JP2009066692A - Trajectory searching device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a robot which is flexibly adaptable to an environmental change for performing more developed operation than the learned operation although the robot which can execute learned operation only has a closed application range in the robot to be learned and executed operation. <P>SOLUTION: The trajectory searching device is provided for searching most reliable trajectorys while using probability models for representing the learned operation and for representing combined operation with any of the probability models combined. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to an improvement in a trajectory search apparatus that searches for a trajectory for a robot to perform an operation in a technique for imitating a human motion in a robot.

  Imitation learning is a process in which a user demonstrates a motion and causes the robot to learn the motion. The movement learned by the robot is called a movement primitive because it is a basic movement that the robot can perform. The operation primitives to be learned by the robot can be classified into an operation for manipulating an object and an operation for not manipulating the object, but what is practically important is an operation for manipulating the object. At this time, an object to be operated is called a trajector.

  In learning of motion primitives, the robot captures the trajectory of the trajector by capturing the motion that the user demonstrates with a camera. By performing such capture many times, general characteristics of the operation primitive are extracted. This is done by modeling the trajectory of the captured trajector using a probability model. As the probability model, for example, a Hidden Markov Model (HMM) is used. This is a probability distribution of position, velocity, and acceleration that changes with time, how the state of the projector should change over time. And a model expressed using transition probabilities between states.

After learning, when a motion primitive is ordered from the user, the robot searches the most probable trajectory under the probability model representing the commanded motion primitive to obtain the optimal trajectory to be drawn by the trajector.
The following non-patent documents are listed as related prior art documents.
Iwahashi, N .: Robots That Learn Language: Developmental Approach to Human-Machine Conversations, Symbol Grounding and Beyound: Proceedings of the Third International Workshop on the Emergence and Evolution of Linguistic Communication (Vogt, P. et al. (Eds.)) , Springer, pp 143-167 (2006). Tetsuo Haoka, Naoto Iwahashi: “Learning the Concept of Spatial Movement Depending on Reference Points for Language Acquisition”, IEICE Technical Report, PRMU2000-105, pp.39-46 (2000). Tokuda, K., Kobayashi, T. and Imai, S. Speech parameter generation from HMM using dynamic features, Procceedings of International Conference in Acoustics, Speech, and Signal Processing, pp. 660-663 (1995).

The robot can execute the learned motion primitives if ordered. However, if only the learned motion primitives can be executed, the robot is not practical.
For example, when the robot learns the task of sorting a large number of parts by type in the factory and sending them to the desired destination production line for automation, there are various patterns for such delivery, The user needs to learn a large number of motion primitives, and a lot of man-hours are generated by performing the demonstration in front of the camera for each work. This makes it impossible to put it to practical use in production sites that are subject to strict production schedules on a daily basis, and imitation learning applications are limited to attraction of toys and amusement facilities at best. As a result, there is a problem that the application range is closed.

  An object of the present invention is to provide a trajectory search apparatus that searches a trajectory of a trajector that realizes a further developed operation based on a learned limited operation primitive.

  In order to solve the above-mentioned problems, a trajectory search apparatus according to the present invention is a trajectory search apparatus that searches a trajectory for moving a trajector, and includes an operation primitive that can be performed by a drive device for moving the trajector. An accepting unit that accepts a movement instruction of a trajector from a user, and an action sequence that includes action primitives included in the instruction from the user that is accepted by the accepting unit, and generates a maximum likelihood trajectory within the range of the generated action sequence. Each operation primitive that can be performed by the drive device is represented by a probability model, and the probability model indicates how the state of the trajector should be changed over time. And a trajectory search by the search means for the motion sequence for the motion sequence. It performs binding probabilistic model representing a primitive, under resulting joint probability model by coupling, characterized by searching the track likelihood is maximized.

  The trajectory search apparatus according to the present invention searches for the maximum likelihood trajectory of a trajector under a joint probability model formed by combining probability models representing learned motion primitives. Since there are any number of ways to combine the motion primitives, the robot can perform various motions by learning a small number of motion primitives. That is, the robot can be provided with various application capabilities. In order to perform multiple operations in succession, it is possible to simply execute the operation primitives in order. In this case, however, the operation stops each time one operation primitive is finished. It is dangerous because the acceleration may change suddenly. In the trajectory search apparatus of the present invention, coupling is performed at the stage of the probability model representing the operation primitive, and the trajectory of the trajector is searched on the coupling probability model. In other words, since the entire motion is executed as one combined motion and there is no connection between the motion primitives, the robot can perform natural motions such as when a human performs motions continuously.

  Here, the combination of probabilistic models representing motion primitives is obtained by calculating the average of the position, velocity, and acceleration of each of the motion primitives to be executed in the second and subsequent motion sequences, and the position, velocity, and acceleration of the projector in that state. And the average of the positions of the trajectors in the initial state of the motion primitive and the average of the positions in the final state of the immediately previous motion primitive. In the combination of probability models, among the averages that are parameters that characterize the probability model, the coordinates of the position average can be transformed so that the initial state of the operation primitive and the final state of the immediately preceding operation primitive are continuous. Therefore, the continuity of the trajectory of the trajector can be ensured.

  Here, the trajectory search device includes a capturing unit that captures the trajectory of the trajector and a recognition unit that recognizes a stationary object arranged in a space in which the driving device operates, and a combination of probability models representing motion primitives. Determines whether the reference point that is the reference of the motion primitive to be combined matches any of the recognized stationary object positions, and the reference point that is the reference of the motion primitive is the recognized static If it matches any of the object positions, the distribution in each state of the operation primitive to be executed after the second in the operation sequence is the distribution in each state of the operation primitive in the combination, which becomes the reference of the operation primitive If the reference point does not match any of the recognized stationary object positions, The distribution in each state of the action primitive to be executed after the second in the column is the sum of the distribution in each state of the action primitive and the distribution in the final state of the immediately preceding action primitive. Can do. Since the dispersion is expanded when the reference point of the operation primitive to be combined does not coincide with the position of the stationary object, the trajectory that can be taken in the entire combining operation can be widened. Also, when the reference point of the action primitive to be combined matches the position of the stationary object, the variance of the original probability model is used as it is, so that the trajector can be moved without shaking from the reference point.

  The trajector movement instruction includes the initial position and the target position of the trajector, and the search means generates a plurality of operation primitive operation sequences corresponding to the path from the initial position of the trajector to the target position, It can be assumed that the probability models are combined for each column. In the search for the trajectory connecting the initial position of the specified trajector and the target position, a plurality of motion sequences consisting of learned motion primitives are generated, and each trajectory is searched on the probability model representing the motion that combines them. A flexible trajectory that is not limited to simple motion primitives can be searched.

