Reproduce “RiskWorld: Object-Centric Latent World Modeling for Autonomous Driving Risk Identification” 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
Jingzheng Li, Yufei Ge, Qianren Mao, Zhijun Chen, Bing Li, Xingyu Peng, Baochang Zhang, Xianglong Liu
RiskWorld, an object-centric latent world model, simulates the imagined future evolution of individual objects relative to an autonomous vehicle to identify specific safety-critical elements. The system achieves a 63.0% F1 score for risk localization with a low 2.1% false-alarm rate on the RiskBench dataset and effectively preserves planning-critical information.
Key insights
- Driving risk is object-specific and temporally evolving, making direct prediction of future ego–object relations crucial for precise, proactive risk localization.
- An object-centric latent world model can effectively identify risk sources by focusing on the imagined evolution of individual objects without requiring full future scene reconstruction, which can be computationally intensive.
- Integrating pretrained predictive video representations with structured historical data, explicit relational modeling, and stochastic latent dynamics significantly enhances risk identification accuracy and reduces false alarms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.