Path Planning Based on Combination of Sampling and Learning-From-Demonstration for Static Obstacle Avoidance of Autonomous Vehicles

This paper presents a sampling-based path planning algorithm for autonomous vehicles with the existence of static obstacles. Conventional sampling-based path planning algorithms pose challenges in balancing driving performance and sample efficiency. To overcome this issue, Learning-from-Demonstration (LfD) approach has been combined with the sampling approach based on expert driving data for efficient path planning with enhanced driver acceptance. First, a neural network is designed to probabilistically estimate proposal points of the ego vehicle’s future path using environmental information and vehicle sensor information. The network is trained with expert driving data. Second, a driving path is planned by sampling path primitives in spatial domain. To generate the primitives, terminal points of the primitives are sampled within reliable ranges centered around the proposal points which are derived from the neural network. The driving path is determined by the evaluation of primitives according to cost function and feasibility constraints. To enhance the driving performance, two-piece path primitives are used to compose the driving path, rather than a single-piece path primitive. Simulation studies have been conducted to compare driving performance and efficiency of the proposed algorithm with other approaches. Vehicle tests have been conducted based on autonomous driving. Test results show that the proposed combination of LfD and sampling approach provides feasible and acceptable path planning results with enhanced computational efficiency.