Deep Learning Based Path-Planning Using CRNN and A* for Mobile Robots

We present a novel approach for solving path-planning problems using a state-of-the-art data-driven deep learning technique. Although machine learning has been previously utilized for path planning, it has proven to be challenging due to the discrete nature of search algorithms. In this study, we propose a deep learning-based algorithm for path planning, which incorporates a Convolutional Recurrent Neural Network (CRNN) to create an end-to-end trainable neural network planner. This planner is then combined with the A* algorithm through an adaptive autonomy concept to autonomously select the best path planning strategy for increasing time efficiency and completeness. To train the CRNN, a labeled data set is generated autonomously from various maps by changing the starting and endpoints. The trained CRNN can find the shortest path from the starting point to the goal point by evaluating map images in one go. Additionally, the CRNN can predict way-points on image inputs. Our simulation results demonstrate that our proposed strategy is capable of finding the shortest path much faster than the A* algorithm in sparse environments, achieving a speed-up of up to 831 in some cases, which is exceptional.