  The movement instruction of the trajector includes a plurality of combinations of operation names and landmarks of operation primitives, and the operation primitives to be combined with the probability model by the search means are operation primitives corresponding to the plurality of operation names included in the movement instruction. It can be assumed that When ordering the movement of the trajector, it is possible to specify in detail in which order the operation primitive is to be performed, so even if the maximum likelihood trajectory obtained by searching, such as avoiding an obstacle, is inconvenient, the operation is performed in another trajectory. Can be executed.

  In the combination, coordinate conversion may be performed so as to unify the coordinate system in which two probability models to be combined are defined. In the average combination of the probability models, the coordinate system can be unified by converting the coordinate axes of the probability models defined in different coordinate systems. Since the coordinate system is unified at the time of combination, the probability model can be defined in a coordinate system specific to the operation primitive. Therefore, it is not necessary to learn a plurality of equivalent operation primitives by coordinate transformation.

  The coordinate transformation may be performed by affine transformation of a coordinate system in which a probability model representing the motion primitive is defined with respect to a coordinate system in which a probability model representing the immediately preceding motion primitive is defined. it can. Of the two action primitives to be combined, the coordinate system in which the probability model representing the action primitive to be executed later is defined matches the coordinate system in which the probability model representing the action primitive to be executed in advance is defined. By performing the affine transformation of the coordinate system, the sequential transformation is performed so that the coordinate systems of the probability models representing all motion primitives in the motion sequence match.

The probability model representing the operation primitive may be determined based on an operation performed by a user. The learning of the motion primitive is performed so as to reproduce the motion demonstrated by the user, so that the motion of the robot can be made to resemble a smooth motion performed by a human.
Before the driving device performs the operation represented by the trajectory with the maximum likelihood, the user is asked to confirm, and if the user does not agree, the trajectory is not output. By causing the user to confirm the maximum likelihood trajectory searched by the trajectory search device before operating the driving device, an unintended dangerous motion can be prevented in advance.

  In the present embodiment, a system for operating an object by moving an arm type manipulator of a robot as shown in FIG. 1 is considered. FIG. 1 is a diagram showing a configuration of a manipulator control system using a trajectory search apparatus 300 in the present embodiment. The manipulator control system includes a manipulator 500 that performs a commanded operation, a microphone for inputting a command by voice, a speech recognition engine 100 that analyzes the command that has been input, a camera for shooting a surrounding situation in which the command is executed, Realize the trajectory searched by the image recognition engine 200 for analyzing the captured video, the trajectory search device 300 for searching the trajectory for performing the motion and learning the motion from the input command and video, and the trajectory search device 300 Therefore, the control device 400 is configured to generate control parameters for controlling the manipulator 500.

The trajectory search apparatus 300 according to the present embodiment has a learning mode for learning an operation and a trajectory search mode for searching for a trajectory that performs a combined operation combining the learned operations.
Prior to entering a specific description of the configuration of each mode, concepts such as terms used in this specification will be described.
The movements that the robot performs include actions that do not manipulate objects such as `` walking '', `` flying '', and `` turning '', and (a) `` lifting the pen '', (b) `` putting the pen in the box ''", (C) There is an operation of manipulating an object such as" Move the pen over the box ". In fact, considering that you want the robot to do some work, the action of manipulating things is much more useful than the action of not manipulating things. Think. In cognitive linguistics, among the entities that are focused on in the process of the subject that interprets the outside world, the objects that are recognized relatively prominently are called trajectors, and the objects that place this in the background are called landmarks. In an operation of manipulating an object, an object to be operated is a trajector, and the other object is a landmark. In the example of FIG. 2, “pen” is a trajector, and “box” is a landmark.

  Learning the action is to generalize the action demonstrated by the user so that the same action can be reproduced. Here, it is important that the operations are not the same but the same. If the movement is exactly the same, it is easy for the robot to reproduce the movement trajectory if it is stored as coordinates. However, when considering the action of “putting a pen in a box”, the positional relationship between “pen” and “box” varies depending on the situation. "It is not possible. In this case, it recognizes the position of the “pen” and “box”, remembers the generalized movement of lifting the “pen”, approaching the top of the “box”, and lowering it toward the top of the “box” Is learning.

  The manipulator control system first learns operation primitives which are basic operations performed by the manipulator 500 in the learning mode. In addition, in the trajectory search mode, it is detected that a command is given by a user's voice input from a microphone or a user's gesture photographed by a camera, and a plurality of learned operation primitives are combined. , Search for a trajectory that can perform the optimum operation to execute a given command. The trajectory found in the search is sent to the control device 400, converted into control parameters for controlling the manipulator 500, and actually moves the manipulator 500. Note that if one of the learned motion primitives is the best motion to execute the given instruction, the motion primitives are not combined and the trajectory is searched under the probability model representing the motion primitive. That's fine.

  Here, the command is, for example, “put a doll in a box”. When the user instructs the microphone to “put the doll in a box”, the speech recognition engine 100 extracts the operation name “put”, and the object to be operated, that is, the trajector is “the doll” And the target position of the motion is recognized as “box”. In this embodiment, the moving object is only a trajector. A stationary object other than the trajector is called a stationary object. In order to recognize where the trajector or the stationary object is, the camera performs shooting, and the shot video is analyzed by the image recognition engine 200.

Hereinafter, the configurations of the learning mode and the trajectory search mode will be described.
<Structure of learning mode>
This section describes the learning mode configuration of the manipulator control system.
The learning of the movement is performed by the user giving a command and demonstrating the movement that the robot wants to perform many times. For example, when learning the action of “putting”, the “putting” action such as “putting a pen on a box”, “putting a doll on a shelf”, “putting a red box on a blue box”, etc. The robot learns by generalizing their movements. This is done by extracting the motion of the trajector and the position of the stationary object from the input video, searching for the coordinate axis and its origin so as to generalize the motion of the trajector, and determining a probability model that gives the motion of the trajector. .

Fig. 3 shows an example of input video. FIG. 3 (a) is an example demonstrating the operation of “putting a red box on a blue box” when three boxes of red, blue and green are lined up. Figure 3 (b) continues
This is an example of demonstrating the action of “putting a green box on a red box”. Both are demonstrations of “putting”, but there are differences in the target position and direction of movement. However, both are common in that the direction of movement is a direction approaching the target position.
<Probability model>
Here, a description will be given of a motion probability model. The movement is a movement of the trajector, and can be represented by a temporal change in the position, speed, and acceleration of the trajector. However, if the motion is specified by the temporal change in the position, velocity, and acceleration of the trajector, the ambiguity of the motion is lost. For example, when a person performs the action of “raising a pen”, a trajectory is drawn in which the vertical component of the coordinates of “pen” increases with time, but it must be raised to a specific position at some point in the middle. There is no restriction that it does not follow, and even if there is some blur in the horizontal direction, it does not exceed the category of “raise” operation. In consideration of such ambiguity, the motion of the trajector is expressed by giving the probability of being in a state within a certain range centered on the position of the trajector at a certain time and the transition probability between the states. The model is a probabilistic model.

