Decision Making for Future Autonomous Vehicles in Complex Scenarios

高水平的车辆自动驾驶保证了更高的安全性和效率，需要高度准确的环境信息。在现实中，由于信息不完整，这种信息具有局限性，无论这是通过外部感知还是本体感知传感器错误，例如噪声、丢失的分组或遮挡，从而车辆可能遮挡传感器对场景的查看。这些限制给自动驾驶车辆带来了挑战，特别是在SAE 4级和5级，不允许进行干预。这项研究的任务是探索自动车辆决策背后的潜在方法，当执行高度动态的操作时，如在单一车道上超车或并入环形交叉路口或高速公路。超车甚至对人类驾驶员来说也是具有挑战性的，但对于某些情况，如缓慢移动的交通或公路建设工作，超车操作对于及时到达目的地或防止由于不采取行动而导致进一步交通堵塞至关重要。传统的行为规划或决策系统使用基于规则的方法，通过对个别场景的彻底系统工程，完成特定操作所需的车辆行为可以在车辆的决策系统中明确定义。这种基于规则的方法的例子包括有限状态机或使用模糊逻辑的方法。例如，如果一辆车在60英里/小时的道路上以每小时40英里的速度跟在一辆领先车辆的后面，它将被指示进行一些预先定义的小动作，以安全有效的方式超车；进入相邻的迎面而来的车道，加速以领先车辆的速度移动，然后最后回到左侧车道并恢复正常驾驶。虽然基于规则的方法允许系统做出决策的透明度，但它们需要广泛的场景知识，而在实践中，由于道路用户的数量和多样性，这些知识永远不会相同。基于学习的方法试图通过使用模拟或真实世界的数据来训练能够生成正确策略的模型，以涵盖一系列不同的场景。这种模式通常是黑箱操作，很难理解它在特定情况下会如何反应。这些方法包括强化学习或使用马尔可夫决策过程来描述系统的行为。本研究的重点是探索如何在任何给定的场景中结合部分可观测性来计算自动驾驶所需的准确策略或动作-状态对。该结果将为使用不确定性生成策略提供洞察力，并描述如何使用来自其他自主控制子系统状态的信息。

High level vehicle autonomy promises improved safety and efficiency requiring highly accurate information of the environment. In reality this information has limitations, owing to incomplete information, whether this is through exteroceptive or proprioceptive sensor errors such as noise, missing packets or occlusions whereby a vehicle may obstruct a sensors view of the scene. These limitations present a challenging situation for autonomous vehicles, particularly at SAE levels 4 and 5 whereby no intervention is permissible. This research mission is to explore the underlying methodologies behind automated vehicle decision making when executing highly dynamic manoeuvres such as overtaking on single carriageway roads or merging onto roundabouts or highways. Overtaking is challenging even for human drivers, yet for certain situations such as slow moving traffic or highway construction work, an overtaking manoeuvre is essential to arrive at a destination in a timely manner or preventing further traffic build up as a result of inaction.Traditional behaviour planning or decision-making systems use rule-based methods whereby through thorough systems engineering of individual scenarios, vehicle behaviours necessary to complete a specific manoeuvre can be explicitly defined in the decision-making system on the vehicle. Such examples of rule-based methodologies include finite state machines or those using fuzzy logic. If, for example, a vehicle was following a single lead vehicle at 40mph on 60mph road, it would be instructed to make a number of small predefined manoeuvres to overtake in a safe and efficient manner; move into the adjacent oncoming traffic lane, accelerate to move ahead of the lead vehicle then finally move back into the left hand lane and resume normal driving. While rule-based methods allow for transparency of the decisions being made by the system, they require extensive knowledge of the scenario which in practice is never the same due to the number and variety of road users.Learning based methods attempt to encompass a range of different scenarios by using either simulated or real-world data to train a model capable of generating the correct policy. This model is often black box and challenging to understand how it will react in a given situation. Such methods include reinforcement learning or using Markov Decision Process in formulating the behaviour of the system.The focus of this research is to explore how to incorporate partial observability in any given scenario can be used to compute accurate policies or action-state pairs required for automated driving. The outcome will provide insight toward using uncertainty for policy generation along with describing how using information from other autonomous control sub-systems states can be used additionally.