Reproduce “KiGRAS: Kinematic-Driven Generative Model for Realistic Agent Simulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jianbo Zhao, Jiaheng Zhuang, Qibin Zhou, Taiyu Ban, Ziyao Xu, Hangning Zhou, Junhe Wang, Guoan Wang and 2 more
KiGRAS introduces a kinematic-driven generative model for realistic agent simulation in autonomous driving by re-formulating trajectory generation as the prediction of control actions. The model achieved the top rank on the Waymo SimAgents Challenge leaderboard with a 0.7597 realism meta-metric, demonstrating superior performance with only 0.7 million parameters, and supports customization of driving habits.
Key insights
- Shifting the modeling focus from 'effect space' (trajectory states) to 'cause space' (control actions) fundamentally simplifies the learning problem and inherently guarantees physical feasibility.
- A compact, kinematic-driven representation of driving behavior can achieve state-of-the-art realism with significantly fewer model parameters, demonstrating extreme efficiency and scalability.
- Integrating previous control actions and a unified spatial representation within a Transformer-based decoder is crucial for enhancing temporal consistency and multi-agent interaction modeling.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.