Reproduce “Traffic Scenario Orchestration from Language via Constraint Satisfaction” 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
Frieda Rong, Chris Zhang, Kelvin Wong, Raquel Urtasun
A neurosymbolic framework integrates a Large Language Model with a Satisfiability Modulo Theory solver to orchestrate traffic scenarios from natural language descriptions. This approach enables precise, controllable, and reactive closed-loop testing for autonomous vehicles, demonstrating improved scenario orchestration success rates and superior precision compared to a learning-based baseline.
Key insights
- Delegating high-level reasoning to an LLM and precise, verifiable low-level reasoning to an SMT solver effectively bridges the gap between the scalability of data-driven methods and the controllability of programmatic approaches.
- Explicitly formalizing scenario constraints allows for guaranteed adherence to user intentions and kinematic feasibility, which is challenging for purely learning-based models that offer only approximate sensitivity.
- Closed-loop replanning is critical for robust autonomous vehicle testing, significantly enhancing the success rate of scenarios that require dynamic interaction and adaptation to the ego vehicle's unpredictable behavior.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.