Automatic Test-Case Generation for Autonomous Vehicles

显然，人们不能仅仅依靠物理测试来确保自动驾驶汽车的正确功能。由于物理测试驱动是昂贵和耗时的，因此伴随它们的是计算机模拟是有利的。然而，由于大多数交通场景并不具有挑战性，即使是模拟也往往过于耗时。该方案的目标是为自动车辆的运动规划提供自动合成具有挑战性的测试用例的方法和工具。这可以被视为对运动规划算法的驾驶测试，需要通过才能在真实车辆中使用。为了获得具有挑战性的测试用例，我们将交通规则形式化，并计算度量来估计交通规则的遵守程度。这使得我们可以控制交通规则的遵守程度，以便自动合成测试用例。我们还将形式化场景的用户规范，以便用户可以控制场景生成过程。下一步，我们将合成初始场景。在从待开发的数据库中提取与场景规范相关的初始场景后，我们优化了其他交通参与者和测试车辆的初始状态。因此，我们针对所测试车辆的符合交通规则的可达集合的期望大小进行优化。从优化的初始交通场景出发，优化周边交通参与者的行为，伪造被测车辆的运动规划。为了进一步测试碰撞缓解的概念，我们还计划让其他交通参与者违反交通规则，导致被测车辆的解空间变空。我们开发的方法将使用我们的运动规划基准套件CommonRoad(CommonRoad(Common).in.um.de)通过数值实验进行评估。为了评估生成的场景的关键程度，我们还计划在驾驶模拟器中进行用户研究，以比较主观感知的风险以及合成场景的真实性等衡量标准。

It is apparent that one cannot rely solely on physical test drives for ensuring the correct functionality of autonomous vehicles. Since physical test drives are costly and time-consuming, it is advantageous to accompany them with computer simulations. However, since most traffic scenarios are not challenging, even simulations are often too time-consuming. The goal of this proposal is to provide methods and tools for automatically synthesizing challenging test cases for motion planning of autonomous vehicles. This can be seen as a driving test for motion planning algorithms that needs to be passed in order to be used in real vehicles. In order to obtain challenging test cases, we formalize traffic rules and compute measures to estimate the degree of traffic rule compliance. This makes it possible to control the degree of traffic rule compliance for our automatic synthesis of test cases. We will also formalize the user specification of the scenario so that users can control the scenario generation process. In a next step, we will synthesize the initial scene. After extracting initial scenes from our to-be-developed database that are relevant for the scenario specification, we optimize the initial states of other traffic participants and the vehicle under test. Thereto, we optimize towards a desired size of the traffic-rule-compliant reachable set of the vehicle under test. Starting from the optimized initial traffic scene, we will optimize the behavior of surrounding traffic participants to falsify the motion planner of the vehicle under test. To additionally test collision mitigation concepts, we also plan to let other traffic participants violate traffic rules causing the solution space of the vehicle under test to become empty.Our developed methods will be evaluated by numerical experiments using our motion planning benchmark suite CommonRoad (commonroad.in.tum.de). To evaluate the criticality of the generated scenarios, we additionally plan to conduct user studies in our driving simulator to compare measures like the subjectively perceived risk as well as the realism of our synthesized scenarios.