Seong-Gon Kim1 and Yong-Gi Kim2 University of Florida 2Gyeongsang National University
Autonomous underwater vehicles should possess intelligent control software that performs intellectual functions such as cognition, decision and action, which originally belong to the ability of domain expert, since the unmanned underwater robot is required to navigate in hazardous environments where humans do not have direct access to. In this paper, we suggest an intelligent system architecture called the RVC model, which can be applied to various kinds of unmanned vehicles. The architecture consists of the collision avoidance system, the navigation system, and the collision-risk computation system. The RVC architecture is devised to make use of artificial intelligence techniques, and to provide the subsystems structural and functional independency.
The collision avoidance system adopts a new heuristic search technique for the autonomous underwater vehicles equipped with obstacle avoidance sonar. The fuzzy relation product between the sonar sections and the properties of real-time environment is used to decide the direction for the vehicle to proceed. The simulation result leads to the conclusion that the heuristic search technique enables the AUV to navigate safely through obstacles and reach its destination goal with the optimal path. The navigation system executes the offline global path planning for the AUV to guarantee the safe and efficient navigation from its start point to the target destination. The system also does the duty of monitoring and controlling the vehicle to navigate following the directed path to destination goal. The collision-risk computation system produces a degree of collision risk for the underwater vehicle against surrounding obstacles using information from the circumstances, obstacles, and positions. The degree is provided to the collision avoidance system as one of the decision tools used for safe avoidance with the obstacles. A 3D simulator is developed to test the AUV navigation system based on the RVC model. The goal of the simulator is to serve as a testing ground for the new technologies and to facilitate the eventual transfer of these technologies to real world applications. The simulation system consists of an environment manager, objects and a 3D viewer. Objects model all physical elements such as the map, obstacles and the AUV. The environment manager plays the role of an intermediary, which allows created objects to interact with each other, and transmits information of the objects to the 3D viewer. The 3D viewer analyzes the received information and visualizes it with 3D graphics by using OpenGL primitives.
The navigation system for autonomous underwater vehicles needs various techniques to be effectively implemented. The autonomous technique usually contains complicated and uncertain factors and thus makes use of some artificial intelligence methods to solve the problems. Artificial intelligence techniques are classified largely into two categories. One is the symbolic AI technique, such as knowledge-based system, which operates in ways similar to the human thought process, and the other is the behaviour-based AI technique such as neural network or fuzzy which behaves much like human sensorial responses. The former is considered a higher-level intelligence but it alone is not enough to make a system conduct intelligently in domains where very sophisticated behaviours are needed.
Research in autonomous navigation systems became very active with the rapid advancement of hardware technologies during the end of the 20th century. Researchers had tried to implement intelligent control for autonomous navigations using symbolic AI techniques but they could not succeed because of the difference in representation methods between the symbolic AI techniques they were attempting to use and the actual information needed to operate the navigation system. The symbolic AI technique is adequate for problems which are well-defined and easy to represent but not for real world problems which are usually ill-defined and in most cases have no limitation. These difficulties made researchers work on the development of AI techniques that were good for solving real world problems. Reactive planning (Agre et al., 1987), computational neuroethology (Cliff, 1991), and task-oriented subsumtion architecture (Brooks, 1986) are the results of the research, and are called behaviour-based AI (Turner et al., 1993). Many researches concluded that symbolic AI or behaviour-based AI techniques alone cannot reach the allowable goal for the navigation system of unmanned underwater vehicles (Arkin, 1989) and recent researches on autonomous navigations are focused on using both AI techniques and improving the performance of the system (Arkin, 1989; Turner, 1993; Scerri & Reed, 1999; Lee et al. 2004; Bui & Kim, 2006). The two AI techniques have different characteristics and thus is hard to combine the two techniques into a single system effectively. In this article, an intelligent system model, called the RVC (Reactive Layer-Virtual World-Considerate Layer), is introduced for the effective combination of symbolic and behaviour-based AI techniques into a system.
Fig. 1. RVC intelligent system
Fig.1 is the schematised RVC intelligent system model. The model is conceptualised for cordial combination of the two different AI techniques, and it also enhances the structural and functional independency of each subsystem, such as collision avoidance system, navigation system, or collision risk computation system. In this model, the reactive layer processes the uncertain problems in the real world and then passes the symbolized results to the considerate layer where the symbolic AI technique makes use of the information for the final decision. For this procedure, the model needs a common information storage space, where the information produced from the reactive layer is represented in real-time before it is consumed by the considerate layer. From the considerate layer's point of view, the information storage space resembles a subset of real world, and thus this storage space will be referred to as a 'Virtual world' henceforth.
