JP2017047505A - Gripping possibility determination method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a gripping possibility determination method in which a gripping success rate is enhanced while increasing the number of components capable of being gripped.SOLUTION: A gripping possibility determination method includes: moving a hand from an approach point to a gripping point, and acquiring sample data in which a 3-dimensional point group in a hand passage area before gripping a gripping object is associated with the gripping success of the gripping object; defining the i-th point of the 3-dimensional point group in the hand passage area as p; defining a distance with the farthest point from the point pas h(p) among intersections of a straight line formed by extending the point pin the approach direction and the boundary of the hand passage area, defining a distance with the closest point from the point pas d(p) among the intersections of the straight line extended from the point pin the grip direction of the hand and the boundary of the hand passage area, assuming a gripping success rate to be high with respect to one having the smaller values of h(p), d(p) and n, and generating a discriminator from a feature vector calculated on the basis of the sample data and the distances h(p) and d(p); and determining the gripping possibility of the gripping object using this discriminator.SELECTED DRAWING: Figure 4

Description

  The present invention relates to a gripping possibility determination method for determining whether or not a gripping object can be gripped by a hand when gripping a plurality of gripping objects stacked in a box by a hand of a robot arm. .

  When gripping a stacked object to be gripped with a robot arm hand, if the robot arm interferes with another object, a device that stops the operation of the robot arm or avoids the object is known. (See Patent Document 1).

JP 2002-331480 A

In the apparatus described above, when the robot arm interferes with another object, the operation of the robot arm is stopped, so that the number of parts that can be gripped may be reduced. On the other hand, if gripping is executed while ignoring other objects, the number of parts that are determined to be grippable increases. For example, there is a possibility that the gripping success will be reduced due to strong collision with other objects.
The present invention has been made in view of such problems, and a main object of the present invention is to provide a gripping possibility determination method capable of increasing the gripping success rate while increasing the number of grippable parts.

One aspect of the present invention for achieving the above object is to determine whether or not the hand can grip the gripping object when gripping a plurality of stacked gripping objects with a hand of a robot arm. A method for determining whether or not a grip is possible, wherein the hand is moved from an approach point that is a predetermined distance away from the gripping object to a gripping point that grips the gripping object by the hand, and the gripping object is moved to the hand at the gripping point. The area through which the hand passes is defined as a hand passage area, and the three-dimensional point cloud data included in the hand passage area is associated with success or failure of gripping the object to be grasped by the hand. obtaining a plurality of sample data, the hand passing the three-dimensional point cloud Fukumaru in the region i-th point of the _{p i (i = 1,2, ···} n) and, the points p _i Among straight line extending in the approach direction of the point of intersection with the boundary of the hand passing area, the distance between the point farthest from the point p _i and h (p _i), extend from the points p _i in the gripping direction of the hand Among the points where the straight line intersects the boundary of the hand passing area, the distance from the point p _i to the closest point is d (p _i ), and the values of the h (p _i ), d (p _i ), and n assuming that that there is a high gripping success rate smaller, from the calculated feature vectors on the basis of said sample data and the distance h (p _i) and d (p _i), for determining the grasping whether the grasped object And a step of determining whether or not the hand can be gripped by the hand using the generated discriminator.
According to this aspect, even when an obstacle is present in the hand passage area, it is not immediately determined that the object to be grasped cannot be grasped, but a determination as to whether or not the object can be grasped is performed using the classifier. For this reason, the number of parts that can be gripped can be increased. When determining whether or not the grip is possible, the grip success rate can be increased by using a discriminator that learns the grip pattern of the hand that has few strong interference points between the hand and the obstacle and has a high grip success rate. That is, the gripping success rate can be increased while increasing the number of parts that can be gripped.
In this aspect, the discriminator may be composed of a random forest, and the feature vector may be calculated using the following equation. According to this aspect, it is possible to increase the gripping success rate while increasing the number of parts that can be gripped using the discriminator configured by a random forest.
In this aspect, the discriminator may be composed of a linear SVM (support vector machine), and the feature vector f _s may be calculated using the following equation. According to this aspect, it is possible to increase the gripping success rate while increasing the number of parts that can be gripped by using the discriminator configured by the linear SVM.

  According to the present invention, it is possible to provide a gripping possibility determination method capable of increasing the gripping success rate while increasing the number of grippable parts.