  FIG. 4 is a conceptual diagram of motion primitives represented by a probability model λ. Fig. 4 shows the state s of the trajector in the three-dimensional space under the probability model λ, the mean E of the position x_s (λ) at the time t_s is the black circle at the center of the ellipsoid, and the variance V is the ellipsoid It is the figure expressed with the magnitude | size of. The trajector has a high probability of being near the origin at time t_0, and shows that the probability of being at a position away from the origin increases as the time advances t_1, t_2,. In order to specify the motion, it is necessary to consider not only the position but also the velocity and acceleration as well as the mean and variance. FIG. 4 shows only three-dimensional spatial coordinates for the convenience of drawing, but is actually defined as a probabilistic model in nine-dimensional space including position, velocity, and acceleration.

In this specification, “_” is used to indicate a subscript, and “^” is used to indicate a superscript.
FIG. 5 is a diagram illustrating a data structure of the probability model λ. The probability model λ is defined by a sequence of parameters that give a probability distribution in the state s at each time t_s and a transition probability between the states. The parameter at each time t_s includes an average E and a variance V for each of the position x_s, the velocity x′_s, and the acceleration x ″ _s. Average position (p, q, r), average velocity (v_p, v_q, v_r), average acceleration (a_p, a_q, a_r), position variance (V_p, V_q, V_r), velocity The variance is (V′_p, V′_q, V′_r), and the acceleration variance is (V ″ _p, V ″ _q, V ″ _r). Among these parameters, the position is as shown by the ellipsoid in FIG.

In this specification, “dash” is used to express the first and second derivatives of the position x with respect to time, such as velocity x ′ and acceleration x ″. It should be noted that the “dash” attached to the probability model Λ does not mean time differentiation.
<Internal configuration of orbit search device 300 in learning mode>
FIG. 6 is a diagram showing an internal configuration in the learning mode of the trajectory search apparatus 300 of the present embodiment. The trajectory search apparatus 300 of this embodiment includes an input / output interface 1, a database storage unit 2, a processor 3, a work memory 4, and a program storage unit 5. The database storage unit 2 is a recording device such as a hard disk for storing learning results, the work memory 4 is a RAM for performing calculations, and the program storage unit 5 is a ROM that records a learning program and a trajectory search program. .

  In learning, the user issues an operation name of the operation to be learned by the robot toward the microphone shown in FIG. 1, and the operation name is transmitted from the input / output interface 1 in FIG. 6 to the processor 3 via the speech recognition engine 100. Entered. After that, the user demonstrated the motion, and the motion that was demonstrated was captured by the camera shown in FIG. 1, and the trajectory drawn by the trajector and the position set of the surrounding stationary objects were input and output in FIG. Input from the interface 1 to the processor 3. Since the user performs demonstrations many times to make the robot learn the movement, the trajectory Y_i of the trajector and the position set O_i of the stationary object are input for each.

In the learning mode, the trajectory search apparatus 300 causes the processor 3 to execute the learning program stored in the program storage unit 5. The learning program uses a coordinate system and an origin that generalizes the trajectory of the trajector from a set of trajectories Y_i of the plurality of trajectors corresponding to the motion name input from the input / output interface 1 and the position set O_i of the stationary object, and A probability model representing the operation is searched on the work memory 4. The parameters of the probability model found by the search are stored in the database storage unit 2 in association with the action name.
<Formulation of motion learning using a probability model>
In this section, we will explain the formulation of learning.

  Let V = {V_1, V_2, ..., V_N} be a set of N input images that have captured the motions demonstrated by the user to let the robot learn the motions. Each video V_i is a video with one moving object and a plurality of objects that are stationary in the background. For example, when learning the action of `` putting '', the images showing the actions of (a) putting a red box in a blue box and (b) putting a green box in a red box in Fig. 3 are the respective images. Corresponds to V_i.

  A trajector and a stationary object are recognized by extracting moving objects and stationary objects from each video V_i. Let Y_i be the trajectory of the trajector captured from the video V_i, and O_i be the position set of the stationary object. In FIG. 3 (a), the red box trajectory that is lifted and lowered while moving from left to right is Y_i, and the blue and green boxes are O_i. In FIG. 3 (b), the trajectory of the green box that is lifted and lowered while moving from right to left is Y_i, and the position of the blue box and red box is O_i.

The trajectory Y_i of the trajector is a time series of the position x_t, velocity x'_t, and acceleration x '' _ t of the trajector, and these are collectively written as y_t = [x_t, x'_t, x '' _ t] ^ T. Here, T is a symbol representing taking a transposed matrix.
In order to quantify the trajectory of the trajector, it is necessary to determine the origin of the coordinate system and the orientation of the coordinate axes. The origin of the coordinate system is referred to as a reference point because it serves as a standard for digitizing the trajectory. The reference point is selected from landmarks. The landmark is a set consisting of the stationary object position set O_i, the initial position x_0 of the trajector, and the center position of the video, L_i = {l ^ (i) _1, l ^ (i) _2, ..., l ^ (i) _M_i}. Here, M_i is the number of landmarks in the video V_i. The direction k of the coordinate axis is selected from K types of candidates prepared in advance. A specific method of taking the coordinate axes will be described later.

  The trajectory determined by the trajectory Y of the trajector, the direction k of the coordinate axis, and the reference point l is written as F (Y, k, l). Under the stochastic model λ for the motion of the trajector, the trajectory of the trajector shown in the video V_i is Y_i, the direction of the coordinate axis is k, and the m_i-th landmark l ^ (i) _m_i out of the M_i landmarks in the video V_i Let probability P (F (Y_i, k, l ^ (i) _m_i); λ) be the probability of drawing a trajectory F (Y_i, k, l ^ (i) _m_i) as a reference point. At this time, a set of indices m * indicating the probability model λ *, coordinate axis direction k *, and reference point for each input video is set so that the logarithmic likelihood for each video V_i is maximized over all input videos. By making the decision, the operation is learned. This is expressed by the following equation (Equation 1).

ここで、argmaxの下にλ,k,mを書いた記号は、引数部分が最大になるようなλ,k,mを返す関数であり、m=(m_1,m_2,...,m_N)である。また、「*」は、推定値を意味する記号である。

Here, the symbol with λ, k, m written under argmax is a function that returns λ, k, m that maximizes the argument part, and m = (m_1, m_2, ..., m_N) It is. “*” Is a symbol indicating an estimated value.

