Academia.eduAcademia.edu

Figure 16 – uploaded by Toshio Fukuda

See full PDFdownloadDownload figure
Fig. 16. An exchanging mutation. The membership functions are exchanged between fuzzy rules.

Figure 16 An exchanging mutation. The membership functions are exchanged between fuzzy rules.

Related Figures (43)

Fig. 1. Emerging synthesis of NN’s, FS, and EC. Each technique plays a specific role in intelligent systems. These techniques are combined to realize greater intelligence.
Fig. 1. Emerging synthesis of NN’s, FS, and EC. Each technique plays a specific role in intelligent systems. These techniques are combined to realize greater intelligence.
Fig. 3. The architecture of a robotic system with structured intelligence. The robotic system makes decisions according to information sensed from its environment. After taking actions, the robotic system receives an evaluation (reward or penalty). The decision-making rules are trained by the learning mechanism. Fig. 2. Architecture of structured intelligence. The structured intelligence is composed of three parts: neurodynamics (NN); skills (FS); and recursive consciousness (EC). The EC as the recursive consciousness makes valid decisions by combining the outputs of NN and FS.
Fig. 3. The architecture of a robotic system with structured intelligence. The robotic system makes decisions according to information sensed from its environment. After taking actions, the robotic system receives an evaluation (reward or penalty). The decision-making rules are trained by the learning mechanism. Fig. 2. Architecture of structured intelligence. The structured intelligence is composed of three parts: neurodynamics (NN); skills (FS); and recursive consciousness (EC). The EC as the recursive consciousness makes valid decisions by combining the outputs of NN and FS.
Fig. 5. External and internal evaluations for motion planning. The external evaluation is received directly from the environment and the internal evaluation quantifies the state of robotic system based on external evaluations. Fig. 4. A sensory network for perceiving environment. Accord- ing to the internal state of the robotic system, the sensed informa- tion is interpreted into qualitative information.
Fig. 5. External and internal evaluations for motion planning. The external evaluation is received directly from the environment and the internal evaluation quantifies the state of robotic system based on external evaluations. Fig. 4. A sensory network for perceiving environment. Accord- ing to the internal state of the robotic system, the sensed informa- tion is interpreted into qualitative information.
Fig. 7. Candidate paths avoiding collision with obstacles. In this case the mobile robot must pass through a narrow area if it selects a target direction reactively. Fig. 6. Sensing direction d; for detecting obstacles. The mo- bile robot has eight range sensors for measuring distance to the obstacles.
Fig. 7. Candidate paths avoiding collision with obstacles. In this case the mobile robot must pass through a narrow area if it selects a target direction reactively. Fig. 6. Sensing direction d; for detecting obstacles. The mo- bile robot has eight range sensors for measuring distance to the obstacles.
Fig. 8. A triangular membership function concerning the dis- tance x; between the mobile robot and obstacles in the zth rule. Hdanger_j iS the grade (degree) of danger and a;,; is central value of a membership function A; ; corresponding to safe, respectively.
Fig. 8. A triangular membership function concerning the dis- tance x; between the mobile robot and obstacles in the zth rule. Hdanger_j iS the grade (degree) of danger and a;,; is central value of a membership function A; ; corresponding to safe, respectively.
Fig. 9. State evaluation of the sensing direction d; by weight coefficients: Wd; = {0.1,0.3,1.0,0.3,0.1}. The collision-free direction is evaluated by the weighted average.
Fig. 9. State evaluation of the sensing direction d; by weight coefficients: Wd; = {0.1,0.3,1.0,0.3,0.1}. The collision-free direction is evaluated by the weighted average.
Fig. 11. Adjusting mechanism of sensory network in the collision avoidance. The velocity and attention range are adjusted according to the sensed information. Fig. 10. Membership functions corresponding to linguistic vari- ables of dangerous and safe based on the attention range. The membership functions are adjusted according to the time series of sensed information concerning the sparseness of obstacles in a work space. Even if the input data (x; and x2) are same, the resulting outputs jr and sro are different in the scaled membership functions of (a) minimal attention range and (b) maximal attention range.
Fig. 11. Adjusting mechanism of sensory network in the collision avoidance. The velocity and attention range are adjusted according to the sensed information. Fig. 10. Membership functions corresponding to linguistic vari- ables of dangerous and safe based on the attention range. The membership functions are adjusted according to the time series of sensed information concerning the sparseness of obstacles in a work space. Even if the input data (x; and x2) are same, the resulting outputs jr and sro are different in the scaled membership functions of (a) minimal attention range and (b) maximal attention range.
