Page - (000104) - in Autonomes Fahren - Technische, rechtliche und gesellschaftliche Aspekte

Implementable Ethics for Autonomous Vehicles92 societal impact of accidents. However, in such cases, the car may take an action that injures the occupant or owner of the vehicle more severely to minimize harm to others. Such self-sacrificing tendencies may be virtuous in the eyes of society but are unlikely to be appreciated by the owners or occupants of the car. In contrast, consider a vehicle that pri- marily considers occupant safety. This has been the dominant paradigm in vehicle design with a few exceptions such as bumper standards and attention to compatibility in pedestrian collisions. A vehicle designed to weight occupant protection heavily might place little weight on protecting pedestrians since a collision with a pedestrian would, in general, injure the vehicle occupant less than a collision with another vehicle. Such cars might not result in the desired reduction in traffic fatalities and would be unlikely to gain societal acceptance. Goodall [3] goes a step further to illustrate how such cost functions can result in unin- tended consequences. He presents the example of a vehicle that chooses to hit a motorcy- clist with a helmet instead of one without a helmet since the chance of survival is greater. Of course, programming automated vehicles to systematically make such decisions dis- courages helmet use, which runs contrary to societal objectives of safety and injury reduc- tion. The analogy could be extended to the vehicle purposefully targeting collisions with vehicles that possess greater crashworthiness, thereby eliminating the benefit to drivers who deliberately choose to purchase the “safer” car. Thus truly understanding the outcomes or consequences of a vehicle’s actions may require considerations well beyond a given accident scenario. Of course for such cases to literally occur, the vehicle must be able to distinguish the make and model of another vehicle or whether or not a cyclist is wearing a helmet and understand how that difference impacts the outcome of a collision. While algorithms for pedestrian and cyclist recognition continue to improve, object classification falls short of 100 percent accuracy and may not include vital information such as posture or relative orientation. As Figure 5.3 indicates, the information available to an automated vehicle from sensors such as a laser scanner is significantly different than that available to human drivers from their eyes and brains. As a result, any ethical decisions made by vehicles will be based on an imperfect understanding of the other objects or road users impacted by that decision. With the objects themselves uncertain, the value of highly detailed calculations of the probability of accident outcomes seems questionable. With all of these challenges to defining an appropriate cost function and obtaining the information necessary to accurately determine the cost of actions, a purely consequentialist approach using a single cost function to encode automated vehicle ethics seems infeasible. Still, the fundamental idea of assigning costs to penalize undesired actions or encourage desired actions can be a useful and vital part of the control algorithm, both for physical con- siderations such as path tracking and issues of ethics. For instance, to the extent that virtues can be captured in a cost function, virtue ethics as proposed by Lin for automated vehicles [12] can be integrated into this framework. This may, for instance, take the form of a more qualitative adjustment of weights for different vehicles. An automated taxi may place a higher weight on the comfort of the passengers to better display its virtues as a chauffeur. An automated ambulance may want to place a wider margin on how close it comes to pedest-