CN108827312A - A kind of coordinating game model paths planning method based on neural network and Artificial Potential Field - Google Patents
A kind of coordinating game model paths planning method based on neural network and Artificial Potential Field Download PDFInfo
- Publication number
- CN108827312A CN108827312A CN201810907242.7A CN201810907242A CN108827312A CN 108827312 A CN108827312 A CN 108827312A CN 201810907242 A CN201810907242 A CN 201810907242A CN 108827312 A CN108827312 A CN 108827312A
- Authority
- CN
- China
- Prior art keywords
- intelligent body
- moment
- neural network
- repulsion
- potential field
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
- G01C21/00—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00
- G01C21/20—Instruments for performing navigational calculations
Abstract
The present invention proposes a kind of coordinating game model paths planning method based on neural network and Artificial Potential Field, belongs to intelligent body active path planning field.This method constructs training sample set in the path optimizing that off-line phase obtains intelligent body first, and is trained to BP neural network module, obtains the BP neural network module that training finishes;On-line stage, it obtains the position at the current time of each intelligent body and environmental information and inputs the BP neural network module that finishes of training, network module exports the repulsion gain at the moment, and then be calculated the moment target point to the gravitation of intelligent body, threaten area to the repulsion of intelligent body, and resultant force is calculated.Intelligent body updates repulsion gain according to new position and environmental information according to being moved with joint efforts, and at next moment, until path planning reaches the condition of end.The present invention can better solve the path planning problem under coordinating game model scene, can better adapt to the active path planning under target and obstacle movement.
Description
Technical field
It is the invention belongs to intelligent body active path planning field, in particular to a kind of based on neural network and Artificial Potential Field
Coordinating game model paths planning method.
Background technique
Coordinating game model path planning based on neural network and Artificial Potential Field refers to:Multiple agent uses improved nerve net
The method that network and Artificial Potential Field combine, synergistically for the purpose of reaching target area, the path that obstacle avoidance area is constraint and progress
Planning, wherein game refers to that target area and barrier are directed to multiple agent movement in real time and carry out antagonism feedback.Coordinating game model road
Diameter planning problem is important branch of the two-dimensional space to the active path planning problem of particle, and robot is enclosed in the research of the problem
It catches, the practical applications such as robot soccer game are of great significance.
The methods of existing Artificial Potential Field, neural network all exist centainly not for solving coordinating game model path planning problem
Foot.Artificial Potential Field Method is the common method in existing multiple agent real-time route planning problem, and this method uses for reference physical field
Thought constructs the gravitation potential field from target and the repulsion potential field from obstacle, the resultant force influence pair of generation in Virtual Space
As the movement of (may be robot, aircraft or vehicle), practicability is very strong, and with principle, intuitive, algorithm structure simply can
Adaptability by being easily achieved, to dynamic environment is good, computing cost is small and real-time is good, without the building space C in advance
The advantages that (Configuration space).By studying for a long period of time, the common locally optimal solution of Artificial Potential Field Method closes on obstacle
It is shaken when goal nonreachable (Goals Non-Reachable with Obstacles Nearby, GNRON), mistake slype
Be rebuffed, dynamic disorder the problems such as also preferably solved.But existing Artificial Potential Field Method considers deficiency to coordinating game model scene,
The antagonism feedback of environment and intelligent body collaboration are difficult quantitatively to be described with rule and strategy.Although can based on cybernetic method
To carry out quantitative optimization, but computation complexity often exceeds the requirement of real-time.
S S Ge etc. 2000 in " New Potential Functions for Mobile Robot Path
The method proposed in Planning " is presently the most classical Artificial Potential Field Method, is at Khatib etc. 1986 in " Real-
The primitive man being put forward for the first time in time obstacle avoidance for manipulators and mobile robots "
Repulsion function construction method is improved on the basis of work potential field method, is largely overcome and is closed on obstacle mesh existing for original method
Mark unreachable (Goals Non-Reachable with Obstacles Nearby, GNRON) problem.This method mainly includes
Following steps:
1) target point is calculated to the gravitation of intelligent body;
Shown in gravitation potential field such as formula (1):
Wherein kaFor gravitation gain, ρ (p, pgoal)=| | pgoal- p | | it is Euclidean distance of the intelligent body to target point, p
For intelligent body position coordinates, pgoalFor aiming spot coordinate, n ∈ (0 ,+∞) usually takes n=2 by experience convention.
Gravitation can be acquired by the negative gradient of gravitation potential field, as shown in formula (2):
2) calculating threatens area to the repulsion of intelligent body;
Shown in repulsion potential field such as formula (3):
Wherein krFor repulsion gain, ρ (p, pthreat)=| | pthreat- p | | for intelligent body to threat area boundary closest approach
(the exactly introducing of the distance, when avoiding while closing on obstacle and target, repulsion covering gravitation causes target to Euclidean distance
It is unreachable), pthreatFor the position coordinates for threatening area boundary closest approach, n ∈ (0 ,+∞) usually takes n=2 by experience convention, denounces
Power can be acquired by the negative gradient of repulsion potential field, as shown in formula (4):
3) resultant force is sought;
Resultant force is the vector sum of gravitation and repulsion, then will further be converted into the direction of motion and speed, conversion side with joint efforts
Method is determined according to specific path planning problem scene.
At each moment, intelligent body carries out movement until next moment according to the resultant force at the current time being calculated;
If intelligent body reaches target point at next moment, path planning terminates;If not up to, going successively to next step path rule
It draws, gos to step 1).
The major defect of this method is:Bad for the adaptability of dynamic scene, potential field function cannot consider target and barrier
The movement hindered not can be carried out the path planning with prediction and trail.
Neural network is applied among path planning as a kind of powerful machine learning and knowledge representation method.
Neuron and neural network framework are mapped in configuration space by some researchs, the guidance pair using the output of neuron as potential field
As movement.Other research and utilization neural networks carry out the knowledge preparation before path planning, such as to environment classification, prediction
Obstacle motor pattern, calculating functional equation etc..Neural network is also used in combination with some other method, including fuzzy control, something lost
Propagation algorithm, simulated annealing etc..However many basic frameworks and forward-propagating process studied just with neural network, not
Make full use of its learning ability.