1 is a block diagram illustrating a schematic system configuration of a transfer robot according to an embodiment of the present invention. It is a figure which shows the picking operation | movement of the hand of a robot arm. It is a figure which shows a hand passage area. Distance is a diagram illustrating a h _{(p i)} and d _{(p i).} It is a figure which shows a learning phase and an execution phase. It is a flowchart which shows the flow of the production | generation method of the holding | grip movement trajectory in a learning phase. It is a flowchart which shows the flow of the production | generation method of the holding | grip movement track | orbit in an execution phase.

Embodiments of the present invention will be described below with reference to the drawings.
The transfer robot according to the present embodiment picks up (holds) parts, which are picking objects (gripping objects) housed in a plurality of boxes, and conveys them.

  FIG. 1 is a block diagram showing a schematic system configuration of a transfer robot according to an embodiment of the present invention. The transfer robot 1 according to the present embodiment includes a robot arm 2, a movable carriage 3, a control device 4 that controls the robot arm 2 and the movable carriage 3, and a detection unit 5 that detects surrounding environmental information. .

  The robot arm 2 is fixed to the moving carriage 3. The robot arm 2 is configured as an articulated arm having a plurality of joint portions such as a wrist joint, an elbow joint, and a shoulder joint, for example. A hand having a pair of finger parts (grippers) (hereinafter simply referred to as a hand) is provided at the tip of the robot arm 2. Each joint is provided with an actuator such as a motor for driving each joint. Each joint is provided with a rotation sensor such as an encoder, a potentiometer, and a resolver that detects the rotation angle of each joint. The control device 4 controls the actuator of each joint by transmitting a control signal to the actuator of each joint. Moreover, the control apparatus 4 controls the opening / closing operation | movement of a finger part by controlling the actuator of a pair of finger part. Thereby, the finger part of the hand grips the component. The configuration of the robot arm 2 is merely an example, and is not limited to this, and any configuration can be applied.

  The movable carriage 3 includes, for example, left and right drive wheels, a motor that drives the drive wheels, a speed reducer, an encoder, and the like. In addition, the moving carriage 3 is provided with a box for placing components. The transport robot 1 places the parts gripped by the robot arm 2 on a box and transports them. The encoder outputs the detected rotation information of the motor to the control device 4. The control device 4 controls the drive wheels of the movable carriage 3 by transmitting a control signal to each motor based on the rotation information output from the encoder.

The control device 4 controls the driving of the robot arm 2 and the movable carriage 3 based on the surrounding environment information (hereinafter referred to as surrounding environment information) output from the detection unit 5.
The control device 4 includes, for example, a CPU (Central Processing Unit) 4a that performs control calculation, calculation processing, and the like, a ROM (Read Only Memory) or a RAM (Random Access Memory) that stores a control program executed by the CPU 4a, a calculation program, and the like. ), And a microcomputer composed of a plurality of interface units (I / F) 4c for inputting / outputting signals to / from the outside. The CPU 4a, the memory 4b, and the interface unit 4c are connected to each other via a data bus or the like.

  The detection unit 5 detects environmental information around the transfer robot 1. The detection unit 5 has a visual sensor 51. As the visual sensor 51, for example, a Kinect (registered trademark) sensor can be used. However, the detection unit 5 is not particularly limited as long as it is a sensor that can detect environmental information around the transfer robot 1. The detection unit 5 outputs the detected surrounding environment information to the control device 4.

  For example, the transfer robot 1 uses the robot arm 2 to hold the parts of the object to be held housed in a plurality of boxes from the boxes, and places them on the passing box of the movable carriage 3 for transfer. Here, when a plurality of parts stacked in a specific box are gripped by the hand of the robot arm 2, it is necessary to determine whether or not the part can be gripped (pickable) by the hand.

  FIG. 2 is a diagram illustrating a picking operation of the hand of the robot arm. When the robot arm 2 grips each component, first, the control device 4 controls the actuator of each joint portion of the robot arm 2 to move the hand 21 to an approach point away from the component by a certain distance (a). Then, the control apparatus 4 controls the actuator of each joint part of the robot arm 2, and moves the hand 21 from the approach point to a gripping point (b). And the control apparatus 4 controls the actuator of the hand 21 so that the finger | toe part 22 of the hand 21 may hold | grip components at the holding point (c).

  By the way, a first passing region through which the hand passes when the hand moves from the approach point to the gripping point, and a second passing region through which the hand passes by the gripping point to complete gripping (first and second). An obstacle may exist in the hand passing area (hatched portion) (hereinafter referred to as “Sweeping Volume”) (FIG. 3).

