JPH05119815A - Obstacle avoiding method using neural network - Google Patents

Abstract

PURPOSE:To facilitate an evaluating calculation to the danger of a collision, to flexibly cope with the addition of a new obstacle and to enable a processing at a high speed by planning an avoiding track by deriving an overall degree of danger with neural networks provided, respectively in accordance with an obstacle having the possibility of a collision at every link of a manipulator. CONSTITUTION:At every link of a manipulator, neural networks 31-34 are provided, respectively in accordance with an obstacle having possibility of a collision with its link. Subsequently, the joint position of each link expressed by respective joint angles theta1-theta3 of the manipulator is converted to a working coordinate at every link and by inputting it to respective neural networks 31-34, the degree of danger for coming into collision with each obstacle is evaluated at every link of the manipulator. Then, in an integrated danger degree deciding part 40, an integrated danger degree is derived by integrating the degree of danger corresponding to each evaluated obstacle and by setting it as an evaluation value, an avoiding track to the obstacle is planned. In such a manner, a collision avoidance processing which does not depend on the number of links and the number of obstacles is available.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an obstacle avoidance method for generating a trajectory of a manipulator from an initial point to a target point while avoiding obstacles scattered in a work space. The present invention relates to an obstacle avoidance method using a neural network adapted to use a network.

[0002]

2. Description of the Related Art In manipulator trajectory planning for planning a trajectory of a manipulator from an initial point to a target point while avoiding obstacles scattered in a work space, conventionally, the following three methods are roughly classified. ..

1) A method in which an obstacle is ignored and a trajectory is determined, and when an obstacle exists on the determined trajectory, a trajectory for avoiding the obstacle is newly generated in the vicinity of the obstacle. 2 ) Assuming the shapes of the manipulator and obstacles to be simple and easy to process, when an obstacle is approached, an evaluation function that adds a penalty according to the distance between the manipulator and the obstacle is given, and the evaluation is performed. Method to generate trajectory based on function so as to minimize penalties 3) Method to memorize free space that is not obstructed by obstacles and generate trajectory from initial point to reach destination point through this free space However, the method of 1) first determines the trajectory by ignoring the obstacle, and only when the obstacle exists on the determined trajectory, the trajectory for avoiding the obstacle is set in the vicinity of the obstacle. Generate new Since it is a method, when multiple obstacles are scattered in the work space, there is a problem that it is difficult to generate an appropriate trajectory considering all the multiple obstacles, and an obstacle is newly added. In such a case, there is a problem that the orbit generation process must be newly performed each time.

In the method 2), it is difficult to properly take in the properties of the manipulator, which is a rigid body, and the shapes of the manipulator and the obstacle are assumed to be simple shapes that can be easily processed. There was a problem that the avoidance plan for an object became local, which created a new collision. To avoid this, unless the above-mentioned simple shape assumption is made, the distance between the manipulator having a complicated shape and the obstacle must be calculated, and the calculation of the evaluation function that adds a penalty according to this distance. Another problem was that it took a lot of time.

Further, the method 3) is suitable for a case where a plurality of obstacles are scattered in the work space because the trajectory is generated by using the free space which is not disturbed by the obstacles stored in advance. There is a problem that it is difficult to generate various trajectories, and an unnecessarily large working space is required to secure this free space.

[0006]

As described above, according to the conventional obstacle avoidance method, as shown in 1), first, the trajectory is decided by ignoring the obstacle, and the obstacle is placed on the decided trajectory. When there is an obstacle, it is a method to newly generate a trajectory to avoid the obstacle in the vicinity of the obstacle, or a free space that is not obstructed by a preset obstacle as seen in 3). Since it is used to generate trajectories, it is difficult to generate proper trajectories when multiple obstacles are scattered in the work space, and new obstacles are added each time new obstacles are added. There was a problem that the orbit generation process had to be performed.

Further, as shown in 2), a trajectory is generated so as to minimize the penalty, based on an evaluation function set so as to apply the penalty according to the distance between the manipulator and the obstacle. However, this method has a problem that it takes a lot of time to calculate the distance between the manipulator having a complicated shape and the obstacle and to calculate the evaluation function corresponding to the distance.

Therefore, the present invention facilitates the evaluation calculation for the risk of collision between the manipulator, which is a rigid body, and an obstacle, and also flexibly responds to the fact that a new obstacle is added, and speeds up. It is an object to provide an obstacle avoidance method using a neural network capable of processing.

