Open Collections

Tian, Yuan

University of British Columbia

Automated driving has been pursued for more than fifty years, but with the widespread adoption of machine learning to train autonomous driving agents, simulators have begun to play an increasing role in the development and deployment of autonomous vehicles because they are not only faster, cheaper, and safer than physical experiments, but also more controllable, observable, and repeatable. In order to create realistic simulations, it is critical to have a good ``social agent'' to control the action of the other vehicles in the simulation. To date, however, no systematic comparison between different social agents has been performed. Here we describe a series of experiments which trained reinforcement learning (RL) agents in the SMARTS (Scalable Multi-Agent RL Training School) traffic simulation environment with three different social agents, and performed a comparison of the performance of the resulting RL agent. Along with SMARTS supported social agents SUMO (Simulation of Urban Mobility) and ZOO (provided by SMARTS), we integrated the DRIVE social agent (provided by Inverted AI) into the simulator for this study. The RL agents were trained and tested in six different task scenarios on five different maps. Overall, the experimental results show that RL agents trained in the environment with the DRIVE social agent performed more consistently on criteria including completion rate, collision rate and off-road rate, indicating DRIVE's heightened behavioral diversity and meaningful interactivity with the ego vehicle. We further conducted an analysis of some of the characteristics of the traffic generated by the different social agents: Traffic density, traffic speed and acceleration distribution, and average neighbor distance.

Text

2024-01-12

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/87226

Computer Science

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/