LMR-Based Controller Learning For Obstacle Avoidance In Mobile Robots

The essential work of a control problem is to approximate the underlying function.There are two basic ways to handle it. Building complicated model or using AIalgorithms can kill this problem. The LSM algorithm much depends on the given model.But AI methods resort to others ways to build the model. The function underlies in theweights for ANN. The function lies in the rules for fuzzy logical algorithm. WhileLocally Weighted Regression finds another way to build the model. Quite simplemodels are introduced to approximate complicated function in LWR algorithm.Traditional regression algorithms have to search for sophisticated models forcomplicated problems, and that’s the hardest part of LSM. LWR doesn’t focus on modelbut on how to find the neighbor points to build the model. Even simple model can dealwith nonlinear tasks through the neighborhood strategy. LWR algorithm is also calledlazy method or memory based method just because it makes the prediction when therecomes a query point and the samples are stored in the memory. Samples are just used inthe process of training for ANN or fuzzy logic algorithm. The prediction was madewithout the samples any more. There are several advantages of LWR, like the simplemodel and it’s extremely easy to adjust the model to the new samples. An obstacleavoidance controller and a wall following controller are designed and executed on themobile robot. The controllers have a good performance in our experiment. The mainwork we completed in this paper is as follows.The meaning of this research is shown for the first, then, the two inner controllersfor obstacle avoidance and wall following are designed as the teacher and the LWRbased controller for obstacle avoidance and wall following are designed. The innercontroller is used for taking samples for LWR. We use a new approach to adjust thereceptive field radius here. The residual of the prediction is used to form the approach inthis approach.We build the experiment environment for the research, and we run the LWRcontroller on the Khepera-II mobile robot. The performance is better than LSM, innercontroller and the global receptive strategy LWR. That means local receptive fieldstrategy LWR can be used to mobile robotics field and LWR is capable of dealinghighly nonlinear problem.