[0009]

To achieve the above object, the present invention provides a plurality of trajectories from an initial point to a target point of an articulated manipulator having a plurality of links and a plurality of joints respectively connecting the links. In a work space in which obstacles are scattered, in an obstacle avoidance method for generating so as to avoid the obstacles, the joint position of each link represented by each joint angle of the manipulator is set to work coordinates for each link. Convert, calculate the risk of colliding with each obstacle for each link of the manipulator by inputting the converted work coordinates into the neural network corresponding to each obstacle, to the calculated each obstacle It is characterized in that the corresponding risk levels are integrated to obtain a total risk level, and an avoidance trajectory for the obstacle is generated using this as an evaluation value.

[0010]

Function: For each link of the manipulator, a neural network is provided corresponding to an obstacle that may collide with the link. Then, the joint position of each link expressed by each joint angle of the manipulator is converted into work coordinates for each link, and this is input to each neural network, and there is a risk of collision with each obstacle for each link of the manipulator. Evaluate the degree. Here, the neural network is able to properly evaluate the risk of obstacles by learning with a finite number of sample points even if the link has a complicated shape, and is further learned by the interpolation function of the neural network. It is possible to evaluate an appropriate degree of risk even at a coordinate that does not exist. Then, the risk degree corresponding to each evaluated obstacle is integrated to obtain a comprehensive risk degree, and the avoidance trajectory for the obstacle is planned using this as an evaluation value. As described above, according to the present invention, the neural networks are provided corresponding to the obstacles that may collide with the link,
Since the collision avoidance processing is performed by utilizing the parallelism of the neural network, the high speed collision avoidance processing independent of the number of links and the number of obstacles becomes possible.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of an obstacle avoidance method using a neural network according to the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing a collision avoidance device for an articulated manipulator constructed by applying the obstacle avoidance method of the present invention.

In FIG. 1, a control unit 10 controls the operation of an articulated manipulator to be controlled, and this control unit 10 is constructed using a so-called recurrent network. That is, the control unit 10, as shown in FIG. 2, is an identifier 13 connected in parallel to the controlled object 12,
The state signal s _{t + 1} corresponding to the state of the controlled object 12 output from the identifier 13 is sent for a predetermined time to output the state signal s
Time delay element 1 to be fed back to the input of the identifier 13 as _t
4, and an operation signal v controller 11 for outputting the _t target state signal s ₀ and a time delay element 14 controlled object 12 by inputting a status signal s _t output from. Here, the controller 11 and the identifier 13 are each composed of a neural network, the identifier 13, the control state signal output from the target 12 s _{t + 1} 'and the state signal s _t output from the identifier 13 It is controlled by the output of the subtracter 15 which takes the deviation from ₊₁ .

Also, in this embodiment, the controlled object 12
Assumes an articulated manipulator having three links 51, 52, 53 as shown in FIG. This articulated manipulator is fixed at the origin (x ₀ , y ₀ ) of the work coordinate x-y at a point P ₀ , and two obstacles 61 and 62 are arranged in the work coordinate x-y. Here, this multi-joint manipulator has a joint angle at each joint point,
That is, the link 51 at the point P ₀ and the work coordinates xy
Angle theta ₁ of the x-axis, the link at the connection point P ₂ of the link 51 and the angle theta ₂ of the link 51 and the link 52 at the connection point P ₁ of the link 52, link 52 and link 53 52
By controlling the angle θ ₃ formed by the link 53 and the link 53, the links 51, 52, 53 are prevented from colliding with the obstacles 61, 62, and the point P ₃ at the tip thereof is controlled to a desired destination point. ..

In FIG. 1, the control unit 10 outputs a signal indicating the joint angles θ ₁ , θ ₂ and θ ₃ of the articulated manipulator to be controlled in each control stage to the coordinate conversion unit 20. The coordinate conversion unit 20 receives the signals indicating the joint angles θ ₁ , θ ₂ , and θ ₃ and inputs them on the work coordinates xy of the points P ₁ , P ₂ , and P ₃ of the articulated manipulator to be controlled. Coordinates (x ₁ , y ₁ ), (x ₂ , y ₂ ), (x ₃ , y ₃ )
Is calculated and the point P of the articulated manipulator to be controlled is calculated.
The coordinates (x ₀ , y) of the origin of the work coordinate x-y where ₀ is fixed
₀ ) Output together.