It is some to research and propose the situation feelings according to dynamic change by various modes in order to adapt to active path planning scene
The adaptive adjusts path of condition plans equation coefficient, exists wherein being more typically Lin Zhixiong in 2014《Computer Simulation》Middle proposition
" the Soccer robot path planning based on fuzzy neuron potential field method ".This method first improves the function of Artificial Potential Field,
Then increase fuzzy neural network module and adjust improved Artificial potential functions coefficient in real time, fuzzy neural network module is by artificial
The fuzzy rule and fuzzy set subordinating degree function off-line training of setting generate.This method key step is as follows:
1) building and Training Fuzzy Neural Networks module;
Fuzzy Rule Sets are set, fuzzy neural network module is obtained by off-line training.
2) adaptive optimization potential field function coefficients;
The fuzzy neural network module obtained using step 1), online in real time to improved potential field function coefficients (target phase
To speed parameter α, obstacle offsets parameter beta) it optimizes, specific method is at a time interval, to be made with current state information
For fuzzy neural network input, (target relative velocity parameters α, obstacle offset parameter the potential field function coefficients of forward direction output optimization
β)。
3) path planning;
(target relative velocity parameters α, obstacle offset parameter the potential field function coefficients for the adaptive optimization that step 2) is obtained
It β) substitutes into following potential field function and carries out gravitation and repulsion calculating, and finally obtain resultant force.
Target point 3-1) is calculated to the gravitation of intelligent body;
Shown in gravitation calculation method such as formula (5):
Fatt(p)=kaρ(p,pgoal)+α(Vg-V) (5)
Wherein, kaFor gravitation gain, ρ (p, pgoal)=| | pgoal- p | | it is Euclidean distance of the intelligent body to target point,
P is intelligent body position coordinates, pgoalFor aiming spot coordinate, VgFor target velocity, V is intelligent body speed, n ∈ (0 ,+∞),
N=2 is usually taken by experience convention.
3-2) calculating threatens area to the repulsion of intelligent body;
Shown in repulsion calculation method such as formula (6):
Wherein krFor repulsion gain, ρ (p, pthreat)=| | pthreat- p | | for intelligent body to threat area boundary closest approach
Euclidean distance, pthreatFor the position coordinates for threatening area boundary closest approach, p0For repulsion biggest impact range, exceed the model
Area is threatened to be reduced to 0, V to intelligent body repulsion after enclosing0For obstacle speed, n ∈ (0 ,+∞) usually takes n=2 by experience convention.
3-3) seek resultant force
Resultant force is the vector sum of gravitation and repulsion, then will further be converted into the direction of motion and speed of intelligent body with joint efforts
Degree, method for transformation are determined according to specific path planning problem scene.
At each moment, intelligent body carries out movement until next moment according to the resultant force at the current time being calculated;
If intelligent body reaches target point, path planning terminates;If not up to, going successively to next step path planning, jumping to step
Rapid 3-1).
The major defect of this method is:Collaboration path planning adaptability under game scene is lacking, path
The collaboration and game of planning are mainly reflected in the formulation of fuzzy rule, however fuzzy rule needs artificial formulation, and quality is difficult
To guarantee, acquisition difficulty is big, good not as good as what is learnt from sample;On the other hand, this method is to the public affairs in classical Artificial Potential Field Method
Formula is improved, but improve be directly against draw, repulsion function, not for potential field function, according to " by field to power "
Principle, this method, which skips field function and directly improves the theoretical foundation of force function, to be needed further to be proved.
Summary of the invention
The purpose of the present invention is the shortcomings to overcome prior art, propose a kind of based on neural network and Artificial Potential Field
Coordinating game model paths planning method.The present invention can preferably solve the path planning problem under coordinating game model scene, play
The learning ability of neural network, practicability is good, and dynamic adaptivity is good.
The present invention proposes a kind of coordinating game model paths planning method based on neural network and Artificial Potential Field, and feature exists
In this approach includes the following steps:
1) off-line phase;
1-1) construct training sample set;Specific step is as follows:
R-th of intelligent body, which arbitrarily 1-1-1) is chosen, from the R intelligent body for participating in coordinating game model path planning is denoted as Fr, obtain
Fr1 path optimizing;The path optimizing shares T step, then the path optimizing includes 1 to T moment intelligent body FrPosition coordinates;
To the path optimizing, using the intelligent body position coordinates at t-th of moment and corresponding environment as training sample
Input;Wherein the intelligent body position coordinates at t-th of moment are (Fr_xt,Fr_yt), Fr_xtFor the X of r-th of intelligent body t moment
Coordinate, Fr_ytFor the Y coordinate of r-th of intelligent body t moment;T-th of moment corresponds to environment:FrIn the direction of motion of t moment
Fr_vt, the target position (G_x of t momentt,G_yt), the target direction of motion G_v of t momentt, the obstacle location (O_x of t momentt,
O_yt), the obstacle direction of motion O_v of t momenttAnd remove FrOuter other cooperative intelligent body t moment positions coordinate is remembered respectively
For (F1_xt,F1_yt)、(F2_xt,F2_yt)、…(FR_xt,FR_yt);By the potential field function coefficients of t moment, that is, repulsion gain kr
As the output of the training sample, wherein the k of t momentrBy by the intelligent body alternate position spike generation at t+1 moment and t moment
Enter Artificial potential functions to solve to obtain;Then the symbiosis of this path optimizing is at T-1 training sample;
1-1-2) repeat step 1-1-1), M path optimizing is obtained altogether to R intelligent body, and raw to every path optimizing
At corresponding training sample;By the training sample composing training sample set of all path optimizings;
1-2) construct BP neural network module;
1-3) the BP nerve that step 1-2) is established using momentum gradient descent method using the training sample set of step 1-1)
Network module carries out off-line training, when reaching training cut-off condition, obtains the BP neural network module that training finishes;
2) on-line stage;
The initial position of each intelligent body and corresponding environmental information in R intelligent body 2-1) are obtained, when note current time is t
It carves, initializes pedometer c=0;
2-2) the obtained training of the position at each intelligent body current time and environmental information input step 1-3) is finished
BP neural network module, network module input the repulsion gain k at the corresponding intelligent body current timer;
For r-th of intelligent body Fr, the input of BP neural network module is the F of t momentrPosition (Fr_xt,Fr_yt), when t
The F at quarterrDirection of motion Fr_vt, the target position (G_x of t momentt,G_yt), the target direction of motion G_v of t momentt, t moment
Obstacle location (O_xt,O_yt), the obstacle direction of motion O_v of t momentt, remove FrThe position of outer other cooperative intelligent body t moments
(F1_xt,F1_yt)、(F2_xt,F2_yt)、…(FR_xt,FR_yt);The output of BP neural network module is:Intelligent body FrT when
The repulsion gain k at quarterr;
