RAIL: Recursive A* and Imitation Learning-Based Approach for Collision-Free Path Estimation
Subhadip Chandra, Kamlesh Kumar Dubey, Gaurav Agarwal, Anumoy Pathak · 2025
This paper discusses collision-free path planning for autonomous vehicles, a crucial aspect of their development. The authors propose a method to detect objects of different sizes, both dynamic and static, and compare algorithms for path planning and avoidance of collision. They use traditional path-finding algorithms like A* Algorithm, which include heuristics and Rapidly Exploring Random Tree (RRT) algorithms for real-time analysis. The authors discuss the challenges of ensuring path-planning algorithms scale to large environments and operate in real-time, and the need for future work on improving the robustness of path-planning algorithms in dynamic and unpredictable scenarios. The methodology involves creating a dataset, analyzing images, and training neural networks to imitate the collision-free path- finding algorithm. Static obstacles change their coordinates with time, making working in dynamic environments difficult. Recursive A* and Imitation Learning (RAIL) based approaches analyze obstacles frame by frame, making them static at a particular time-instant. The Generative Adversarial Network (GAN) and Inverse Reinforcement Learning are used to train an imitation learning model for faster trajectory finding in autonomous vehicles. Future development could involve using the Internet of Things (IoT) to share the trained trajectory-finding model results.