Project Portfolio

Trolley Mod (Kansas State University, January 2018-Present)

The Trolley Mod is an open-sourced mod for the CARLA driving simulator. It’s purpose is to address an important problem involving moral decision-making for AI in autonomous vehicles: a self-driving car will eventually be presented with a situation where there is a crash with unavoidable casualties. What is the “moral” decision for the AI in this case? Who or what must it collide with in order to achieve the “best” possible outcome?

This mod seeks to answer these questions by enabling the easy creation of “trolley problems” in the CARLA simulator. A researcher (or any other curious individual) can create their own set of trolley problems consisting of various people and objects, and set up the simulator so a human controls the car and must make the decision on how to handle the crash. The mod then stores the results of the collisions, and then the data can be utilized for applications such as teaching a machine-learning algorithm a moral framework, conducting surveys for sociology, and other related fields.

The mod utilizes the Unreal Engine 4 on a Linux OS and supports a Python API. It is currently in the beta stage, and ideally will be fully complete by mid-October. A technical paper on the mod will also be submitted in November for the AAAI 2019 conference in Honolulu, Hawaii.