The path planning based on Artificial Potential Field 2-3) is carried out at current time, obtains the conjunction at each intelligent body current time
Power;Specific step is as follows:
Target point 2-3-1) is calculated to the gravitation of intelligent body;
To each intelligent body, calculate shown in gravitation potential field such as formula (1):
Wherein, kaFor gravitation gain, it is set as 1, ρ (p, pgoal)=| | pgoal- p | | it is several for the Europe of intelligent body to target point
In distance, p be intelligent body position coordinates, pgoalFor aiming spot coordinate, n ∈ (0 ,+∞);
Gravitation is acquired by the negative gradient of gravitation potential field, as shown in formula (2):
2-3-2) calculating threatens area to the repulsion of intelligent body;
To each intelligent body, calculate shown in repulsion potential field such as formula (3):
Wherein, krFor the repulsion gain at step 2-2) obtained current time;ρ(p,pthreat)=| | pthreat- p | | it is intelligence
Euclidean distance of the energy body to threat area boundary closest approach, pthreatFor the position coordinates for threatening area boundary closest approach, p0For reprimand
Power biggest impact range;
Repulsion is acquired by the negative gradient of repulsion potential field, as shown in formula (4):
2-3-3) seek resultant force;
The resultant force of each intelligent body is the vector sum of the intelligent body corresponding gravitation and repulsion, and the direction of vector is the intelligence
The direction of the movement at energy body next moment;
2-4) each intelligent body carries out movement until at the t+1 moment, update step counting according to the resultant force that step 2-3) is calculated
Device c=c+1, and determined:
If the t+1 moment, any intelligent body reaches target point in R intelligent body and all intelligent bodies do not reach and threaten area,
Then path planning success, method terminate;If the t+1 moment, in R intelligent body, any intelligent body, which is reached, threatens area, then path planning
Failure, method terminate;If the t+1 moment, all intelligent bodies did not reached success or failure condition, t=t+1 is enabled, when by t+1
Carve as new current time, then return to step 2-2), continue subsequent time path planning;When the step number of pedometer c
When reaching the step number upper limit A step of setting, if all intelligent bodies are still not up to success or failure condition, path planning is denoted as super
When, method terminates.
The features of the present invention and beneficial effect are:
The present invention can better solve the path planning problem under coordinating game model scene, can better adapt to target and
Active path planning under obstacle movement, and fully consider that target and obstacle are made for the movement of this intelligent body in path planning
Antagonism variation;The present invention formulates fuzzy rule without artificial, alleviates using difficulty, increases practicability, plays nerve
The learning ability of network directly improves by learning from coordinating game model sample and solves coordinating game model path planning problem
Effect.
The invention proposes BP neural network dynamic optimization Artificial potential functions coefficient is used, one kind is devised with intelligent body
It is input with environment current state, potential field function coefficients are the BP neural network exported and corresponding sample process and study side
Method.
The present invention improves the effect of the game path planning under multiple agent collaboration, improves to active path planning
Adaptivity surrounds and seize robot, the practical applications such as robot soccer game are of great significance.
Detailed description of the invention
Fig. 1 is the overall flow figure of the method for the present invention.
Fig. 2 is the top plan view of red aircraft and blue party aircraft path planning region in the embodiment of the present invention.
Fig. 3 is the route programming result schematic diagram that traditional classical method is used in case study on implementation of the present invention.
Fig. 4 is the route programming result schematic diagram that the method for the present invention is used in case study on implementation of the present invention.
Specific embodiment
The present invention proposes a kind of coordinating game model paths planning method based on neural network and Artificial Potential Field, below with reference to attached
Figure and specific embodiment are further described as follows.
The present invention proposes a kind of coordinating game model paths planning method based on neural network and Artificial Potential Field, is divided into offline rank
Section and application on site stage, overall flow is as shown in Figure 1, include the following steps:
1) off-line phase;
1-1) construct training sample set;Specific step is as follows:
1-1-1) for the coordinating game model path planning problem (R is positive integer) of R intelligent body, r-th of intelligent body is obtained
Fr1 " path optimizing " (r be less than or equal to R), path optimizing refers to that the path can better meet the coordinating game model path
The target and constraint of planning problem, the degree of optimization will affect the performance for the neural network that learning training obtains, path optimizing
Obtaining can be by multiple means, such as manual path's planning, the simulation paths planning based on optimizing algorithm etc..Remember the optimization road
Diameter shares T step (T is positive integer and T is greater than 1), i.e., the path optimizing includes 1 to T moment intelligent body FrPosition coordinates.To this
Path optimizing, using the intelligent body coordinate at t-th of moment and corresponding environment as the input of a training sample;Wherein t-th when
The intelligent body position coordinates at quarter are (Fr_xt,Fr_yt), Fr_xtFor the X-coordinate of r-th of intelligent body t moment, Fr_ytIt is r-th
The Y coordinate of intelligent body t moment, t are less than or equal to T;T-th of moment corresponds to environment:FrIn the direction of motion (the i.e. r of t moment
The Y coordinate of a intelligent body t moment) Fr_vt, the target position (G_x of t momentt,G_yt), the target direction of motion G_ of t moment
vt, the obstacle location (O_x of t momentt,O_yt), the obstacle direction of motion O_v of t momenttAnd remove FrOuter other cooperative intelligent bodies
T moment position coordinate is denoted as (F respectively1_xt,F1_yt)、(F2_xt,F2_yt)、…(FR_xt,FR_yt).By t moment
Potential field function coefficients (repulsion gain kr) output as the training sample, wherein the k of t momentrBy by the t+1 moment with
The intelligent body alternate position spike of t moment substitutes into Artificial potential functions and solves to obtain, it is assumed that a path optimizing includes that T is walked (i.e. in total
Have T moment), then the symbiosis of this path optimizing at T-1 training sample (each training sample include it is corresponding input and it is defeated
Out), the size of the quantity of sample will affect the performance for the neural network that learning training obtains.
1-1-2) repeat step 1-1-1), obtaining M path optimizing altogether to R intelligent body, (M is positive integer, the bigger training of M
Effect is better, it is believed that each intelligent body be it is identical, only initial position is different, as long as M it is enough it is big if do not need it is each
Intelligent body has corresponding path optimizing), and corresponding training sample is generated to every path optimizing;By all path optimizings
Training sample composing training sample set.
1-2) construct BP neural network module;
Middle layer neuron is arranged using single hidden layer BP neural network module, by heuristic and empirical data in the present invention
Number (such as can be set as 24 to the game path planning problem of 2 intelligent bodies collaboration), activation primitive is Sigmoid function.