The calculation in the coordinate conversion unit 20 is performed as follows when the lengths of the links 51, 52 and 53 are L ₁ , L ₂ and L ₃ , respectively.

X ₁ = x ₀ + L ₁ cos θ ₁ y ₁ = y ₀ + L ₁ sin θ ₁ x ₂ = x ₁ + L ₂ cos (θ ₁ + θ ₂ ) y ₂ = y ₁ + L ₂ sin (θ ₁ + θ ₂ ) X ₃ = x ₂ + L ₃ cos (θ ₁ + θ ₂ + θ ₃ ) y ₃ = y ₂ + L ₃ sin (θ ₁ + θ ₂ + θ ₃ ) The risk determination units 31, 32, 33, and 34 are respectively illustrated in FIG. The degree of risk of collision between the links 51, 52, 53 and the obstacles 61, 62 of the multi-joint manipulator, which is the control target, is determined. This risk determination unit 31, 3
Reference numerals 2, 33, and 34 are each composed of a neural network, and learning is possible and high-speed parallel processing is possible.

By the way, in this embodiment, as is apparent from FIG. 3, the link 1 may collide with the obstacle 61, and the link 2 may collide with the obstacle 61 and the obstacle 62. The link 3 may collide with the obstacle 62.

Therefore, in this embodiment, the link 1
And the obstacle 61, the risk determination unit 31 for determining the risk of collision, the risk determination unit 32 for determining the risk of collision between the link 2 and the obstacle 61, the link 2 and the obstacle 62 There are provided four risk degree determining sections, namely, a risk degree determining section 33 for determining the risk degree of collision with the vehicle and a risk degree determining section 34 for determining the risk degree of collision between the link 3 and the obstacle 62. There is.

Each of the risk degree judging sections 31, 32, 33, 34
Inputs the coordinates on the work coordinates xy sufficient to specify the position of the link to be the collision determination target. In this embodiment, assuming that the articulated manipulator to be controlled moves in a work space on a plane, that is, work coordinates x-y, position coordinates of both ends of a link to be a collision determination target, that is, a total of four inputs are used for the link. Specify the work position, and measure the distance between the link and the obstacle that may cause a collision.
The degree of risk corresponding to that is assigned to the state and learned.

That is, the risk determining section 31 has the coordinates (x ₀ , y ₀ ) of the origin of the work coordinates xy from the coordinate converting section 20.
Then, the coordinates (x ₁ , y ₁ ) of the point P ₁ are input, the distance between the link 1 and the obstacle 61 is measured, and a degree of risk corresponding to the distance is learned by being assigned to that state. Also, the risk determination unit 32, the coordinates (x _1, y ₁₎ the point P ₁ and point P ₂
Enter the coordinates (x ₂ , y ₂ ) of the link 2 and obstacle 6
The distance to 1 is measured, and the degree of risk is assigned to the state for learning. In addition, the risk determination unit 33
Enter the point P ₁ of the coordinates (x _1, y ₁₎ and point P ₂ of the coordinates (x _2, y _2), by measuring the distance between the link 2 and the obstacle 62, the risk accordingly Assign degrees to the state and learn. Further, the risk determination unit 34 determines that the point P ₂
Coordinates (x ₂ , y ₂ ) of the point P ₃ and coordinates (x ₃ ,
y ₃ ) is input, the distance between the link 3 and the obstacle 62 is measured, and a degree of danger corresponding to the distance is learned for that state.

In this structure, when a new condition is added, the evaluation value is changed and new learning is repeated.

Here, the degree of risk used for learning is zero when the distance between the link subject to collision determination and the obstacle is sufficiently large, and the distance between the link subject to collision determination and the obstacle is short. If there is a high possibility of collision, assign a high risk.

For example, the distance between the link and the obstacle, which is the object of collision determination, is defined as the closest distance between these two objects, and when the value is a certain value or more, the risk is set to zero, and In the following cases, the risk is determined according to the shape of the articulated manipulator that is the control target, that is, the shape of the link that is the collision determination target. That is, a representative sample position according to the shape of the link that is the collision determination target is selected, the risk is assigned, and learning is performed for each link and obstacle set.

