LitSim introduces a novel approach to long-term interactive traffic simulation by addressing the shortcomings of existing methods. The key focus is on maximizing realism while avoiding unrealistic collisions. By utilizing a conflict-aware policy, LitSim intervenes only when unrealistic conflicts are predicted, ensuring interactions among agents and reducing the likelihood of collisions. The model is trained and validated on real-world datasets, outperforming popular approaches in terms of realism and reactivity.
LitSim aims to bridge the gap between simulation and reality by accurately predicting future agent trajectories within a specified region of interest (ROI). The method involves joint motion prediction with interaction, conflict detection, and conflict-aware control policy. Through experiments on NGSIM data, LitSim demonstrates superior performance in terms of realism, reactivity, progress, and relevant rate compared to traditional simulators like IDM and GAIL.
The ablation study highlights the importance of joint prediction and control policy components in enhancing simulation performance. Different ROI ranges impact ADE, collision rate, progress, and relevant rate metrics. Despite some limitations related to deep learning predictors and observed states availability, LitSim strikes a commendable balance between realism and reactivity in long-term simulations.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Haojie Xin,X... at arxiv.org 03-08-2024
https://arxiv.org/pdf/2403.04299.pdfDeeper Inquiries