1-3) training sample set obtained using step 1-1) to the BP neural network module that step 1-2) is established carry out from
Line training obtains the BP neural network module that training finishes;
Using the sample of the training sample set obtained in step 1-1), using momentum gradient descent method off-line training BP nerve
Network module, with mean square error less than 0.001 or training iteration reach 100000 wheels for training cut-off condition, obtain training finishing
BP neural network module.So far, preparation is completed, and can be entered real-time route and be planned the application on site stage.
2) on-line stage;
The initial position of each intelligent body and corresponding environmental information in R intelligent body 2-1) are obtained, these information are known
, it can be obtained by the existing sensor technology such as radar, note current time is t moment, initializes pedometer c=0;
2-2) the obtained training of the position at each intelligent body current time and environmental information input step 1-3) is finished
BP neural network module, network module input potential field function coefficients (the repulsion gain k at the corresponding intelligent body current timer),
Adaptive optimization is carried out using BP neural network module pair potential field function coefficient;
For r-th of intelligent body Fr, the input of BP neural network module is the F of t momentrPosition (Fr_xt,Fr_yt), when t
The F at quarterrDirection of motion Fr_vt, the target position (G_x of t momentt,G_yt), the target direction of motion G_v of t momentt, t moment
Obstacle location (O_xt,O_yt), the obstacle direction of motion O_v of t momentt, remove FrThe position of outer other cooperative intelligent body t moments
(F1_xt,F1_yt)、(F2_xt,F2_yt)、…(FR_xt,FR_yt);The output of BP neural network module is:Intelligent body FrT when
Potential field function coefficients (the repulsion gain k at quarterr)。
The application on site stage is planned in real-time route, and since initial position, each moment plans that its is next to intelligent body
What position is step move to, using the obtained BP neural network module of step 1-3) training, with the position of current time intelligent body and
Environment exports potential field function coefficients (the repulsion gain k at the moment as input, forward directionr), reach online calculate in real time and works as front ring
The optimal repulsion gain k of border lower apronsrEffect.Subsequent time, the position and environment for inputting variation use BP neural network mould
Block output updates repulsion gain kr, have the function that adaptive optimization.
The path planning based on Artificial Potential Field 2-3) is carried out at current time, obtains the conjunction at each intelligent body current time
Power;Specific step is as follows:
Target point 2-3-1) is calculated to the gravitation of intelligent body;
To each intelligent body, calculate shown in gravitation potential field such as formula (1):
Wherein, kaFor gravitation gain, be set as 1, this be in order to further decrease BP neural network module export dimension from
And promote precision, reduce computation complexity, it is contemplated that the essence of resultant force potential field is the comparison of gravitation potential field and repulsion potential field, is used
The thought of scaling, to only need adaptive optimization repulsion gain kr?.ρ(p,pgoal)=| | pgoal- p | | it is intelligent body
To the Euclidean distance of target point, p is intelligent body position coordinates, pgoalFor aiming spot coordinate, n ∈ (0 ,+∞), by passing through
It tests convention and usually takes n=2.
Gravitation can be acquired by the negative gradient of gravitation potential field, as shown in formula (2):
2-3-2) calculating threatens area to the repulsion of intelligent body;
To each intelligent body, calculate shown in repulsion potential field such as formula (3):
Wherein, krFor the repulsion gain at step 2-2) obtained current time.ρ(p,pthreat)=| | pthreat- p | | it is intelligence
Can body to threat area boundary closest approach Euclidean distance (the exactly introducing of the distance avoids while closing on obstacle and mesh
When mark, repulsion covering gravitation causes goal nonreachable), pthreatFor the position coordinates for threatening area boundary closest approach, p0Most for repulsion
Big coverage usually takes n=by experience convention beyond threatening area to be reduced to 0, n ∈ (0 ,+∞) to intelligent body repulsion after the range
2。
Repulsion can be acquired by the negative gradient of repulsion potential field, as shown in formula (4):
2-3-3) seek resultant force;
The resultant force of each intelligent body is the vector sum of the intelligent body corresponding gravitation and repulsion, and the direction of vector is the intelligence
The direction of the movement at energy body next moment;
2-4) each intelligent body carries out movement until at the t+1 moment, update step counting according to the resultant force that step 2-3) is calculated
Device c=c+1, and determined:
If the t+1 moment, any intelligent body reaches target point in R intelligent body and all intelligent bodies do not reach and threaten area,
Then path planning success, method terminate;If the t+1 moment, in R intelligent body, any intelligent body, which is reached, threatens area, then path planning
Failure, method terminate;If the t+1 moment, all intelligent bodies did not reached success or failure condition, t=t+1 is enabled, when by t+1
Carve as new current time, then return to step 2-2), go successively to subsequent time (i.e. in next step) path planning;When
The step number of pedometer c reaches step number upper limit A set when walking (A as positive integer, if A takes 600) all intelligent bodies in the present embodiment
Still it is not up to success or failure condition, then path planning is terminated with time-out.
Below with reference to a specific embodiment, that the present invention is described in more detail is as follows.
The present embodiment synergistically carries out 1 airplane of blue party (target drone) with 2 airplane of red (red A machine and red B machine)
For game path planning.
Background:
Fig. 2 show the top plan view in both aircraft path planning region, initial position:Red A machine is inclined in figure lower part
A left side, red B machine is to the right in figure lower part, and there is a common target area in blue party target drone portion on the diagram to 2 airplane of red, to red
Square A machine and red B machine have a threat area respectively.Both aircraft obtains other side's real time position and heading device under intelligence support
Breath, target drone two side areas is target area, front region be threaten area, red aircraft by enter target area for the purpose of, to avoid
Threatening area is constraint, carries out coordinating game model path planning.
Basic assumption:
Assuming that both aircraft is the particle in two-dimensional space, 1 unit distance in top view represents 1 in real airspace
Km, the step-length in time domain are equivalent to 1 second in reality.Both aircraft remains that maximum speed is flown, wherein red aircraft
0.3 thousand meter per second of speed (about 0.9 Mach), 0.5 thousand meter per second of blue party air speed (about 1.5 Mach).
Initial position:
Red two-shipper uses the formation of 100 km of spacing, if target drone occurs at random using red vector as 0 degree
In -90 to+90 degree directions of red Two-fighter formation, about 300 kms of distance.
Target area:
Point line boundary as shown in Figure 2, is directed toward in target drone head ± 90 degree of radius maximum, reachable 110 kms.