In this way, when there is an obstacle in the hand passing area S (when the hand 21 interferes with the obstacle), if it is simply determined that the part cannot be gripped, the gripping success rate when actually gripping is performed. However, the number of parts that can be gripped may decrease.
On the other hand, if the grip is executed while ignoring the obstacle in the hand passing area S, the number of parts that are determined to be grippable increases. However, for example, there is a possibility that the success rate of gripping when the hand 21 strongly collides with an obstacle and actually grips is lowered.

  On the other hand, in the grasping possibility determination method according to the present embodiment, the discriminator that learns which position in the hand passing area S has failed (or succeeded) when the hand 21 grasps the component. And using this discriminator, it is determined whether or not the hand 21 can hold the component. As a result, even when an obstacle exists in the hand passing area S, it is not immediately determined that the part cannot be gripped, but a determination as to whether or not the part can be gripped is performed using a discriminator. For this reason, the number of parts that can be gripped can be increased. When determining whether or not gripping is possible, a discriminator is used which learns at which position in the hand passage area S there is an obstacle (or success) when gripping a part by the hand 21 fails. For this reason, even if there is an obstacle in the hand passing area S, the success rate of gripping can be increased. That is, the gripping success rate can be increased while increasing the number of parts that can be gripped.

For example, if the hand 21 comes into contact with other parts that are stacked or lightly pressed (in such a case, it is estimated that the gripping success rate is high), the part 21 is pushed and the gripping target is In some cases, a part can be gripped. Therefore, it is possible to grip such a component that has been determined to be impossible to grip because the hand 21 interferes with other components.
The gripping possibility determination method according to the present embodiment appropriately allows contact between the hand 21 and a component adjacent to the gripping target component in the gripping operation of the parts stacked by the robot arm 2 as described above. Thus, the gripping success rate is increased while increasing the number of grippable parts.

  In the gripping possibility determination method according to the present embodiment, a gripping success condition when there is an obstacle in the hand passing area S (the position of the obstacle in the hand passing area S is successful in gripping the part by the hand 21). Is generated by the following learning. That is, a discriminator learning a grip pattern having a high grip success rate is generated.

  The control device 4 associates the point cloud (three-dimensional point cloud data) representing the obstacle included in the hand passing area S in the past hand gripping operation with the gripping success / failure of the part to be gripped by the hand 21. Sample data obtained as learning data. And the control apparatus 4 produces | generates the discriminator for determining a holding | gripping success or failure using the acquired learning data. The point cloud is obtained by excluding points near the part to be grasped.

In the present embodiment, a discriminator is generated in which a feature vector represents how far an obstacle (point cloud) is from the hand 21, in other words, how much the obstacle interferes with the hand 21. Here, the i-th point included in the hand passage area is defined as p _i (i = 1, 2,..., N), and the following variables are defined (FIG. 4).

h (p _i): Among the straight line extended from the point p _i in the approach point direction of the point of intersection between the hand passing area boundary distance of a point farthest from p _i. The distance by which the point p _i is pushed in the approach direction.
d (p _i): Of the point p _i straight line extending in the closing direction (gripping direction) of the finger portion of the point of intersection between the hand passing area boundary distance of a point closest from p _i. The distance at which the point p _i enters the inside of the hand passing area (the closing direction of the finger portion 22).

When h (p _i ) and d (p _i ) are large, this indicates that the obstacle strongly interferes with the hand 21.
Here, it is assumed that if the number of points of strong interference between the hand 21 and the obstacle is reduced, the success rate of gripping parts by the hand 21 is increased. That is, it is assumed that the gripping success rate is higher when the values of h (p _i ), d (p _i ), and n are smaller. Based on this assumption, the classifier is generated using the feature vector generated when the grasping success rate is high.

As described above, in the present embodiment, as described above, the distance h (p _i ) and the distance d (p _i ) of the point pi are defined, and h (p _i ), d (p _i ), and n Assuming that the smaller the value is, the higher the success rate of gripping, and the identification for determining whether or not gripping is possible from the feature vector calculated based on the sample data and the distance h ( _pi ) and d ( _pi ) Create a container. By using this discriminator, even when an obstacle is present in the hand passage area S (when the hand 21 is in contact with the obstacle), there are few strong interference points between the hand 21 and the obstacle, and the gripping success rate is high. Since the grip pattern of the hand 21 can be selected, the gripping success rate can be increased.

  The discriminator is composed of, for example, a linear SVM (support vector machine) or a random forest. The discriminator is not limited to the above example, and can be configured by an arbitrary learning algorithm.

