Datasets:

taesiri
/

arxiv_audio

+ Simulation plays a vital role in the development of self-driving systems. Simulators are used to develop, test, and enhance driving systems without risking the safety of humans, vehicles, or the environment. However, simulators face a challenge in generating realistic, scalable, and interesting content. Recent advancements in rendering and scene reconstruction have improved the creation of static scene assets, but modeling their layout, dynamics, and behaviors still remains difficult. This study proposes the use of language as a source of guidance for generating dynamic traffic scenes. The proposed model, LCTGen, combines a large language model with a transformer-based decoder architecture to select map locations from a dataset of maps and generate initial traffic distribution and vehicle dynamics. LCTGen achieves superior results in both unconditional and conditional traffic scene generation in terms of realism and fidelity. Code and a video demonstration are available on the project's website.