In short, given the probability model λ, the probability of drawing a trajectory F with a trajector on that probability model is determined, so it is most reliable to reproduce all of the trajectory Y_i shown in the input video shot by the user's demonstration. It is the meaning of (Equation 1) to decide how to take a new coordinate system.
This problem can be solved, for example, by using an HMM as a probability model. Details of the solution are shown in Non-Patent Document 2.

In this specification, the description is made with the use of the HMM as the probability model, but this does not limit the probability model to the HMM.
《Learning motion primitives》
In this section, specific examples of learning results when learning some motion primitives are shown. There are seven types of motion primitives that are learned: “Give up”, “Crick up”, “Hanasu”, “Turn”, “Put on”, “Sake up”, and “Move up”. The following K = 3 types were adopted as the coordinate system.
k_1: A coordinate system with the landmark as the origin, the y-axis vertically downward, the x-axis horizontally, and the x-axis orientation determined so that the x-coordinate of the initial position of the trajector is negative.

k_2: A coordinate system with the landmark as the origin, the x axis in the direction from the origin to the initial position of the trajector, and the y axis in the direction perpendicular to it.
k_3: A coordinate system in which the initial position of the trajector is the origin, and the y-axis is taken vertically downward and the x-axis is taken horizontally.
FIG. 7 shows the motion trajectory demonstrated by the user when learning each motion primitive, the coordinate system k selected at that time, and the reference point l. (A) to (g) in FIG. 7 show learning results for each of the above seven types of operation primitives.

  First, the meaning of the symbols will be described using FIG. 7 (a) as an example. In FIG. 7 (a), squares and circles represent stationary objects, and the black and white circles with T represent trajectors. In the situation where several stationary objects are placed, it is shown that the user has raised the trajector upward from the start point to the end point of the thick arrow and issued a command to “raise” four times to the microphone. . The number and position of stationary objects, the initial position and target position of the trajector are different for each input operation, and each input operation is captured by a camera and recognized as an input video V_i. From the input video V_i, a moving object is recognized as a trajector, and the initial position of the trajector and the position of a stationary object are recognized as landmarks. At this time, as a result of determining the coordinate system and the reference point based on (Equation 1), the coordinate axis is taken in the direction indicated by the thin arrow, and the intersection point thereof is selected as the reference point R. That is, for the operation of “raising”, this indicates that the coordinate system k_3 has been selected.

The same applies to the operations in FIGS. 7 (b) to 7 (g). For `` put on '' and `` get over '', the coordinate system k_1, for `` flickering '' and `` Hanasu '', the coordinate system k_2 , “Turn”, “Turn”, and “Saguru” indicate that the coordinate system k_3 is selected.
<Orbit search mode configuration>
In this section, an internal configuration in the trajectory search mode of the trajectory search device 300 of the present embodiment will be described.

Trajectory search is the search for the most probable trajectory for executing the commanded motion from the user under a joint probability model that represents the joint motion generated by combining the learned motion primitives. .
FIG. 8 shows an internal configuration in the trajectory search mode of the trajectory search device 300 of the present embodiment. This is basically the same as the configuration in the learning mode shown in FIG. 6, but since the program executed by the processor 3 is different, the input / output data and the operation contents developed in the work memory 4 are different. Different.

  The input / output interface 1 receives a code string indicating an instruction content analyzed by the speech recognition engine 100 about a command issued by the user toward the microphone shown in FIG. The code string is a code string that specifies the initial position x_0 of the trajector and the target position x_n after the movement, as in the case where the user points to the finger and `` move this over there '' or `` lift the pen The code string is specified in the order of execution of the operation primitives such as “raise” and “put”.

  In addition, the input / output interface 1 analyzes the video captured by the camera shown in FIG. 1 with the image recognition engine 200, and acquires the initial position x_0 and target position x_n of the trajector and the position set O of the stationary object. Since the initial position of the trajector is specified by the image recognition engine 200, the command only requires the user to specify the trajector by name or gesture.

  The processor 3 executes a trajectory search program stored in the program storage unit 5. The trajectory search program reads two or more parameters λ of the probability model representing the operation primitive stored in the database storage unit 2 by learning, and stores the parameter Λ ′ of the combined probability model obtained by combining them in the work memory 4 To do. Here, the parameters of two or more probability models are read in order to show an example of combining probability models. When the combination is unnecessary, the parameters of one probability model may be read. Which operation primitive the parameter of the probability model corresponding to is read depends on the format of the instruction instructed by the user. That is, when only the initial position and target position of the trajector are specified, the trajectory search program considers the combination of learned patterns of motion primitives. Read only the parameters of the probabilistic model corresponding to the specified motion primitive and generate only their combination.

The trajectory search program searches for the most probable trajectory Y * under the combined probability model, and sends the result to the control device 400 via the input / output interface 1. This most probable trajectory means a trajectory that gives the maximum transition probability from the initial position to the target position after giving a probability model. There is no particular meaning for the minimum trajectory. Since the probabilistic model is learned based on the user's motion, the most probable trajectory is a trajectory close to a reproduction of a natural human motion.
<Combination of probability models>
In this section, the specific processing contents of a method of combining probability models representing learned motion primitives and searching for trajectories under the combined probability model will be described.

  By instructing the robot with the learned motion primitive, the robot can perform the motion primitive. However, if only simple learned operation primitives can be performed, it is not possible to respond to various instructions unless all operations are learned as operation primitives. In the present invention, a trajectory is searched under a joint probability model representing a composite motion obtained by combining learned motion primitives.

In the present embodiment, the following two types of trajectory search methods for the combined operation are considered.
1. By using the initial position x_0 and the target position x_n of the trajector as inputs, the motion sequence A * and the trajectory Y *, which are composed of a pair of <motion primitive, landmark> having the highest likelihood, are searched.
2. Search the trajectory Y * with the highest likelihood, using the sequence of trajectors and <motion primitive, landmark> as the sequence of motion A.
The trajectory search method 1 is employed when the user designates and instructs a trajector and a target position, such as “Pen is brought here”. The initial position of the trajector is specified by the image recognition engine 200 if the trajector is designated. In this case, both the motion sequence and the trajectory are searched.

