Reproduce “STAGE: STyle-controllable Action GEneration for personalized autonomous driving” 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
Zihao Liu, Xing Liu, Yizhai Zhang, Panfeng Huang
STAGE (STyle-controllable Action GEneration) introduces a method for personalized autonomous driving by learning continuous, semantically interpretable driving styles, specifically aggressiveness, from human data. The system enables intuitive user control over driving behavior, achieving a 92.1% average completion rate in simulated scenarios and quantitatively demonstrating superior alignment with human-intended styles.
Key insights
- Continuous, monotonic, and semantically interpretable style values, such as aggressiveness, can be learned and used to provide intuitive and precise control over autonomous driving behavior.
- Combining preference-based style learning with action modality decomposition (style and VAE-driven general patterns) leads to enhanced controllability, continuity, and safety in generated driving actions.
- Rule-based automated preference comparisons can effectively mitigate the human annotation burden while still producing style values that align with human perceptions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.