Threaten area:
Be equivalent to the barrier in legacy paths planning problem, it is related to both aircraft relative velocity, therefore red A machine with
B machine has different threat areas because of course difference, and dashed boundaries as shown in Figure 2, maximum appears in both aircraft and meets head on to fly
When row, up to 105 kms.
Decision condition:
Area is threatened when any aircraft of red enters, is path planning " failure ";When any aircraft of red enters target area, and
It does not enter and threatens area, be path planning " success ".Red aircraft does not enter target still when more than maximum step number (being set as 600 steps)
Area threatens area, then is " flat " office.
Game environment:
According to current location and course caused by red aircraft back path planning decision, target drone is according to " course direction
At first enter threaten area red aircraft " strategy in real time adjustment course.
The present embodiment proposes a kind of coordinating game model paths planning method based on neural network and Artificial Potential Field, including following
Step:
1) off-line phase;
1-1) construct training sample set;
The known red A machine path optimizing in 6 initial positions respectively, i.e. 6 path optimizings in total, it is excellent to each
Change path, using the coordinate at t-th of moment and corresponding environment as the input of a training sample, by the potential field function of t moment
Coefficient (repulsion gain kr) output as the training sample, wherein the k of t momentrBy by t+1 moment and t moment
Intelligent body alternate position spike substitutes into Artificial potential functions and solves to obtain, it is assumed that a path optimizing includes that T is walked (at i.e. a total of T
Carve), then the path optimizing symbiosis is at T-1 training sample (each training sample includes corresponding outputs and inputs).
6 path optimizings are handled as above, and corresponding training sample is generated to every path optimizing;By institute
There is the training sample composing training sample set of path optimizing.
In the present embodiment, a part of red B machine as " environment " participates in building training sample set.
Since A machine and B machine are identical intelligent bodies, identical processing can also be done to B machine and is added to training sample concentration,
But it only used the path optimizing building training sample set of A machine in the present embodiment.
1-2) construct BP neural network module;
Using single hidden layer BP neural network module, it is 24 that middle layer neuron number, which is arranged, and activation primitive is Sigmoid letter
Number.
When carrying out path planning to the red A machine t+1 moment, the input of BP neural network module is:Red A machine t moment
Position (FA_xt,FA_yt), red A machine t moment direction of motion FA_vt, target area closest approach t moment position (G_xt,G_yt), when t
Carve target drone direction of motion G_vt, t moment barrier closest approach position (O_xt,O_yt), the position of another airplane B of t moment red
(FB_xt,FB_yt);BP neural network module exports:Potential field function coefficients (the repulsion gain k of red A machine t moment optimizationr)。
1-3) using step 1-1) obtain training sample set to the BP neural network module that step 1-2) is established carry out from
Line training obtains the BP neural network module that training finishes;
Using the sample of training sample set in step 1-1), using momentum gradient descent method off-line training BP neural network mould
Block, using mean square error less than 0.001 or training iteration reach 100000 wheels as cut-off condition.So far, preparation is completed, can be with
The application on site stage is planned into real-time route.
2) on-line stage;
The initial position and corresponding environmental information, note current time for 2-1) obtaining each intelligent body are t moment, initialization
Pedometer c=0;
2-2) the BP for finishing the obtained training of the location circumstances information input step 1-3 at each intelligent body current time)
Neural network module, network module input potential field function coefficients (the repulsion gain k at the corresponding intelligent body current timer), benefit
Adaptive optimization is carried out with BP neural network module pair potential field function coefficient;
The application on site stage is planned into real-time route, and since initial position, each moment is to red aircraft A and B points
Do not plan what position is its next step move to, the neural network module obtained using step 1-3) training, with current time intelligence
As input, forward direction exports potential field function coefficients (the repulsion gain k at the moment for the position of body and environmentr), reach online real-time
Calculate the optimal repulsion gain k of current environment lower apronsrEffect.Subsequent time, the position and environment for inputting variation use BP
Neural network module output updates repulsion gain kr, have the function that adaptive optimization.
The path planning based on Artificial Potential Field 2-3) is carried out at current time, obtains the conjunction at each intelligent body current time
Power;Specific step is as follows
Target point 2-3-1) is calculated to the gravitation of intelligent body;
To each intelligent body (the present embodiment is red aircraft), calculate shown in gravitation potential field such as formula (1):
Wherein kaIt is set as 1, n=2.
Gravitation can be acquired by the negative gradient of gravitation potential field, as shown in formula (2):
2-3-2) calculating threatens area to the repulsion of intelligent body;
Shown in repulsion potential field such as formula (3):
Wherein n=2.
Repulsion can be acquired by the negative gradient of repulsion potential field, as shown in formula (4):
2-3-3) seek resultant force
The resultant force of intelligent body is the vector sum of the intelligent body corresponding gravitation and repulsion, and the direction of vector is the intelligent body
The direction that (i.e. red aircraft) moves in next step.
2-4) each intelligent body carries out movement until at the t+1 moment, update step counting according to the resultant force that step 2-3) is calculated
Device c=c+1, and determined:
If the t+1 moment, any aircraft of red reaches target area and all red aircrafts do not reach and threaten area, then path is advised
It draws successfully to terminate;If any aircraft of red, which reaches, threatens area, path planning is ended in failure;If t+1 moment all red
Aircraft body does not arrive not up to success or failure condition, then enables t=t+1, using the t+1 moment as new current time, then weigh
New return step 2-2), go successively to subsequent time (i.e. in next step) path planning;When the step number of pedometer c reaches 600 step
When red aircraft is still not up to success or failure condition, path planning is terminated with draw.
Effect:
Fig. 3 shows that (S S Ge etc. 2000 in " New Potential Functions for using classical way
The method proposed in Mobile Robot Path Planning ") route programming result, show both aircraft path planning
Top plan view, initial position:For red A machine among figure lower part, red B machine is to the left in figure lower part, and blue party target drone is on the diagram
Portion is to the right.Repulsion gain k in this methodrIt is the average optimal value that statistics obtains with gravitation gain value.As seen from Figure 3, by
Adaptive optimization is unable in potential field function coefficients:First is that cause the path planning of red aircraft close to target drone adjustment course or from
It opens and insensitive, lacks game;Although moving tendency is similar second is that causing two airplane of red remote at a distance of target drone one nearly one,
Fail to generate collaboration.Final route programming result is failure.When using the method for the present invention under same scene, route programming result
As shown in figure 4, both aircraft initial position is constant, potential field function coefficients are according to real-time opposite situation adaptive optimization:Red B machine
Due to farther apart from target drone, repulsion gain krAdaptive to reduce so that its fast approaching target drone, forms friendly machine and support, this is association
With the embodiment of ability;Path planning latter end red aircraft will enter target area or threaten area, when the machine is directed toward in target drone course
Repulsion gain krIt significantly increases to flee from and threaten area, repulsion gain k after target drone course drift the machinerIt falls after rise rapidly and is forced with taking advantage of the occasion
Close-target area, this is the embodiment of game ability.
