A Novel Learning Framework for Sampling-Based Motion Planning in Autonomous Driving

• DOI: 10.1609/aaai.v34i01.5473
• Publication Date: 2020-06-03T08:15:35Z
• Authors: Yifan Zhang; Jinghuai Zhang; Jindi Zhang; Jianping Wang; Kejie Lu; Jeff Hong
• Keywords: Smoothness; Sample (material); Sampling bias

Sampling-based motion planning (SBMP) is a major trajectory approach in autonomous driving given its high efficiency practice. As the core of SBMP schemes, sampling strategy holds key to whether smooth and collision-free can be found real-time. Although some bias strategies have been explored literature accelerate SBMP, generated under existing may lead sharp lane changing. To address this issue, we propose new learning framework for SBMP. Specifically, develop novel automatic labeling scheme 2-Stage prediction model improve accuracy predicting intention surrounding vehicles. We then an imitation generate sample points based on experience human drivers. Using results, design algorithm by strategically selecting necessary that avoid Data-driven experiments show proposed outperforms strategies, terms computing time, traveling smoothness trajectory. The results also our even better than