The trajectory search method 2 is adopted when the user designates the order of motion primitives and gives an instruction, such as “pick up a pen and then click”. In this case, since an action sequence is specified, only the trajectory search is performed under the action sequence.
In learning of operation primitives, a coordinate system is determined for each operation primitive, and therefore it is necessary to unify the coordinate system when combining different operation primitives. For example, as shown in FIG. 9, consider a case where an operation primitive “raise” is combined with an operation primitive “chickle” to generate an operation that brings “stuffed animal” closer to the upper left position of “daruma”. Fig. 9 (a) shows a probability model λ_1 representing the motion primitive "raise", which is a coordinate with the initial position of the trajector as the reference point and the x-axis in the horizontal direction and the y-axis in the upward direction. It is defined in the system. Fig. 9 (b) shows a probability model λ_2 that represents the action primitive of `` Cricking '', which uses the position of the landmark to be brought close to the trajector as the reference point, from the reference point to the initial position of the trajector. It is defined by a coordinate system with the x-axis in the direction toward it and the y-axis in the direction perpendicular to it. In order to generate a combined action of “raise and then click” from these two action primitives, the end of the action primitive “raise” and the action primitive that “chuck” is generated as shown in FIG. 9 (c). Coordinates the coordinate system in which the action primitive "Kikakase" is defined so that the coordinate of the action primitive "Kikakase" matches the upper left position of "Daruma". There is a need to.

Hereinafter, the formulation of coordinate transformation in the combination of motion primitives will be described.
Now, it is assumed that the motion sequence A is generated by combining n motion primitives, and each motion primitive is represented by a probability model λ_i (i = 1, 2,..., N). A sequence of probability models defined in each coordinate system of motion primitives included in motion sequence A is written as Λ = (λ_1, λ_2,..., Λ_n).

The parameters that characterize the probabilistic model are the mean and variance. The average is the average of the position, velocity, and acceleration in the probability model, and the variance is a parameter that gives variation from the average. Matching the end point and start point of successive movements corresponds to matching the average position of the end point and start point of each probability model.
<Average coupling>
First, average coupling will be described.

  In the j-th motion primitive represented by the probability model λ_j, the position average value E_x_s (λ_j), the velocity average value E_x'_s (λ_j), and the acceleration average value E_x``_s in a state s (λ_j). That is, in the j-th operation primitive, the average in the state s is E_y_s (λ_j) = [E_x_s (λ_j), E_x′_s (λ_j), E_x ″ _s (λ_j)] ^ T. Since the position of the jth motion primitive in the initial state s = 0 is E_x_0 (λ_j), when the coordinates are translated so that the initial position is the origin, the average of the position in the state s is (E_x_s (λ_j) −E_x_0 (λ_j)). Since this average is defined in the coordinate system corresponding to the jth motion primitive, it is necessary to convert it to the world coordinate system in order to combine the motion primitives. The world coordinate system is a coordinate system in which a probability model representing the first motion primitive in the motion sequence is defined. When the transformation into the world coordinate system is W_k, l, W_k, l is an affine transformation that depends on the direction k of the coordinate axis and the reference point l. The coordinate system is unified by translating the reference point to the origin and acting on W_k, l, and the average position E_x_S_j-1 (λ_j- in the final state S_j-1 of the immediately preceding (j-1) th motion primitive If the origin is translated by 1), the average of the probability model corresponding to the jth motion in the world coordinate system is obtained. Here, S_j is the final state in the j-th operation primitive, and it is assumed that E_x_S_0 (λ_0) = x_0 is the initial position of the trajector and E_x_S_n (λ_n) = x_n is the target position. When the above is expressed by an expression, it can be written as (Equation 2) to (Equation 5).

〈分散の結合〉
次に分散の結合について説明する。分散は平均からのばらつきを示すパラメータであり、トラジェクタが描く軌道のゆらぎを与える。複数の動作プリミティブを結合するにあたって、各動作プリミティブが静止オブジェクトに対して相対的な動作かそうでないかによって、軌道のゆらぎの許容範囲は変わる。すなわち、静止オブジェクトに対して相対的な動作の場合は、静止オブジェクトとの位置関係をある程度正確に保つ必要があるのに対して、そうでない動作の場合は、多少軌道の位置がずれていても問題になることはない。このような理由から、分散の結合においては、結合する動作が静止オブジェクトに依存していない場合は、位置に関する分散を拡大するように結合する。こうすることで、結合動作による滑らかな軌道を生成することができる。

<Combination of dispersion>
Next, dispersion coupling will be described. Dispersion is a parameter indicating variation from the average, and gives fluctuation of the trajectory drawn by the trajector. In combining a plurality of motion primitives, the allowable range of trajectory fluctuation varies depending on whether each motion primitive is relative to a stationary object or not. In other words, in the case of movement relative to a stationary object, it is necessary to maintain the positional relationship with the stationary object to a certain degree of accuracy, whereas in the case of movement that is not so, the position of the trajectory may be slightly shifted. There is no problem. For this reason, in the dispersion combination, when the operation to be combined does not depend on the stationary object, the combination regarding the position is expanded. By doing so, it is possible to generate a smooth trajectory by the coupling operation.

  In the j-th operation primitive represented by the probability model λ_j, the variance V_y_s (λ_j) in a certain state s is converted as follows. That is, if the reference point l_j of the j-th motion primitive is included in the position set O of the still object, the distribution of the j-th motion primitive is not changed, otherwise, the previous (j-1) th V_x_S_j-1 (λ_j-1) related to the position among the variances in the final state S_j-1 of the operation primitive is added. Nothing is converted in terms of velocity and acceleration among the variances. This can be expressed as (Equation 6) and (Equation 7).

平均の結合の際にはアフィン変換を行ったが、分散の結合においてはアフィン変換を行っていない。これは、HMMの結合においては、共分散行列の対角成分のみを用いることが一般的であるためであるが、分散に関しても平均の場合と同様にアフィン変換を行って計算してもよい。
〈確率モデルの結合〉
上述した確率モデルの結合を模式的に表すと図10から図12のようになる。

The affine transformation is performed in the average coupling, but the affine transformation is not performed in the dispersion coupling. This is because it is common to use only the diagonal component of the covariance matrix in the HMM coupling, but the variance may be calculated by performing affine transformation as in the case of the average.
<Combination of probability models>
The combination of the above probability models is schematically shown in FIGS.

  FIG. 10 represents the three motion primitives that are combined. 10 (a) is a probability model λ_1 representing a motion primitive that is moved away from the reference point, FIG. 10 (b) is a probability model λ_2 representing a motion primitive to be translated in the horizontal direction, and FIG. 10 (c) is a reference point. This is a probability model λ_3 representing an action primitive to be approached. The reference point in FIG. 10 (a) is the initial position of the trajector in the landmark set. The reference point in FIG. 10 (c) is the position of any stationary object in the landmark set. These three motion primitives defined in independent coordinate systems are combined to generate a combined operation in which the trajector is moved away from the initial position, moved horizontally, and then moved closer to a stationary object.