First, a method for defining the feature vector of the random forest will be described.
A point distribution in the hand passing area is used as the feature vector. For this reason, a two-dimensional histogram using h (p _i ) and d (p _i ) is defined as follows. In the definition below, binwidth is the bin width, and bin_h and bin_d are bin upper limit values (threshold values for determining unsuccessful grasping failure) of each axis.
Distance from the feature vector = _{(bottom: min ((h (p i} ) / binwidth), bin_h),
Distance from side: min ((d (p _i ) / binwidth), bin_d))

  Enumerate objective variables (successful gripping, gripping failure). The control device 4 calculates a gripping success rate for an arbitrary feature vector by calculating a gripping success rate when reaching each leaf node from the sample data. When the discriminator is composed of a random forest, it is preferable to use random data with no region bias as the sample data. The control device 4 may determine whether or not gripping is possible by setting a threshold for the gripping success rate according to the situation. When it is desired to increase the gripping success rate, this threshold value is set large.

Next, a method for defining a feature vector of linear SVM will be described.
The feature vector is defined as following a two-dimensional vector f _s.

  The control device 4 calculates a gripping success rate in each region based on the sample data. The control device 4 calculates a gripping success rate by identifying a region belonging to an arbitrary feature vector. In addition, the control device 4 determines whether or not it can be gripped by setting a threshold value according to the situation. When it is desired to increase the gripping success rate, this threshold value is set large.

  As shown in FIG. 5, the control device 4 according to the present embodiment executes a learning phase for generating the discriminator described above and an execution phase for performing an actual gripping operation using the discriminator generated in the learning phase. To do.

  In the learning phase, the control device 4 acquires preconditions such as component information and geometric information of the transfer robot 1 (1). Next, the control device 4 derives a grippable position / posture for each part and constructs a gripping position / posture database (2). Thereafter, the detection unit 5 detects the position and orientation of the component (3). Further, the control device 4 controls the robot arm 2 so as to actually pick the part based on the gripping position / posture database and the detected position / posture of the part (4). At this time, as described above, the control device 4 uses the sampling data associating the point cloud included in the hand passing area and the gripping success / failure of the gripping target part by the hand 21 as learning data to determine the gripping success / failure. Learning and generating a classifier (5).

  Subsequently, in the execution phase, the detection unit 5 detects the position and orientation of the component (1). Next, the control device 4 selects a gripping position / posture with a high gripping success rate using the discriminator, and controls the robot arm 2 to pick a component based on the detected position / posture of the component ( 2). Hereinafter, the learning phase and the execution phase described above will be described in detail.

<< Learning Phase >>
(1) Acquisition of precondition First, the precondition of the learning phase mentioned above is demonstrated.
(Premise 1) It is assumed that the parts held by the robot arm 2, the geometric model of the robot arm 2 itself, and the weights of the parts are given in advance.
(Assumption 2) The types of parts contained in each box are known, but the position of the parts in the box and the number of parts are unknown.
For example, the memory 4b stores in advance information such as the above-described geometric model of the part gripped by the robot arm 2, the geometric model of the robot arm 2 itself, the weight of the part, and the type of part contained in each box. It is remembered.

(2) Construction of a database of gripping positions and orientations Subsequently, before the robot arm 2 starts a gripping operation, a database of positions and postures (hereinafter referred to as gripping positions and postures) at which the hand 21 can grip each component is constructed. The control device 4 constructs a gripping position and orientation database for each component, for example, according to the following <Phases 1 to 6>.

<Phase 1>
First, the control device 4 derives an evaluation index for determining the optimum number of divisions for the part shape (geometric model). More specifically, the control device 4 merges clusters until the following condition 1 is satisfied by the area division number determination method.

但し、vol_maxは最大クラスタ領域のバウンディングボックスの体積、overlap_voliは最大クラスタ領域とクラスタｉのバウンディングボックスが重なった領域の体積である。 <Condition 1>
Error> 0.01 and cluster overlap degree <0.2, and the cluster overlap degree is derived from the following equation.

Where vol _max is the volume of the bounding box of the maximum cluster region, and overlap _voli is the volume of the region where the maximum cluster region and the bounding box of cluster i overlap.

<Phase 2>
The control device 4 uses a general method (for example, non-patent literature: Sakai, Harada, Yamanobe, Nagata, Nakamura, Hasegawa, etc.) so as to minimize the evaluation index derived in <Phase 1>. Dividing (clustering) based on "Grip posture conversion for general object gripping", Robotics Symposia Lectures, pp.219-224, 2013). As a result, a plurality of quadric surfaces are assigned to the geometric model of the part.