2.2 Autonomous navigation architecture based on RVC intelligent system model
Autonomous navigation system based on the RVC intelligent system model uses the concept of information production/consumption and client/server for transferring the collected information from the real world to each module of the system in real-time. For this purpose, the intelligent navigation system contains functions such as memory management, data communication, and scheduling. Data communication in the system adopts the TCP/IP protocol, and this makes the system platform-independent and thus makes load balancing smooth. The scheduling function synchronizes the exchanging of real-time data among the modules, and it also processes possible errors in the system. The RVC intelligent system model guarantees independency among the modules in the system, and this enables the parallel development of each system module. Fig. 2 is the autonomous navigation architecture based on the RVC system model.
Fig. 2. Autonomous navigation architecture based on RVC system model
Fig. 2. Autonomous navigation architecture based on RVC system model
3. Subsystems for autonomous navigation system 3.1 Collision avoidance system
Relational representation of knowledge makes it possible to perform all the computations and decision making in a uniform relational way, by mean of special relational compositions called triangle and square products. These were first introduced by Bandler and Kohout and are referred to as the BK-products in the literature. Their theory and applications have made substantial progress since then (Bandler & Kohout, 1980a, 1980b; Kohout & Kim, 1998, 2002; Kohout et al., 1984).
There are different ways to define the composition of two fuzzy relations. The most popular extension of the classical circular composition to the fuzzy case is so called max-min composition (Kohout et al., 1984). Bandler and Kohout extended the classical circular products to BK-products as sub-triangle (<l, "included in"), super-triangle ([>, "includes"), and square (□, "are exactly the same"). Assume the relations R and S are fuzzy relations, then the R-afterset of x, xR and the S-foreset of z, Sz, obviously are fuzzy sets in Y. The common definition of inclusion of the fuzzy set xR in Y in the fuzzy set Sz in Y is given by (1).
A fuzzy implication is modeled by means of a fuzzy implication operator. A wide variety of fuzzy implication operators have been proposed, and their properties have been analyzed in detail (Bandler & Kohout, 1980c; Lee et al., 2002). For this study, we make use only of operator 5 as shown in (2).
Using (2), with n the cardinality of Y, we easily obtain the definitions for the sub-triangle and supper-triangle products in (3), (4) while the square product using the intersection and the minimum operator is shown in (5) and (6) respectively.
X(R < S)Zj = - V min(1, 1 - xR(y) + SZj (y)) (3)
X (R > S)zj = -^ min(l, 1 + xR(y) - Sz} (y)) (4)
Xi(R □ S)zj = Xi(R < S)zj n Xi (R > S)zj (5)
Xi(R □ S)zj = min(xi(R < S)zj , Xi (R > S)zj) (6)
Along with the above definitions, a-cut and Hasse diagram are also the two important features of this method. The a-cut transforms a fuzzy relation into a crisp relation, which is represented as a matrix (Kohout & Kim, 2002; Kohout et al., 1984). Let R denotes a fuzzy relation on the X x Y , the a-cut relation of R is defined as the equation (7).
The Hasse diagram is a useful tool, which completely describes the partial order among the elements of the crisp relational matrix by a Hasse diagram structure. To determine the Hasse diagram of a relation, the following three steps should be adopted (Lee & Kim, 2001). Step 1. Delete all edges that have reflexive property. Step 2. Eliminate all edges that are implied by the transitive property.
Step 3. Draw the diagraph of a partial order with all edges pointing upward, and then omit arrows from the edges.
In this study it is required that obstacle avoidance sonar range can be partitioned into several sub-ranges. One of these represents for the successive heading candidate for AUVs to go ahead. Whenever obstacle is detected, the sonar return is clustered and the sections in which obstacles present can be identified. The sonar model is illustrated as in Fig.3. Domain experts who have wide knowledge about ocean science could give the properties about the environmental effects to the of AUVs navigation.
A forward looking obstacle avoidance sonar whose coverage range can be divided into multi-sections is used to determine a heading candidate set S. Otherwise, a property set P describes the effects of AUVs toward the real time environment. The fuzzy rule base and membership function for the corresponding property can be estimated subjectively by the expert knowledge. With the set of the candidateS = (si,s^,S3,...,s} and the set of environmental properties P = (p1,p2,...,p}, the relation R is built as (8). The elements rij of this relation mean the possibility the section si can be characterized by the property pj. The value of rij is calculated by means of the rule bases with the membership functions.
Was this article helpful?