RIA: Closed-Loop LLM Decision Making with World Models for Autonomous Driving
A team of researchers has introduced the Reason–Imagine–Act (RIA) framework, designed for online safety verification in autonomous driving by integrating a large language model (LLM) reasoner with an action-conditioned world model. In this system, the LLM generates an action template along with potential sub-actions at each stage, while the world model conducts short-horizon rollouts. A safety scorer then identifies the safest action to execute, providing feedback for the subsequent reasoning phase. Utilizing a standardized CARLA point-goal protocol across 1000 episodes, RIA achieves a route completion rate of 80.05%, an arrival rate of 51.10%, and a collision rate of 0.20%.
Key facts
- RIA stands for Reason–Imagine–Act
- It is a closed-loop framework for autonomous driving
- Combines an LLM reasoner with an action-conditioned world model
- Performs online safety verification at each decision step
- LLM proposes action templates and candidate sub-actions
- World model runs short-horizon rollouts
- Safety scorer selects the safest executable action
- Tested on CARLA point-goal protocol with 1000 episodes
- Achieves 80.05% route completion
- Achieves 51.10% arrival rate
- Achieves 0.20% collision rate
Entities
—