Motion Planning Integrating Human Driving Style in Autonomous Driving

Project: Research

View graph of relations


In recent years, autonomous driving has demonstrated its feasibility but there are still many issues to be solved before the large-scale adoption of self-driving vehicles. One of the key issues is how to address the co-existence of human driven vehicles and self-driving vehicles, which will be the norm in the coming decades. Such a co-existence brings challenges and opportunities to motion planning in autonomous driving. On the one hand, a human driver usually maintains a consistent driving style, which offers the opportunity for an autonomous vehicle to recognize the driving style of a surrounding vehicle and then predict its trajectory. Thus, an autonomous vehicle can have a better environment perception and take safer driving behaviors. On the other hand, the driving behaviors of an autonomous vehicle shall make both passengers in the car and other nearby drivers feel comfortable, which can be accommodated by training an autonomous vehicle to have a human-like driving style. To address the above issue, this proposal will thoroughly consider the impact of human driving style on the motion planning of an autonomous vehicle. Specifically, this proposal aims to develop a novel motion-planning framework that organically integrates two types of motion planning, namely, algorithmic motion planning, which can generate collision-free and energy-efficient trajectory, and imitation learning-based motion planning, which emulates the human driving style and generates more interpretable trajectory. In the proposed motion-planning framework, state-of-the-art learning techniques will be adopted to provide more accurate inputs for algorithmic motion planning. In particular, we will first cluster trajectories in available datasets according to different driving styles and driving behaviors. For each given driving behavior and driving style pair, a prediction model is trained. The trained models will help an autonomous vehicle predict the future trajectory of a surrounding vehicle, which will provide better environment perception for motion planning. To incorporate the comfort of passengers in the autonomous vehicle, preferred human driving style will be utilized to select the driving behavior, e.g., car-following or lane-changing, which determines the goal point for motion planning. At last, algorithmic motion planning algorithms will be developed to generate a collision-free and human-decided trajectory, where a penalty is introduced to minimize the deviation from the preferred human-decided trajectory. The project will enhance the safety in autonomous driving, build the trust bond between passengers and the autonomous vehicle, and reduce the disruption of autonomous vehicle’s driving behavior to other human driven vehicles. 


Project number9042950
Grant typeGRF
Effective start/end date1/01/21 → …