<Phase 3>
The control device 4 derives these bounding boxes for those to which an ellipse or a circle is assigned as a quadric surface. Note that cluster regions whose three side lengths of the bounding box are smaller than 0.15 times the whole side bounding box length (however, the multiple can be changed as appropriate) are excluded from the gripping point search target. .

<Phase 4>
The control device 4 uses the bounding box derived in <Phase 3> as, for example, non-patent literature: Harada, Kaoru, Kaneko, Kanehiro, Maruyama, "Multi-fingered hand grasp plan based on a rectangular parallelepiped model", Journal of the Japan Society of Mechanical Engineers. (Part C), vol.76, no.762, pp.331-339, GRC (Grasping Rectangular Convex) defined in 2010, assuming discrete points on the GRC surface and approaching this point The hand 21 is virtually moved.

<Phase 5>
The control device 4 virtually closes the finger portion 22 from the posture of the hand 21 in <Phase 4> until the finger portion 22 of the hand 21 comes into contact with a component.

<Phase 6>
When all the finger parts 22 are virtually in contact with the parts, the control device 4 is, for example, a non-patent document: Harada, Sakai, Utsu, Yamanobe, Nagata, “soft finger type gripping stability analysis”, automatic measurement Grip stability is checked by the method disclosed in Proceedings of the Society of Control Engineers System Integration Division, 2013. The control device 4 checks each gripping position / posture, and excludes the gripping position / posture determined to not satisfy the gripping stability condition.

  The control device 4 can construct a gripping position / posture database for each component by repeating <Phase 5> and <Phase 6> on the discrete points described above. Each data in the gripping position / posture database includes the position data of the wrist joint of the robot arm 2 as viewed from the component coordinate system, the posture matrix data of the wrist joint of the robot arm 2 as viewed from the component coordinate system, and the finger of the hand 21 The joint angle vector data of the joint of the unit 22 is included.

  The control device 4 performs the above-described <Phase 1> to <Phase 6> for all the pairs of the component and the hand 21 and holds the gripping position of the hand 21 with respect to all the pairs of the component and the hand 21. A posture database is constructed and stored in the memory 4b. Note that the database of the grip position / posture of the hand 21 may be stored in the memory 4b in advance.

(3) Measurement of position and orientation of part The position and orientation (hereinafter referred to as position and orientation) of the part in the box that is the gripping object are measured.
The transfer robot 1 starts its operation, and the control device 4 controls the motor to move the moving carriage 3 to the position of the robot arm 2 included in the gripping position / posture of the database constructed as described above. And the detection part 5 measures the position and orientation of components using the visual sensor 51 in the position.

  A plurality of parts are contained in the box, and the position and orientation of one of these parts is measured. The detection unit 5 acquires the position and orientation of a part using, for example, a PCL (Point Cloud Library) algorithm. The PCL algorithm is, for example, non-patent literature: A. Aldoma, F. Tombari, RBRusu, and M. Vincze, “OUR-CVFH-Oriented, Unique and Repeatable Clustered Viewpoint Feature Histogram for Object Recognition and 6DOF Pose Estimation. "A. Pinz et al. (Eds.): DAGM / OAGM 2012, LNCS, pp. 113-122, 2012, which is described in detail.

  The detection unit 5 uses the visual sensor 51 to detect a point cloud (three-dimensional point cloud data) of the components in the box. The detection unit 5 detects the position and orientation of the component from the point cloud by matching the geometric model of the component with the detected point cloud using a PCL algorithm. Note that the position and orientation of the components in the box may be stored in advance in the memory 4b.

(4) Grasping operation (Picking operation)
The control device 4 actually operates the robot arm 2 and generates a gripping motion trajectory of the robot arm 2 at that time as follows.
As described above, the visual sensor 51 measures the position and orientation of the components contained in the box. At this time, the three-dimensional position vector indicating the position of the part viewed from the absolute coordinate system is set as Po, the 3 × 3 rotation matrix indicating the position and orientation of the part is set as Ro, and the i-th gripping object among the recognized parts is set. They are defined as Po _i and Ro _i , respectively.

As described above, the database of the gripping position and orientation of the hand 21 for each component is stored in the memory 4b. Of the recognized component, of the i-th object to be held, the position of the OPW ij wrist joint as viewed from the object coordinate system included in the j-th gripping position and orientation _data, a matrix representing the posture to ORW _ij. At this time, the target position and orientation of the wrist joint viewed from the absolute coordinate system can be defined by the following expression.
Pw _ij = Po _i + Ro _i * oPw _ij
Rw _ij = Ro _i * oRw _ij

  The control device 4 calculates a target joint angle in the gripping posture of the robot arm 2 by performing inverse kinematics calculation on the position and posture of the wrist joint shown in the above formula. The control device 4 sequentially performs the inverse kinematics operation described above for each component in the grip position / posture database of the hand 21 and calculates a target joint angle for each component.