In this way, when the learning is completed for all obstacles, the risk degree judging units 31, 32, 33, 34
Thus, the risk levels H ₁ , H ₂ , H ₃ , and H ₄ of collision with each obstacle with respect to the current position of each link of the articulated manipulator to be controlled can be calculated.

The total risk determining section 40 specifies weights for the entire risk determining sections 31, 32, 33 and 34, and the respective risk determining sections 31, 32, 33 and 34 calculate the respective weights. Collision risk levels H ₁ , H ₂ , H ₃ , H ₄
Is input to calculate the overall risk, that is, the total risk H.

The total risk level H calculated by the total risk level determination unit 40 is added to the control unit 10, and the control unit 10 is shown in FIG. The work space is controlled by controlling the neural network of the controller 11 and the identifier 13 to calculate the sensitivity of each joint angle that minimizes the overall risk H, and controlling each joint angle with the calculated sensitivity. The trajectory of the articulated manipulator to be controlled from the initial point to the target point is planned while avoiding the obstacles scattered around.

FIG. 4 is a flow chart showing the operation of the controller 10.

That is, first, the current joint angles θ ₁ , θ ₂ , θ ₃ of the articulated manipulator to be controlled are output to the coordinate conversion unit 20 (step 101). Then, corresponding to the output current joint angles θ ₁ , θ ₂ , θ ₃ , the coordinate conversion unit 20, the risk determination units 31, 32, 33, 34,
The comprehensive risk H calculated by the comprehensive risk determination unit 40 is imported from the comprehensive risk determination unit 40 (step 10
2) Calculate the sensitivities of the joint angles θ ₁ , θ ₂ , and θ ₃ by the descent method from the taken comprehensive risk level H (step 10).
3), the calculated respective joint angles theta ₁ and, theta _2, the joint angle theta ₁ by the sensitivity of the theta _3, theta _2, controls the theta ₃ (Step 1
04). Then, it is judged whether or not the predetermined end condition is satisfied (step 105). If the predetermined end condition is not satisfied, the process returns to step 102 and the above operation is repeated until the predetermined end condition is satisfied. When it is determined in step 105 that the predetermined termination condition is satisfied,
This process ends.

This makes it possible to plan an appropriate trajectory of the articulated manipulator to be controlled from the initial point to the target point while avoiding obstacles scattered in the work space.

[0031]

As described above, according to the present invention,
For each link of the manipulator, a neural network is provided for each obstacle that may collide with that link, and this neural network evaluates the risk of collision with each obstacle for each link. , Since the risk corresponding to each of the evaluated obstacles is integrated to obtain the total risk, and the avoidance trajectory for the obstacle is planned using this as an evaluation value, the manipulator can be treated as a rigid body, In addition, even if a new obstacle is added, it is possible to flexibly respond by simply adding a new neural network, and since parallel processing in the neural network is possible, high speed processing is required. It also has the effect of being able to withstand.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a collision avoidance device for an articulated manipulator configured by applying the obstacle avoidance method of the present invention.

FIG. 2 is a block diagram showing a configuration example of a control unit of the embodiment shown in FIG.

FIG. 3 is a configuration diagram illustrating a structure of an articulated manipulator that is a control target of the embodiment illustrated in FIG.

FIG. 4 is a flowchart for explaining the operation of the embodiment shown in FIG.

[Explanation of symbols]

 10 Control part 11 Controller 12 Control object 13 Identifier 14 Time delay element 15 Subtractor 20 Coordinate conversion part 31-34 Danger degree judgment part 40 Comprehensive risk judgment part 51-53 Link 61,62 Obstacle

Claims (1)

[Claims] 1. A trajectory of an articulated manipulator having a plurality of links and a plurality of joints respectively connecting the links from an initial point to a target point, in a work space in which a plurality of obstacles are scattered, In an obstacle avoidance method of generating so as to avoid, the joint position of each link expressed by each joint angle of the manipulator is converted into working coordinates for each link, and the conversion is performed into a neural network corresponding to each obstacle. Calculate the risk of colliding with each obstacle for each link of the manipulator by inputting the work coordinates, calculate the risk corresponding to each of the calculated obstacles, and obtain the overall risk, An obstacle avoidance method using a neural network is characterized in that an avoidance trajectory for the obstacle is generated with the evaluation value as.