Research · Python · Autonomous Driving
REVO
A research framework for evaluating the reasoning and decision behavior of LLM-based autonomous agents operating in the CARLA driving simulator. REVO provides structured scenario testing, behavior logging, and a visualization pipeline to surface how agents respond to complex traffic situations.
Overview
REVO was built to answer a specific research question: how do LLM-based agents actually make decisions when placed in realistic driving scenarios? Rather than just measuring outcomes, the framework captures the reasoning chain behind each agent decision, making it possible to analyze where and why agents fail.
Scenarios are run inside CARLA, a high-fidelity open-source driving simulator. Agents receive scene context through structured prompts and respond with driving actions. REVO logs each decision alongside the agent's reasoning and visualizes behavior patterns across scenario runs.
Build Details
- CARLA simulator for realistic driving scenarios
- Structured prompting pipeline for LLM agent control
- Decision and reasoning logging per scenario step
- Visualization pipeline for behavior analysis across runs