渐优随机采样算法在结构化道路无人驾驶中的应用

Asymptotically Sampling-based Algorithm with Applications to Autonomous Urban Driving on Structured Road

为了解决随机采样算法在结构化道路无人驾驶应用中无法优化收敛的问题,采用渐进优化的采样算法框架设计符合驾驶需求的规划算法。针对渐进优化算法的耗时问题,首先选择不需要Steer(转向函数)的SST算法作为基础框架以规避求解边界值问题。其次,算法融入"Anytime"策略以提高优化解的利用率。再次,改进的闭环控制策略能减少车辆的实际轨迹与规划路径的误差。在设计的闭环策略中,应用4-D车辆运动模型以保证规划路径符合车辆的实际运动轨迹。为了保证驾驶的安全和舒适,设计了一个综合四重因素的代价函数,且根据不同的驾驶场景调整相应的权重参数。最后,利用真实的无人车在无人驾驶城市测试道路上进行测试,测试场景包括前方静态障碍物躲避、前方动态障碍物跟随以及超车和复合动静态障碍物。测试中,采用车辆的速度和转向数据代表算法的优化收敛特性和运动平稳性。研究结果表明:设计的算法能在时速30km·h-1下完成避障、跟车、超车等机动;无人车在跟驰决策下可保持30km·h-1的最高速度,在避障过程中可实现最高15km·h-1的速度,在跟车决策下可根据前车速度变换自身速度以保持合理的车距和运动平滑性。

To solve the optimization and convergence problems during urban driving on the structured road, this study adopts an asymptotically optimal sampling-based framework for designing a planning algorithm applied to the autonomous vehicle. To improve the real-time computational problem, we chose an asymptotically optimal sampling-based method, called Stable Sparse RRT （SST）, as the algorithmic frame, which does not require a time-consuming steering function to avoid solving the BVP. Furthermore, the ＂anytime＂ strategy was used to alleviate the time-consumption problem and improve the utilization efficiency of the optimal solution. In the algorithm, an improved close-loop control strategy was integrated to reduce the error between the planner and the real motion. In the closed loop system, a four-dimensional dynamic model was implemented to confirm the feasibility of the planned trajectory for the vehicle movement. We also designed an objective function to consider four weighted cost terms to ensure safety and comfort and tuned the weight coefficients under the driving conditions. Finally, we used an autonomous vehicle to conduct a field test along an open urban road as specified for the testing of autonomous vehicles. The testing scenario included avoiding a front static obstacle, following and overtaking a front dynamic obstacle, and applying a complex scenario with both dynamic and static obstacles. During the test, the velocity and steering data were collected to show the convergence properties and smoothness of the motion. The results demonstrate that the proposed algorithm can perform well when driving at a maximum of 30 km·h^-1 during an obstacle- avoidance task, as well as with the following and overtaking tests. The vehicle can maintain the lane at 30 km·h^-1 , and avoid obstacles at a maximum speed of 15 km·h^-1. When following the front car, the autonomous vehicle can accommodate a velocity required to maintain a reasonable distance and smooth driving.