Fig. 12. The architecture of perception, decision making, and action for the mobile robot. The sensory network receives the distance x; to obstacles in the sensing direction d; and perceives the state of its environment as the degree of sparseness sprs(t) of obstacles. According to x; anc sprs(t), the mobile robot decides the velocity V(t) and A6(t) using the fuzzy rules for collisior avoidance, the control rules for the target trace, and the adjusting mechanism.
Fig. 12. The architecture of perception, decision making, and action for the mobile robot. The sensory network receives the distance x; to obstacles in the sensing direction d; and perceives the state of its environment as the degree of sparseness sprs(t) of obstacles. According to x; anc sprs(t), the mobile robot decides the velocity V(t) and A6(t) using the fuzzy rules for collisior avoidance, the control rules for the target trace, and the adjusting mechanism.
Fig. 13. A candidate solution representing fuzzy rules. A fuzzy tule is composed of condition and consequence parts. A member- ship function A; ; is represented by the central value a;,; and the width ;,;. In the consequence part, a singleton w,,; is used as each output parameter.
Fig. 13. A candidate solution representing fuzzy rules. A fuzzy tule is composed of condition and consequence parts. A member- ship function A; ; is represented by the central value a;,; and the width ;,;. In the consequence part, a singleton w,,; is used as each output parameter.
Fig. 14. A one-point crossover operation. In this example, the membership functions of x, in the third rules are exchanged between two candidate solutions.
Fig. 14. A one-point crossover operation. In this example, the membership functions of x, in the third rules are exchanged between two candidate solutions.
Fig. 15. A simple mutation. The validity of the second rule is changed from one to zero.
Fig. 15. A simple mutation. The validity of the second rule is changed from one to zero.
Fig. 17. The initial state of the mobile robot and the work space including several obstacles. The objective of the mobile robot is to reach the target point while avoiding collision with obstacles.
Fig. 17. The initial state of the mobile robot and the work space including several obstacles. The objective of the mobile robot is to reach the target point while avoiding collision with obstacles.
Fig. 18. Coding of membership functions corresponding to lin- guistic variables. The linguistic variables of dangerous and safe are encoded into 1 and 2, respectively. If neither of them is used in a fuzzy rule, 0 is used for representing the invalidity of the input variable xy.
Fig. 18. Coding of membership functions corresponding to lin- guistic variables. The linguistic variables of dangerous and safe are encoded into 1 and 2, respectively. If neither of them is used in a fuzzy rule, 0 is used for representing the invalidity of the input variable xy.
Fig. 20. Direction to the target point. The mobile robot moves toward the target point, while avoiding collision with obstacles.
Fig. 20. Direction to the target point. The mobile robot moves toward the target point, while avoiding collision with obstacles.
Fig. 21. The variation of the attention range of the mobile robot. The sensory network reduces the attention range when it detects obstacles in the attention range.
Fig. 21. The variation of the attention range of the mobile robot. The sensory network reduces the attention range when it detects obstacles in the attention range.
Simulation Results of the Mobile Robot with/Without Sensory Network Concerning the Performance Indexes in Fig. 19 (In the Column of the Sensing Range, SN Indicates a Simulation Result of the Mobile Robot with the Sensory Network)
Simulation Results of the Mobile Robot with/Without Sensory Network Concerning the Performance Indexes in Fig. 19 (In the Column of the Sensing Range, SN Indicates a Simulation Result of the Mobile Robot with the Sensory Network)
Fig. 23. Evaluation values of SSGA for optimizing fuzzy con- troller (case 1). SSGA reduces the evaluation values over iterations. Recall that “fitness” here is error and the task is to minimize.
Fig. 23. Evaluation values of SSGA for optimizing fuzzy con- troller (case 1). SSGA reduces the evaluation values over iterations. Recall that “fitness” here is error and the task is to minimize.
Fig. 24. Simulation results of collision avoidance of the mobile robot (case 2). (a) A trajectory before learning. (b) A trajectory after leaning by SSGA and delta rule.
Fig. 24. Simulation results of collision avoidance of the mobile robot (case 2). (a) A trajectory before learning. (b) A trajectory after leaning by SSGA and delta rule.
Fig. 25. Simulation results of collision avoidance of the mobile robot (case 3). (a) A trajectory before learning. (b) A trajectory after learning by SSGA and delta rule.
Fig. 25. Simulation results of collision avoidance of the mobile robot (case 3). (a) A trajectory before learning. (b) A trajectory after learning by SSGA and delta rule.