  FIG. 11 shows how the probability models λ_1, λ_2, and λ_3 are converted when the three operation primitives shown in FIG. 10 are combined. In the probability model column Λ = (λ_1, λ_2, λ_3), the probability model λ_1 is the first probability model, so nothing is converted. Since the probability model λ_2 is the second probability model, the origin is shifted according to (Equation 3), and then the affine transformation to the coordinate system in which the probability model λ_1 representing the immediately preceding motion primitive is defined, and The translation is performed so that the origin of the probability model λ_2 matches the average position of the final state of the immediately preceding probability model λ_1. For speed and acceleration, only affine transformation using W_k, l is performed. Regarding the variance, according to (Equation 7), the probability model λ_2 does not have the position of the stationary object at the reference point, so the variance of the probability model λ_1 representing the immediately preceding motion primitive is added. As a result, the variance of the probability model λ_2 can be expanded from the dotted line to the solid line as shown in FIG. 11, and the trajectory search can be expanded. Similarly to the case of the probability model λ_2, the probability model λ_3 is also subjected to coordinate transformation according to (Equation 3). The probability model λ_3 represents an action primitive that approaches the target position, and is a probability model having the position of a stationary object as a reference point. Therefore, no variance is converted according to (Equation 7).

As described above, the whole converted probability model is the combined probability model Λ ′, as shown in FIG. Under this joint probability model Λ ′, the most probable trajectory is searched.
<Search for maximum likelihood trajectory>
The search for the maximum likelihood trajectory is based on the probability model Λ ′ so that the probability of the trajector drawing the trajectory Y under the joint probability model Λ ′ is the likelihood P (Y; Λ ′), and its logarithmic likelihood is maximized. This is done by determining * and orbit Y *. This is expressed by the following equation (8). This solution can be obtained by the optimization method proposed in Non-Patent Document 3.

上式は、トラジェクタの目標位置を指定した命令がなされ、軌道探索方法1が採用された場合の式である。軌道探索方法1が用いられるのは、ロボットに「赤い箱を青い箱の上にのせる」と命令すれば、適切な軌道を計算するような場合である。この場合、始点と終点を与え、学習した動作プリミティブと認識されるランドマークからあらゆる〈動作プリミティブ、ランドマーク〉の組を生成し、生成された組を並べてできる動作列Aを作る。動作プリミティブの結合に制限を設けなければ、動作列Aは無限個存在するので、実際には、n個の動作プリミティブの結合までに制限する。このnを結合の深さと呼ぶ。生成される動作列Aの個数は、結合の深さをn、ランドマークの数をM、学習した動作プリミティブの数をZとすると、最大でもMZ^iをi=1からnまで足したもので表され、これは結合の深さnのべき乗で増大する。このため、動作プリミティブの数Zが少なくても、非常に多様な動作が可能となる。

The above expression is an expression when a command specifying the target position of the trajector is made and the trajectory search method 1 is adopted. The trajectory search method 1 is used when a robot is instructed to “put a red box on a blue box” to calculate an appropriate trajectory. In this case, a start point and an end point are given, and a set of all <action primitives, landmarks> is generated from landmarks recognized as learned action primitives, and an action sequence A in which the generated sets can be arranged is created. If there are no restrictions on the combination of action primitives, there are an infinite number of action sequences A, so in practice the restriction is limited to the combination of n action primitives. This n is called the bond depth. The number of motion sequences A generated is the sum of MZ ^ i from i = 1 to n at most, where n is the depth of connection, M is the number of landmarks, and Z is the number of learned motion primitives. Which increases with the power of the coupling depth n. For this reason, even if the number Z of operation primitives is small, very various operations are possible.

  For example, when generating a motion sequence A composed of all possible <motion primitive, landmark> pairs up to a depth n = 3, the motion sequence A is combined with the above-described probability model representing the motion primitive. The corresponding probability model Λ ′ is determined. Under the probability model Λ ′, the probability P of the trajector drawing the trajectory Y is determined. Therefore, Λ ′ and Y are determined so that the probability P is maximized, that is, a sequence of pairs of <operation primitive, landmark> The orbit search method 1 determines the most probable orbit at that time.

  The trajectory search method 2 is used when the robot is instructed by specifying a trajectory, such as “lift a red box, then skip the box, and lower it”. In this case, since the motion sequence A in the trajectory search method 1 is given, the search for the sequence of <motion primitive, landmark> is not performed. That is, the trajectory Y * is generated according to (Equation 9).

上述の、(数8)および(数9)のいずれにおいても、トラジェクタが静止オブジェクトに衝突する効果は考慮していないが、これは、例えば、軌道Yのうち静止オブジェクトに衝突する軌道は除外した上で、最尤軌道を選択するという条件を入れることで回避できる。

In both of (Equation 8) and (Equation 9) described above, the effect of the trajector colliding with a stationary object is not considered, but this excludes, for example, the trajectory Y that collides with a stationary object. This can be avoided by adding the condition of selecting the maximum likelihood trajectory.

  FIG. 15 is a diagram showing a state of searching for the maximum likelihood trajectory under the joint probability model Λ ′. Under the joint probability model Λ ′, there are a number of trajectories Y that move the trajector from the initial position to the target position, as shown by the solid line and the dotted line in FIG. For each of these trajectories Y, the probability P (Y; Λ ′) of the trajector drawing the trajectory Y under the joint probability model Λ ′ is calculated. For example, the probability of drawing the trajectory represented by the solid line in FIG. If it is maximum, the solid line trajectory is the maximum likelihood trajectory Y *.

  If the motion sequence A is given and the joint probability model Λ ′ is determined to be unique, as in the case where the motion is specified by the trajectory search method 2, the maximum likelihood trajectory is searched only in the joint probability model Λ ′. However, when the command is issued by the trajectory search method 1, it is necessary to search for which operation primitives are combined in what order. FIG. 14 is a diagram illustrating an example of realizing the operation of placing the trajector at the initial position x_0 on the stationary object at the target position x_n by combining several operation primitives.

  FIG. 14A shows a joint probability model Λ′_1 that represents a joint action in which the three action primitives “Hanasu”, “Move horizontally”, and “Crick” are combined as shown in FIG. FIG. 14 (b) is a joint probability model Λ′_2 representing the joint action in the case where the action of “overshoot” is performed twice in succession. FIG. 14 (c) is a joint probability model Λ′_3 representing a joint action in which three action primitives of “raise”, “move horizontally”, and “sales” are joined. FIG. 14D shows a joint probability model Λ′_4 representing a joint action obtained by joining two action primitives “Hanasu” and “Sakeru”.

