■ 제목: AI-Powered Human-Like Driving Behavior Research

■ 연사: 신동훈 교수(서울대 미래모빌리티기술센터(FMTC))

■ 일시: 12.01(목) 16:00

■ URL: https://kaist.zoom.us/j/82464453823?pwd=akVSdlhvc2dnUXhtb0M5THA0S2FRQT09  

■ 초록: There has been increased interest in Autonomous Driving recently. Large corporations, such as Waymo, Tesla, Cruise, and Uber, are investing heavily into it. Current widely adopted rule-based algorithm make it possible to compute paths for autonomous vehicles, however, in many cases they do not convey all the information required for a comfortable ride. In order to achieve and reproduce proactive driving behaviors, a trade-off between safety, speed and ride comfort needs to be investigated, which is difficult to model without using human data. Our group concentrates on extracting driving behaviors that result in a safe and comfortable ride for the passengers. We use the data from human expert drivers who can drive smoothly while proactively avoiding potential hazards. We also define a comfortable drive as providing an experience that puts passengers at ease and relaxed while in the vehicle, while avoiding situations causing unease or discomfort. This can be achieved by limiting the amount of discomfort due to velocity, lateral acceleration and jerk by having our agent imitate expert drivers. The proactive driving behaviors performed by these experts alleviates these causes. I will implement a deep autoencoder network to process large volumes of data to extract latent features. We clustered these latent features into behaviors and created velocity profiles. This allows us to create an autonomous agent to correctly select the proper behavior to use depending on the environment it was in. Our research aims to show comparable results to that of expert drivers in the same environment. With a successfully trained model, these extracted behaviors are applicable to other similar environments. We will show our agent was able to use the professional experience, gained from years of human driving instructing, in order to drive proactively in several urban environments.