  Based on the calculated target joint angle for each component, the control device 4 avoids an obstacle by the hand 21 from the initial posture with respect to each component to the gripping position posture and further from the gripping position posture to a posture in which each component is placed in the box. A gripping motion trajectory that moves while moving is generated. This trajectory generation method is described in, for example, non-patent literature: G. Sanchez and J.-C. Latombe, “A Single-Query Bi-Directional Probabilistic Roadmap Planner with Lazy Collision Checking”, Robotics Research, Springer Tracts in Advanced It is disclosed in detail in Robotics, vol.6, pp.403-417, 2003, and can be used for this.

FIG. 6 is a flowchart showing the flow of the method for generating the gripping motion trajectory in the learning phase.
The control device 4 acquires the position Po _i and the posture Ro _i of the part to be gripped from the detection unit 5 (step S101).

  The control device 4 segments the part point cloud and calculates the bounding box of each segment. The control device 4 assigns the position and orientation of the component to the calculated bounding box whose shape is close to the shape of the bounding box of the component to be grasped.

The control device 4 performs initialization by setting parameters i = 1, j = 1, and k = 1 (step S102).
The control device 4 calculates the target position and orientation candidates Pw _ij and Rw _ij of the gripped component by using the gripping position and orientation database with respect to the position and orientation of the component detected by the detection unit 5 (step) S103).

The control device 4 determines whether or not the inverse kinematic problem of the calculated target position and orientation candidates Pw _ij and Rw _ij of the part to be gripped has a solution (step S104).
When the control device 4 determines that the inverse kinematics problem of the calculated target position / posture candidates Pw _ij and Rw _ij of the component to be gripped has a solution (YES in step S104), It is determined whether or not the robot arm 2 does not interfere with the obstacle (step S105). On the other hand, if the control device 4 determines that the inverse kinematics problem of the calculated target position and orientation candidates Pw _ij and Rw _ij of the part to be gripped has no solution (NO in step S104), it will be described later (step S110). Move on to processing.

  When the control device 4 determines that the robot arm 2 does not interfere with the obstacle in the target position / posture candidates (YES in step S105), the control device 4 determines the gripping position / posture from the initial posture and the gripping position / posture for each component. It is determined whether or not a trajectory up to the posture of placing the part in the box can be generated (step S106). On the other hand, when the control device 4 determines that the robot arm 2 is interfering with an obstacle in the target position / posture candidates (NO in step S105), the control device 4 proceeds to the processing in (step S110) described later.

  If the control device 4 determines that it is possible to generate a trajectory from the initial posture with respect to each part to the gripping position posture, and from the gripping position posture to a posture in which each component is placed in the box (YES in step S106), the control device 4 sets the kth gripping motion trajectory. It memorize | stores in the memory 4b (step S107), and transfers to the process of below-mentioned (step S110). On the other hand, when the control device 4 determines that the trajectory from the initial posture to each component to the gripping position posture and from the gripping position posture to the posture in which each component is placed on the box cannot be generated (NO in step S106), the control device 4 will be described later (step The process proceeds to S110).

The control device 4 determines whether or not determination of all gripping objects and all gripping position / postures has been completed (i and j are maximum values) (step S108).
If the control device determines that the determination of all gripping objects and all gripping positions and orientations has been completed (YES in step S108), the control process ends. On the other hand, if the control device determines that the determination of all gripping objects and all gripping positions and postures has not been completed (NO in step S108), k is incremented (k = k + 1 (provided that the database of the i-th gripping target object). If the number is exceeded, i = i + 1 and j = 1 are set.) (Step S109), the process proceeds to (Step S110) described later.

The control device increments the parameter j (however, if the number of databases of the i-th gripping object is exceeded, i = i + 1, j = 1 is set) (step S110), and the process returns to the above (step S103).
The control device 4 generates a discriminator in the following (5) using the generated gripping motion trajectory.
(5) Generation of discriminator As described above, the description is duplicated and will be omitted.

<< Execution Phase >>
(1) Measurement of the position and orientation of a component Since this is the same as the measurement of the position and orientation of a component in the learning phase, description thereof is omitted.
(2) Grasping operation (Picking operation)
FIG. 7 is a flowchart showing a flow of a method for generating a gripping motion trajectory in the execution phase.

