Enable faster and smoother spatio-temporal trajectory planning for autonomous vehicles in constrained dynamic environment

Abstract

Trajectory planning is of vital importance to decision-making for autonomous vehicles. Currently, there are three popular classes of cost-based trajectory planning methods: sampling-based, graph-search-based, and optimization-based. However, each of them has its own shortcomings, for example, high computational expense for sampling-based methods, low resolution for graph-search-based methods, and lack of global awareness for optimization-based methods. It leads to one of the challenges for trajectory planning for autonomous vehicles, which is improving planning efficiency while guaranteeing model feasibility. Therefore, this paper proposes a hybrid planning framework composed of two modules, which preserves the strength of both graph-search-based methods and optimization-based methods, thus enabling faster and smoother spatio-temporal trajectory planning in constrained dynamic environment. The proposed method first constructs spatio-temporal driving space based on directed acyclic graph and efficiently searches a spatio-temporal trajectory using the improved A* algorithm. Then taking the search result as reference, locally convex feasible driving area is designed and model predictive control is applied to further optimize the trajectory with a comprehensive consideration of vehicle kinematics and moving obstacles. Results simulated in four different scenarios all demonstrated feasible trajectories without emergency stop or abrupt steering change, which is kinematic-smooth to follow. Moreover, the average planning time was 31 ms, which only took 59.05%, 18.87%, and 0.69%, respectively, of that consumed by other state-of-the-art trajectory planning methods, namely, maximum interaction defensive policy, sampling-based method with iterative optimizations, and Graph-search-based method with Dynamic Programming.

A Sufficient Condition for Convex Hull Property in General Convex Spatio-Temporal Corridors

“…There is an additional 5-meter safety margin behind the cut-in vehicle. The closeloop longitudinal dynamics of the ego vehicle is simulated by a first-order dynamics with a time constant of 0.3 s. The corridors are generated with the algorithm presented in [13] and [12], respectively for general convex shape and trapezoidal shape corridors.…”

This is the Way: Differential Bayesian Filtering for Agile Trajectory Synthesis

“…There is existing work on trajectory synthesis for autonomous driving [9]- [11]. However, these works only consider a "0 th " order view of the problem i.e.…”

End-to-End Autonomous Driving Through Dueling Double Deep Q-Network

“…Traditional hierarchical methods always decompose the problem into several parts like environment perception, path planning, and motion control, which are too complicated to design [1,2,17]. Instead, the structure of end-to-end autonomous driving is relatively simple and straightforward in that they directly map raw sensor data to vehicle control signals through deep neural network.…”