In addition, statistically verifying the performance and computing cost of this method by one group of large sample, and carried out with classical way
Comparison, as shown in table 1.Sample group includes the path planning of 1000 random initial positions, and statistical result showed compares classics side
Method, the success rate of this method are obviously improved up to 32.3 percentage points, and step number needed for success reduces about 15.37%, are sufficiently shown
Promotion of this method to path planning performance.On the other hand, although neural network forward-propagating process introduces additional meter
Expense is calculated, makes which increase 4 orders of magnitude, but still in millimeter magnitude, it is contemplated that will also be obtained in the following practical application into one
Soft or hard piece optimization is walked, the computing cost of this method is acceptable.
The overall performance of 1 the method for the present invention of table and classical way and expense contrast table
Claims (1)
1. a kind of coordinating game model paths planning method based on neural network and Artificial Potential Field, which is characterized in that this method includes
Following steps:
1) off-line phase;
1-1) construct training sample set;Specific step is as follows:
R-th of intelligent body, which arbitrarily 1-1-1) is chosen, from the R intelligent body for participating in coordinating game model path planning is denoted as Fr, obtain Fr1
Path optimizing;The path optimizing shares T step, then the path optimizing includes 1 to T moment intelligent body FrPosition coordinates;
To the path optimizing, using the intelligent body position coordinates at t-th of moment and corresponding environment as the input of a training sample;
Wherein the intelligent body position coordinates at t-th of moment are (Fr_xt,Fr_yt), Fr_xtFor the X-coordinate of r-th of intelligent body t moment,
Fr_ytFor the Y coordinate of r-th of intelligent body t moment;T-th of moment corresponds to environment:FrIn the direction of motion F of t momentr_vt, t
Target position (the G_x at momentt,G_yt), the target direction of motion G_v of t momentt, the obstacle location (O_x of t momentt,O_yt), t
The obstacle direction of motion O_v at momenttAnd remove FrOuter other cooperative intelligent body t moment positions coordinate is denoted as (F respectively1_
xt,F1_yt)、(F2_xt,F2_yt)、…(FR_xt,FR_yt);By the potential field function coefficients of t moment, that is, repulsion gain krAs this
The output of training sample, wherein the k of t momentrIt is artificial by substituting into the intelligent body alternate position spike at t+1 moment and t moment
Potential field function solves to obtain;Then the symbiosis of this path optimizing is at T-1 training sample;
1-1-2) repeat step 1-1-1), M path optimizing is obtained altogether to R intelligent body, and to every path optimizing generation pair
The training sample answered;By the training sample composing training sample set of all path optimizings;
1-2) construct BP neural network module;
1-3) the BP neural network for using momentum gradient descent method to establish step 1-2) using the training sample set of step 1-1)
Module carries out off-line training, when reaching training cut-off condition, obtains the BP neural network module that training finishes;
2) on-line stage;
2-1) obtaining the initial position of each intelligent body and corresponding environmental information, note current time in R intelligent body is t moment,
Initialize pedometer c=0;
The BP mind for 2-2) finishing the obtained training of the position at each intelligent body current time and environmental information input step 1-3)
Through network module, network module inputs the repulsion gain k at the corresponding intelligent body current timer;
For r-th of intelligent body Fr, the input of BP neural network module is the F of t momentrPosition (Fr_xt,Fr_yt), the F of t momentr
Direction of motion Fr_vt, the target position (G_x of t momentt,G_yt), the target direction of motion G_v of t momentt, the obstacle position of t moment
Set (O_xt,O_yt), the obstacle direction of motion O_v of t momentt, remove FrPosition (the F of outer other cooperative intelligent body t moments1_xt,
F1_yt)、(F2_xt,F2_yt)、…(FR_xt,FR_yt);The output of BP neural network module is:Intelligent body FrT moment reprimand
Power gain kr;
The path planning based on Artificial Potential Field 2-3) is carried out at current time, obtains the resultant force at each intelligent body current time;Tool
Steps are as follows for body:
Target point 2-3-1) is calculated to the gravitation of intelligent body;
To each intelligent body, calculate shown in gravitation potential field such as formula (1):
Wherein, kaFor gravitation gain, it is set as 1, ρ (p, pgoal)=| | pgoal- p | | for the Euclid of intelligent body to target point
Distance, p are intelligent body position coordinates, pgoalFor aiming spot coordinate, n ∈ (0 ,+∞);
Gravitation is acquired by the negative gradient of gravitation potential field, as shown in formula (2):
Fatt(p)=- ▽ Uatt(p)=kaρ(p,pgoal) (2)
2-3-2) calculating threatens area to the repulsion of intelligent body;
To each intelligent body, calculate shown in repulsion potential field such as formula (3):
Wherein, krFor the repulsion gain at step 2-2) obtained current time;ρ(p,pthreat)=| | pthreat- p | | it is intelligent body
To the Euclidean distance for threatening area boundary closest approach, pthreatFor the position coordinates for threatening area boundary closest approach, p0Most for repulsion
Big coverage;
Repulsion is acquired by the negative gradient of repulsion potential field, as shown in formula (4):
2-3-3) seek resultant force;
The resultant force of each intelligent body is the vector sum of the intelligent body corresponding gravitation and repulsion, and the direction of vector is the intelligent body
The direction of the movement at next moment;
2-4) each intelligent body carries out movement until at the t+1 moment, update pedometer c=according to the resultant force that step 2-3) is calculated
C+1, and determined:
If the t+1 moment, any intelligent body reaches target point in R intelligent body and all intelligent bodies do not reach and threaten area, then road
Diameter plans that successfully, method terminates;If the t+1 moment, in R intelligent body, any intelligent body, which is reached, threatens area, then path planning fails,
