Toward Efficient Trajectory Planning based on Deterministic Sampling and Optimization

The solution of optimization-based planners depends heavily on a good initialization and their run-time is often non-deterministic, especially in dense obstacle fields. Sampling-based planning, whether probabilistic or deterministic, is a well-established method for exploring the search space. However, the downsides are also obvious: potentially intractable computational overhead, the curse of dimensionality and the sub-optimality due to discretization. Motivated by this observation, this paper introduces a real-time trajectory planning algorithm based on the combination of sampling and optimization approaches, which is applicable to autonomous vehicles operating in highly constrained environments. A maximum corridor width region and initial drivable path are firstly extracted from deterministic sampling. Then the initial path is further optimized through a splined-based quadratic programming and appended with a speed profile. This planner is scalable to both high-speed off-road scenarios and structured urban driving.

I. Introduction

In July 2019, an all-terrain vehicle Mission-Terminator (Fig. 1) completed a task-based off-road Unmanned System Challenge fully autonomously and won the first place[1]. The vehicle demonstrated the ability to complete an integrated task in desert grassland environment, while safely interacting with other human driven vehicles.