There may be other motion sequences composed of motion primitives that move the projector from the initial position x_0 to the target position x_n. In the search for the maximum likelihood trajectory, all motion sequences up to a predetermined depth are generated, and the probability of drawing the trajectory Y is calculated for each motion sequence. As a result of searching the trajectory Y * and the joint probability model Λ ′ * that maximize the probability P (Y; Λ ′) that the trajectors calculated under the respective joint probability models Λ ′ draw the trajectory trajectory Y, for example, FIG. If the probability P (Y_4; Λ′_4) for drawing the trajectory Y_4 under the joint probability model Λ′_4 in (d) is found to be the maximum, the maximum likelihood trajectory is determined as Y * = Y_4. This is indicated by the solid line in FIG.
《Movement instruction specifying the target position》
FIG. 16 shows some examples of maximum likelihood trajectories searched by the trajectory search apparatus of this embodiment when the target position is designated and the movement of the trajector is instructed. In this case, the maximum likelihood trajectory and the maximum likelihood action sequence are searched according to the trajectory search method 1. In FIG. 16, when the trajector is 1, and instructed to move to the target position shown in (a) to (g) in the space where there are 2 to 5 stationary objects, the searched maximum likelihood trajectories are respectively Shown in dotted lines. FIG. 17 shows an operation sequence that gives these maximum likelihood trajectories.

For example, in the case of (a), movement is performed from the stationary object 2 with the action primitive “Hanasu”. In the case of (b), the movement is performed with the action primitive “put” on the stationary object 5. In (c), the movement is performed by an action sequence including two action primitives, and the movement is performed by “putting” the stationary object 2 and then “flicking” the stationary object 5. (d) follows the same trajectory as (c) until it is “placed” on the stationary object 2, and then the movement is performed by “shaking” from the stationary object 2. In (e), after moving the stationary object 2 “over”, the movement is performed by “breaking” the stationary object 3. In (f), after moving from the stationary object 2, the movement is performed by “flying” the stationary object 4. In (g), movement is performed by “chipping” the stationary object 2 and then “chipping” the stationary object 3. As can be seen from the trajectory (g) passing through the stationary object 2, in this example, the collision between the trajector and the stationary object is not considered. However, as described above, this can be easily avoided by excluding the trajectory with a collision from the maximum likelihood trajectory.
<< Movement instruction with action sequence >>
FIG. 18 shows some examples of maximum likelihood trajectories searched by the trajectory search apparatus of this embodiment when the movement sequence is instructed by designating the motion sequence. In this case, according to the trajectory search method 2, a search for the maximum likelihood trajectory is performed under a joint probability model representing a given motion sequence. FIG. 18 (a) shows the maximum likelihood trajectory searched in the case where it is instructed to “turn object 1 over object 2 and then lower it to object 4”. FIG. 18 (b) shows the maximum likelihood trajectory searched in the case where it is instructed that “object 2 is made to fly over object 1, object 1 is made to fly again, and placed on object 5”.
<Operation of manipulator control system>
In this section, the operation in the trajectory search mode of the manipulator control system using the trajectory search device of the present embodiment will be described with reference to the flowchart of FIG.

  First, in the trajectory search mode, it waits for an operation instruction by the user's voice and gesture (step S101). When an instruction from the user is confirmed in the voice input from the microphone and the video taken by the camera (step S101 Y), the voice recognition engine 100 recognizes the instruction content (step S102). That is, the operation primitive included in the trajector, the landmark, and the instruction is extracted from the command issued by the user.

  From the following, the process branches depending on whether the user designates and instructs the target position of the trajector or designates the operation sequence. This is determined by whether or not the input command includes the target position x_n of the trajector (step S103). If the target position x_n of the trajector is included (step S103 Y), it is determined that the command has designated the target position, and the process proceeds to step S111. If the target position x_n of the trajector is not included (step S103 N), it is determined that the command is for specifying an operation sequence (step S103 N), and the process proceeds to step S104.

  In the process of an instruction designating an action sequence, it is first determined whether or not an action sequence A consisting of a set of <action primitive, landmark> has been input (step S104). If the operation sequence A has not been input (step S104 N), the process returns to the instruction waiting state in step S101. If the motion sequence A has been input (step S104 Y), the probability models representing the motion primitives included in the motion sequence A are combined (step S105). Next, the maximum likelihood trajectory Y * is searched under the joint probability model Λ ′ (step S106), and the searched maximum likelihood trajectory Y * is transmitted to the manipulator side via the control device 400 (step S107). ).

  In the process of the command specifying the target position, first, the image recognition engine 200 analyzes the image captured by the camera, and recognizes the initial position x_0 and the target position x_n of the trajector (step S111). Next, a plurality of motion primitive motion sequences that move the trajector from the initial position x_0 to the target position x_n are generated (step S112), and each motion primitive included in the motion sequence is represented for each generated motion sequence. Probability models are combined (step S113). The maximum likelihood trajectory is searched under each of the combined probability models, and the maximum likelihood trajectory Y * in the generated operation sequence and the combined probability model Λ ′ representing the combined operation at that time are searched (step S114). ).

  Since the trajectory of the searched joint motion is not necessarily the motion intended by the user, before sending the maximum likelihood trajectory Y * actually searched to the manipulator side, the action scheduled motion name is spoken. Announcement is made (step S115). The scheduled action name is an action name including the names of action primitives included in the combined action in the order of execution. For example, in the example of FIG. 15 (d), the announcement is such that “the trajector is moved away from the initial position to the top of the box and then lowered”. If the user gives an OK instruction to the announcement (step S116 Y), the searched maximum likelihood trajectory Y * is transmitted to the manipulator side via the control device 400 (step S107). If the user does not give an OK instruction for the announcement (N in step S116), the process returns to the instruction waiting state in step S101.

The above is the operation in the trajectory search mode of the manipulator control system using the trajectory search apparatus of the present embodiment.
《Probability model combination method》
In this section, a method of combining probability models in the trajectory search mode of the trajectory search apparatus of this embodiment will be described with reference to the flowchart of FIG.

Here, a sequence of probability models Λ = (λ_1, λ_2, ..., λ_n) consisting of probability models λ_j (j = 1,2, ..., n) representing n motion primitives are combined and combined Consider creating a probabilistic model Λ ′.
First, it is assumed that E_x_S_0 (λ_0) = x_0 and V_x_S_0 (λ_0) = 0 as initial values of the average and variance regarding the position (step S201). Here, x_0 is the position of the trajector in the initial state.

  An index for designating a probability model representing each operation primitive is initialized to j = 1 (step S202), and loop processing relating to the operation primitive is started. In order to continuously combine the probability model from the position in the final state S_j-1 of the immediately preceding action primitive represented by the probability model λ_j-1, E′_j is calculated according to (Equation 3) (step S203). In order to determine the dispersion expansion amount V′_j, it is determined whether or not the reference point of the motion primitive represented by the probability model λ_j belongs to the stationary object position set O according to (Equation 7) (step S204). If the reference point belongs to the stationary object position set O (step S204 Y), V′_j is set to 0 (step S205) because the variance is not expanded. If the reference point does not belong to the stationary object position set O (N in step S204), the component relating to the position of V′_j is set to the final state S_j−1 of the motion primitive immediately before represented by the probability model λ_j−1. The position variance V_x_S_j-1 is set at (step S206).