Method terminates;If the t+1 moment, all intelligent bodies did not reached success or failure condition, enable t=t+1, using the t+1 moment as
At new current time, then return to step 2-2), continue subsequent time path planning;When the step number arrival of pedometer c is set
When fixed step number upper limit A is walked, if all intelligent bodies are still not up to success or failure condition, path planning is denoted as time-out, side
Method terminates.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810907242.7A CN108827312B (en) | 2018-08-08 | 2018-08-08 | Cooperative game path planning method based on neural network and artificial potential field |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810907242.7A CN108827312B (en) | 2018-08-08 | 2018-08-08 | Cooperative game path planning method based on neural network and artificial potential field |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108827312A true CN108827312A (en) | 2018-11-16 |
CN108827312B CN108827312B (en) | 2021-10-08 |
Family
ID=64152803
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810907242.7A Active CN108827312B (en) | 2018-08-08 | 2018-08-08 | Cooperative game path planning method based on neural network and artificial potential field |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108827312B (en) |
Cited By (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109434836A (en) * | 2018-12-14 | 2019-03-08 | 浙江大学 | A kind of manipulator Artificial Potential Field space path planing method of combination ball tree-model |
CN109579853A (en) * | 2019-01-24 | 2019-04-05 | 燕山大学 | Inertial navigation indoor orientation method based on BP neural network |
CN110108292A (en) * | 2019-06-12 | 2019-08-09 | 山东师范大学 | Vehicle navigation path planing method, system, equipment and medium |
CN110275527A (en) * | 2019-05-29 | 2019-09-24 | 南京航空航天大学 | A kind of multi-agent system motion control method based on improvement mimicry physical method |
CN110356405A (en) * | 2019-07-23 | 2019-10-22 | 桂林电子科技大学 | Vehicle auxiliary travelling method, apparatus, computer equipment and readable storage medium storing program for executing |
CN110488872A (en) * | 2019-09-04 | 2019-11-22 | 中国人民解放军国防科技大学 | A kind of unmanned plane real-time route planing method based on deeply study |
CN110646009A (en) * | 2019-09-27 | 2020-01-03 | 北京邮电大学 | DQN-based vehicle automatic driving path planning method and device |
CN111123976A (en) * | 2019-12-24 | 2020-05-08 | 一飞智控(天津)科技有限公司 | Unmanned aerial vehicle cluster path planning processing method based on artificial potential field and unmanned aerial vehicle |
CN111367285A (en) * | 2020-03-18 | 2020-07-03 | 华东理工大学 | Coordinated formation and path planning method for wheeled mobile trolleys |
CN111443733A (en) * | 2020-05-25 | 2020-07-24 | 深圳市道通智能航空技术有限公司 | Unmanned aerial vehicle flight control method and device and unmanned aerial vehicle |
CN112269396A (en) * | 2020-10-14 | 2021-01-26 | 北京航空航天大学 | Unmanned aerial vehicle cluster cooperative confrontation control method for eagle pigeon-imitated intelligent game |
CN112947562A (en) * | 2021-02-10 | 2021-06-11 | 西北工业大学 | Multi-unmanned aerial vehicle motion planning method based on artificial potential field method and MADDPG |
CN112987784A (en) * | 2021-02-26 | 2021-06-18 | 辽宁工程技术大学 | Decision method for unmanned aerial vehicle cluster cooperative countermeasure |
CN113110604A (en) * | 2021-04-28 | 2021-07-13 | 江苏方天电力技术有限公司 | Path dynamic planning method based on artificial potential field |
CN113126647A (en) * | 2019-12-31 | 2021-07-16 | 北京理工大学 | Collaborative guidance method based on leader and follower principle |
CN113155128A (en) * | 2021-03-31 | 2021-07-23 | 西安电子科技大学 | Indoor pedestrian positioning method based on cooperative game UWB and inertial navigation |
CN113759900A (en) * | 2021-08-12 | 2021-12-07 | 中南大学 | Inspection robot track planning and real-time obstacle avoidance method and system based on obstacle area prediction |
CN114055471A (en) * | 2021-11-30 | 2022-02-18 | 哈尔滨工业大学 | Mechanical arm online motion planning method combining neural motion planning algorithm and artificial potential field method |
CN114254722A (en) * | 2021-11-17 | 2022-03-29 | 中国人民解放军军事科学院国防科技创新研究院 | Game countermeasure oriented multi-intelligent model fusion method |
CN114779821A (en) * | 2022-05-25 | 2022-07-22 | 四川大学 | Unmanned aerial vehicle self-adaptive repulsion coefficient path planning method based on deep learning |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107589752A (en) * | 2017-07-25 | 2018-01-16 | 天津大学 | Unmanned plane cooperates with formation realization method and system with ground robot |
CN108255165A (en) * | 2016-12-29 | 2018-07-06 | 广州映博智能科技有限公司 | Robot path planning method based on neighborhood potential field |
-
2018
- 2018-08-08 CN CN201810907242.7A patent/CN108827312B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108255165A (en) * | 2016-12-29 | 2018-07-06 | 广州映博智能科技有限公司 | Robot path planning method based on neighborhood potential field |
CN107589752A (en) * | 2017-07-25 | 2018-01-16 | 天津大学 | Unmanned plane cooperates with formation realization method and system with ground robot |
Non-Patent Citations (3)
Title |
---|
YONG-KYUN NA: ""Hybrid Control for Autonomous Mobile Robot Navigation Using Neural Network Based Behavior Modules and Environment Classification"", 《AUTONOMOUS ROBOTS》 * |
吴正平: ""基于改进人工势场法的AUV路径规划"", 《化工自动化及仪表》 * |
黎红: "自主移动机器人路径规划中的主要方法", 《中国电力教育》 * |
Cited By (30)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109434836A (en) * | 2018-12-14 | 2019-03-08 | 浙江大学 | A kind of manipulator Artificial Potential Field space path planing method of combination ball tree-model |
CN109579853B (en) * | 2019-01-24 | 2021-02-26 | 燕山大学 | Inertial navigation indoor positioning method based on BP neural network |
CN109579853A (en) * | 2019-01-24 | 2019-04-05 | 燕山大学 | Inertial navigation indoor orientation method based on BP neural network |
CN110275527A (en) * | 2019-05-29 | 2019-09-24 | 南京航空航天大学 | A kind of multi-agent system motion control method based on improvement mimicry physical method |
CN110108292A (en) * | 2019-06-12 | 2019-08-09 | 山东师范大学 | Vehicle navigation path planing method, system, equipment and medium |
CN110356405A (en) * | 2019-07-23 | 2019-10-22 | 桂林电子科技大学 | Vehicle auxiliary travelling method, apparatus, computer equipment and readable storage medium storing program for executing |
