Robotic Path Planning Using Recurrent Neural Networks

Abstract

Every autonomous vehicle or any other application which requires reaching a destination, path planning is an important task. Once a destination is determined there can be various ways to reach there, but optimum use of resources to reach there is important. Hence path planning is an important part of navigation. There have been many developments over the years to realize the best path using different techniques. Recent developments on Neural Networks made it an important research topic for path planning. Many classifiers, learning algorithms have been tested and experimented for path planning. Deep Neural Networks have recently developed and seems to given better results compared to other machine learning techniques and there is larger space for research as well. CNN (Convolutional Neural Network) under the deep neural networks are observed and regarded to be usually applicable for path planning. But in this study Recurrent Neural Network technique (RNN) is being proposed. This is because RNN uses past inputs for present output which makes it applicable for temporal data. Hence in dynamic obstacle avoidance environment this method can help to predict path quicker than CNN. LSTM (Long Short Term Memory) is the algorithm that is mainly focused in this paper. LSTM is an extended version of RNN and this makes it suitable for congested environment. In this paper LSTM is applied for path planning in environment with obstacles and compared with the results of highly popular A* algorithm.