A Hybrid CNN-LSTM Architecture for Path Planning of Mobile Robots in Unknow Environments

Abstract

Path planning algorithms generally require several steps including mapping, localization, sensor data processing, etc. Deep learning-based approach has been proposed to achieve end-to-end path planning, alleviating human design tasks and saving the cost of building maps. In this paper, a CNN-LSTM model which combines convolutional neural network (CNN) with Long Short-Term Memory (LSTM) is constructed to accomplish the path planning task of mobile robots. The CNN structure extracts environment features from the sensor data and calculates the proper command velocity while the LSTM module aims to learn the relationship between continuous actions and smooth the velocity of the robot. The input data of the model is acquired by a lidar, which abstracts the environment information and increases the generalization capability of different environments. In the Robot Operating System (ROS)-based simulator, the proposed model is tested in both the training map and a more complex new map. It performs well in the path planning task even in the untrained map. Compared to CNN model, the CNN-LSTM model has a great advantage in continuous planning and velocity smoothing. Furthermore, both the linear and angular velocity calculated by this model is within limited range and real time performance of this algorithm is satisfactory.