Autonomous driving system based on dual process theory and deliberate practice theory.

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Autonomous driving faces challenges in open environments due to poor hazard perception and generalization.
• Existing systems struggle with few-shot learning and corner cases, impacting reliability.

Purpose of the Study:

• To introduce CogniDrive, a novel framework enhancing autonomous driving robustness and generalization.
• To leverage Large Language Models (LLMs) and dual-process theories for improved driving systems.

Main Methods:

• CogniDrive employs two modes: InstinctNav (behavioral cloning, retrieval augmented generation) and ReflectPlan (self-reflection, reward internalization).
• A vision-language model is integrated for multimodal environmental understanding and corner-case hazard detection.
• Deliberate practice and dual-process theories inform the framework's design for enhanced learning and reasoning.

Main Results:

• Demonstrated state-of-the-art performance in extensive open-loop and closed-loop experiments.
• CogniDrive shows significant improvements in hazard perception, few-shot generalization, and corner-case handling.
• The proposed evaluation framework includes safety, comfort, and energy efficiency metrics.

Conclusions:

• CogniDrive effectively addresses key limitations in current autonomous driving technology.
• The framework offers a promising approach to developing more reliable and generalizable self-driving systems.
• LLM-driven contextual reasoning and multimodal self-reflection are crucial for advancing autonomous driving capabilities.