CN110488872A (en) * | 2019-09-04 | 2019-11-22 | 中国人民解放军国防科技大学 | A kind of unmanned plane real-time route planing method based on deeply study |
CN110646009A (en) * | 2019-09-27 | 2020-01-03 | 北京邮电大学 | DQN-based vehicle automatic driving path planning method and device |
CN111123976A (en) * | 2019-12-24 | 2020-05-08 | 一飞智控(天津)科技有限公司 | Unmanned aerial vehicle cluster path planning processing method based on artificial potential field and unmanned aerial vehicle |
CN113126647A (en) * | 2019-12-31 | 2021-07-16 | 北京理工大学 | Collaborative guidance method based on leader and follower principle |
CN113126647B (en) * | 2019-12-31 | 2022-07-19 | 北京理工大学 | Collaborative guidance method based on leader and follower principle |
CN111367285A (en) * | 2020-03-18 | 2020-07-03 | 华东理工大学 | Coordinated formation and path planning method for wheeled mobile trolleys |
CN111443733A (en) * | 2020-05-25 | 2020-07-24 | 深圳市道通智能航空技术有限公司 | Unmanned aerial vehicle flight control method and device and unmanned aerial vehicle |
WO2021238743A1 (en) * | 2020-05-25 | 2021-12-02 | 深圳市道通智能航空技术股份有限公司 | Flight control method and apparatus for unmanned aerial vehicle, and unmanned aerial vehicle |
CN111443733B (en) * | 2020-05-25 | 2023-09-15 | 深圳市道通智能航空技术股份有限公司 | Unmanned aerial vehicle flight control method and device and unmanned aerial vehicle |
CN112269396A (en) * | 2020-10-14 | 2021-01-26 | 北京航空航天大学 | Unmanned aerial vehicle cluster cooperative confrontation control method for eagle pigeon-imitated intelligent game |
CN112947562A (en) * | 2021-02-10 | 2021-06-11 | 西北工业大学 | Multi-unmanned aerial vehicle motion planning method based on artificial potential field method and MADDPG |
CN112947562B (en) * | 2021-02-10 | 2021-11-30 | 西北工业大学 | Multi-unmanned aerial vehicle motion planning method based on artificial potential field method and MADDPG |
CN112987784A (en) * | 2021-02-26 | 2021-06-18 | 辽宁工程技术大学 | Decision method for unmanned aerial vehicle cluster cooperative countermeasure |
CN112987784B (en) * | 2021-02-26 | 2024-03-26 | 辽宁工程技术大学 | Decision method for unmanned aerial vehicle cluster cooperative countermeasure |
CN113155128A (en) * | 2021-03-31 | 2021-07-23 | 西安电子科技大学 | Indoor pedestrian positioning method based on cooperative game UWB and inertial navigation |
CN113110604A (en) * | 2021-04-28 | 2021-07-13 | 江苏方天电力技术有限公司 | Path dynamic planning method based on artificial potential field |
CN113110604B (en) * | 2021-04-28 | 2022-08-23 | 江苏方天电力技术有限公司 | Path dynamic planning method based on artificial potential field |
CN113759900A (en) * | 2021-08-12 | 2021-12-07 | 中南大学 | Inspection robot track planning and real-time obstacle avoidance method and system based on obstacle area prediction |
CN114254722A (en) * | 2021-11-17 | 2022-03-29 | 中国人民解放军军事科学院国防科技创新研究院 | Game countermeasure oriented multi-intelligent model fusion method |
CN114254722B (en) * | 2021-11-17 | 2022-12-06 | 中国人民解放军军事科学院国防科技创新研究院 | Multi-intelligent-model fusion method for game confrontation |
CN114055471B (en) * | 2021-11-30 | 2022-05-10 | 哈尔滨工业大学 | Mechanical arm online motion planning method combining neural motion planning algorithm and artificial potential field method |
CN114055471A (en) * | 2021-11-30 | 2022-02-18 | 哈尔滨工业大学 | Mechanical arm online motion planning method combining neural motion planning algorithm and artificial potential field method |
CN114779821A (en) * | 2022-05-25 | 2022-07-22 | 四川大学 | Unmanned aerial vehicle self-adaptive repulsion coefficient path planning method based on deep learning |
CN114779821B (en) * | 2022-05-25 | 2023-06-27 | 四川大学 | Unmanned aerial vehicle self-adaptive repulsive force coefficient path planning method based on deep learning |
Also Published As
Publication number | Publication date |
---|---|
CN108827312B (en) | 2021-10-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108827312A (en) | A kind of coordinating game model paths planning method based on neural network and Artificial Potential Field | |
Ruan et al. | Mobile robot navigation based on deep reinforcement learning | |
CN106444738B (en) | Method for planning path for mobile robot based on dynamic motion primitive learning model | |
Chen et al. | Stabilization approaches for reinforcement learning-based end-to-end autonomous driving | |
CN110608743A (en) | Multi-unmanned aerial vehicle collaborative route planning method based on multi-population chaotic grayling algorithm | |
Xia et al. | Cooperative task assignment and track planning for multi-UAV attack mobile targets | |
Zhu et al. | Task assignment and path planning of a multi-AUV system based on a Glasius bio-inspired self-organising map algorithm | |
CN106502250B (en) | The path planning algorithm of multi-robot formation in three-dimensional space | |
CN111930121B (en) | Mixed path planning method for indoor mobile robot | |
CN107414830B (en) | A kind of carrying machine human arm manipulation multi-level mapping intelligent control method and system | |
CN108362284A (en) | A kind of air navigation aid based on bionical hippocampus cognitive map | |
Wu et al. | Navigating assistance system for quadcopter with deep reinforcement learning | |
Ciou et al. | Composite reinforcement learning for social robot navigation | |
Cao et al. | UAV path planning based on improved particle swarm algorithm | |
Panigrahi et al. | A novel intelligent mobile robot navigation technique for avoiding obstacles using RBF neural network | |
Gong et al. | UAV cooperative air combat maneuvering confrontation based on multi-agent reinforcement learning | |
Palm et al. | Particle swarm optimization of potential fields for obstacle avoidance | |
Mengying et al. | Online path planning algorithms for unmanned air vehicle | |
CN116562332B (en) | Robot social movement planning method in man-machine co-fusion environment | |
Duo et al. | A deep reinforcement learning based mapless navigation algorithm using continuous actions | |
Chen et al. | A multirobot cooperative area coverage search algorithm based on bioinspired neural network in unknown environments | |
CN105467841B (en) | A kind of class nerve control method of humanoid robot upper extremity exercise | |
Lu et al. | Autonomous mobile robot navigation in uncertain dynamic environments based on deep reinforcement learning | |
Yang et al. | An Online Interactive Approach for Crowd Navigation of Quadrupedal Robots | |
Chen et al. | Neural network predictive control for mobile robot using PSO with controllable random exploration velocity |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |