Conceptual voicebot in the context of a passenger information system in an automated bus

Abstract: Automated driving is seen as an essential part of the mobility of the future. This does not only affect the private sector but the public one as well. One consequence of this automation, in the public sector, is, that there won’t be any vehicle driver. This will lead to a need for additional systems, taking over further tasks on top of driving, which were previously executed by the vehicle driver too. One of those tasks is to answer the passengers’ questions, e.g. regarding future stops or alternative routes. This is particularly important in the bus sector. In this sector, there is greater passenger uncertainty, because the vehicle is not bound to fixed routes, due to the use of public roads, like in the case of a (suburban) train. This usage can lead to route changes, caused by e.g. road works or traffic jams. Consequently, an automated vehicle needs to be able to answer questions asked by the passengers. In order to address the passenger’s uncertainty, these answers should be easy to understand and personalized/ fitted to the questions asked.In this paper, three different approaches to automatically answer questions in the context of an automated bus are proposed. These approaches are 1.) a rule-based system, 2.) a system based on a large language model (LLM) like GPT-4 or LaMDA, and 3.) a hybrid system of rules and an LLM. The different approaches are being conceptualized and discussed. In the scope of the conceptualization, requirements as well as further challenges are derived. The discussion focuses on the capabilities to answer questions correctly and handle different languages as well as bad language. Additionally, the remaining challenges are further addressed. This includes e.g. handling emergency calls, vandalism reports, and the distribution of responsibilities regarding and the interaction with a passenger emergency management system. Following the discussion, the arguments are used to rate the approaches. Using this rating, the hybrid approach seems to be the most suited one. The reasons for this conclusion are, on the one hand, the capability to restrictions using the rules, including a defined course of action in certain situations, and on the other hand, the LLM’s ability to answer in a natural manner. Lastly, possible extensions for the hybrid approach are discussed.

Keywords: Automated public transportation, Passenger information system, Large language model

DOI: 10.54941/ahfe1004553

Cite this paper:

​