A comprehensive learning framework for sampling-based motion planning in autonomous driving

Abstract

Motion Planning serves as the key for self-driving vehicle to achieve full autonomy. Recently, Sampling-based motion planning (SBMP) has become a major motion planning approach in the field given its high efficiency in practice and its capability of potentially finding the optimal path. Although plenty of SBMP algorithms have been proposed in the literature, most of them fail to generate desirable collision-free trajectories in a dynamic environment. Besides, even though some algorithms are designed to solve the motion planning problem in some specific scenarios such as urban area, they cannot generalize the performance and the efficiency as well as the outcome depends largely on the sampled points and provided templates. In this work, we firstly propose five criteria that the motion planning algorithm should follow, which include human-like driving behaviors, safety in a dynamic environment, consistency, efficiency, and interaction with environments respectively. Then, we design a comprehensive learning framework for real-time motion planning, aiming to generate collision-free trajectory for self-driving vehicle on par with or even better than human drivers by taking these criteria into consideration. Specifically, we develop a novel automatic labeling scheme and a 2-Stage prediction model to improve the accuracy in predicting the intention of surrounding vehicles. Utilizing the prediction results, we design a new biased sampling strategy to generate collision-free trajectories and avoid undesirable driving behaviors in a dynamic environment. Then, to accelerate the planning phase of SBMP algorithm, we conduct a comprehensive investigation on the lane-changing cases of NGSIM dataset and propose sampling states dynamically according to the complexity of the environment. Finally, we focus on the time lag problem, which will be introduced in this work, and come up with a feasible solution to emphasize the consistency between successive replanning. Extensive experiments will be conducted to show that the proposed framework can achieve the state-of-the-art performance.

Access is unrestricted.