Risk-Aware Net: An Explicit Collision-Constrained Framework for Enhanced Safety Autonomous Driving

Motion planning is a vital part of autonomous driving. To ensure the safety of autonomous vehicles, motion planning algorithms need to precisely model potential collision risks and execute essential driving maneuvers. Drawing inspiration from the human driving process, which involves making preliminary decisions based on an initial assessment of the driving scenario and subsequently generating refined trajectories, this study presents a two-step framework to replicate this process. In the first step, we use a learning-based network to create an initial plan. To model potential collision risks, we introduce a corridor-based constraint that considers the environment and trajectory information. By combining imitation loss with corridor-constrained loss through supervised learning, we incorporate driving heuristics into the plan. In the second step, we refine the plan using an optimization-based pose refinement module. This module optimizes the trajectory at a point level to create the final, improved path. To validate our approach, we trained and evaluated our methodology using the large-scale Lyft dataset. The results show that our method outperforms other baseline approaches by reducing collision rates and discomfort rates.