The control device 4 acquires the position Po _i and the posture Ro _i of the part to be gripped from the detection unit 5 (step S201).
The control device 4 segments the part point cloud and calculates the bounding box of each segment. The control device 4 assigns the position and orientation of the component to the calculated bounding box whose shape is close to the shape of the bounding box of the component to be grasped.

The control device 4 performs initialization by setting parameters i = 1, j = 1, and k = 1 (step S202).
The control device 4 calculates the target position and orientation candidates Pw _ij and Rw _ij of the gripped component by using the gripping position and orientation database with respect to the position and orientation of the component detected by the detection unit 5 (step) S203).

The control device 4 determines whether or not the target position / posture candidates Pw _ij and Rw _ij can be gripped by using the discriminator (step S204).
When the control device 4 determines that the target position / orientation candidates Pw _ij and Rw _ij can be gripped by using the discriminator (YES in step S204), the calculated target position / posture candidate Pw _ij of the part to be gripped is determined. , Rw _ij determines whether or not the inverse kinematics problem has a solution (step S205). On the other hand, when the control device 4 determines that the target position / posture candidates Pw _ij and Rw _ij cannot be gripped by using the discriminator (NO in step S204), the control device 4 proceeds to the processing of (step S211) described later.

When the control device 4 determines that the inverse kinematics problem of the calculated target position / posture candidates Pw _ij and Rw _ij of the component to be gripped has a solution (YSE in step S205), It is determined whether or not the robot arm 2 is not interfering with an obstacle (step S206). On the other hand, when the control device 4 determines that the inverse kinematic problem of the calculated target position and orientation candidates Pw _ij and Rw _ij of the part to be gripped has no solution (NO in step S205), it will be described later (step S211). Move on to processing.

  When the control device 4 determines that the robot arm 2 does not interfere with the obstacle in the target position / posture candidates (YES in step S206), the control device 4 determines the gripping position / posture from the initial posture and the gripping position / posture for each component. It is determined whether or not a trajectory up to the posture of placing the part in the box can be generated (step S207). On the other hand, when the control device 4 determines that the robot arm 2 is interfering with an obstacle in the target position / posture candidate (NO in step S206), the control device 4 proceeds to the processing in (step S211) described later.

  When the control device 4 determines that a trajectory from the initial posture to each component to the gripping position posture and from the gripping position posture to the posture in which each component is placed in the box can be generated (YES in step S207), the k-th gripping motion trajectory is determined. It memorize | stores in the memory 4b (step S208), and transfers to the process of below-mentioned (step S211). On the other hand, if the control device 4 determines that the trajectory from the initial posture to each component to the gripping position posture and from the gripping position posture to the posture in which each component is placed on the box cannot be generated (NO in step S207), the control device 4 will be described later (step The process proceeds to S211).

The control device 4 determines whether determination of all gripping objects and all gripping position / postures has been completed (i and j are maximum values) (step S209).
If the control device 4 determines that the determination of all gripping objects and all gripping position / postures has been completed (YES in step S209), the control device 4 proceeds to the processing in (step S211) described later. On the other hand, if the control device 4 determines that the determination of all gripping objects and all gripping positions and orientations has not been completed (NO in step S209), k is incremented (k = k + 1 (however, the database of the i-th gripping object). (I = i + 1, j = 1))) (step S210), the process proceeds to (step S211) described later.

The control device 4 increments the parameter j (however, if the number of databases of the i-th gripping object is exceeded, i = i + 1, j = 1 is set) (step S211), and the process returns to the above (step S203). .
When a plurality of gripping motion trajectories are obtained, the control device 4 calculates an evaluation index I for each gripping motion trajectory. As described above, it is assumed that the success rate of component gripping by the hand 21 is increased if the number of points of strong interference between the gripping motion trajectory of the hand 21 and the obstacle is reduced. That is, it is assumed that the gripping success rate is higher when the values of h (p _i ), d (p _i ), and n are smaller.

The control device 4 calculates the evaluation index I of the gripping motion trajectory using the following formula. In the following formulas, alpha, beta, is γ is a weighting factor that takes positive values, an index value of the gripping stability for gripping position and orientation I _g was given. Incidentally, the index value I _g of this gripping stability, for example, Non-Patent Document: Harada, Tsuji, Uto, Yamanobe, Nagata, "Stability Analysis of soft finger-type grasping", Instrument and Control Engineers System Integration Division Annual Conference Proceedings Shu, 2013. It is disclosed in this document and can be used.

