[论文解读] Decentralised Cooperative Collision Avoidance with Reference-Free Model Predictive Control and Desired Versus Planned Trajectories

C. E. Wartnaby, Daniele Bellan|arXiv (Cornell University)|Apr 15, 2019

Traffic control and management被引用 4

一句话总结

本文提出了一种去中心化的协作避碰框架，采用无参考模型预测控制（MPC）并结合期望轨迹与规划轨迹。车辆基于非协作车辆的预测路径计算各自的期望轨迹，并调整以避开他人的期望路径，从而实现无领导者、鲁棒的协作，无需显式协调，在多样化场景中实现有效的避碰，且计算成本呈线性增长。

ABSTRACT

Connected and automated vehicles provide a new opportunity for highly advanced collision avoidance, in which several cars cooperate to reach an optimal overall outcome, that no single car acting in isolation could achieve. For example, one car may automatically swerve to allow another to avoid an obstacle. However, this requires solving the challenging problem of deciding what joint trajectories an ad-hoc group of cooperating vehicles should follow, with no obvious leader known in advance. To avoid the complexities of agreeing what plan to follow in an ever-evolving situation, a protocol requiring no leader and no explicit inter-vehicle agreement is desirable, which nevertheless yields cooperative, robust behaviour. One method is demonstrated here, in simulation. This uses the notion of "desired" versus "planned" trajectories, allowing vehicles to influence each other for mutual benefit, without requiring a leader or explicit agreement protocol. Essentially the desired trajectory is that which the vehicle would choose if other cooperating vehicles were not present, avoiding the predicted paths of non-cooperating actors. The planned trajectory additionally accounts for the planned trajectories of other cooperating vehicles, giving the safest currently available path. Both trajectories are broadcast. As each vehicle attempts to (weakly) avoid the desired trajectories of other vehicles, cooperative behaviour emerges. A simple form of model predictive control is used. The cost function penalises predicted collisions, accounting for severity. There is a weak preference for maintaining the current road lane, but no explicit reference trajectory. This decentralised planning and simple optimisation scheme results in effective handling of a wide range of collision scenarios, with no hard limit to the number of cooperating vehicles. The computing cost is linear in the number of vehicles.

研究动机与目标

• 在无需领导者或车辆间显式协议的情况下，实现联网自动驾驶车辆之间的协作避碰。
• 解决在无中心协调的动态、临时性车辆群体中去中心化轨迹规划的挑战。
• 开发一种可扩展的解决方案，在最小化计算开销的同时保持安全性和鲁棒性。
• 通过轨迹广播使车辆能够相互影响行为，促进相互安全，而无需显式交换计划。

提出的方法

• 每辆车辆基于避开非协作车辆预测路径，使用无参考MPC公式计算其'期望'轨迹。
• 通过避免其他车辆的期望轨迹，生成'规划'轨迹，以确保协作安全性。
• 将期望轨迹和规划轨迹广播给邻近车辆，实现相互感知。
• MPC代价函数对预测碰撞施加严重性加权，优先选择更安全的路径。
• 包含对保持车道的弱偏好，但不使用显式参考轨迹。
• 该方案以去中心化方式运行，计算成本随车辆数量线性增长。

实验结果

研究问题

• RQ1 如何在无指定领导者的情况下实现去中心化的协作避碰？
• RQ2 期望轨迹与规划轨迹在实现无需显式协调的隐式协作中起到何种作用？
• RQ3 无参考MPC方法能否在动态多车场景中保持安全性和可扩展性？
• RQ4 缺乏参考轨迹如何影响避碰的鲁棒性和性能？
• RQ5 随着协作车辆数量的增加，所提方法的计算可扩展性如何？

主要发现

• 所提方法在广泛复杂场景中实现了有效的避碰，无需领导者或显式协调。
• 通过相互避开其他车辆的期望轨迹，协作行为自然涌现，无需直接交换计划。
• 系统在车辆数量增加时保持鲁棒性能，计算成本呈线性增长。
• 缺乏参考轨迹简化了控制设计，同时保持了安全性和稳定性。
• 仿真结果表明，该方法能够有效应对动态、不可预测的环境，处理多个相互作用的智能体。