LLM reasoning orchestrates PPO and PID controllers in CARLA rather than driving the car directly

arXiv·low signalXBlueskyLinkedInCopy link

An August 20 arXiv paper builds a hybrid autonomous driving stack where a centralized coordinator dispatches between PPO-trained RL agents and PID control, using LLM common-sense reasoning for situational context and iteratively refining the RL reward function for dynamic environments. Evaluation ran on highly randomized CARLA scenarios across varied weather and traffic. The design point worth stealing is keeping the LLM out of the latency-critical control loop and using it as a supervisor, which is the same tradeoff that shows up in any agent controlling real-time systems.