sugawara 3-4-1 Okubo, Shinjuku-ku, Tokyo F-term in Waseda University Faculty of Science and Technology (reference) 3C007 AS36 CS08 KS39 KT01 LV12 WA03 WA13 WB05 WB07 WB17 WB19

Claims (4)

[Claims] In a humanoid bipedal walking robot, walking pattern change information from a visual sensor and / or a voice input sensor is recognized, and a walking pattern can be changed in real time according to the walking pattern change information. A motion trajectory control method for a biped walking robot. 2. A walking pattern is created in real time during walking by calculating a trajectory by dividing a moment compensation amount into each part of the body by using a moment compensation trajectory calculation algorithm by coordination of body parts. The method of controlling a motion trajectory of a biped walking robot according to claim 1. 3. The motion trajectory control method for a bipedal walking robot according to claim 2, wherein a computer for controlling each part of the body and a computer for generating a pattern are separately performed when creating the moment compensation trajectory calculation algorithm. 4. A robot sensor according to claim 1, further comprising a visual sensor used for designating a foot trajectory necessary for creating a walking pattern and / or a voice input sensor using a microphone sound. And determining a stride for toe trajectory generation, wherein the voice input sensor responds only to a preset command, and selects a preset stride in association with a word by a voice input word. The method of controlling a motion trajectory of a bipedal walking robot according to claim 1.