Simulation Results of Collision Avoidance Conceming the Perception Coefficient (7): ( Performance Indexes in Case 1 Where 7 = {0.5, 0.7, 0.8, 0.9, 0.92, 0.93, 0.94, 0.95, 0. a 0.97}; (b) Performance Indexes in Case 2 Where r = {0.93,0.94, 0.95, 0.96, 0.97}; (c) Performance Indexes in Case 3 Where 7 = {0.93, 0.94, 0.95, 0.96, 0.97} EC-based planning [25], [38], [49], [58], but we have not used the already learned control rules of robotic systems in path planning. The robotic system should adapt to dynamic environmental conditions, and therefore it should perform path planning on the basis of the acquired knowledge and skill. Consequently, this section discusses path planning of the mobile robot in unknown work spaces without map building. network becomes 12 and the mobile robot requires many more fuzzy rules than the fuzzy controller with the sensory network. The above simulation results show the mobile robot with the learned fuzzy controller can reach the target point without colliding with obstacles. The adjusting mechanism of the sensory network scales the output of the fuzzy controller according to the density of obstacles in the work space. Thus, the high perception capability is very important for robotic systems. Also, the closer linkage of perception, decision making, and action can realize the higher intelligence structured as a whole system.
Simulation Results of Collision Avoidance Conceming the Perception Coefficient (7): ( Performance Indexes in Case 1 Where 7 = {0.5, 0.7, 0.8, 0.9, 0.92, 0.93, 0.94, 0.95, 0. a 0.97}; (b) Performance Indexes in Case 2 Where r = {0.93,0.94, 0.95, 0.96, 0.97}; (c) Performance Indexes in Case 3 Where 7 = {0.93, 0.94, 0.95, 0.96, 0.97} EC-based planning [25], [38], [49], [58], but we have not used the already learned control rules of robotic systems in path planning. The robotic system should adapt to dynamic environmental conditions, and therefore it should perform path planning on the basis of the acquired knowledge and skill. Consequently, this section discusses path planning of the mobile robot in unknown work spaces without map building. network becomes 12 and the mobile robot requires many more fuzzy rules than the fuzzy controller with the sensory network. The above simulation results show the mobile robot with the learned fuzzy controller can reach the target point without colliding with obstacles. The adjusting mechanism of the sensory network scales the output of the fuzzy controller according to the density of obstacles in the work space. Thus, the high perception capability is very important for robotic systems. Also, the closer linkage of perception, decision making, and action can realize the higher intelligence structured as a whole system.
Simulation Results of the Collision Avoidance Where Obstacles Are Randomly Located in Work Spaces (Case 4 and 5); Each Value Indicates the Average of 50 Trials
Simulation Results of the Collision Avoidance Where Obstacles Are Randomly Located in Work Spaces (Case 4 and 5); Each Value Indicates the Average of 50 Trials
Fig. 27. Simulation results of collision avoidance of the mobile robot in the work space including obstacles randomly located (case 4). (a) A trajectory before learning. (b) A trajectory after learning by SSGA and delta rule.
Fig. 27. Simulation results of collision avoidance of the mobile robot in the work space including obstacles randomly located (case 4). (a) A trajectory before learning. (b) A trajectory after learning by SSGA and delta rule.
Fig. 30. Representation of an intermediate point for path plan- ning. IPyaiiaij and (IPzi,;,IP,4,;) ave the validity and the position of an intermediate point j of a candidate solution :, respectively. Fig. 29. Path planning of a mobile robot in an unknown work space. In the figure, x is an intermediate point. The mobile robot generates and traces several intermediate points.
Fig. 30. Representation of an intermediate point for path plan- ning. IPyaiiaij and (IPzi,;,IP,4,;) ave the validity and the position of an intermediate point j of a candidate solution :, respectively. Fig. 29. Path planning of a mobile robot in an unknown work space. In the figure, x is an intermediate point. The mobile robot generates and traces several intermediate points.
Fig. 31. Procedure of path planning in an unknown work space. Failure means the number of the failure to reach the target point in the simulation on SSGA for fuzzy controller. Initially, the mobile robot does not perform the path planning. After several simulations in SSGA for fuzzy controller, if it cannot reach the target point, the mobile robot gradually searches for a path in SSGA for path planning by adding intermediate points.
Fig. 31. Procedure of path planning in an unknown work space. Failure means the number of the failure to reach the target point in the simulation on SSGA for fuzzy controller. Initially, the mobile robot does not perform the path planning. After several simulations in SSGA for fuzzy controller, if it cannot reach the target point, the mobile robot gradually searches for a path in SSGA for path planning by adding intermediate points.
