Toward Safe, Feasible and Human-Like Motion Generation for Urban Autonomous Driving

Motion prediction and motion planning problems are deeply interconnected for urban autonomous driving. In this project, we target the problem of safe, feasible and human-like motion generation. A non-conservatively defensive strategy (NCDS) is proposed, which is defensive under the worst case to guarantee safety, but not conservative so as to degrade driving quality and generate behaviors which are not human-like. A planning framework based on NCDS is constructed, which generates desirable tentative actions by incorporating decision-making and planning in a variety of challenging urban driving scenarios. Also, a prediction framework based on NCDS is proposed, which can obtain most possible future motions of others under each intention, and the intention probabilities without training a behavioral model with collected data. Preliminary results of safe and feasible motion generation via deep net are also illustrated in the report. 1 Overview Motion prediction of other vehicles and motion planning of the host vehicle are closely related to each other. Motion prediction is a stochastic version of motion planning, and the mechanism of motion planning should be considered to predict the motion of others. Correspondingly, motion planning is to choose the best among motions predicted by others, and how others expect the motion of the host vehicle should be considered. Due to the inherent connection of prediction and planning, we handle the two problems together as motion generation. Motions generated by prediction and planning should be 1) safe to avoid collisions, 2) feasible according to the vehicle kinematics and dynamics, and 3) human-like to be not overly conservative to enhance driving quality. In order to achieve safe, feasible and human-like motion generation, a non-conservatively defensive strategy (NCDS) is proposed. It is a driving strategy which is defensive under the worst case to guarantee safety, but is not conservative so as to degrade driving quality. NCDS can be applied to decision-making and motion planning, as well as intention recognition and motion prediction. When NCDS is applied to decision-making and planning, a unified planning framework under uncertainty can be constructed in various kinds of urban driving scenarios. In the planning framework, feasibility and safety are guaranteed by providing the limits steering angle, tire friction and engine traction, as well as by checking collisions under the worst case, such as violations or aggressive behaviors of others. However, overly conservative actions are avoided by exploiting the probabilities of possible intentions of others to form an expected cost, which comprehensively considers time-efficiency, comfort, road structure and traffic rules under each case. The probabilities are obtained by behavioral models trained by real world motion data. In practical, we cannot collect large amounts of motion data and training behavioral models for all scenarios which might be encountered by autonomous vehicles. An intention recognition and motion prediction framework is required to handle uncommon scenarios so that probabilities and most possible predicted motion of each possible intention of others can be obtained. The NCDS we proposed can also be used to approximate the intention probability and predict most possible motion without training behavioral model with motion data.