Reproduce “HMVLM: Multistage Reasoning-Enhanced Vision-Language Model for Long-Tailed Driving Scenarios” 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
Daming Wang, Yuhao Song, Zijian He, Kangliang Chen, Xing Pan, Lu Deng, Weihao Gu
The authors present HMVLM, a vision-language driving planner that interprets five camera views and recent vehicle motion, states a driving decision, and predicts a path. It refines waypoints using outlier removal and adaptive Savitzky–Golay smoothing. In the 2025 Waymo challenge, the submission placed second with an RFS of 7.7367, 2.77% above the public baseline; component effects remain untested.
Key insights
- Vision-Language Models can effectively perform complex strategic planning and high-level semantic understanding for autonomous driving in challenging scenarios.
- Multi-stage Chain-of-Thought prompting provides interpretable intermediate reasoning steps, improving planning performance and model transparency.
- Integrating selective multi-view camera data with historical ego vehicle kinematics offers rich contextual input, enhancing motion prediction and decision fidelity, while trajectory refinement ensures kinematic feasibility.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.