Fig. 32. (a) Random generation of an intermediate point. In the figure, a new intermediate point x is generated in the work space without considering other intermediate points. (b) Interpolated generation of an intermediate point. In the figure, a new intermediate point x is generated whithin the circle of which the center is defined as the middle point between two randomly selected points. (c) Neighborhood generation of an intermediate point. In the figure, a new intermediate point x is generated within the circle whose center is a randomly selected point.
Fig. 32. (a) Random generation of an intermediate point. In the figure, a new intermediate point x is generated in the work space without considering other intermediate points. (b) Interpolated generation of an intermediate point. In the figure, a new intermediate point x is generated whithin the circle of which the center is defined as the middle point between two randomly selected points. (c) Neighborhood generation of an intermediate point. In the figure, a new intermediate point x is generated within the circle whose center is a randomly selected point.
Fig. 33. A trajectory of mobile robot by the fuzzy controller without path planning (case 6). It cannot escape from a local area surrounded by obstacles.
Fig. 33. A trajectory of mobile robot by the fuzzy controller without path planning (case 6). It cannot escape from a local area surrounded by obstacles.
Fig. 34. Intermediate points generated by SSGA for path plan- ning (case 6). One intermediate point is generated between the starting point and the target point. The directly connected path collides with obstacles, but we need not consider the collision, since the mobile robot avoids the collision with obstacles by using the fuzzy controller. Fig. 35. A trajectory of the mobile robot traced by reactive motions (case 6). It reaches the target point without becoming trapping in local area.
Fig. 34. Intermediate points generated by SSGA for path plan- ning (case 6). One intermediate point is generated between the starting point and the target point. The directly connected path collides with obstacles, but we need not consider the collision, since the mobile robot avoids the collision with obstacles by using the fuzzy controller. Fig. 35. A trajectory of the mobile robot traced by reactive motions (case 6). It reaches the target point without becoming trapping in local area.
Fig. 36. A trajectory of mobile robot only by reactive motions in a work space including many obstacles and dead ends (case 7). The size of work space is 800 x 800 and the number of obstacles is 26. The mobile robot cannot escape from a local area surrounded by obstacles (first trap).
Fig. 36. A trajectory of mobile robot only by reactive motions in a work space including many obstacles and dead ends (case 7). The size of work space is 800 x 800 and the number of obstacles is 26. The mobile robot cannot escape from a local area surrounded by obstacles (first trap).
Fig. 37. A trajectory of mobile robot using the first intermediate point (case 7). The robot can avoid the first trap, but it cannot escape from the local area surrounded by obstacles near the target point (second trap).
Fig. 37. A trajectory of mobile robot using the first intermediate point (case 7). The robot can avoid the first trap, but it cannot escape from the local area surrounded by obstacles near the target point (second trap).
Fig. 38. An actual trajectory of the mobile robot traced by reactive motions through four intermediate points (case 7). It reaches the target point avoiding in the first and second traps.
Fig. 38. An actual trajectory of the mobile robot traced by reactive motions through four intermediate points (case 7). It reaches the target point avoiding in the first and second traps.
Fig. 39. (a) Evaluation values of the best candidate solution in the path planning (case 7). (b) TI number of intermediate points of the best candidate solution (case 7). The number of intermedia points increases in the beginning, but finally it is four. (c) The minimal distance to the target poi of the best candidate solution (case 7). The mobile robot succeeds to reach the target point | the interation 71.
Fig. 39. (a) Evaluation values of the best candidate solution in the path planning (case 7). (b) TI number of intermediate points of the best candidate solution (case 7). The number of intermedia points increases in the beginning, but finally it is four. (c) The minimal distance to the target poi of the best candidate solution (case 7). The mobile robot succeeds to reach the target point | the interation 71.
Simulation Results of the Path Planning (Case 6 and 7); Each Value Indicates the Average of 30 Trials
Simulation Results of the Path Planning (Case 6 and 7); Each Value Indicates the Average of 30 Trials
aaa MI aaa Dr. Kubota received the IECON’ 96 Best Paper Award and the CIRA’97 Best Paper Award.
aaa MI aaa Dr. Kubota received the IECON’ 96 Best Paper Award and the CIRA’97 Best Paper Award.
Related topics:
Biomedical EngineeringCollision AvoidanceMobile RobotFuzzy SystemFuzzy ControllerEvolutionary SystemElectrical and Electronic Engineering
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved