Neural Network Based Research On Autonomous Behavior Of Evolutionary Robot

The ability of autonomously behaving of robot had been the target of robot researcher for a long time. The traditional design method of "Sense-Plan-Action" had many difficulties. Behavior based robotics gave a new answer to it. But it also left the problem of manually designing the behaviors. Instead, evolutionary robotics used the evolutionary algorithm to get the control parameters rather than designing it manually.Firstly, this paper gave a fitness evaluation method based on fuzzy reasoning. At first, it dealt with the evolution items with the fuzzy method. Then, it introduced the evaluation method of human beings by fuzzy reasoning. At last, it got the fitness of the robot to the environment and the work by the process of clarification. But, as the number of the items of evaluation increased, the rules needed increased very fast. In addition, it was needed to adjust and modify the fuzzy rules when new evaluation items were added or the importance of the items was changed. In order to deal with this problem, this paper used the weight way of fuzzy synthesis evaluation to simplify the fuzzy reasoning based method and got the new method named:simplified fuzzy evaluation.Secondly, this paper evolved the wander behavior using the evolutionary robotics method. The fuzzy reasoning based and simplified fuzzy evaluation was used to get the fitness separately. The simulation results showed that the motion smoothness and fastness item of the wander behavior converged fast, collision item converged slowly. So the two fitness evaluation methods were reasonable and effective.Thirdly, the paper illustrated a method of designing a targets collection robot. It divided the work process of the robot into two complex behaviors:target searching and homing. With the behavior dividing and coordinating method, this paper raised the layered method of designing complex behavior:the low layer was composed by the rule based field method; the high lower was composed by a group of sub-behaviors. The relationship of the behaviors was sorted into conflict and cooperative. The conflict behaviors were triggered by priority, the behavior with the higher priority was triggered first. The cooperative behaviors produce an output by coordinating the outputs of all of them. This method was used to design the targets collection and homing behavior. The simulation showed that the effect was good. At last, the targets collection and homing behavior were tested in an integration scene, which showed that the method worked well. In addition, in order to test the flexibility of the robot in other environment, a new environment was build. The test in the new environment showed that the robot could adapt the new environment fairly well.