The control device 4 selects a gripping motion trajectory that minimizes the evaluation index I from the plurality of gripping motion trajectories obtained (step S212). As described above, when there is an obstacle in the hand passing region of the gripping motion trajectory, there are various situations such as the hand 21 colliding with the obstacle or slightly touching. Therefore, when multiple gripping motion trajectories are obtained, finally, the gripping motion trajectory with the smallest evaluation index indicating how much the obstacle interferes with the gripping motion trajectory (minimum interference) is selected. Further, the gripping success rate can be further increased.
As described above, in the grasping possibility determination method according to the present embodiment, a plurality of sample data in which the three-dimensional point cloud data included in the hand passage area and the success / failure of the grasping of the grasped object by the hand 21 are acquired. Assuming that the smaller the values of distance h (p _i ), distance d (p _i ), and n, the higher the success rate of gripping, and based on the sample data and distances h (p _i ) and d (p _i ) A discriminator for determining whether or not the gripping target object can be gripped is generated from the calculated feature vector, and whether or not the gripping target object can be gripped by the hand is determined using the generated classifier.
As a result, even when an obstacle exists in the hand passing area S, it is not immediately determined that the part cannot be gripped, but a determination as to whether or not the part can be gripped is performed using a discriminator. For this reason, the number of parts that can be gripped can be increased. When determining whether or not the gripping is possible, the gripping success rate can be increased by using a discriminator that learns the gripping pattern of the hand 21 that has few strong interference points between the hand 21 and the obstacle and has a high gripping success rate. it can. That is, the gripping success rate can be increased while increasing the number of parts that can be gripped.

  Note that the present invention is not limited to the above-described embodiment, and can be changed as appropriate without departing from the spirit of the present invention.

In addition, the present invention can be realized by causing the CPU 4a to execute a computer program, for example, the processes shown in FIGS.
The program may be stored using various types of non-transitory computer readable media and supplied to a computer. Non-transitory computer readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (for example, flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (for example, magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R / W and semiconductor memory (for example, mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM, RAM (random access memory)) are included.

  The program may also be supplied to the computer by various types of transitory computer readable media. Examples of transitory computer readable media include electrical signals, optical signals, and electromagnetic waves. The temporary computer-readable medium can supply the program to the computer via a wired communication path such as an electric wire and an optical fiber, or a wireless communication path.

  DESCRIPTION OF SYMBOLS 1 Transfer robot, 2 Robot arm, 3 Moving trolley, 4 Control apparatus, 5 Detection part, 21 Hand, 22 Finger part, 51 Visual sensor

Claims (3)

A gripping possibility determination method for determining whether or not the hand can grip the gripping object when gripping a plurality of gripping objects stacked by the robot arm,
The hand is moved from an approach point that is a predetermined distance away from the gripping object to a gripping point for gripping the gripping object by the hand, and until the gripping object is gripped by the hand at the gripping point. A step of acquiring a plurality of sample data in which a region through which the hand passes is set as a hand passage region, and three-dimensional point cloud data included in the hand passage region is associated with success / failure of gripping the gripping object by the hand When,
The i-th point of the three-dimensional point group included in the hand passing area is defined as p _i (i = 1, 2,... N), and a straight line obtained by extending the point p _i in the approach direction is the hand passing area. of the point of intersection with the boundary, the distance of a point farthest from the point p _i and h (p _i), a straight line extending in the gripping direction of the hand from the point p _i intersects with the boundary of the hand passing area It is assumed that the distance from the point p _i to the point closest to the point p _i is d (p _i ), and that the smaller the values of h (p _i ), d (p _i ) and n, the higher the gripping success rate. Generating a discriminator for determining whether or not the object to be grasped is grasped from a feature vector calculated based on the sample data and the distances h (p _i ) and d (p _i );
Determining whether or not the hand can be gripped by the hand using the generated discriminator;
A gripping availability determination method characterized by comprising: A method for determining whether or not gripping is possible according to claim 1,
The classifier is composed of a random forest,
The method of determining whether or not the gripping is possible, wherein the feature vector is calculated using the following equation.
The feature vector = (min ((h (p i) / binwidth), bin_h),
_{min ((d (p i)} / binwidth), bin_d))
However, the above formula, binwidth is a bin width, and bin_h and bin_d are bin upper limit values of each axis. A method for determining whether or not gripping is possible according to claim 1,
The discriminator is composed of a linear SVM (support vector machine),
The feature vector f _s is calculated by using the following equation, and a gripping possibility determination method is characterized.

Images