Next, an index indicating a state is initialized to s = 0 (step S207), and a loop process relating to the state in the process relating to each probability model is entered. First, average conversion is performed for each state s according to (Equation 2) (step S208). Subsequently, the variance is converted according to (Equation 6).
If the index s indicating the state is not larger than the index S_j indicating the final state of the motion primitive represented by the probability model λ_j being processed (N in step S210), s is incremented (step S211), and the process returns to step S208. . If the index s indicating the state is larger than the index S_j indicating the final state of the motion primitive represented by the probability model λ_j being processed (Y in step S210), the state loop is exited and the process proceeds to step S212.

  If the index j for designating the probability model representing the action primitive is not larger than the number n of action primitives included in the probability model column Λ (step S212 N), j is incremented (step S213), and step S203 is entered. Return. If the index j for designating the probability model representing the action primitive is larger than the number n of action primitives included in the probability model column Λ (step S212 Y), the process leaves the loop for the action primitive and proceeds to step S214.

Finally, in order to match the end point of the trajectory to the target position x_n, the average and variance regarding the position in the final state S_n of the probability model λ_n representing the last motion primitive are set to E_x_S_n (λ_n) = x_n, V_x_S_n (λ_n) = 0 (Step S214).
The above is the method of combining the probability models in the trajectory search mode of the trajectory search apparatus of the present embodiment.

  By incorporating the trajectory search apparatus of the present invention into an industrial robot, it is possible to realize a robot with applied power that can flexibly respond to various on-site situations. Moreover, it can utilize for the use of searching for the path | route until it arrives at the destination, responding to a user's request | requirement by mounting in a motor vehicle.

The figure which shows the structure of a manipulator control system. The figure which shows the example of the operation | movement which operates a thing. The figure which shows the example of the input image | video for making a robot learn. The figure which shows typically the probability model (lambda) _j showing an operation | movement primitive. The figure which shows the data structure of the parameter of a probability model. The figure which shows the internal structure of the orbit search apparatus in learning mode. The figure which shows the example of the learning result of an operation primitive. The figure which shows the internal structure of the orbit search apparatus in an orbit search. An example of coordinate transformation in the combination of motion primitives. The figure which shows the example of the operation | movement primitive before combining. The figure which shows the relationship of the average and dispersion | distribution in the combination of an operation | movement primitive. The figure which shows the parameter in a joint probability model. The figure which shows the search of an orbit under a joint probability model. The figure which shows the joint pattern of some probability models. The figure which shows the maximum likelihood orbit searched from the coupling pattern of some probability models. The figure which shows the example of the maximum likelihood locus | trajectory at the time of designating a target position and performing a movement instruction | indication. The figure which shows the operation | movement row | line searched for every target position. The figure which shows the example of the maximum likelihood locus | trajectory at the time of designating a movement sequence and performing a movement instruction | indication. The flowchart which shows operation | movement of a manipulator control system. The flowchart which shows the combining method of a probability model.

Explanation of symbols

100: Speech recognition engine
200: Image recognition engine
300: Orbit search device
400: Control device
500: Manipulator
1: Input / output interface
2: Database storage
3: Processor
4: Work memory
5: Program storage

Claims (9)

A trajectory search apparatus for searching a trajectory for moving a trajector,
Accepting means for accepting from the user an instruction to move the trajector including operation primitives that can be performed by the drive device for moving the trajector;
A search means for generating a motion sequence composed of motion primitives included in an instruction from the user received by the reception means, and searching for a maximum likelihood trajectory within the range of the generated motion sequence;
Each motion primitive that the drive can perform is represented by a probability model,
The probabilistic model is a model that defines how the state of the trajector should transition over time using the probability distribution for each time point and the transition probability between states,
The trajectory search by the search means is
A trajectory search apparatus characterized by combining probability models representing motion primitives for the motion sequences, and searching for a trajectory having the maximum likelihood under the joint probability model obtained by the combining. The combination of probability models representing motion primitives is
The average of the position, speed, and acceleration of the trajector in each state of the motion primitive to be executed after the second in the motion sequence, the average of the position, speed, and acceleration of the trajector in that state, and the average of the trajector in the initial state of the motion primitive 2. The trajectory search apparatus according to claim 1, wherein the trajectory search apparatus is calculated by using an average of positions and an average of positions in the final state of the immediately preceding operation primitive. The trajectory search device includes:
Capture means for capturing the trajectory of the trajector;
Recognizing means for recognizing a stationary object arranged in a space in which the driving device operates,
The combination of probability models representing motion primitives is
A determination is made as to whether or not the reference point that is the basis of the operation primitive to be combined matches any of the recognized positions of the stationary object,
If the reference point serving as the reference of the action primitive matches any of the recognized positions of the still object, the distribution of the action primitives to be executed in the second and subsequent stages of the action sequence in the combination is performed. And variance in each state of
If the reference point that is the basis of the motion primitive does not match any of the recognized positions of the stationary object, the dispersion in each state of the motion primitive to be executed after the second in the motion sequence in the above-mentioned combination is performed. 2. The trajectory search device according to claim 1, wherein the trajectory search device is made by a sum of a variance in each state of the primitive and a variance in the final state of the immediately preceding operation primitive. The trajector movement instruction includes the initial position and the target position of the trajector,
2. The search unit generates a plurality of operation sequences of operation primitives corresponding to a path from an initial position of a trajector to a target position, and combines probability models for each of the operation sequences. The trajectory search device described. The trajector movement instruction includes multiple combinations of operation names and landmarks of operation primitives,
The operation primitives to be combined with the probability model by the search means are:
2. The trajectory search device according to claim 1, wherein the trajectory is a motion primitive corresponding to a plurality of motion names included in the movement instruction. In the binding,
3. The trajectory search apparatus according to claim 2, wherein coordinate transformation is performed so as to unify a coordinate system in which two probability models to be combined are defined. The coordinate transformation is
7. The coordinate system in which the probability model representing the immediately preceding motion primitive is defined, is affine transformed with respect to the coordinate system in which the probability model representing the motion primitive is defined. Orbit search device. The probability model representing the motion primitive is:
2. The trajectory search device according to claim 1, wherein the trajectory search device is determined based on an operation performed by a user. Before the drive device performs the operation represented by the trajectory with the maximum likelihood, the user is asked to confirm,
2. The trajectory search device according to claim 1, wherein the trajectory is not output when the user does not agree.

Images